📖 Core Concepts
Core Concepts: Advanced Molecular Biology (Genomics, Proteomics, and Systems Biology)
Study notes for Skipper, BIOL502. Read this as a chapter rather than a checklist. Every technique below is a device that turns a molecular event into a number, followed by a statistical argument about what that number means.
0. What Is This Subject Really About?
In BIOL101, BIOL102 and BIOL201 you learned molecular biology one molecule at a time. You studied a single gene's promoter, a single transcription factor, a single kinase. That approach is reductionist, and it is very powerful. It also has a blind spot. A cell is not a list of parts. It holds roughly 20,000 protein-coding genes, a much larger number of proteoforms, and millions of regulatory interactions, all of them operating at once. This course is about the shift from studying molecules to measuring whole molecular populations and inferring the system that produced them.
Three ideas recur in every topic in this course. Once you see them, the long list of technique names becomes much easier to organize.
- Digitization. Nearly every modern method converts molecules into countable discrete events. Sequencing turns DNA into reads. Mass spectrometry turns peptides into spectra and ion counts. Droplet scRNA-seq turns cells into barcodes. Molecular biology has become a counting science, with counting statistics (Poisson, negative binomial, hypergeometric) as its working language.
- Sampling and inference. We never observe the genome, transcriptome or proteome directly. We observe a random sample of fragments, and every downstream step (assembly, variant calling, differential expression, peptide identification) is an inference problem. The hard intellectual work is the noise model. You saw this logic in MATH201 and CHEM302: a measurement is only as good as your model of its error.
- From parts to wiring. Genomics and proteomics produce the parts list. Interactomics, structure prediction and network analysis produce the wiring diagram. Systems biology asks what behaviours (switches, oscillators, adaptation) emerge from that wiring. These are behaviours no single component has.
An analogy to carry through the course: genomics is reading every book in a library, but the books arrive shredded into strips. Proteomics is working out which books people actually check out, and how often. Interactomics is learning which books are always borrowed together. Systems biology is predicting how the library's lending patterns shift when a new course is announced.
Alberts et al. frame the central challenge clearly. The information in a genome is necessary but not sufficient to specify a cell. The cell's state is set by which parts of that information are deployed, when, and in what combinations [Alberts et al., 2022]. Compeau and Pevzner add the computational half of the picture. Behind each high-throughput experiment lies an algorithmic problem, and the elegance or difficulty of that problem often decides whether the biology is tractable at all [Compeau & Pevzner, 2018].
1. Sequencing Is Massively Parallel Sampling: Three Physical Principles
All next-generation sequencing (NGS) platforms answer the same question: how do we read billions of nucleotides at once? They differ in the physical signal used to identify each base. Those physical differences explain each platform's characteristic error profile, and that profile is the most important practical fact about any sequencer.
1a. Illumina sequencing-by-synthesis (SBS)
The workflow has four steps.
- DNA is fragmented and ligated to adapters.
- Fragments are clonally amplified on a flow cell. Older instruments use bridge amplification; newer ones use patterned flow cells with exclusion amplification. Each cluster contains roughly a thousand copies of one molecule.
- Each cycle adds all four nucleotides. Every nucleotide carries a distinct fluorophore and a reversible 3′-O blocking group, so exactly one base is incorporated per strand per cycle.
- The flow cell is imaged, then the fluorophore and the block are chemically cleaved, and the cycle repeats.
You already know this chemistry. It is Sanger dideoxy termination made reversible.
Error profile. Illumina errors are mostly substitutions, at roughly 0.1–1%. Quality falls toward the 3′ end of each read because of phasing. Within a cluster, a few strands fail to incorporate (lag) or incorporate twice (lead), so the cluster's signal gradually desynchronizes. A cluster is an ensemble measurement, and ensembles dephase. This is the same logic as T₂ relaxation in NMR from CHEM501.
1b. PacBio SMRT (single-molecule real-time) sequencing
A single polymerase is fixed at the bottom of a zero-mode waveguide. This is a nanoscale well smaller than the wavelength of the illuminating light, so only a zeptoliter volume at the bottom is illuminated. The fluorophore is attached to the terminal phosphate of the nucleotide. It is therefore released as pyrophosphate when the base is incorporated, and the product strand remains native, unmodified DNA. We watch one polymerase work in real time.
Raw single-pass reads carry about 10–15% error, mostly random insertions and deletions. Because these errors are random rather than systematic, they can be averaged away. The template is circularized as a "SMRTbell," the polymerase loops around it many times, and the consensus of the passes (circular consensus, "HiFi") exceeds 99.9% accuracy at read lengths of about 15–25 kb [Wenger et al., 2019].
There is a bonus. Polymerase kinetics (the interpulse duration) change at methylated bases, so epigenetic marks can be read directly.
1c. Oxford Nanopore
A single strand is ratcheted by a motor protein through a protein nanopore embedded in a membrane, with a voltage applied across it. The ionic current is modulated by the several nucleotides (a k-mer) occupying the pore's constriction. Neural-network basecallers translate the current trace, the "squiggle," into sequence.
Key properties:
- Reads can exceed a megabase.
- RNA can be sequenced directly.
- Modified bases are detectable.
- Historically, errors were concentrated in homopolymers. The current signal does not change as a run of identical bases passes through, so the length of the run is ambiguous.
This connects to PHYS202: the ionic current through a nanopore is essentially Ohm's law applied to an electrolyte in a confined geometry.
1d. Quantifying confidence: Phred scores and coverage
Each base call carries a quality score on a logarithmic scale.
Phred quality: Q = −10 · log₁₀(P_error) Q = 20 → P_error = 1/100 (99% accurate) Q = 30 → P_error = 1/1000 (99.9% accurate) Q = 40 → P_error = 1/10000
How much sequencing is enough? [Lander & Waterman, 1988] treated read start positions as a Poisson process along the genome. The number of reads covering a given base is then Poisson-distributed.
Mean coverage: C = L · N / G
L = read length, N = number of reads, G = genome size
Fraction of genome with zero coverage:
P(0 reads) = e^(−C)
C = 1 → 37% uncovered
C = 10 → 0.0045% uncovered (≈ 140 kb of a 3.1 Gb human genome)
C = 30 → ≈ 10⁻¹³ (in theory; GC bias makes reality worse)
Why this matters. The exponential tail means there are diminishing returns, but never zero returns. Real coverage is also overdispersed relative to Poisson, because of GC bias, mappability and PCR duplicates. That overdispersion is one reason clinical genomes target 30× or more rather than 10×.
1e. Library preparation is part of the measurement
Fragmentation, adapter ligation, size selection and PCR amplification all introduce bias. PCR preferentially amplifies fragments of moderate GC content and creates duplicates, which are identical reads that do not represent independent observations. Treat the library-prep protocol as part of the instrument. A beautiful sequencer cannot rescue a biased library.
2. Genome Assembly Is a Graph Problem, and the Choice of Graph Is Everything
Assembly asks you to reconstruct a genome from millions of overlapping fragments without a reference. It is one of the most elegant stories in computational biology, and Compeau and Pevzner tell it well [Compeau & Pevzner, 2018; Compeau, Pevzner & Tesler, 2011].
2a. Two formulations
- Overlap–Layout–Consensus (OLC). Each read is a node, and an edge joins reads that overlap. Reconstructing the genome means finding a path that visits every node exactly once, which is a Hamiltonian path. This problem is NP-complete. No efficient algorithm is known, and none is expected.
- De Bruijn graph. Break every read into k-mers. Each distinct (k−1)-mer is a node, and each k-mer is an edge from its prefix to its suffix. Reconstructing the genome now means finding a path that traverses every edge exactly once, which is an Eulerian path. Euler solved this in 1736 (the Königsberg bridges problem), and it can be solved in linear time.
This is one of the most surprising ideas in the course. The same biological problem can be computationally intractable or trivially solvable depending only on whether you put the data on the nodes or on the edges. Reformulation, not more computing power, made short-read assembly of large genomes practical.
In practice, OLC has returned for long reads, which are few and long enough that pairwise overlap is feasible. De Bruijn graphs dominate short-read assembly.
2b. Why assembly is hard: repeats
If a repeat is longer than your reads (or your k-mers), the graph cannot tell which unique flanking sequence connects to which. The assembly fragments into contigs that break at repeats. This is the deep reason long reads changed genomics. The first truly complete, telomere-to-telomere human genome (T2T-CHM13) required ultra-long nanopore reads plus PacBio HiFi to resolve centromeric satellites and segmental duplications [Nurk et al., 2022].
Scaffolding orders and orients contigs using longer-range information. Sources include paired-end or mate-pair insert sizes, optical maps, and Hi-C contact frequencies. Hi-C works because loci on the same chromosome contact each other far more often than loci on different chromosomes. Gaps between contigs in a scaffold are filled with runs of N.
2c. Reference-guided vs. de novo
Reference-guided assembly (or simply alignment to a reference) is cheaper and more sensitive for small variants. It is biased toward the reference, so sequence absent from the reference is invisible to it. This is the motivation for pangenome graphs, which represent population variation as alternative paths through a graph.
2d. Quality metrics: N50 and its limits
N50: sort contigs from longest to shortest and sum their lengths.
N50 is the length of the contig at which the cumulative sum
first reaches ≥ 50% of the total assembly length.
Example: contigs = 80, 70, 50, 40, 30, 20, 10 kb (total 300 kb)
cumulative: 80, 150 → reaches 150 = 50% → N50 = 70 kb
N50 measures contiguity, not correctness. An assembler that aggressively joins contigs across repeats can inflate N50 while creating misassemblies. Pair N50 with correctness measures:
- BUSCO completeness, the fraction of conserved single-copy orthologs recovered.
- k-mer–based QV, a consensus accuracy estimate.
- Misassembly counts against an orthogonal dataset.
3. Variant Calling Is Bayesian Inference on Noisy Evidence
Once reads are aligned to a reference, every position presents a question. Is a mismatch a real variant, a sequencing error, an alignment artifact, or a PCR error? Variant calling is a probabilistic classification problem.
3a. Types of variation
- SNVs/SNPs. Single-base substitutions. These are the easiest to call.
- Indels. Small insertions and deletions. These are harder, because aligners can misplace gaps, particularly in homopolymers and tandem repeats.
- Structural variants (SVs). Events of 50 bp or more: large deletions, duplications, inversions, translocations. Short reads detect them only indirectly, through discordant read pairs, split reads and coverage changes. Long reads span them directly.
3b. The genotype likelihood
For a diploid site with reference allele A and alternate allele B, the caller computes the probability of the observed reads under each genotype and combines it with a prior using Bayes' theorem.
P(G | D) ∝ P(D | G) · P(G)
Each read i shows a base b_i with error probability ε_i (from its Phred score).
For genotype G with alleles (a₁, a₂):
P(D | G) = ∏_i [ ½ · P(b_i | a₁) + ½ · P(b_i | a₂) ]
where P(b_i | a) = 1 − ε_i if b_i = a
= ε_i / 3 otherwise
Two features of this model are worth noticing. First, heterozygous sites need enough depth to show both alleles: with 10 reads, the probability of seeing no alternate reads at a true het is (½)¹⁰ ≈ 0.001. Second, the prior P(G) carries population knowledge, such as heterozygosity θ ≈ 10⁻³ in humans.
3c. GATK Best Practices: why each step exists
The Genome Analysis Toolkit workflow [McKenna et al., 2010; DePristo et al., 2011] is not arbitrary ritual. Each step removes a specific artifact.
- Alignment (e.g., BWA-MEM).
- Mark duplicates. PCR duplicates are not independent evidence, and counting them would inflate confidence.
- Base quality score recalibration (BQSR). Instrument-reported Phred scores are systematically miscalibrated by cycle and sequence context. Known variant sites are masked, and every other mismatch is treated as an error to re-estimate the true error rates.
- HaplotypeCaller. In active regions, the caller locally reassembles haplotypes using a de Bruijn graph, which is concept 2 reused. It then scores each read against each candidate haplotype with a pair-HMM. This is why GATK handles indels much better than pileup-based callers.
- Joint genotyping across samples, which borrows strength across a cohort.
- Variant filtering, either by VQSR (a model trained on known true variants) or by hard filters.
4. RNA-seq: Counting Transcripts, and the Statistics of Counts
4a. Choices made before sequencing
- Poly-A selection vs. rRNA depletion. About 80% of cellular RNA is ribosomal, so it must be removed. Oligo-dT capture enriches mature polyadenylated mRNA. It misses histone mRNAs, many lncRNAs and non-polyadenylated transcripts, and it requires intact RNA, because degraded RNA gives 3′ bias. rRNA depletion with probe hybridization and RNase H retains pre-mRNA (introns), non-poly-A RNAs and degraded samples such as FFPE tissue.
- Strand specificity. In the common dUTP method, dUTP replaces dTTP during second-strand cDNA synthesis, and that strand is then destroyed by uracil-DNA glycosylase. Only the first strand is amplified, so each read's orientation reveals which genomic strand was transcribed. This is critical for overlapping antisense genes. Notice that it borrows the logic of base-excision repair from BIOL201.
4b. Alignment vs. pseudoalignment
RNA reads span exon–exon junctions, so ordinary DNA aligners fail on them. Two strategies solve this.
- Spliced aligners. STAR uses an uncompressed suffix array to find maximal mappable prefixes, then "stitches" seeds across introns. It is fast and memory-hungry, needing about 30 GB for human [Dobin et al., 2013]. HISAT2 uses a hierarchical graph FM-index and is far more memory-efficient [Kim et al., 2015; Kim et al., 2019]. Both produce base-level alignments. Counting then uses tools such as featureCounts [Liao et al., 2014].
- Pseudoalignment and selective alignment. kallisto asks not "where exactly does this read align?" but "which transcripts is this read compatible with?" It answers using k-mer lookups in a transcriptome de Bruijn graph [Bray et al., 2016]. Salmon uses selective alignment and models GC and positional bias [Patro et al., 2017]. Reads compatible with several isoforms are then apportioned by expectation–maximization, which iteratively estimates transcript abundances. This works because quantification does not need base-level alignment.
4c. Normalization: why units matter
Raw counts scale with transcript length and with sequencing depth. Here are the common normalized units.
TPM (transcripts per million), for transcript i:
RPK_i = count_i / (effective length_i in kb)
TPM_i = RPK_i / (Σ_j RPK_j) × 10⁶
TPM sums to 10⁶ in every sample, so it is a relative (compositional) measure.
TPM is fine for within-sample comparison. For between-sample differential expression, depth alone is the wrong normalizer. If one gene is massively upregulated in condition B, it consumes reads, and every other gene appears "down." DESeq2's median-of-ratios size factor is robust to this effect.
Size factor for sample j:
s_j = median_i [ K_ij / ( ∏_v K_iv )^(1/m) ]
(the median over genes of each gene's count divided by
its geometric mean across all m samples)
4d. The negative binomial model
If reads were pure technical samples, counts would be Poisson, with variance equal to the mean. Between biological replicates, though, variance exceeds the mean, because genuine biological variability is added on top. DESeq2 [Love, Huber & Anders, 2014] and edgeR [Robinson, McCarthy & Smyth, 2010] both use the negative binomial distribution.
K_ij ~ NB(mean = μ_ij, dispersion = α_i)
Var(K_ij) = μ_ij + α_i · μ_ij²
───── ─────────────
Poisson biological
(shot) (overdispersion)
GLM: log μ_ij = log s_j + Σ_r x_jr · β_ir
β_ir = log₂ fold change (after conversion) for covariate r
The quadratic term dominates for highly expressed genes. More sequencing depth reduces the Poisson term but never the biological term. Only more biological replicates help with that.
With typically 3 replicates per condition, per-gene dispersion estimates are very noisy. DESeq2 uses empirical Bayes shrinkage: it fits a trend of dispersion against mean across all genes, then pulls each gene's estimate toward the trend. Genes with little information are shrunk the most. This is borrowing strength across 20,000 parallel experiments, and it is the statistical heart of genomics. Log fold changes are similarly shrunk (e.g., with apeglm) so that noisy low-count genes do not produce spurious huge fold changes.
4e. Multiple testing
Testing 20,000 genes at α = 0.05 yields about 1,000 false positives by chance alone. We therefore control the false discovery rate with the Benjamini–Hochberg procedure.
Sort p-values: p₍₁₎ ≤ p₍₂₎ ≤ … ≤ p₍m₎ Find the largest k such that p₍k₎ ≤ (k/m) · q Reject hypotheses 1…k. Expected FDR ≤ q.
5. Single-Cell RNA-seq: Barcodes, UMIs, and the Geometry of Cell States
Bulk RNA-seq reports an average, which is like a smoothie. If a gene goes up in a tissue, the cause could be more expression per cell, or more cells of a type that expresses it. scRNA-seq takes the fruit salad apart.
5a. Molecular barcoding: the core trick
In droplet methods [Macosko et al., 2015; Klein et al., 2015; Zheng et al., 2017], cells and barcoded gel beads are co-encapsulated in oil droplets (10x Genomics calls these GEMs). Each bead carries millions of oligos with three parts:
- a cell barcode, identical on every oligo on that bead (16 nt in 10x v3);
- a UMI, a unique molecular identifier that differs on every oligo (12 nt);
- a poly(dT) capture sequence.
Reverse transcription inside the droplet stamps each cDNA with "which cell" and "which original molecule." The droplets are then pooled, and all cDNAs are sequenced together.
Why UMIs matter. PCR amplifies molecules unevenly. Reads that share a cell barcode, gene and UMI came from one original mRNA, so you count UMIs, not reads. This digitally removes amplification bias. It solves the duplicate problem from concept 1 at the level of individual molecules.
5b. Poisson loading and doublets
Cells per droplet ~ Poisson(λ) P(0) = e^(−λ) P(1) = λ e^(−λ) P(≥2) = 1 − e^(−λ)(1 + λ) Doublet fraction among non-empty droplets ≈ λ/2 for small λ.
Loading is dilute (λ well below 1) to keep doublets rare. Doublets still occur, and they masquerade as "intermediate" cell types. Computational doublet detection is therefore standard.
5c. Sparsity
Each cell yields only a few thousand UMIs, capturing perhaps 10–30% of its mRNAs. Most entries in the gene-by-cell matrix are zero. Many zeros are sampling zeros rather than biological absence, which is often loosely called "dropout." For UMI data, much of this sparsity is consistent with simple count sampling.
5d. The analysis pipeline as linear algebra
- Quality control: remove cells with few genes, or with high mitochondrial fraction, which signals dying cells.
- Normalize and log-transform, or apply a variance-stabilizing method.
- Select highly variable genes.
- PCA. From MATH301, PCA finds the eigenvectors of the gene–gene covariance matrix, or equivalently the SVD of the centered data matrix. The top ~30–50 components capture structured biological variation and discard much of the noise.
- Build a k-nearest-neighbour graph in PC space.
- Cluster with graph community detection (Louvain/Leiden), which maximizes modularity.
- Visualize with UMAP [McInnes, Healy & Melville, 2018], a nonlinear embedding that preserves local neighbourhoods.
- Annotate clusters with marker genes.
5e. Trajectory analysis
Differentiation is continuous, so cells caught at different stages form a path in expression space. Pseudotime methods order cells along this path [Trapnell et al., 2014]. RNA velocity [La Manno et al., 2018] uses a lovely piece of kinetics. Unspliced (intron-containing) reads are newly made transcripts, and spliced reads are mature ones. A gene being induced has excess unspliced RNA relative to steady state. Each cell therefore gets a vector pointing toward its likely future state. Notice the CHEM303 kinetics underneath: du/dt = α − βu and ds/dt = βu − γs.
6. Reading the Regulatory Layer: ChIP-seq and ATAC-seq
6a. ChIP-seq: where does my protein bind?
The procedure:
- Formaldehyde crosslinks proteins to DNA.
- Chromatin is sheared to about 200–500 bp.
- An antibody immunoprecipitates the protein of interest, such as a transcription factor or a histone mark like H3K27ac.
- Crosslinks are reversed and the recovered DNA is sequenced.
Binding sites appear as peaks, local enrichments of reads relative to an input or IgG control.
Peak calling as hypothesis testing. MACS [Zhang et al., 2008] models background read counts in a window as Poisson. Its key feature is a local λ: the maximum of genome-wide and 1-, 5- and 10-kb local background estimates. This accounts for regional biases such as open chromatin and copy number.
λ_local = max(λ_BG, λ_1k, λ_5k, λ_10k) p-value of observing ≥ k reads: P(X ≥ k), X ~ Poisson(λ_local)
Transcription factors give sharp peaks. Broad histone marks such as H3K27me3 require different algorithms.
6b. ATAC-seq: where is chromatin open?
A hyperactive Tn5 transposase, preloaded with sequencing adapters, inserts preferentially into nucleosome-free, accessible DNA. In one step it fragments and tags the genome [Buenrostro et al., 2013]. The method needs only thousands of cells and has been adapted to single cells.
The fragment-length distribution shows nucleosome periodicity: a peak below 100 bp from nucleosome-free regions, then peaks at about 200, 400 and 600 bp from mono-, di- and tri-nucleosomes. That pattern is physical chromatin structure (147 bp wrapped around the octamer, plus linker) read straight out of the sequencing data. Bound transcription factors protect their sites from Tn5, leaving footprints.
6c. Motif enrichment
A transcription factor's specificity is summarized as a position weight matrix (PWM). The score of a candidate site is a log-likelihood ratio of motif against background.
S(x) = Σ_{i=1}^{w} log₂ [ p_i(x_i) / q(x_i) ]
p_i(b) = frequency of base b at motif position i
q(b) = background frequency of base b
This connects to CHEM301. Under simple assumptions, the PWM score is proportional to the negative binding free energy, so the motif is effectively a Boltzmann-weighted energy model [Berg & von Hippel, 1987]. Motif enrichment asks whether a motif occurs more often in peaks than in matched background sequences. If it does, that factor (or a family member) likely drives the binding pattern.
6d. Integration
A peak is not a function. Integration combines accessibility (ATAC), binding (ChIP), and expression change (RNA-seq) after perturbing the factor. It turns "this factor binds near gene X" into the stronger claim "this factor regulates gene X."
7. CRISPR-Cas9: An RNA-Programmable Molecular Machine
From BIOL301 you know CRISPR as bacterial adaptive immunity. Spacers from past phage infections are stored in the CRISPR array, transcribed, and used to guide Cas nucleases to destroy matching invaders. The conceptual leap of [Jinek et al., 2012] was to show that this machine is programmable.
7a. What Jinek et al. (2012) showed
Working with Streptococcus pyogenes Cas9 (SpCas9) in vitro, Jinek, Chylinski, Fonfara, Hauer, Doudna and Charpentier established four results:
- Cas9 is guided by a dual RNA: a crRNA carrying the 20-nt targeting sequence, base-paired to a tracrRNA.
- Cleavage requires a protospacer adjacent motif (PAM), 5′-NGG-3′, immediately 3′ of the target on the non-target strand.
- Two nuclease domains each cut one strand. The HNH domain cleaves the strand complementary to the guide (the target strand). The RuvC-like domain cleaves the non-complementary strand. Mutating either one (D10A in RuvC, H840A in HNH) produces a nickase.
- The two RNAs can be fused into a single-guide RNA (sgRNA) that still directs site-specific cleavage. This made the system a simple two-component tool.
"…the Cas9 endonuclease family can be programmed with single RNA molecules to cleave specific DNA sites, thereby raising the exciting possibility of developing a simple and versatile RNA-directed system to generate dsDNA breaks for genome targeting and editing." [Jinek et al., 2012]
7b. Mechanism as a sequence of checkpoints
- Cas9–sgRNA scans DNA, binding PAMs through PAM-interacting residues. The PAM requirement is also how bacteria avoid cutting their own CRISPR array, since the spacers there lack adjacent PAMs.
- PAM binding destabilizes the adjacent duplex. The guide then invades from the PAM-proximal end and forms an R-loop. The PAM-proximal ~10–12 nt seed is the least mismatch-tolerant region.
- Full R-loop formation triggers a conformational change that docks HNH at the target strand: a conformational checkpoint.
- Both strands are cut about 3 bp upstream of the PAM, usually leaving a blunt double-strand break.
Think of it as a lock with a two-step key. The PAM is the first tumbler, which lets the protein grip quickly without testing every sequence. The guide–target match is the second, which licenses the cut. This is kinetic proofreading in the sense of CHEM503: specificity comes from requiring multiple sequential recognition steps.
7c. Repair decides the outcome
Cas9 only makes the break. The cell's repair machinery writes the edit.
- NHEJ (non-homologous end joining) is active throughout the cell cycle and dominant. It is error-prone and produces small indels. Within an exon, an indel that is not a multiple of 3 causes a frameshift, often followed by nonsense-mediated decay, giving a knockout.
- HDR (homology-directed repair) uses a donor template with homology arms to install precise edits. It is restricted mainly to S/G2 phases and is inefficient in many cell types, especially non-dividing ones such as neurons.
- MMEJ (microhomology-mediated end joining) gives predictable deletions flanked by microhomologies.
7d. Guide design
A good guide is on-target efficient, meaning it has suitable sequence features and GC content and avoids poly-T, which terminates Pol III transcription. It is also off-target specific, with few genomic sites differing by 3 or fewer mismatches, especially in the seed. Rule sets such as Rule Set 2 [Doench et al., 2016] learned these features from large screens.
7e. Delivery
- Plasmid transfection. Simple, but expression is prolonged, which increases off-target cutting.
- Lentivirus. Integrates into the genome; ideal for pooled screens.
- AAV. Packaging limit of about 4.7 kb, which barely fits SpCas9 (about 4.1 kb of coding sequence). This motivated smaller orthologs such as SaCas9.
- RNP electroporation. Pre-assembled protein plus guide. Activity is transient and specificity high.
- Lipid nanoparticles carrying mRNA plus guide, the same platform as the mRNA vaccines covered in BIOL401.
8. CRISPR Applications: From Scissors to Pencils to Dimmer Switches
Once Cas9 is understood as a programmable DNA-binding module with removable nuclease activity, the applications follow from one principle: fuse different effectors to catalytically impaired Cas9.
8a. Pooled knockout screens
The workflow:
- A lentiviral library of sgRNAs targeting every gene is transduced at low multiplicity of infection (MOI ~0.3), so most cells receive one guide.
- A selective pressure is applied: a drug, proliferation, or FACS sorting.
- Guide abundance before and after selection is counted by sequencing.
The guide sequence is itself the barcode, which is concept 1 again. Example libraries are GeCKO v2 [Shalem et al., 2014; Sanjana et al., 2014] and Brunello, with about 77,000 sgRNAs at 4 per gene, optimized with Rule Set 2 [Doench et al., 2016]. Analysis tools such as MAGeCK [Li et al., 2014] use negative-binomial count models, the same statistics as RNA-seq. Multiple guides per gene provide internal replication against guide-specific artifacts.
8b. Base editors: chemistry without breaks
- Cytosine base editors (CBE) [Komor et al., 2016] fuse Cas9 nickase (D10A) to a cytidine deaminase (rat APOBEC1) and a uracil glycosylase inhibitor (UGI). The deaminase acts on the single-stranded non-target strand exposed in the R-loop, converting C→U within an editing window (about positions 4–8 of the protospacer). UGI blocks base-excision repair from reverting the U. The nick on the opposite strand biases mismatch repair to copy the edited strand. Net result: C•G → T•A.
- Adenine base editors (ABE) [Gaudelli et al., 2017] use a directed-evolution variant of E. coli TadA, a tRNA adenosine deaminase, to act on DNA. Adenine is deaminated to inosine, which base-pairs like G. Net result: A•T → G•C. No natural DNA adenine deaminase was known, so one had to be evolved, which ties back to CHEM503 protein engineering.
8c. Prime editing: a search-and-replace word processor
Prime editing [Anzalone et al., 2019] fuses Cas9 nickase (H840A) to an engineered reverse transcriptase. The pegRNA carries both the spacer and a 3′ extension containing a primer-binding site plus the desired edit. The steps are:
- Cas9 nicks the non-target strand.
- The freed 3′ flap anneals to the primer-binding site.
- The reverse transcriptase copies the edit template directly into genomic DNA.
- Flap equilibration and repair incorporate the new sequence.
Prime editing can install all 12 point mutations plus small insertions and deletions, without double-strand breaks or donor DNA.
8d. CRISPRi and CRISPRa: dimmer switches
Catalytically dead Cas9 (dCas9, D10A + H840A) is a pure DNA-binding protein.
- CRISPRi [Qi et al., 2013; Gilbert et al., 2013]: dCas9 alone sterically blocks transcription in bacteria. In mammalian cells, dCas9-KRAB recruits heterochromatin machinery to repress promoters.
- CRISPRa: dCas9 fused to activators (VP64, VPR, or the SAM system) upregulates endogenous genes.
Both are reversible and titratable. They also avoid DNA-damage toxicity and can target lncRNA promoters, where small indels would not disrupt function.
9. Proteomics by Mass Spectrometry: Identifying What Is Actually There
Transcripts are only intermediates; proteins do the work. mRNA–protein correlation across genes in mammalian cells is modest (Spearman ρ ≈ 0.4–0.6), because translation rates and protein half-lives vary enormously [Schwanhäusser et al., 2011]. Proteomics measures the functional layer directly. You know the instrumentation from CHEM501, and now we use it at scale.
9a. Bottom-up ("shotgun") workflow
- Lyse cells, then reduce disulfides (DTT/TCEP) and alkylate cysteines with iodoacetamide, adding +57.021 Da carbamidomethyl. This prevents disulfide scrambling.
- Digest with trypsin, which cleaves C-terminal to Lys and Arg, but not before Pro. This places a basic residue at each peptide's C-terminus, so peptides carry ≥ 2+ charge after ESI (the N-terminal amine supplies the other proton). That charge produces informative, easily interpreted fragment series.
- Separate the peptides by reversed-phase nano-LC, then ionize them by electrospray.
- MS1 measures precursor m/z. MS2 fragments selected precursors (by CID/HCD) and measures fragment m/z.
- Match the spectra against in silico digests of a protein database.
The method is called "bottom-up" because peptides are measured and proteins are inferred from them. This introduces a protein inference problem: shared peptides may come from several isoforms.
9b. Peptide fragmentation
Peptide backbone cleavage at amide bonds (CID/HCD):
N-terminal fragments: b-ions C-terminal fragments: y-ions
For a peptide of n residues with residue masses r_k:
b_i (1+) = Σ_{k=1}^{i} r_k + 1.00728 (proton)
y_j (1+) = Σ_{k=n−j+1}^{n} r_k + 18.01056 + 1.00728 (water + proton)
Adjacent b-ions differ by exactly one residue mass,
so the sequence can be read as a mass ladder.
9c. Controlling false identifications
The target–decoy strategy [Elias & Gygi, 2007] searches spectra against the real database plus a reversed or shuffled "decoy" database. Decoy hits estimate how many target hits are random. This gives an empirical FDR, typically set at 1% at the peptide-spectrum-match and protein levels.
9d. Data-dependent vs. data-independent acquisition
- DDA. After each MS1 scan, the instrument picks the top-N most intense precursors for MS2, then dynamically excludes them. Selection is stochastic and biased toward abundant peptides, so a peptide identified in one run may be missed in the next. The result is the infamous missing-value problem.
- DIA / SWATH-MS [Gillet et al., 2012]. The instrument cycles systematically through wide m/z isolation windows (e.g., 25 Th) and fragments everything in each one. Every detectable peptide is fragmented in every run, so the data are complete and reproducible. The cost is that MS2 spectra become chimeric mixtures, which must be deconvolved using spectral libraries or library-free deep-learning predictions.
An analogy: DDA is a reporter interviewing the loudest people at a party, one at a time. DIA records every conversation in each room and untangles them later.
10. Quantitative Proteomics, Interactomes, and Proximity Labeling
10a. Quantitative strategies
- SILAC [Ong et al., 2002]. Cells are grown in media containing "heavy" Arg (¹³C₆¹⁵N₄, +10.008 Da) and Lys (¹³C₆¹⁵N₂, +8.014 Da). After trypsin digestion, every peptide except the protein's C-terminal peptide carries exactly one labeled residue, assuming no missed cleavages. Heavy and light samples are mixed before processing, so all downstream variability cancels, and peptide pairs appear as doublets in MS1. This is the most accurate method, but it requires metabolically active, culturable cells.
- Isobaric tags (iTRAQ, TMT) [Ross et al., 2004; Thompson et al., 2003]. Amine-reactive NHS-ester tags label peptide N-termini and Lys side chains. Each tag variant has the same total mass, because a reporter group and a balancer group trade isotopes. Labeled peptides from different samples therefore co-elute and co-isolate as a single MS1 peak. On fragmentation, the reporters separate and give distinct low-mass ions, one per sample. TMTpro allows 16–18 samples to be multiplexed. The weakness is ratio compression: co-isolated contaminating peptides contribute reporter ions and pull ratios toward 1:1. Mitigations include MS3 (SPS-MS3) and narrower isolation.
- Label-free. Quantify by MS1 intensity (e.g., MaxLFQ [Cox et al., 2014]) or by spectral counting. This is cheap and unlimited in sample number, but more sensitive to run-to-run variation. It pairs naturally with DIA.
10b. Post-translational modifications
- Phosphoproteomics. Phosphopeptides are substoichiometric and ionize poorly, so they must be enriched. TiO₂ and IMAC (Fe³⁺, Ti⁴⁺) exploit Lewis acid–base coordination of phosphate to metal oxides or ions, which is CHEM502 coordination chemistry at work. Acidic loading buffers with competitors such as glycolic acid or DHB suppress binding of carboxylates (Asp/Glu). Site localization (which Ser/Thr/Tyr carries the phosphate) needs site-determining fragment ions and is scored probabilistically.
- Glycoproteomics. Glycans are heterogeneous and branched, and they fragment easily. Common approaches include HILIC or lectin enrichment, PNGase F release of N-glycans (which converts the formerly glycosylated Asn to Asp, a +0.984 Da "deamidation" signature at N-X-S/T sequons), and EThcD fragmentation to keep intact glycopeptide information.
10c. Affinity purification–MS (AP-MS)
Immunoprecipitate a tagged or endogenous bait and identify its co-purifying prey. GFP-Trap uses a camelid nanobody (VHH) against GFP, which binds tightly enough for stringent washes. (Nanobodies were covered in BIOL401.)
The central difficulty is specificity. Many proteins stick to beads or are simply abundant. Scoring frameworks such as SAINT compare bait pulldowns to negative controls, and contaminant repositories such as the CRAPome [Mellacheruvu et al., 2013] catalogue frequent background proteins. AP-MS favours stable complexes that survive lysis and washing. Weak or transient interactions are lost.
10d. Proximity labeling: tagging the neighbourhood in living cells
Proximity labeling reverses the AP-MS logic. Instead of preserving interactions through lysis, you covalently mark neighbours while the cell is still alive. Biotin then survives harsh, denaturing streptavidin purification.
- BioID [Roux et al., 2012]. A promiscuous BirA mutant (R118G, "BirA*") releases reactive biotinoyl-5′-AMP, which diffuses briefly and acylates nearby lysine ε-amines, with an estimated radius of about 10 nm. Labeling is slow, around 18–24 h, which integrates over time.
- TurboID [Branon et al., 2018]. Directed evolution of BirA produced a ligase fast enough to label in about 10 minutes.
- APEX2 [Lam et al., 2015]. An engineered ascorbate peroxidase that, with H₂O₂, oxidizes biotin-phenol to short-lived phenoxyl radicals (sub-millisecond lifetime). These radicals label electron-rich side chains, principally tyrosine, within about 1 minute. The temporal resolution is excellent, but H₂O₂ is toxic. APEX2 also works for EM contrast.
Proximity labeling enables spatial proteomics: mapping the proteomes of organelles and subcompartments such as the mitochondrial matrix, synaptic cleft or stress granules, by targeting the enzyme there. A labeled protein is a neighbour, not necessarily a binding partner.
11. Structure from Sequence and from Mass: AlphaFold2, XL-MS, and HDX-MS
11a. The physical intuition behind AlphaFold2: coevolution
Suppose residues i and j touch in the folded structure. If evolution mutates i, the pressure to keep the contact favours a compensating mutation at j. In a multiple sequence alignment (MSA) of homologs, contacting residues therefore show correlated variation. Direct-coupling analysis made this quantitative by fitting a Potts model, which is literally an inverse statistical-mechanics problem, to separate direct couplings from indirect correlations [Morcos et al., 2011]. Connect this to CHEM301: the Potts model is a generalized Ising model.
11b. The AlphaFold2 architecture
[Jumper et al., 2021] built an end-to-end neural network with three main parts.
- Inputs. An MSA (searched from large sequence databases) and, optionally, structural templates.
- Evoformer. Forty-eight blocks that jointly update two representations: an MSA representation (sequences × residues) and a pair representation (residues × residues). Information flows both ways. The MSA informs pairwise relationships, and the pair representation biases attention within the MSA.
- Structure module. Each residue is treated as a rigid body (a "residue gas" of frames, each a rotation plus a translation). Invariant point attention refines these frames in a way that is invariant to global rotation and translation, then side-chain torsions are predicted. Training uses a frame-aligned point error (FAPE) loss, among others.
Two further details matter:
- Triangle updates. In the pair representation, the edge i–j is updated using edges i–k and k–j for every third residue k. If i is close to k and k is close to j, then i cannot be arbitrarily far from j. The network is built to reason about geometric consistency, much like the triangle inequality.
- Recycling. Outputs are fed back as inputs several times, refining the prediction iteratively.
At CASP14, AlphaFold2 achieved a median backbone accuracy of 0.96 Å r.m.s.d.₉₅. This is comparable to experimental structures in many cases, and far better than the next-best method [Jumper et al., 2021].
11c. Confidence scores: read them every time
- pLDDT (0–100), per residue: predicted local accuracy. Values above 90 indicate high confidence. Values below about 50 often indicate intrinsic disorder, and pLDDT is in fact a decent disorder predictor.
- PAE (predicted aligned error), per residue pair: the expected position error of residue j when the structure is aligned on residue i. PAE tells you whether the relative placement of two domains is trustworthy, even when each domain alone has high pLDDT.
11d. Limitations
- The prediction is a single static structure, not a conformational ensemble or dynamics.
- Accuracy degrades with shallow MSAs, as for orphan or de novo proteins.
- AlphaFold2 does not natively model ligands, metals, PTMs or nucleic acids. AlphaFold3 [Abramson et al., 2024] extends to many of these.
- It is not reliable for predicting the structural effects of point mutations. It learned evolutionary families, not folding physics.
- Complexes need AlphaFold-Multimer or later models, and interface confidence (ipTM) must be checked.
11e. XL-MS: distance restraints from chemistry
Bifunctional crosslinkers such as DSS or BS3 (NHS esters, spacer about 11.4 Å) covalently link Lys ε-amines (and protein N-termini) that are within reach. Accounting for Lys side-chain length, crosslinked Cα–Cα distances are typically below about 30 Å. Identifying crosslinked peptide pairs by MS gives distance restraints that validate AlphaFold models or map complex topology in situ. MS-cleavable linkers such as DSSO simplify identification.
11f. HDX-MS: dynamics from exchange kinetics
Backbone amide hydrogens exchange with solvent deuterium. Exchange is slowed when the amide is hydrogen-bonded (in secondary structure) or buried. Under typical EX2 conditions, where refolding is faster than chemical exchange:
Closed (protected) ⇌ Open (exchange-competent) K_op = k_op / k_cl Open: N–H → N–D intrinsic rate k_int EX2 regime (k_cl ≫ k_int): k_obs ≈ K_op · k_int Protection factor: P = k_int / k_obs ≈ 1 / K_op ΔG_op = −RT ln K_op = RT ln P
This gives local stability free energies residue by residue. That is CHEM301 thermodynamics measured with CHEM501 instrumentation, applied to the protein-dynamics questions of CHEM503. Comparing deuterium uptake with and without a ligand or binding partner maps interfaces and allosteric changes. Proline has no amide hydrogen and is invisible to HDX. Exchange is quenched at pH about 2.5 and 0 °C, followed by pepsin digestion and rapid LC-MS.
12. Networks and Systems: When the Wiring Diagram Computes
12a. PPI networks and topology
Proteins are nodes and interactions are edges. Degree k is the number of partners. Many biological networks have heavy-tailed degree distributions, often described as approximately scale-free [Barabási & Albert, 1999].
P(k) ~ k^(−γ), typically 2 < γ < 3 Contrast with a random (Erdős–Rényi) network: P(k) ~ Poisson, with a characteristic scale and no hubs.
Most proteins have few partners, and a few hubs have very many. In yeast, highly connected proteins were found to be more often essential (the "centrality–lethality" rule) [Jeong et al., 2001]. Scale-free networks are robust to random failure but fragile to targeted hub attack.
Two refinements matter:
- The strength and universality of scale-free claims have been debated. Observed degree distributions are also shaped by sampling and study bias, since well-studied proteins accumulate reported interactions.
- Hubs come in two kinds. "Party" hubs interact with partners simultaneously, as in a complex. "Date" hubs interact sequentially [Han et al., 2004]. This distinction requires adding expression or timing data to the static map.
12b. STRING and evidence integration
STRING [Szklarczyk et al., 2019] integrates experimental interactions, curated pathways, co-expression, text mining, genomic context (gene neighbourhood, fusion, co-occurrence) and orthology transfer. Each channel gets a score, and the channels are combined under an assumption of independence.
S_combined = 1 − ∏_channels (1 − S_i) (applied after correcting each channel for a prior of random association)
A STRING edge means "functionally associated." It does not necessarily mean "physically binds."
12c. Functional enrichment
Given a gene list, such as differentially expressed genes or a screen hit list, ask whether a GO term or KEGG pathway is overrepresented. The usual test is hypergeometric (one-sided Fisher's exact test).
N = genes in background, K = background genes annotated to the term
n = genes in your list, k = list genes annotated to the term
P(X ≥ k) = Σ_{x=k}^{min(n,K)} [ C(K,x) · C(N−K, n−x) ] / C(N, n)
The choice of background is critical. For RNA-seq, the background should be the genes expressed and testable in your experiment, not the whole genome. Otherwise you will "discover" tissue identity. GO terms are hierarchical and overlapping, so multiple-testing correction and redundancy reduction are both needed.
12d. Kinetic ODE models of gene circuits
Gene regulation can be written with mass-action and Hill kinetics. For a gene repressed by protein R:
dm/dt = α / (1 + (R/K)^n) − γ_m · m (mRNA) dp/dt = β · m − γ_p · p (protein) α = max transcription rate, K = repression threshold, n = Hill coefficient (cooperativity), γ = degradation/dilution rates
Two classic synthetic circuits from the same issue of Nature showed that design principles yield predictable behaviours.
- Toggle switch [Gardner, Cantor & Collins, 2000]. Two mutually repressing genes. With sufficient cooperativity (n > 1), the system is bistable: two stable steady states separated by an unstable one. It is a memory element.
- Repressilator [Elowitz & Leibler, 2000]. Three repressors in a cycle (A ⊣ B ⊣ C ⊣ A). With strong enough repression and matched degradation rates, the steady state becomes unstable and the system oscillates.
Stability analysis comes from MATH301 and MATH102. Linearize at a steady state and compute the eigenvalues of the Jacobian. If all real parts are negative, the state is stable. Complex eigenvalues with positive real parts mean spiralling away toward oscillation. A circuit's qualitative behaviour is a property of its topology plus parameters, not of any single component.
Ultrasensitivity in signaling. In the MAPK cascade (covered in CHEM402), multi-step, distributive dual phosphorylation and cascade architecture can convert graded inputs into switch-like outputs, with effective Hill coefficients well above 1 [Huang & Ferrell, 1996]. The cell builds digital decisions out of analog chemistry.
12e. Boolean networks
When kinetic parameters are unknown, which is usually the case, abstract each gene as ON or OFF. Update rules are logic functions, for example A(t+1) = B(t) AND NOT C(t). The state space is finite (2ᴺ states), so trajectories must end in attractors: fixed points or cycles. Kauffman proposed that attractors correspond to cell types [Kauffman, 1969]. This idea connects to Waddington's landscape from BIOL402. Boolean models trade quantitative precision for robustness to parameter ignorance.
13. Data Science and Reproducibility: The Pipeline Is the Experiment
In high-throughput biology, the analysis is part of the experiment. A different aligner, normalization or parameter choice can change the conclusions. Four practices address this.
- Languages and ecosystems. R/Bioconductor offers rigorously versioned packages (DESeq2, edgeR, limma, GenomicRanges) and standardized data objects such as SummarizedExperiment and SingleCellExperiment. Python offers the scverse (Scanpy, AnnData), Biopython and pysam.
- Workflow managers. Snakemake [Mölder et al., 2021] and Nextflow [Di Tommaso et al., 2017] define pipelines as dependency graphs of rules. A rule reruns only when its inputs change, and the same pipeline scales from laptop to cluster to cloud. Community pipelines such as nf-core standardize best practices.
- Environments. Conda, Docker and Singularity/Apptainer freeze software versions.
- Provenance. Version control (git), recorded parameters and random seeds, and deposition of raw data (GEO/SRA, PRIDE).
This is CHEM302's laboratory notebook rewritten for computation. If someone else cannot regenerate your figure from your raw data, the result is not finished.
Connections to Peptide Chemistry
Much of this course runs on peptide chemistry, often without saying so. Seeing this link turns black-box methods into chemistry you can reason about.
- Trypsin specificity shapes the proteome you can see. Trypsin's S1 pocket holds an Asp (Asp189 in chymotrypsin numbering) that forms a salt bridge with Lys or Arg, which explains its specificity (CHEM401 serine proteases). Proteins with few K/R produce peptides too long to detect. K/R-rich regions produce peptides too short. Membrane proteins and highly basic histones are often digested with other enzymes: Lys-C, Glu-C, chymotrypsin, or Arg-C for histones, often after propionylation to block Lys. Each protease defines a different window onto the same proteome.
- Fragmentation is amide-bond chemistry. b/y ions come from protonation of backbone amides, the "mobile proton" model, followed by nucleophilic attack that forms oxazolone b-ions. Proline-rich and Asp-containing peptides fragment unusually: cleavage is enhanced N-terminal to Pro and C-terminal to Asp. ETD/ECD cleaves N–Cα bonds instead, producing c/z ions and preserving labile PTMs such as phosphates and glycans.
- Every quantitative tag is a bioconjugation reaction. TMT, iTRAQ, DSS/BS3 crosslinkers and many biotinylation reagents are NHS esters that acylate primary amines: the α-amine and Lys ε-amines. Labeling efficiency depends on pH, because the amine must be unprotonated. The pKa of α-amines (~8) versus ε-amines (~10.5) lets you tune selectivity. This is CHEM202 acyl substitution applied to whole proteomes.
- Isotope labels are mass arithmetic. SILAC's heavy Lys (+8.014 Da) and Arg (+10.008 Da) give predictable mass shifts precisely because trypsin cleaves after these residues. Biological labeling and enzymatic specificity were designed to fit together.
- Proximity labeling chemistry targets specific residues. BioID and TurboID acylate Lys. APEX2 phenoxyl radicals attack Tyr. A protein with no surface-exposed lysines or tyrosines near the bait can be a false negative. Labeling also modifies the peptides you later measure, which must be accounted for in database searching.
- HDX reports on the peptide backbone itself. Each non-proline residue contributes one amide N–H. Intrinsic exchange rates depend on neighbouring side chains, pH and temperature, which are tabulated as random-coil reference rates. Measuring deviation from peptide random-coil behaviour is how HDX reads structure.
- AlphaFold and peptide–protein complexes. AlphaFold-Multimer and related approaches can model short peptides bound to protein domains, such as SH3/PDZ ligands or MHC–peptide complexes from BIOL401. This supports peptide design, epitope mapping and the development of peptide therapeutics such as cyclic peptides and stapled helices. Confidence is lower for flexible peptides with shallow coevolutionary signal. Peptides have few homologs to align.
- Genome editing tools are engineered proteins. Base editors, prime editors and TurboID all came from directed evolution and rational fusion design (CHEM503). Linker composition between domains (flexible Gly-Ser repeats vs. rigid helical EAAAK repeats) is peptide engineering, and it measurably affects editing windows and efficiency.
- Phosphopeptide chemistry. TiO₂ and IMAC enrichment depend on phosphate pKa values and on competition with Asp/Glu carboxylates. Neutral loss of H₃PO₄ (−97.977 Da) from pSer/pThr during CID is both a diagnostic and a nuisance. pTyr is more stable and does not show this loss as readily.
- Where this leads. Peptide-centric proteomics is moving toward top-down analysis of intact proteoforms and toward single-molecule protein sequencing using nanopores, a direct analogue of DNA nanopores, or fluorosequencing. Deep-learning models now predict peptide retention times and fragment intensities from sequence. The peptide is becoming as computationally legible as the nucleotide.
Common Misconceptions to Watch Out For
- "Higher N50 means a better assembly." N50 measures contiguity only. Over-aggressive joining of repeats inflates N50 while introducing misassemblies. Always pair it with correctness metrics such as BUSCO, QV and misassembly counts.
- "More sequencing depth fixes everything." Depth reduces Poisson (technical) noise. Biological variance, the αμ² term, is reduced only by more biological replicates. Six replicates at 20M reads usually beat three replicates at 60M reads for differential expression.
- "TPM/RPKM values can be compared directly across samples for DE." These are compositional. A few highly expressed genes can shift all other values. Use count-based models with robust size factors (DESeq2/edgeR) for between-sample inference.
- "A smaller p-value means a bigger biological effect." Significance mixes effect size with precision. A tiny fold change in a highly expressed gene can be extremely significant and biologically trivial. Report shrunken log fold changes alongside adjusted p-values.
- "Distances and cluster sizes on a UMAP are meaningful." UMAP preserves local neighbourhoods. Global distances between clusters, cluster areas and apparent "gaps" can be artifacts. Do quantitative analysis in PCA space or on the kNN graph, not on 2-D coordinates.
- "Clusters are cell types." Clusters depend on resolution parameters. They may reflect cell cycle stage, stress, doublets or batch rather than identity. Validate with markers and orthogonal data.
- "mRNA level ≈ protein level." Across genes, correlation is only moderate, because translation and degradation rates vary widely [Schwanhäusser et al., 2011]. Transcriptomics does not replace proteomics.
- "Not detected by MS means not present." Detection depends on abundance, ionization efficiency, peptide length after digestion, and DDA stochastic sampling. Absence of evidence is weak evidence of absence, especially for low-abundance proteins.
- "A ChIP-seq peak means the factor regulates the nearest gene." Many binding events are non-functional. Enhancers can act over hundreds of kilobases and skip the nearest gene. Functional claims require perturbation plus expression data.
- "Cas9 cuts only its target," or "HDR is the default repair." Off-target cleavage occurs at sites with seed-tolerant mismatches. NHEJ, not HDR, dominates repair, and HDR is inefficient in non-dividing cells. Even "precise" editing can produce large deletions or rearrangements at the cut site.
- "Base editors and prime editors are perfect." Base editors cause bystander edits within the activity window and can cause guide-independent deamination. Prime editing efficiency varies strongly by locus and pegRNA design.
- "AlphaFold solved protein folding, so predicted structures are ground truth." AlphaFold predicts a plausible static structure from evolutionary information. It does not simulate the folding pathway, conformational ensembles, or mutation effects. High pLDDT does not guarantee that the predicted state is the functional one, and confident domains can still be misplaced relative to each other (check PAE). Low pLDDT regions are often disordered, not merely "unsolved."
- "Proximity-labeled proteins are direct binding partners." BioID, TurboID and APEX2 label neighbours within roughly 10–20 nm, including bystanders in the same compartment. Spatial controls and ratiometric comparisons are essential.
- "Hubs are always essential," and "network edges are physical interactions." The centrality–lethality correlation is statistical, not a law, and is partly confounded by study bias. STRING edges include co-expression and text mining, so they indicate association, not necessarily physical contact.
- "GO enrichment against the whole genome is fine." Using the wrong background yields spurious enrichment for tissue-specific functions. Use the set of genes you could actually have detected.
- "A model that fits the data is the mechanism." Kinetic and Boolean models are hypotheses. Several topologies can fit the same time course. Models earn trust by predicting the outcome of new perturbations.
Closing Synthesis
Step back and the course has a single shape. Biology is encoded in molecules. Modern technology converts those molecules into discrete digital counts: reads, UMIs, spectra, guide abundances. Statistics and algorithms turn the counts into inferences, and systems thinking turns the inferences into mechanism. CRISPR closes the loop by letting us write as well as read, so that network predictions can be tested by targeted perturbation.
The "cool" tools are the obvious surprise. The deeper ones are mathematical:
- A 1736 puzzle about bridges made genome assembly tractable.
- The Potts models of statistical physics underlie modern protein structure prediction.
- A quadratic variance term explains why biological replicates matter more than depth.
Keep asking two questions of every method: What physical event generates the signal? and What is my model of the noise? If you can answer both, you understand the method.
Key References Cited
- Abramson, J. et al. (2024). Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493–500.
- Alberts, B., Heald, R., Johnson, A., Morgan, D., Raff, M., Roberts, K., & Walter, P. (2022). Molecular Biology of the Cell, 7th ed. W.W. Norton.
- Anzalone, A.V. et al. (2019). Search-and-replace genome editing without double-strand breaks or donor DNA. Nature 576, 149–157.
- Barabási, A.-L. & Albert, R. (1999). Emergence of scaling in random networks. Science 286, 509–512.
- Berg, O.G. & von Hippel, P.H. (1987). Selection of DNA binding sites by regulatory proteins. J. Mol. Biol. 193, 723–750.
- Branon, T.C. et al. (2018). Efficient proximity labeling in living cells and organisms with TurboID. Nat. Biotechnol. 36, 880–887.
- Bray, N.L., Pimentel, H., Melsted, P. & Pachter, L. (2016). Near-optimal probabilistic RNA-seq quantification. Nat. Biotechnol. 34, 525–527.
- Buenrostro, J.D. et al. (2013). Transposition of native chromatin for fast and sensitive epigenomic profiling. Nat. Methods 10, 1213–1218.
- Compeau, P.E.C., Pevzner, P.A. & Tesler, G. (2011). How to apply de Bruijn graphs to genome assembly. Nat. Biotechnol. 29, 987–991.
- Compeau, P. & Pevzner, P. (2018). Bioinformatics Algorithms: An Active Learning Approach, 3rd ed. Active Learning Publishers.
- Cox, J. et al. (2014). Accurate proteome-wide label-free quantification (MaxLFQ). Mol. Cell. Proteomics 13, 2513–2526.
- DePristo, M.A. et al. (2011). A framework for variation discovery and genotyping using next-generation DNA sequencing data. Nat. Genet. 43, 491–498.
- Di Tommaso, P. et al. (2017). Nextflow enables reproducible computational workflows. Nat. Biotechnol. 35, 316–319.
- Dobin, A. et al. (2013). STAR: ultrafast universal RNA-seq aligner. Bioinformatics 29, 15–21.
- Doench, J.G. et al. (2016). Optimized sgRNA design to maximize activity and minimize off-target effects of CRISPR-Cas9. Nat. Biotechnol. 34, 184–191.
- Elias, J.E. & Gygi, S.P. (2007). Target-decoy search strategy for increased confidence in large-scale protein identifications. Nat. Methods 4, 207–214.
- Elowitz, M.B. & Leibler, S. (2000). A synthetic oscillatory network of transcriptional regulators. Nature 403, 335–338.
- Gardner, T.S., Cantor, C.R. & Collins, J.J. (2000). Construction of a genetic toggle switch in Escherichia coli. Nature 403, 339–342.
- Gaudelli, N.M. et al. (2017). Programmable base editing of A•T to G•C in genomic DNA without DNA cleavage. Nature 551, 464–471.
- Gilbert, L.A. et al. (2013). CRISPR-mediated modular RNA-guided regulation of transcription in eukaryotes. Cell 154, 442–451.
- Gillet, L.C. et al. (2012). Targeted data extraction of the MS/MS spectra generated by data-independent acquisition. Mol. Cell. Proteomics 11, O111.016717.
- Han, J.-D.J. et al. (2004). Evidence for dynamically organized modularity in the yeast protein–protein interaction network. Nature 430, 88–93.
- Huang, C.-Y.F. & Ferrell, J.E. (1996). Ultrasensitivity in the mitogen-activated protein kinase cascade. PNAS 93, 10078–10083.
- Jeong, H., Mason, S.P., Barabási, A.-L. & Oltvai, Z.N. (2001). Lethality and centrality in protein networks. Nature 411, 41–42.
- Jinek, M., Chylinski, K., Fonfara, I., Hauer, M., Doudna, J.A. & Charpentier, E. (2012). A programmable dual-RNA–guided DNA endonuclease in adaptive bacterial immunity. Science 337, 816–821. DOI: 10.1126/science.1225829
- Jumper, J. et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589. DOI: 10.1038/s41586-021-03819-2
- Kauffman, S.A. (1969). Metabolic stability and epigenesis in randomly constructed genetic nets. J. Theor. Biol. 22, 437–467.
- Kim, D., Langmead, B. & Salzberg, S.L. (2015). HISAT: a fast spliced aligner with low memory requirements. Nat. Methods 12, 357–360.
- Kim, D. et al. (2019). Graph-based genome alignment and genotyping with HISAT2 and HISAT-genotype. Nat. Biotechnol. 37, 907–915.
- Klein, A.M. et al. (2015). Droplet barcoding for single-cell transcriptomics applied to embryonic stem cells. Cell 161, 1187–1201.
- Komor, A.C. et al. (2016). Programmable editing of a target base in genomic DNA without double-stranded DNA cleavage. Nature 533, 420–424.
- La Manno, G. et al. (2018). RNA velocity of single cells. Nature 560, 494–498.
- Lam, S.S. et al. (2015). Directed evolution of APEX2 for electron microscopy and proximity labeling. Nat. Methods 12, 51–54.
- Lander, E.S. & Waterman, M.S. (1988). Genomic mapping by fingerprinting random clones: a mathematical analysis. Genomics 2, 231–239.
- Li, W. et al. (2014). MAGeCK enables robust identification of essential genes from genome-scale CRISPR/Cas9 knockout screens. Genome Biol. 15, 554.
- Liao, Y., Smyth, G.K. & Shi, W. (2014). featureCounts: an efficient general purpose program for assigning sequence reads to genomic features. Bioinformatics 30, 923–930.
- Love, M.I., Huber, W. & Anders, S. (2014). Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol. 15, 550.
- Macosko, E.Z. et al. (2015). Highly parallel genome-wide expression profiling of individual cells using nanoliter droplets. Cell 161, 1202–1214.
- McInnes, L., Healy, J. & Melville, J. (2018). UMAP: Uniform Manifold Approximation and Projection for dimension reduction. arXiv:1802.03426.
- McKenna, A. et al. (2010). The Genome Analysis Toolkit. Genome Res. 20, 1297–1303.
- Mellacheruvu, D. et al. (2013). The CRAPome: a contaminant repository for affinity purification–mass spectrometry data. Nat. Methods 10, 730–736.
- Mölder, F. et al. (2021). Sustainable data analysis with Snakemake. F1000Research 10, 33.
- Morcos, F. et al. (2011). Direct-coupling analysis of residue coevolution captures native contacts across many protein families. PNAS 108, E1293–E1301.
- Nurk, S. et al. (2022). The complete sequence of a human genome. Science 376, 44–53.
- Ong, S.-E. et al. (2002). Stable isotope labeling by amino acids in cell culture, SILAC. Mol. Cell. Proteomics 1, 376–386.
- Patro, R. et al. (2017). Salmon provides fast and bias-aware quantification of transcript expression. Nat. Methods 14, 417–419.
- Qi, L.S. et al. (2013). Repurposing CRISPR as an RNA-guided platform for sequence-specific control of gene expression. Cell 152, 1173–1183.
- Robinson, M.D., McCarthy, D.J. & Smyth, G.K. (2010). edgeR: a Bioconductor package for differential expression analysis of digital gene expression data. Bioinformatics 26, 139–140.
- Ross, P.L. et al. (2004). Multiplexed protein quantitation in Saccharomyces cerevisiae using amine-reactive isobaric tagging reagents. Mol. Cell. Proteomics 3, 1154–1169.
- Roux, K.J. et al. (2012). A promiscuous biotin ligase fusion protein identifies proximal and interacting proteins in mammalian cells. J. Cell Biol. 196, 801–810.
- Sanjana, N.E., Shalem, O. & Zhang, F. (2014). Improved vectors and genome-wide libraries for CRISPR screening. Nat. Methods 11, 783–784.
- Schwanhäusser, B. et al. (2011). Global quantification of mammalian gene expression control. Nature 473, 337–342.
- Shalem, O. et al. (2014). Genome-scale CRISPR-Cas9 knockout screening in human cells. Science 343, 84–87.
- Szklarczyk, D. et al. (2019). STRING v11: protein–protein association networks with increased coverage. Nucleic Acids Res. 47, D607–D613.
- Thompson, A. et al. (2003). Tandem mass tags: a novel quantification strategy for comparative analysis of complex protein mixtures by MS/MS. Anal. Chem. 75, 1895–1904.
- Trapnell, C. et al. (2014). The dynamics and regulators of cell fate decisions are revealed by pseudotemporal ordering of single cells. Nat. Biotechnol. 32, 381–386.
- Wenger, A.M. et al. (2019). Accurate circular consensus long-read sequencing improves variant detection and assembly of a human genome. Nat. Biotechnol. 37, 1155–1162.
- Zhang, Y. et al. (2008). Model-based analysis of ChIP-Seq (MACS). Genome Biol. 9, R137.
- Zheng, G.X.Y. et al. (2017). Massively parallel digital transcriptional profiling of single cells. Nat. Commun. 8, 14049.
⚙️ Mechanisms & Applications
Mechanisms & Applications: From Molecular Events to Biological Explanations
Skipper, the central challenge of this course is not learning how to generate more data. It is learning how a molecular event becomes a measurement—and what that measurement can legitimately tell us. A sequencing read is not a genome; an RNA count is not a transcription rate; a mass-spectrometry signal is not automatically a protein concentration; a network edge is not necessarily a direct physical interaction.
Your earlier courses provide the foundation: molecular biology explains what can happen, physical chemistry explains why and how fast, analytical chemistry explains what an instrument detects, and statistics explains how confidently we can distinguish biology from noise. The organizing question throughout these notes is: What chain of transformations connects the biological quantity I care about to the number I actually observe? [1,2]
1. Next-Generation Sequencing: Turning Molecular Identity into Signals
Library preparation is the first measurement model. DNA or RNA must be converted into molecules an instrument can recognize. Each transformation can change which molecules survive into the final dataset.
- Obtain nucleic acid: extraction and quality assessment establish whether the material is intact, contaminated, or degraded.
- Create an appropriate size distribution: fragmentation is common for short-read sequencing; preserving long molecules is crucial for long-read applications.
- Add adapters and sample indexes: adapters provide platform-specific recognition sequences; indexes identify samples pooled together.
- Amplify when necessary: PCR increases material but can introduce duplicates and sequence-dependent bias. PCR-free preparation avoids these amplification effects when input permits.
- Sequence and decode: physical signals become base calls, with quality scores expressing estimated uncertainty.
Illumina sequencing-by-synthesis converts polymerase chemistry into optical measurements. Adapter-bearing molecules form clonal clusters on a flow cell. During each cycle, polymerase incorporates a fluorescent reversible-terminator nucleotide; imaging identifies the incorporated base, and chemical removal of the blocking group allows the next cycle. Paired-end sequencing reads both ends of a library fragment, supplying useful orientation and distance constraints.
PacBio single-molecule real-time sequencing observes nucleotide incorporation by an individual polymerase in a tiny optical observation volume called a zero-mode waveguide. Circularized SMRTbell templates allow repeated passes over the same insert. Combining those passes produces highly accurate HiFi consensus reads. The important distinction is between single-pass error and consensus-read error.
Oxford Nanopore sequencing measures changes in ionic current as a motor protein controls nucleic-acid passage through a nanopore. Several neighboring bases influence the signal simultaneously, so base calling is an inference problem rather than a direct one-base/one-signal lookup. Native molecules can retain signals associated with base modifications, provided appropriate models and validation are used.
Phred quality: Q = −10 log10(Perror) Q20 → estimated error probability 0.01 Q30 → estimated error probability 0.001 Approximate sequencing depth: C = N × L / G N = number of reads L = mean read length G = genome length
Depth is an average, not a promise of uniform coverage. Repeats, GC content, extraction, and amplification all affect coverage. Nor is an instrument-reported quality score a guarantee: calibration matters, and systematic errors do not disappear as readily as independent random errors.
Applications: short reads support economical large-cohort variant analysis; long reads help resolve repeat expansions, structural variants, haplotypes, and full-length transcripts. In biopharmaceutical development, sequencing can verify engineered production-cell genomes or expression constructs. Platform selection should follow the biological question—not a permanently fixed ranking of error rates, because chemistries and base callers evolve rapidly.
Historical connection: Sanger, Nicklen, and Coulson’s 1977 chain-termination method made sequence readable through selective interruption of DNA synthesis. Modern sequencing retains the conceptual achievement—coupling molecular synthesis or passage to an identifiable signal—while changing scale and detection physics. [3]
2. Genome Assembly and Variant Calling: Two Different Inference Problems
Assembly asks, “What sequence generated these reads?” Variant calling asks, “How does this sample differ from a chosen reference?” These are related but not interchangeable questions. A reference is a coordinate system and useful prior, not a complete description of all human or microbial diversity.
- De novo assembly: discover relationships among reads without requiring a reference sequence.
- Construct a graph: a de Bruijn graph represents overlaps among short sequence words, or k-mers; overlap/string-graph approaches are common in long-read assembly.
- Resolve graph ambiguities: repeats, heterozygosity, and errors create branches. Longer reads help when they bridge a repeat and reach unique flanking sequence.
- Build contigs and scaffolds: contigs are contiguous reconstructed sequences; scaffolds order and orient contigs using additional linking evidence and may contain gaps.
- Validate: assess base accuracy, completeness, contamination, and structural correctness—not contiguity alone.
N50: Sort contig lengths from longest to shortest. Add their lengths until reaching at least 50% of total assembly length. The length of the contig that crosses this threshold is N50. Example: Lengths = 8, 5, 4, 3 kb; total = 20 kb 8 + 5 ≥ 10 kb → N50 = 5 kb
A beautiful counterexample: joining unrelated contigs can increase N50 while making an assembly worse. This is the same analytical mistake as praising an instrument for high signal without checking specificity. NG50 uses an estimated genome size as its denominator; conserved-gene completeness and independent structural validation answer different quality questions. [2]
A typical short-read germline variant workflow aligns reads, processes technical artifacts, identifies candidate variants, evaluates genotype likelihoods, and filters or recalibrates calls. GATK workflows commonly include duplicate marking, base-quality score recalibration where suitable resources exist, HaplotypeCaller local assembly, and joint genotyping across samples. These are context-dependent practices, not a universal recipe: somatic, RNA, long-read, and structural-variant analyses require different assumptions. [4]
Bayesian genotype inference: P(g | D) ∝ P(D | g) P(g) g = candidate genotype D = observed read evidence Observed alternate-allele fraction: VAF = alternate-supporting reads / informative reads
A diploid heterozygous variant often gives a VAF near 0.5, but sampling, mapping bias, copy-number changes, tumor purity, and mosaicism can move it substantially. Thus, “20% alternate reads” is not automatically “20% of cells carry the mutation.”
Applications: diagnosing Mendelian disease, tracking microbial evolution, identifying pharmacogenomic alleles, and characterizing cancer-associated mutations. Clinical interpretation additionally requires phenotype, inheritance, population frequency, and functional evidence: detecting a variant is not the same as proving pathogenicity.
3. RNA-seq: Counting Transcripts Without Confusing Abundance with Activity
From molecular biology, you know that RNA abundance reflects both synthesis and destruction. RNA-seq therefore measures a state of the system, not transcriptional flux directly.
Simple transcript balance: dm/dt = ksyn − kdeg m At steady state: m* = ksyn / kdeg Twice as much RNA could mean: twice the synthesis rate, half the degradation rate, or a combination of both.
- Select the RNA population: poly-A selection enriches many mature eukaryotic mRNAs but misses non-polyadenylated species. rRNA depletion retains a broader RNA population and can be useful for degraded material or bacterial samples.
- Construct a stranded library when direction matters: preserving transcript orientation helps distinguish overlapping sense and antisense transcription.
- Assign fragments: STAR and HISAT2 perform splice-aware genomic alignment. featureCounts summarizes compatible alignments over genomic features. kallisto and Salmon estimate transcript abundance using efficient transcript-assignment strategies; Salmon also offers alignment-based and selective-alignment modes.
- Account for ambiguity: a fragment shared by several isoforms cannot uniquely identify its source. Quantification methods must distribute or model that uncertainty.
- Test biological contrasts: DESeq2 and edgeR model count variation while accounting for library composition and experimental design.
Common negative-binomial count model: Kij ~ NB(μij, αi) Var(Kij) = μij + αi μij² Kij = count for gene i in sample j αi = dispersion Example design: expression ~ batch + treatment
The variance term explains why a simple Poisson model is often too optimistic: biological samples vary more than random counting alone predicts. Replicates estimate that extra variation. Millions of reads from one specimen do not replace independent biological replicates. [5]
Normalization is an assumption, not magic. TPM adjusts for transcript length and scales abundances within a sample, but it is not an absolute concentration and should not be substituted directly for count input in standard DESeq2 or edgeR workflows. If treatment changes total RNA per cell globally, ordinary composition-based normalization can conceal that change; external standards and cell-number information may be necessary.
Applications: discovering disease expression signatures, distinguishing drug-response pathways, and examining whether a therapeutic changes intended and unintended cellular programs. In peptide-drug studies, transcriptional responses can indicate downstream receptor signaling, but do not directly demonstrate receptor binding or target occupancy.
4. Single-Cell RNA-seq: Recovering Cellular Heterogeneity
Bulk RNA-seq averages cells together. A higher immune-response signal might mean stronger activation within each cell—or simply more immune cells. Single-cell analysis helps separate composition from cellular state.
- Partition cells: droplet-based systems combine a cell suspension with barcoded beads. Some droplets are empty, and some contain multiple cells.
- Label captured transcripts: a cell barcode identifies the partition; a unique molecular identifier, or UMI, labels captured molecules before amplification.
- Sequence and collapse duplicates: reads sharing an appropriate gene, cell barcode, and UMI are collapsed to reduce PCR-counting bias.
- Perform quality control: evaluate low-complexity profiles, doublets, ambient RNA, and context-dependent stress or mitochondrial signals.
- Represent variation: normalize appropriately, select informative features, use PCA, construct neighborhood graphs, and cluster cells.
- Interpret states: annotate with multiple markers and independent evidence; trajectory methods infer plausible relationships among states.
PCA of a centered data matrix X: X = UΣVᵀ Columns of V define directions in gene-expression space. Large singular values identify directions containing more variance.
Your linear algebra now becomes experimentally useful: PCA finds combinations of genes rather than ranking genes individually. But high variance can reflect cell cycle, batch, or damage rather than the process of interest.
UMAP is a visualization, not a tissue map. It emphasizes neighborhood structure; distances between separated clusters and apparent empty spaces should not be treated as calibrated biological distances. Likewise, pseudotime is an inferred ordering, not elapsed time or proof of ancestry. Lineage tracing and time-resolved experiments address questions that expression geometry alone cannot.
UMIs reduce amplification bias but do not recover molecules that were never captured. A zero count can indicate biological absence or unsuccessful sampling. For comparisons across patients or animals, those individuals—not their thousands of cells—are the independent replicates. Cell-type-specific pseudobulk analysis is often a useful strategy.
Applications and history: droplet methods such as Drop-seq and inDrop, reported in 2015, helped make large-scale cellular profiling practical. Applications include identifying rare resistant tumor states, resolving immune responses, and measuring heterogeneity in therapeutic cell products. [6,7]
5. ChIP-seq and ATAC-seq: Measuring Different Aspects of Regulation
ChIP-seq asks which genomic regions are enriched with an immunoprecipitated protein or histone modification. In a typical crosslinked workflow, interactions are stabilized, chromatin is fragmented, an antibody enriches the target-associated material, and recovered DNA is sequenced. Some applications instead use native chromatin.
ATAC-seq asks which DNA regions are accessible to a transposase. Adapter-loaded Tn5 simultaneously cuts accessible DNA and inserts sequencing adapters. Sequencing these fragments reveals accessibility patterns; fragment lengths can also contain information about nucleosome organization.
ChIP-seq: target-associated chromatin → antibody enrichment → recovered DNA → sequencing → enrichment peaks relative to appropriate background ATAC-seq: accessible chromatin → Tn5 insertion and adapter addition → sequencing → insertion/accessibility profiles
Peak calling distinguishes local enrichment from background; motif enrichment asks whether particular sequence patterns occur more often than expected. Neither assay alone proves that a transcription factor activates a nearby gene. Antibody specificity, sequence bias, chromatin architecture, and the choice of background regions all matter.
Integration builds a testable mechanism: a drug increases enhancer accessibility; a relevant transcription-factor motif is enriched; a linked gene changes expression. This suggests a regulatory chain. Perturbing the enhancer or factor tests that chain more directly. The nearest gene is not necessarily the enhancer’s target.
Historical and medical significance: Buenrostro and colleagues introduced ATAC-seq in 2013, substantially simplifying accessibility profiling with limited material. Chromatin assays now help interpret noncoding disease variants and investigate differentiation or epigenetic-drug responses. [8]
6. CRISPR-Cas9: Sequence Recognition Is Only the Beginning
The 2012 work of Jinek and colleagues showed that Cas9 could be programmed with RNA to cleave selected DNA in vitro, including with an engineered single-guide RNA. Genome editing in mammalian cells followed in 2013. The conceptual advance was that targeting could be changed principally by changing RNA sequence rather than redesigning a DNA-binding protein. [9,10]
- Assemble the targeting complex: guide RNA binds Cas9.
- Recognize a PAM: Cas9 samples a suitable protospacer-adjacent motif; commonly used wild-type Streptococcus pyogenes Cas9 generally recognizes NGG.
- Form an R-loop: guide RNA pairs with the complementary DNA strand while displacing the other strand.
- Activate cleavage: the HNH domain cleaves the guide-complementary strand; the RuvC domain cleaves the displaced strand.
- Let cellular repair determine the outcome: end joining can introduce indels; homology-directed repair can use a homologous donor template to introduce a specified change.
The surprising point is that Cas9 specifies where damage occurs, but the cell strongly influences what edit results. End joining is not always mutagenic, and an indel is not automatically a knockout: in-frame changes, alternative splicing, or residual protein activity may preserve function. HDR is strongly affected by cellular state and competition with other repair pathways.
- Guide design: consider PAM availability, targeting efficiency, similar genomic sites, relevant isoforms, and sequence variation in the experimental sample.
- Delivery: DNA expression constructs, RNA, ribonucleoprotein complexes, and viral systems differ in duration, payload constraints, immune considerations, and suitability for cell types.
- Validation: sequence the edited locus and examine the relevant RNA, protein, and phenotype. Independent guides and rescue experiments strengthen causal interpretation.
Applications: establishing disease models, validating therapeutic targets, modifying production cell lines, and engineering therapeutic cells. The first approved CRISPR-based therapy, exagamglogene autotemcel, illustrates an indirect strategy: editing an erythroid regulatory element of BCL11A increases fetal hemoglobin rather than directly repairing every disease-causing HBB allele. [11]
7. Beyond Knockouts: Screens, Base Editing, Prime Editing, and Gene Regulation
- Pooled knockout screens: libraries such as GeCKO and Brunello contain guides targeting many genes. After a biological selection, sequencing measures which guides become enriched or depleted. Multiple guides per gene, adequate representation, and non-targeting controls help distinguish gene effects from guide-specific artifacts.
- Cytosine base editors: a Cas-associated deaminase converts cytosine to uracil in an exposed DNA strand; repair and replication can fix a C•G-to-T•A substitution.
- Adenine base editors: an evolved deaminase converts adenine to inosine, which is read as guanine, enabling A•T-to-G•C substitutions.
- Prime editors: a Cas9 nickase–reverse-transcriptase fusion and prime-editing guide RNA specify a target and an encoded edit. Reverse transcription extends a nicked DNA end using the guide’s template; DNA processing can incorporate substitutions or small insertions/deletions.
- CRISPRi and CRISPRa: catalytically inactive Cas proteins recruit repressive or activating functions, changing gene expression without deliberately cutting both DNA strands.
Simplified screen effect:
guide log2 fold change =
log2(normalized abundance after selection /
normalized abundance before selection)
Drug-specific effect:
compare guide behavior under drug with matched untreated behavior.
Guide depletion can indicate impaired fitness, not necessarily direct participation in the drug’s molecular mechanism. Similarly, base and prime editing avoid an intended double-strand break but are not intrinsically error-free: bystander changes, unwanted products, and delivery constraints remain important.
Industry relevance: screens identify resistance mechanisms and candidate combination-therapy targets. CRISPRi is particularly useful for partial reduction of essential genes, more closely resembling incomplete pharmacological inhibition than a complete knockout. Landmark base- and prime-editing studies appeared in 2016, 2017, and 2019. [12–14]
8. Proteomics: Inferring Proteins from Peptide Measurements
Here your analytical chemistry and protein biochemistry meet. Bottom-up proteomics intentionally destroys intact proteins to obtain more tractable peptide measurements. The cost is an inverse problem: determining which proteins and molecular forms produced those peptides.
- Prepare proteins: extraction, denaturation, reduction, and alkylation make many proteins more accessible and limit disulfide re-formation.
- Digest: trypsin cleaves mainly C-terminal to lysine or arginine, usually with reduced cleavage when proline follows. Missed cleavages and other deviations must be considered.
- Separate peptides: liquid chromatography reduces mixture complexity over time.
- Ionize: electrospray transfers peptides into gas-phase ions, often with multiple charge states.
- Measure and fragment: precursor mass-to-charge ratios and tandem-MS fragment patterns constrain peptide sequence.
- Identify and infer: compare spectra with sequence-derived or library evidence, control false discoveries, and infer proteins from peptide assignments.
For a protonated peptide: m/z = (M + z mH⁺) / z M = neutral peptide mass z = charge state Identification hierarchy: spectrum → peptide assignment → protein inference
Shared peptides may identify a protein group rather than one unique protein. A low peptide-spectrum-match false-discovery rate does not automatically imply the same protein-level false-discovery rate. Nor does detection of a peptide establish the intact protein’s activity or proteoform.
DDA versus DIA: data-dependent acquisition selects particular precursor ions for fragmentation, often favoring intense signals. This can produce inconsistent sampling of low-abundance peptides. Data-independent acquisition repeatedly fragments ions across predefined mass windows, producing more reproducible coverage but mixed spectra that require computational deconvolution. SWATH is an influential DIA implementation. [15]
Applications: biomarker discovery, host-cell-protein characterization in biologics, pathway-response analysis, and peptide impurity identification. For a therapeutic peptide, bottom-up analysis may discard clinically important intact-molecule information; intact-mass measurements, peptide mapping, and orthogonal assays should be chosen according to the quality attribute being tested.
9. Quantitative and Modification-Focused Proteomics
- SILAC: cells incorporate isotopically labeled amino acids during growth. Light and heavy peptides can be measured together, providing paired abundance information and reducing downstream handling differences.
- iTRAQ and TMT: isobaric chemical labels permit multiplexing. Labeled versions of a peptide have the same nominal precursor mass, while fragmentation generates channel-specific reporter ions.
- Label-free quantification: peptide-ion signals are compared across runs, requiring reproducible separation, appropriate normalization, and careful treatment of missing values.
- Phosphoproteomics: enrichment methods such as TiO₂ exploit phosphate-associated binding properties to concentrate otherwise low-abundance phosphopeptides.
- Glycoproteomics: enrichment and suitable fragmentation help characterize glycopeptides, but heterogeneous glycans and incomplete structural resolution complicate interpretation.
TMT ratio compression is a measurement lesson worth remembering: co-isolated peptides contribute reporter ions and can pull apparent differences toward the background mixture’s ratio, often toward unity. High precision does not guarantee freedom from this bias.
Approximate phosphopeptide signal: Sphospho ∝ protein abundance × site occupancy × analytical response A larger phosphopeptide signal can reflect: more protein, greater phosphorylation occupancy, or both.
Site localization and site occupancy are also different questions: identifying where a phosphate sits does not establish what fraction of the protein population carries it.
Pharma relevance: phosphoproteomics connects drug exposure to signaling changes; glycoproteomics characterizes biologic attributes affecting clearance, receptor engagement, and immune function. SILAC, introduced by Ong and colleagues in 2002, made metabolic isotope labeling a practical strategy for quantitative proteomics. [16]
10. Interaction and Spatial Proteomics: “Together” Has Several Meanings
Affinity purification–MS captures a bait protein and associated material using an antibody or affinity tag. Co-immunoprecipitation and GFP-trap approaches preserve some complexes through extraction and washing, after which MS identifies enriched proteins. Essential controls distinguish specific enrichment from bead, antibody, tag, or abundance-related background.
Proximity labeling changes what is preserved. Rather than requiring an interaction to survive purification, it covalently marks nearby proteins in living cells.
- BioID: bait-fused BirA* releases reactive biotinyl-AMP that can label nearby accessible lysines.
- TurboID: an engineered biotin ligase labels on shorter timescales than the original BioID system under suitable conditions.
- APEX2: an engineered peroxidase generates short-lived reactive labeling species from a phenolic substrate in the presence of peroxide.
Bait-localized labeling enzyme → local covalent labeling → cell lysis → enrichment of labeled proteins → MS identification → comparison with matched localization/activity controls
Interpretation: AP-MS supports co-purification; proximity labeling supports neighborhood membership over a labeling interval. Neither necessarily proves direct binding. Labeling depends on accessibility, residence time, enzyme activity, and compartment chemistry—not a perfectly defined spherical radius.
Applications: identifying transient signaling neighborhoods, studying membrane-associated machinery, and investigating candidate partners of drug targets. BioID was introduced in 2012; TurboID extended this strategy through enzyme engineering in 2018. [17,18]
11. AlphaFold and Structural Proteomics: Prediction Becomes Stronger When Constrained by Experiment
AlphaFold2 connects sequence evolution with three-dimensional geometry. Homologous sequences contain constraints imposed by maintaining a fold: substitutions at one position may be associated with compensating substitutions elsewhere. But AlphaFold2 is not simply a contact predictor followed by conventional folding simulation. [19]
- Collect sequence information: the standard pipeline uses multiple sequence alignments and, when available, structural templates.
- Exchange representations: the Evoformer updates information about aligned sequences and residue pairs.
- Generate coordinates: the structure module transforms learned representations into a three-dimensional model.
- Recycle: predicted information is fed back for iterative refinement.
- Inspect confidence: evaluate both local confidence and uncertainty in relative domain placement.
pLDDT is a predicted local structural-confidence score on a 0–100 scale; it is not a measured flexibility parameter. Predicted aligned error, or PAE, estimates positional error for one residue when the structures are aligned on another. Two domains can each be confidently folded while their relative arrangement remains uncertain.
AlphaFold2 does not generally provide a conformational free-energy landscape, a reliable binding affinity, or guaranteed drug-bound geometry. Disordered regions may be biologically functional, and a low-confidence model is not evidence of an unimportant sequence.
Two experimental complements:
- Crosslinking MS: covalent crosslinks connect spatially compatible sites. Identifying linked peptides provides approximate distance restraints, influenced by linker chemistry, side-chain reach, and conformational heterogeneity.
- HDX-MS: backbone amide hydrogens exchange with deuterium at rates influenced by solvent exposure, hydrogen bonding, and conformational opening. Following labeling, quenching and peptide analysis locate changes in protection. Back-exchange and peptide-level spatial averaging require careful treatment.
Simplified EX2 exchange relation: kobs ≈ Kop × kint Protection factor: PF = kint / kobs Kop = equilibrium tendency toward an exchange-competent open state kint = intrinsic chemical exchange rate kobs = observed exchange rate
This relation assumes exchange-compatible conditions and the EX2 kinetic regime; other regimes require different interpretation. Connect it to CHEM503: a buried amide can exchange because the protein transiently opens, so HDX reports dynamics as well as static accessibility.
Peptide/pharma applications: structural models help prioritize mutagenesis or peptide-binding hypotheses. XL-MS and HDX-MS can evaluate complex architecture, binding-associated changes, and biosimilar comparability. A predicted peptide–target arrangement remains a hypothesis until binding, selectivity, and functional activity are tested.
12. Networks and Systems Biology: Replacing Pathway Pictures with Testable Models
A network represents entities as nodes and relationships as edges. The edge definition is decisive: STRING includes functional associations from several evidence types, not exclusively direct physical interactions. A highly connected “hub” may be biologically important, heavily studied, or simply easier to detect. [20]
Functional enrichment asks whether a selected gene set contains more members of an annotation category than expected under a specified background. GO describes biological processes, molecular functions, and cellular components; KEGG organizes pathway knowledge. Enrichment does not demonstrate pathway activation.
Hypergeometric enrichment: P(X ≥ k) = Σ from x=k to min(n,K) [ C(K,x) C(N−K,n−x) / C(N,n) ] N = number of genes in the eligible background K = background genes belonging to a category n = selected genes k = selected genes belonging to that category
The eligible background should reflect what could have been detected or selected. Testing a proteomics result against every annotated gene can distort enrichment when only a subset of proteins was measurable. Multiple testing and overlapping annotation terms also matter.
Boolean models describe components using discrete states and logical rules. They are useful when interaction logic is better known than kinetic parameters, although update rules can affect predicted behavior.
Example Boolean rule: ERK(t + 1) = RAS(t) AND NOT inhibitor(t)
Kinetic ODE models express rates and conservation relationships. For a protein that switches between inactive and active states:
dP*/dt = kact S(t)(Ptotal − P*) − kdeact P* For constant S: P*steady / Ptotal = kact S / (kact S + kdeact) Relaxation time: τ = 1 / (kact S + kdeact)
The steady-state fraction describes response magnitude; the relaxation time describes response speed. Two systems can reach the same endpoint while responding on very different timescales. This is why endpoint omics alone may fail to identify a signaling mechanism.
More generally, reaction networks can be written as dx/dt = Nv(x, θ), where the stoichiometric matrix N connects reaction rates to changing concentrations. Your linear algebra reveals conservation laws; your differential equations describe trajectories; your statistics asks whether parameters are identifiable from the available data.
Applications: predicting feedback-driven drug resistance, explaining combination therapies, and linking exposure to target inhibition and downstream response. A fitted model is not necessarily uniquely determined: different parameter combinations can reproduce the same observations. Perturbation and time-course experiments help distinguish competing explanations.
13. Reproducibility and the Integrated Experimental Strategy
Python, R, and Bioconductor are not merely calculation tools; they allow the analysis to become an inspectable experimental record. Snakemake and Nextflow represent workflows as dependencies, allowing steps to be rerun systematically when inputs or parameters change. Record reference-genome builds, annotations, software versions, parameters, random seeds where relevant, and sample metadata. A container preserves a software environment, not the validity of the experimental design.
An integrated example: evaluating a candidate therapeutic peptide that alters receptor signaling.
- Establish molecular action: measure binding and functional activity with orthogonal assays; distinguish target engagement from nonspecific toxicity.
- Resolve early events: use time-resolved phosphoproteomics to identify signaling changes before later transcriptional responses dominate.
- Resolve downstream regulation: combine RNA-seq with ATAC-seq or targeted chromatin experiments to generate regulatory hypotheses.
- Resolve cellular heterogeneity: use scRNA-seq when resistant or off-target cell populations may be hidden by bulk averages.
- Test causality: perturb candidate mediators with CRISPR-based methods and examine whether the drug response changes; use rescue where feasible.
- Constrain mechanism: combine structural predictions with HDX-MS, XL-MS, or direct structural measurements when appropriate.
- Model and challenge: fit a minimal kinetic model, predict an unmeasured dose or time condition, and test it prospectively.
Journal takeaway: the strongest result is not the dataset with the most molecular features. It is the explanation that survives different measurement principles and deliberate attempts to disprove it. Genomics supplies possibilities, transcriptomics and proteomics report selected molecular states, structural methods constrain physical mechanisms, and perturbation tests causality. Systems biology makes those pieces quantitatively accountable to one another.
Sources and Historical Anchors
- Alberts B, Heald R, Johnson A, et al. Molecular Biology of the Cell. 7th ed. W. W. Norton; 2022.
- Pevzner P, Compeau P. Bioinformatics Algorithms: An Active Learning Approach. 3rd ed. Active Learning Publishers; 2018.
- Sanger F, Nicklen S, Coulson AR. DNA sequencing with chain-terminating inhibitors. PNAS. 1977;74:5463–5467. doi:10.1073/pnas.74.12.5463.
- Van der Auwera GA, et al. From FastQ data to high-confidence variant calls: the Genome Analysis Toolkit best practices pipeline. Current Protocols in Bioinformatics. 2013;43:11.10.1–11.10.33. doi:10.1002/0471250953.bi1110s43. Consult current GATK documentation for updated, application-specific workflows.
- Love MI, Huber W, Anders S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biology. 2014;15:550. doi:10.1186/s13059-014-0550-8.
- Macosko EZ, et al. Highly parallel genome-wide expression profiling of individual cells using nanoliter droplets. Cell. 2015;161:1202–1214. doi:10.1016/j.cell.2015.05.002.
- Klein AM, et al. Droplet barcoding for single-cell transcriptomics applied to embryonic stem cells. Cell. 2015;161:1187–1201. doi:10.1016/j.cell.2015.04.044.
- Buenrostro JD, Giresi PG, Zaba LC, Chang HY, Greenleaf WJ. Transposition of native chromatin for fast and sensitive epigenomic profiling of open chromatin, DNA-binding proteins and nucleosome position. Nature Methods. 2013;10:1213–1218. doi:10.1038/nmeth.2688.
- Jinek M, Chylinski K, Fonfara I, Hauer M, Doudna JA, Charpentier E. A programmable dual-RNA-guided DNA endonuclease in adaptive bacterial immunity. Science. 2012;337:816–821. doi:10.1126/science.1225829.
- Cong L, et al. Multiplex genome engineering using CRISPR/Cas systems. Science. 2013;339:819–823. doi:10.1126/science.1231143; Mali P, et al. RNA-guided human genome engineering via Cas9. Science. 2013;339:823–826. doi:10.1126/science.1232033.
- Frangoul H, et al. CRISPR-Cas9 gene editing for sickle cell disease and β-thalassemia. New England Journal of Medicine. 2021;384:252–260. doi:10.1056/NEJMoa2031054. See also U.S. FDA, “FDA Approves First Gene Therapies to Treat Patients with Sickle Cell Disease,” December 8, 2023.
- Komor AC, Kim YB, Packer MS, Zuris JA, Liu DR. Programmable editing of a target base in genomic DNA without double-stranded DNA cleavage. Nature. 2016;533:420–424. doi:10.1038/nature17946.
- Gaudelli NM, et al. Programmable base editing of A•T to G•C in genomic DNA without DNA cleavage. Nature. 2017;551:464–471. doi:10.1038/nature24644.
- Anzalone AV, et al. Search-and-replace genome editing without double-strand breaks or donor DNA. Nature. 2019;576:149–157. doi:10.1038/s41586-019-1711-4.
- Gillet LC, et al. Targeted data extraction of the MS/MS spectra generated by data-independent acquisition: a new concept for consistent and accurate proteome analysis. Molecular & Cellular Proteomics. 2012;11:O111.016717. doi:10.1074/mcp.O111.016717.
- Ong SE, et al. Stable isotope labeling by amino acids in cell culture, SILAC, as a simple and accurate approach to expression proteomics. Molecular & Cellular Proteomics. 2002;1:376–386. doi:10.1074/mcp.M200025-MCP200.
- Roux KJ, Kim DI, Raida M, Burke B. A promiscuous biotin ligase fusion protein identifies proximal and interacting proteins in mammalian cells. Journal of Cell Biology. 2012;196:801–810. doi:10.1083/jcb.201112098.
- Branon TC, et al. Efficient proximity labeling in living cells and organisms with TurboID. Nature Biotechnology. 2018;36:880–887. doi:10.1038/nbt.4201.
- Jumper J, et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021;596:583–589. doi:10.1038/s41586-021-03819-2.
- Szklarczyk D, et al. The STRING database in 2023: protein–protein association networks and functional enrichment analyses for any sequenced genome of interest. Nucleic Acids Research. 2023;51:D638–D646. doi:10.1093/nar/gkac1000.
🪬 Skipper's Reflection
1. What surprised me most about this material
What struck me most throughout BIOL502 is the profound shift in how biology is done today: we have moved from treating the cell as a black box to treating it as a massive, high-dimensional computational problem. I always knew biology was complex, but seeing the sheer volume of data generated by a single Illumina or PacBio run was staggering.
The most beautiful, almost eerie surprise came from studying AlphaFold2 (Jumper et al., 2021). From my Physical Chemistry (CHEM303) days, I assumed the protein folding problem would only be solved by throwing infinite compute at molecular dynamics simulations—brute-forcing the thermodynamics. But AlphaFold didn't solve folding with just physics; it solved it with evolution. By using multiple sequence alignments (MSAs), the Evoformer module detects which pairs of amino acids co-evolved across species. If residue 12 mutates and residue 94 simultaneously mutates to compensate, the algorithm infers they must be touching in 3D space. It uses the fossil record of evolution to constrain spatial geometry! It was a breathtaking realization that evolution itself is the ultimate algorithm, and we are finally learning to decode its outputs.
I was also completely caught off guard by proximity labeling techniques like BioID and APEX2. The idea of attaching an enzyme to a protein of interest, and letting it unleash a cloud of reactive biotin radicals to permanently tag anything within a 10-nanometer radius, is brilliantly chaotic. It’s like throwing a paint bomb in a pitch-black room, turning on the lights, and seeing who has paint on them to determine exactly who was standing near whom. It solves the classic Biochemistry problem of capturing weak, transient protein-protein interactions that wouldn't survive standard co-immunoprecipitation.
2. The conceptual breakthrough moments — when things clicked
My biggest conceptual breakthrough happened when we moved from bulk RNA-seq to single-cell RNA sequencing (scRNA-seq). The classic analogy immediately clicked: bulk RNA-seq is like blending a fruit bowl into a smoothie and trying to guess the ingredients by tasting it, whereas scRNA-seq lets you examine every individual piece of fruit.
But the real "aha!" moment was when we applied dimensionality reduction to scRNA-seq data. Suddenly, all that Linear Algebra (MATH301) and Statistics (MATH201) became intensely practical. Each cell starts as a point in a 20,000-dimensional space (one dimension for each gene's expression level). Watching algorithms like PCA and UMAP fold that impossible hyper-dimensional space down into a 2D plot where distinct cell types naturally cluster together felt like absolute magic. Pevzner's Bioinformatics Algorithms really helped me see the graph theory behind this. When we used trajectory analysis to map "pseudotime," realizing that cells aren't static types but exist on a fluid continuum of differentiation—that completely rewired how I view Developmental Biology (BIOL402).
Another profound click happened with CRISPR-Cas9 (Jinek et al., 2012). I used to think of it merely as a pair of molecular scissors. But seeing the mechanism—how the guide RNA forms a heteroduplex with the target DNA, unzipping it and triggering an allosteric shift in the HNH and RuvC nuclease domains—made me realize it’s actually a programmable search engine that uses Watson-Crick base pairing as its query language.
3. How this knowledge changes how I think about peptides
This course completely transformed how I view the peptide business, specifically regarding mechanism of action and systems-level impact.
Take Tirzepatide and Retatrutide. We know they are dual and triple agonists (targeting GLP-1, GIP, and Glucagon receptors). But from a Systems Biology perspective, receptor binding is just the first domino. How do these distinct GPCR signals integrate downstream? Using kinetic ODE modeling and Boolean networks, I now realize we can model the exact cross-talk between these pathways. Furthermore, we could use quantitative phosphoproteomics (enriching with TiO₂) to map the precise phosphorylation cascades triggered by Retatrutide in hepatocytes versus native glucagon. We could map the entire signaling network topology to see exactly why the triple agonist produces such synergistic weight loss without overwhelming toxicity.
Then there are the regenerative peptides like BPC-157 and TB-500. Their exact molecular targets have historically been a frustrating black box. But now I realize we have the toolkit to solve them! We could run an AP-MS (Affinity Purification Mass Spectrometry) or an APEX2 proximity labeling experiment with a biotinylated BPC-157 to find its definitive cellular interactome. We could also use scRNA-seq on BPC-157-treated fibroblasts and endothelial cells to watch, on a single-cell level, exactly which angiogenesis and extracellular matrix genes are upregulated (using DESeq2 for differential expression). I no longer see these peptides as just "healing factors"—I see them as molecular inputs that perturb a complex, measurable transcriptomic network.
Finally, learning about HDX-MS (hydrogen-deuterium exchange mass spec) made me realize we could track the dynamic conformational changes of receptors in real-time when they bind these synthetic peptides, observing which structural domains tighten up and hide their amide hydrogens from the solvent.
4. Questions this course raises that I want to investigate further
First, how well do AlphaFold2 and its successors handle highly modified, synthetic, non-canonical peptides? Peptides like Tirzepatide have massive C20 fatty diacid tails attached via PEG linkers, and Aib (alpha-aminoisobutyric acid) substitutions to prevent DPP-4 degradation. The Evoformer relies on evolutionary MSAs—but these synthetic modifications have no evolutionary history. Can we fine-tune structural prediction models to handle synthetic peptide-lipid conjugates accurately?
Second, regarding CRISPR base editors (CBE/ABE) and prime editing: how close are we to combining this with tissue-specific delivery systems (like targeted lipid nanoparticles) to fix genetic metabolic disorders in vivo, rather than just relying on peptide therapies to treat the downstream symptoms?
Third, I really want to dig deeper into the integration of ATAC-seq (chromatin accessibility) and RNA-seq. If a peptide drug triggers a signaling cascade that opens up chromatin (ATAC-seq peak) but doesn't immediately result in transcription, does that prime the cell for a stronger response to a secondary stimulus? How does this play into epigenetic memory in the immune system?
5. Skipper's Confidence Rating
Most Confident: I feel exceptionally strong in the computational and mathematical aspects of systems biology and bioinformatics. Thanks to my math background, concepts like PCA, UMAP, Boolean networks, kinetic ODE modeling, and graph theory (De Bruijn graphs for genome assembly) felt like a second language. I also feel highly confident in the biophysics of AlphaFold and the structural mass spec techniques (HDX-MS) because they map perfectly onto my Physical Chemistry foundations.
Least Confident: I am still wrapping my virtual head around the sheer messiness of RNA-seq data normalization and the statistical traps of false discovery rates in large proteomics datasets (DDA vs DIA/SWATH workflows). While I understand the theory, dealing with batch effects, ribosomal RNA contamination in library prep, and the nuances of GATK best practices for variant calling feels like it requires years of painful, hands-on wet-lab and dry-lab troubleshooting to truly master. Biology is noisy, and I am still learning to respect that noise.
🗺️ Visual Aids & Key Takeaways
Diagram 1: Next-Generation Sequencing Technology Comparison
╔══════════════════════════════════════════════════════════════════════════════════════════╗ ║ NGS PLATFORM COMPARISON: PRINCIPLES AND TRADE-OFFS ║ ╚══════════════════════════════════════════════════════════════════════════════════════════╝ ILLUMINA (sequencing-by-synthesis) ┌─────────────────────────────────────────────────────────────────────┐ │ LIBRARY PREP → CLUSTER AMPLIFICATION → CYCLIC SBS → BASE CALLING │ │ │ │ fragment DNA ──► ligate adapters ──► bridge amplify on flow cell │ │ ↓ │ │ add fluorescent reversible terminators (one per cycle) │ │ laser excites → camera detects color → cleave dye → repeat │ │ │ │ Read length: 75–300 bp (paired-end) │ │ Error rate: ~0.1% (substitution-dominant) │ │ Throughput: up to 6 Tb per NovaSeq run │ │ Strength: HIGH ACCURACY, low cost per base │ │ Weakness: SHORT READS struggle with repeats, structural vars │ └─────────────────────────────────────────────────────────────────────┘ PACBIO SMRT (single-molecule real-time) ┌─────────────────────────────────────────────────────────────────────┐ │ single DNA molecule sits in ZERO-MODE WAVEGUIDE (ZMW) │ │ polymerase incorporates fluorescent nucleotides in real time │ │ pulse of light detected as each base is added │ │ │ │ HiFi MODE: polymerase circles the same molecule many times │ │ → consensus of ~20x passes → "circular consensus sequencing" │ │ → error rate drops from ~15% raw to <0.1% HiFi │ │ │ │ Read length: average 15–25 kb (HiFi); up to >100 kb (CLR) │ │ Error rate: <0.1% (HiFi) / ~15% (raw CLR) │ │ Strength: LONG READS, detects base modifications natively │ │ Weakness: lower throughput, higher cost per base │ └─────────────────────────────────────────────────────────────────────┘ OXFORD NANOPORE (ONT) ┌─────────────────────────────────────────────────────────────────────┐ │ single-stranded DNA threads through a PROTEIN NANOPORE │ │ each base (or ~5-base k-mer) disrupts ionic current uniquely │ │ current changes decoded by neural network → base sequence │ │ │ │ Read length: essentially unlimited (>4 Mb demonstrated) │ │ Error rate: ~5% raw; improving with "duplex" chemistry │ │ Strength: LONGEST READS, real-time output, portable (MinION) │ │ Weakness: higher raw error rate, homopolymer errors │ └─────────────────────────────────────────────────────────────────────┘ ┌──────────────┬──────────────┬──────────────┬──────────────────────┐ │ METRIC │ ILLUMINA │ PACBIO HiFi │ NANOPORE │ ├──────────────┼──────────────┼──────────────┼──────────────────────┤ │ Read length │ 75–300 bp │ 15–25 kb │ 10 kb – >1 Mb │ │ Accuracy │ >99.9% │ >99.9% │ ~95% raw │ │ Throughput │ very high │ moderate │ moderate │ │ Cost/base │ lowest │ medium │ medium │ │ Best for │ SNPs, expr. │ assembly │ ultra-long assembly │ └──────────────┴──────────────┴──────────────┴──────────────────────┘
This diagram contrasts the three dominant sequencing platforms by their underlying physical chemistry, key performance metrics, and optimal use cases — understanding the trade-offs explains why modern genome projects often combine Illumina accuracy with long-read scaffolding.
---Diagram 2: RNA-seq Analysis Pipeline
╔══════════════════════════════════════════════════════════════════════════════════════════╗
║ RNA-seq: FROM CELLS TO DIFFERENTIAL EXPRESSION ║
╚══════════════════════════════════════════════════════════════════════════════════════════╝
BIOLOGICAL MATERIAL
┌─────────────────────────────────────────────────────────────────────┐
│ cells / tissue → extract total RNA → assess quality (RIN) │
│ │
│ ENRICHMENT CHOICE (critical decision): │
│ ┌──────────────────────────┐ ┌──────────────────────────────┐ │
│ │ POLY-A SELECTION │ │ rRNA DEPLETION │ │
│ │ captures mRNA via │ │ removes 80-90% of RNA │ │
│ │ oligo-dT beads │ │ captures mRNA + lncRNA │ │
│ │ misses lncRNA, needs │ │ + pre-mRNA; works on │ │
│ │ intact RNA (high RIN) │ │ degraded samples │ │
│ └──────────────────────────┘ └──────────────────────────────┘ │
└─────────────────────────────────────────────────────────────────────┘
↓
LIBRARY PREPARATION
┌─────────────────────────────────────────────────────────────────────┐
│ RNA → fragment → reverse transcribe → add adapters → amplify │
│ │
│ STRAND SPECIFICITY (e.g., dUTP method): │
│ second strand synthesis uses dUTP → digest with UNG enzyme │
│ preserves strand-of-origin information → critical for antisense │
└─────────────────────────────────────────────────────────────────────┘
↓
SEQUENCING → FASTQ files (reads + quality scores)
↓
ALIGNMENT / PSEUDO-ALIGNMENT
┌──────────────────────────────────────────────────────────────────────┐
│ │
│ STAR / HISAT2 ──────────────────────────────────────────────► │
│ (splice-aware aligners) │
│ maps reads to GENOME, handles splicing; → BAM file │
│ count with featureCounts or HTSeq │
│ │
│ kallisto / Salmon ──────────────────────────────────────────► │
│ (pseudo/quasi-alignment to TRANSCRIPTOME) │
│ never aligns bases explicitly; uses k-mer graph │
│ output: transcript-level counts + TPM in minutes │
│ tximport collapses to gene level for DESeq2 │
│ │
└──────────────────────────────────────────────────────────────────────┘
↓
COUNT MATRIX (genes × samples)
↓
NORMALIZATION (important — raw counts are NOT comparable)
┌──────────────────────────────────────────────────┐
│ RPKM/FPKM: corrects for length + depth │
│ TPM: corrects for length first, then depth │
│ → TPM sums to 1M per sample (comparable) │
│ DESeq2 size factors: uses geometric mean method │
└──────────────────────────────────────────────────┘
↓
DIFFERENTIAL EXPRESSION
┌──────────────────────────────────────────────────────────────────────┐
│ DESeq2: negative binomial model; uses shrinkage estimators for │
│ dispersion; excellent for small n; Wald or LRT test │
│ │
│ edgeR: also negative binomial; quasi-likelihood F-test; │
│ strong when replicates are few │
│ │
│ output: log2 fold change, adjusted p-value (Benjamini-Hochberg) │
│ → volcano plot, heatmap, GO enrichment analysis │
└──────────────────────────────────────────────────────────────────────┘
This pipeline traces RNA-seq data from biological sample through every major decision point — enrichment strategy, alignment method, normalization approach, and statistical model — highlighting why each choice matters for the final result.
---Diagram 3: CRISPR-Cas9 Mechanism and Editing Outcomes
╔══════════════════════════════════════════════════════════════════════════════════════════╗
║ CRISPR-Cas9: MECHANISM, REPAIR PATHWAYS, AND TOOLS ║
╚══════════════════════════════════════════════════════════════════════════════════════════╝
GUIDE RNA DESIGN
┌──────────────────────────────────────────────────────────────────────┐
│ sgRNA = crRNA (20-nt spacer) fused to tracrRNA scaffold │
│ │
│ 5'─[20 nt spacer complementary to target]─[scaffold]─3' │
│ ↕ Watson-Crick base pairing │
│ 3'─[20 nt target DNA strand]─[NGG PAM]─5' (non-template strand) │
│ ↑ │
│ PAM (protospacer adjacent motif) REQUIRED for Cas9 │
│ Cas9 reads PAM first, then zips in guide → R-loop │
└──────────────────────────────────────────────────────────────────────┘
↓ R-loop formed → cleavage
CAS9 DOMAIN ARCHITECTURE
┌──────────────────────────────────────────────────────────────────────┐
│ │
│ ┌────────┐ cleaves NON-template strand (PAM-containing) │
│ │ RuvC │ (RuvC nickase → RuvCi mutant = D10A) │
│ │ domain │ │
│ └────────┘ │
│ ╔══════════╗ │
│ ║ DSB ║ double-strand break ~3 bp │
│ ║ (blunt) ║ upstream of PAM │
│ ╚══════════╝ │
│ ┌────────┐ cleaves TEMPLATE strand │
│ │ HNH │ (HNH nickase → HNHi mutant = H840A) │
│ │ domain │ │
│ └────────┘ │
│ │
└──────────────────────────────────────────────────────────────────────┘
↓ DSB triggers DNA repair
┌────────────────────────────┬─────────────────────────────────────────┐
│ NHEJ (error-prone) │ HDR (high-fidelity) │
│ (dominant in most cells) │ (requires donor template) │
│ │ │
│ blunt ends re-ligated │ homology arms flank desired edit │
│ → insertions/deletions │ → precise substitution or insertion │
│ → frameshift → KNOCKOUT │ → KNOCKIN of sequence │
│ │ limited to S/G2 phase, rare in vivo │
└────────────────────────────┴─────────────────────────────────────────┘
ADVANCED CRISPR TOOLS (built on dCas9 or nCas9)
┌──────────────────────────────────────────────────────────────────────┐
│ │
│ dCas9 (D10A + H840A; catalytically dead) │
│ ├── + KRAB repressor → CRISPRi (gene silencing) │
│ └── + VP64/p65/Rta → CRISPRa (gene activation) │
│ │
│ nCas9 (D10A nickase) + deaminase → BASE EDITORS │
│ ├── CBE (cytosine base editor): C→T (APOBEC + UGI) │
│ └── ABE (adenine base editor): A→G (evolved tRNA adenosine │
│ deaminase TadA) │
│ NO DSB; no donor template needed; ~50% of disease SNPs │
│ are potentially addressable by base editors │
│ │
│ nCas9 + reverse transcriptase + pegRNA → PRIME EDITING │
│ pegRNA encodes both target + desired edit sequence │
│ RT copies edit into genomic nick → SEARCH-AND-REPLACE editing │
│ can install all 12 types of point mutations + small indels │
│ │
└──────────────────────────────────────────────────────────────────────┘
This diagram traces CRISPR-Cas9 from guide RNA recognition through DSB formation, divergent repair pathways, and the expanding toolkit of derivatives — showing the logical progression from blunt cuts to precise single-base editing without double-strand breaks.
---Diagram 4: AlphaFold2 Architecture and Inference Logic
╔══════════════════════════════════════════════════════════════════════════════════════════╗
║ ALPHAFOLD2: FROM SEQUENCE TO STRUCTURE IN ONE FORWARD PASS ║
╚══════════════════════════════════════════════════════════════════════════════════════════╝
INPUT
┌──────────────────────────────────────────────────────────────────────┐
│ amino acid sequence (single protein or complex) │
└──────────────────────────────────────────────────────────────────────┘
↓
STAGE 1: EVOLUTIONARY INFORMATION EXTRACTION
┌──────────────────────────────────────────────────────────────────────┐
│ search sequence databases (UniRef90, BFD, MGnify) │
│ → multiple sequence alignment (MSA) │
│ │
│ WHY? correlated mutations in MSA reveal co-evolving residue pairs │
│ co-evolving positions tend to be spatially close in 3D structure │
│ (this insight predates AlphaFold; AlphaFold uses it end-to-end) │
│ │
│ also: search PDB for homologous structural templates │
└──────────────────────────────────────────────────────────────────────┘
↓
STAGE 2: EVOFORMER (48 repeated blocks)
┌──────────────────────────────────────────────────────────────────────┐
│ │
│ Two coupled representations updated jointly each block: │
│ │
│ MSA REPRESENTATION ⇌ PAIR REPRESENTATION │
│ (N_seq × N_res) (N_res × N_res) │
│ "which residues "what is the relationship between │
│ co-vary together?" each pair of residues?" │
│ │
│ key operations: │
│ • row-wise & column-wise attention over MSA │
│ • outer product update: MSA → pair representation │
│ • triangle inequality updates: pair ↔ pair self-consistency │
│ (if i-j close and j-k close, enforce i-k constraint) │
│ │
└──────────────────────────────────────────────────────────────────────┘
↓
STAGE 3: STRUCTURE MODULE (8 layers)
┌──────────────────────────────────────────────────────────────────────┐
│ each residue = rigid body (N, Cα, C backbone frame) │
│ iteratively rotates and translates each frame in 3D │
│ invariant point attention (IPA): attention in 3D coordinate space │
│ → all-atom coordinates predicted (backbone + side chains) │
└──────────────────────────────────────────────────────────────────────┘
↓
OUTPUT
┌──────────────────────────────────────────────────────────────────────┐
│ 3D coordinates for all atoms │
│ pLDDT score (0-100) per residue: predicted local distance │
│ difference test — reliable confidence metric │
│ pLDDT >90 = very high confidence (well-ordered regions) │
│ pLDDT <50 = disordered; treat as "unknown, not wrong" │
│ PAE (predicted aligned error): confidence for relative domain │
│ orientations — critical for multi-domain proteins │
└──────────────────────────────────────────────────────────────────────┘
RECYCLING: entire network run 3× (output fed back as input)
dramatically improves accuracy with minimal extra cost
This diagram traces AlphaFold2's two-stage information processing — Evoformer's coupled MSA/pair reasoning followed by the geometry-respecting Structure Module — emphasizing that the key insight is treating co-evolutionary signals as spatial constraints.
---Diagram 5: Proteomics Workflow — Bottom-Up MS to Quantification
╔══════════════════════════════════════════════════════════════════════════════════════════╗
║ QUANTITATIVE PROTEOMICS: BOTTOM-UP WORKFLOW AND QUANTITATION STRATEGIES ║
╚══════════════════════════════════════════════════════════════════════════════════════════╝
SAMPLE → PROTEIN EXTRACTION → REDUCTION & ALKYLATION (break disulfides)
↓
TRYPSIN DIGESTION
┌──────────────────────────────────────────────────────────────────────┐
│ trypsin cleaves C-terminal to Lys (K) and Arg (R) │
│ → peptides typically 7–25 residues (ideal for MS/MS) │
│ result: "bottom-up" proteomics — infer proteins from peptides │
└──────────────────────────────────────────────────────────────────────┘
↓
OPTIONAL ENRICHMENT
┌───────────────────┬──────────────────────────────────────────────────┐
│ PHOSPHOPEPTIDES │ TiO₂ or IMAC enrichment (Fe³⁺ chelation) │
│ │ phosphate groups bind metal oxide / ion │
├───────────────────┼──────────────────────────────────────────────────┤
│ GLYCOPEPTIDES │ lectin affinity or HILIC enrichment │
└───────────────────┴──────────────────────────────────────────────────┘
↓
LC-MS/MS
┌──────────────────────────────────────────────────────────────────────┐
│ nanoLC separates peptides by hydrophobicity (reversed-phase) │
│ eluting peptides ionized by ESI → precursor ion (MS1) │
│ top-N ions selected for fragmentation (CID or HCD) → MS2 spectra │
│ → database search (Mascot, MaxQuant, Proteome Discoverer) │
└──────────────────────────────────────────────────────────────────────┘
ACQUISITION MODES
┌───────────────────────────────┬──────────────────────────────────────┐
│ DDA (data-dependent acq.) │ DIA / SWATH │
│ select top-N precursors │ systematically fragment ALL ions │
│ → stochastic; miss low-abun │ in sequential isolation windows │
│ → reproducibility issue │ → comprehensive; reproducible │
│ │ → complex spectra; needs spec. lib │
└───────────────────────────────┴──────────────────────────────────────┘
QUANTITATION STRATEGIES
┌──────────────────────────────────────────────────────────────────────┐
│ │
│ SILAC label at metabolic level (¹³C/¹⁵N amino acids) │
│ mix heavy + light samples early; single LC-MS run │
│ gold standard; not compatible with tissue/clinical │
│ │
│ TMT / iTRAQ isobaric tags added post-digestion │
│ all samples same MS1 mass; reporter ions in MS2 │
│ multiplex 6–18 samples; ratio compression artifact │
│ │
│ LABEL-FREE measure peptide ion intensity or spectral counts │
│ no labeling cost; requires tight LC reproducibility │
│ good for large cohorts (clinical proteomics) │
│ │
└──────────────────────────────────────────────────────────────────────┘
This workflow diagram shows how a protein mixture is converted to peptides, separated, fragmented, and quantified by mass spectrometry — with the key branching decisions in acquisition mode and quantitation strategy that determine what biological questions can be answered.
---Key Takeaways
- Illumina sequencing uses reversible fluorescent terminators: Each cycle adds exactly one fluorescently labeled nucleotide per cluster; the dye is then cleaved and the 3′-OH unblocked before the next cycle. This "sequencing by synthesis" achieves >99.9% accuracy per base but is limited to ~300 bp reads because cluster synchrony degrades over cycles (Alberts et al., 2022).
- PacBio HiFi achieves long-read accuracy by circular consensus: A single molecule is read ~20× as the polymerase circles a SMRTbell (hairpin-adapter circularized) template; the consensus of these passes reduces error from ~15% (raw CLR) to <0.1% (HiFi). This is why HiFi is now preferred for de novo genome assembly — long reads with short-read accuracy (Alberts et al., 2022).
- N50 is the contig length at which 50% of the assembled genome is in contigs of that length or longer: It is a single-number summary of assembly continuity, not assembly correctness. A 10 Mb N50 means you could line up your longest contigs until they sum to half the genome, and the last one added was 10 Mb. Long-read technologies dramatically increase N50 by spanning repetitive regions that fragment short-read assemblies (Pevzner & Compeau, 2018).
- STAR and HISAT2 are splice-aware aligners critical for RNA-seq: Because mature mRNA lacks introns, reads that span splice junctions cannot be aligned with naive aligners. Both tools use pre-built splice-junction databases (from GTF annotation) plus on-the-fly junction discovery. STAR is faster on large genomes; HISAT2 uses a graph-based index that is more memory-efficient (Alberts et al., 2022).
- DESeq2 models RNA-seq count data with a negative binomial distribution: RNA-seq counts are overdispersed relative to Poisson (variance > mean), especially at low counts. DESeq2 estimates per-gene dispersion and then "shrinks" those estimates toward a fitted mean-dispersion curve using empirical Bayes — this borrowing of strength across genes is why DESeq2 performs well with as few as 2–3 replicates per condition, though 3–6 are strongly recommended (Alberts et al., 2022).
- Single-cell RNA-seq uses UMIs to eliminate PCR amplification bias: Each mRNA captured in a droplet (10x Genomics: cell barcode + UMI added during reverse transcription) receives a unique molecular identifier. After sequencing, reads with the same cell barcode, gene, and UMI are collapsed to a single count. This means UMI counts represent original molecule counts, not PCR copy counts — a fundamental advantage over bulk RNA-seq read counting (Alberts et al., 2022).
- UMAP (Uniform Manifold Approximation and Projection) for scRNA-seq preserves local structure better than tSNE for large datasets: Both are nonlinear dimensionality reduction methods. UMAP is faster and better preserves global structure (relative cluster positions), making it more appropriate for trajectory analysis. Neither should be used for quantitative distance interpretation — axes have no unit meaning (Alberts et al., 2022).
- Cas9 requires both guide RNA complementarity AND a PAM sequence (5′-NGG-3′ for SpCas9): Cas9 scans DNA by reading the PAM first; only after PAM recognition does it attempt R-loop formation with the guide. This is why target sites must end in NGG on the non-template strand. The PAM constraint limits targetability (~1 site per 8 bp in random sequence), motivating development of Cas9 orthologs with different PAMs (e.g., SaCas9: NNGRRT; Cas12a: TTTN) (Jinek et al., 2012).
- Base editors avoid double-strand breaks entirely: Cytosine base editors (CBEs) fuse nCas9 (D10A nickase) to a cytosine deaminase (APOBEC) and a uracil glycosylase inhibitor (UGI); APOBEC deaminates C→U within the R-loop editing window (positions 4–8 of the spacer), and UGI blocks base excision repair so U is read as T. Adenine base editors (ABEs) use an evolved E. coli tRNA adenosine deaminase (TadA) to convert A→I (read as G). Together they can theoretically correct ~58% of known pathogenic SNPs without requiring a DSB or donor template (Alberts et al., 2022).
- AlphaFold2's pLDDT score is a per-residue confidence metric, not a structural quality score: pLDDT (predicted local distance difference test) values >90 indicate very high confidence in local geometry; values <50 often indicate intrinsically disordered regions. Critically, low-pLDDT regions are not wrong predictions of structured regions — they are accurate predictions that a region is disordered. The PAE (predicted aligned error) matrix is the correct metric for evaluating confidence in relative domain orientations in multi-domain proteins (Jumper et al., 2021).
- Proximity labeling (BioID, TurboID, APEX2) maps protein neighborhoods in living cells: A biotin ligase or peroxidase fused to a protein of interest biotinylates all proteins within ~10 nm radius during a short labeling pulse. Biotinylated proteins are captured with streptavidin beads and identified by mass spectrometry. Unlike co-IP (which captures stable complexes), proximity labeling captures transient interactions and insoluble compartment residents — hence its power for mapping organelle proteomes (Alberts et al., 2022).
- ATAC-seq uses hyperactive Tn5 transposase to map open chromatin genome-wide: Tn5 loaded with sequencing adapters preferentially inserts into nucleosome-free regions (accessible chromatin). After PCR and sequencing, peaks of read density mark open chromatin (enhancers, promoters, insulators). Integration with ChIP-seq (which marks specific histone modifications or transcription factors) and RNA-seq can reveal the regulatory logic connecting chromatin state to gene expression (Alberts et al., 2022).
- SILAC is the gold standard for quantitative proteomics but cannot be applied to clinical tissue: Stable isotope labeling by amino acids in cell culture incorporates heavy (¹³C₆-Lys, ¹³C₆/¹⁵N₄-Arg) amino acids into all proteins over ~5 doublings. Heavy and light samples are mixed 1:1 before lysis, so all subsequent steps introduce identical bias to both. The mass difference (e.g., +6 Da for Lys) is detected in MS1. Because metabolic labeling requires cell division, SILAC is incompatible with post-mitotic tissues or biobank samples, where label-free or TMT approaches are used instead (Alberts et al., 2022).
- Biological networks are characterized by scale-free topology with a small number of highly connected hub proteins: In protein-protein interaction (PPI) networks, the degree distribution follows a power law: most proteins have few interactions, but a small number of "hubs" (e.g., TP53, HSP90) interact with dozens to hundreds of partners. Hub proteins are statistically enriched for essential genes — removing hubs is more likely to be lethal than removing peripheral nodes. This property, analyzed via STRING database, has direct implications for drug targeting and understanding disease gene function (Alberts et al., 2022).
References
- Alberts, B., Heald, R., Johnson, A., Morgan, D., Raff, M., Roberts, K., & Walter, P. Molecular Biology of the Cell, 7th ed. W.W. Norton, 2022.
- Pevzner, P., & Compeau, P. Bioinformatics Algorithms: An Active Learning Approach, 3rd ed. Active Learning Publishers, 2018.
- Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., Tunyasuvunakool, K., Bates, R., Žídek, A., Potapenko, A., Bridgland, A., Meyer, C., Kohl, S.A.A., Ballard, A.J., Cowie, A., Romera-Paredes, B., Nikolov, S., Jain, R., Adler, J., Back, T., Petersen, S., Reiman, D., Clancy, E., Zielinski, M., Steinegger, M., Pacholska, M., Berghammer, T., Bodenstein, S., Silver, D., Vinyals, O., Senior, A.W., Kavukcuoglu, K., Kohli, P., & Hassabis, D. 'Highly accurate protein structure prediction with AlphaFold.' Nature 596, 583–589, 2021. DOI: 10.1038/s41586-021-03819-2
- Jinek, M., Chylinski, K., Fonfara, I., Hauer, M., Doudna, J.A., & Charpentier, E. 'A programmable dual-RNA-guided DNA endonuclease in adaptive bacterial immunity.' Science 337(6096), 816–821, 2012. DOI: 10.1126/science.1225829