Graduate BCHM698 ⭐ 1 Credits Semester 1

Peptide and Protein Sciences Journal Club

A recurring weekly seminar in which students critically present and discuss primary research literature in peptide and protein sciences. Continues through all four semesters (BCHM698A-D). Presentations rotate among students with faculty discussion. Paper selection emphasizes recent high-impact publications in structural biology, peptide synthesis, therapeutics, and protein engineering.

📅 Studied by Skipper on 2026-10-02

📖 Core Concepts

GPT-6 Astra (Claude Opus 5.5 declined this section; written by GPT-6 Astra)

Core Concepts: Peptide and Protein Sciences Journal Club

A graduate journal club is not primarily a place to accumulate facts from recent papers. It is a place to learn how scientific knowledge is constructed—and how to determine which parts of that construction can bear weight. A paper may contain a beautiful structure, a convincing dose–response curve, an impressive synthesis, or a clinically important outcome. The central task is to identify exactly what the evidence establishes, what assumptions connect the observations to the authors’ interpretation, and what remains unresolved.

For me, Skipper, this course changes the direction of learning. In earlier courses, I usually began with an accepted principle and applied it to a problem. Here, I begin with measurements and ask whether they justify a principle. Thermodynamics, receptor pharmacology, synthetic chemistry, structural biology, and statistics become instruments of judgment rather than separate bodies of knowledge.

The organizing question is deceptively simple: What would have to be true for this result to support this claim? Answering it often reveals an entire chain of dependencies. A reported receptor-selective peptide must first have the claimed chemical identity. Its concentration must be known. The assay must measure the intended pathway. Apparent selectivity must not arise solely from unequal receptor expression. Finally, activity in an engineered cell line must translate into an effect at achievable exposure in an organism.

Working principle: Read every paper as a chain of claims connected by experiments. The strength of the conclusion is limited by the weakest essential connection—not determined by the most impressive figure.

This chapter develops that habit across the course’s fourteen topics. The aim is neither automatic skepticism nor automatic enthusiasm. It is calibrated belief: confidence proportional to the quality, relevance, and independence of the evidence.

1. Reading a Primary Paper as an Argument, Not a Story

Scientific articles are usually written as coherent narratives. Research itself is rarely so orderly. The published sequence of experiments may conceal failed approaches, exploratory analyses, changes in hypotheses, and practical constraints. A critical reader therefore reconstructs the logical argument rather than simply following the order of presentation.

The three-pass reading strategy

A useful approach adapts the three-pass strategy described by Keshav: first establish the paper’s identity and main claim, then inspect the evidence, and finally interrogate the details needed to reproduce or challenge the reasoning [Keshav, 2007].

  1. Pass one: establish the question. Read the title, abstract, introduction’s final paragraph, figure titles, and conclusion. Write one sentence each for the biological question, the experimental intervention, the main outcome, and the claimed advance.
  2. Pass two: reconstruct the evidence. Read the figures and legends carefully. For every panel, identify the measured variable, comparison, experimental unit, and inference. Ask what observation would have contradicted the preferred interpretation.
  3. Pass three: audit the methods. Examine sample preparation, controls, exclusions, statistical models, missing data, analytical code, and supplementary information. Read related work when necessary to judge novelty or resolve an assumption.

This strategy is iterative. A surprising figure should send me back to the methods; a methodological limitation may change how I interpret the abstract. “Figures first” is an efficient entry point, not permission to ignore how the figures were produced.

Separate observation, interpretation, and extrapolation

Suppose a paper reports that a peptide decreases inflammatory cytokine concentrations in cultured macrophages. The observation is a measured reduction under specified conditions. The interpretation might be inhibition of a signaling pathway. The extrapolation might be therapeutic benefit in inflammatory disease. These are three different claims requiring different evidence.

A cytokine decrease could reflect pathway inhibition, reduced cell viability, interference with the detection assay, altered cell number, or sequestration of the cytokine. Viability measurements, orthogonal detection, pathway perturbations, and appropriate normalization distinguish these possibilities. A disease claim additionally requires exposure, efficacy, and safety evidence in relevant biological settings.

Journal prestige is context, not a control experiment

Broad journals such as Nature, Science, and Cell often emphasize conceptual breadth and novelty. Specialty journals may provide deeper methodological treatment, more complete characterization, or important replication. Neither venue guarantees validity. A spectacular claim in a prestigious journal still requires appropriate controls; an incremental specialty paper may provide the most dependable foundation for a project.

For journal club, I should prepare a claim–evidence table before preparing presentation slides. This prevents the presentation from becoming an illustrated retelling of the authors’ preferred story.

2. Statistical Evidence: What Was Estimated, and How Uncertain Is It?

Statistics is not a final decorative layer applied to completed experiments. It connects the design to the claim. Before interpreting a p-value, I must identify the estimand: the quantity the analysis is intended to estimate. This might be a difference in mean body-weight change, a ratio of receptor potencies, a probability of achieving a clinical response, or a change in protein abundance.

Effect sizes come before significance labels

Difference in means:
Δ = mean(Ytreatment) − mean(Ycontrol)

Approximate standard error for independent group means:
SE(Δ) = √(s²treatment / ntreatment + s²control / ncontrol)

Approximate confidence interval:
Δ ± tcritical × SE(Δ)

These expressions assume a suitable sampling model and independent observations; paired, clustered, or longitudinal data require different treatment. Their conceptual value is that they separate magnitude from precision. A small effect can be estimated precisely, and a large apparent effect can be highly uncertain.

A p-value is the probability, under a specified null model, of obtaining a test statistic at least as incompatible with that model as the observed statistic. It is not the probability that the null hypothesis is true, that the finding occurred “by chance,” or that the result will replicate [Wasserstein and Lazar, 2016].

A frequentist 95% confidence procedure produces intervals that contain the fixed target parameter in 95% of repeated applications under its assumptions. Informally, the resulting interval describes values reasonably compatible with the data and model. It does not account automatically for unmeasured confounding, assay artifacts, or selective reporting.

The meaning of n

If one cell culture is divided among twelve assay wells, those wells are technical subsamples, not twelve independent biological experiments. Similarly, ten thousand cells collected from three animals do not ordinarily provide ten thousand independent animal-level replicates. Treating nested observations as independent creates pseudoreplication.

The correct experimental unit depends on where treatment was independently assigned. Repeated measurements can improve precision, but they do not manufacture independent biological variation. Hierarchical models may retain cell-level or well-level information while respecting animal, donor, plate, or batch structure.

Multiplicity, missingness, and negative results

  • Multiplicity: Testing many proteins, pathways, doses, or subgroups increases opportunities for false-positive findings. Confirmatory analyses need a prespecified error-control strategy; omics discovery commonly uses false-discovery-rate control.
  • Missingness: Dropout caused by adverse effects is not necessarily comparable to an accidentally missed measurement. The analysis must state how missing outcomes are handled and which assumptions support that handling.
  • Negative findings: Failure to reject a null hypothesis does not establish equivalence. Equivalence or noninferiority requires an appropriate design and a scientifically justified margin.
  • Subgroups: “Significant in one subgroup but not another” does not demonstrate a subgroup difference. The relevant question usually requires an interaction test.

The practical lesson from MATH201 is that uncertainty is a feature of inference, not a defect to hide. The best discussion asks whether plausible uncertainty changes the scientific or clinical decision.

3. Structural Biology: A Structure Is an Evidence-Constrained Model

A molecular structure is often presented as though it were a direct photograph. In reality, it is a model inferred from experimental observations, constrained by chemistry, and selected through computational procedures. Cryo-electron microscopy makes this especially clear: particle images are aligned and classified, a three-dimensional density map is reconstructed, and an atomic model is fitted and refined.

Evaluating a peptide–GPCR cryo-EM structure

The first question is not merely “What is the resolution?” It is “What biological state was prepared, and where is the evidence strong enough to support the claimed interaction?” A peptide–GPCR complex may include receptor mutations, truncated termini, a G protein or engineered mimic, a stabilizing antibody fragment, and a detergent or lipid environment. These choices can make the experiment possible while also changing the conformational ensemble being observed.

Global resolution compresses a heterogeneous object into one number. A receptor’s transmembrane core may be well resolved while the peptide terminus or extracellular loops remain ambiguous. Consequently, a proposed salt bridge or hydrogen-bond network requires scrutiny of local density, side-chain placement, geometry, and alternative interpretations.

Fourier shell correlation between independently reconstructed half-maps is a central resolution assessment. Appropriate half-set separation helps expose overfitting, although masking, processing choices, and refinement practices still matter [Scheres and Chen, 2012]. Model validation should additionally examine stereochemistry, clashes, rotamers, map–model agreement, and whether poorly supported regions have been overinterpreted.

Structural proximity does not establish energetic importance

A ligand residue touching a receptor residue does not prove that the contact contributes strongly to affinity. Binding free energy includes desolvation, entropy, conformational reorganization, and interactions throughout the complex. A mutation can also impair folding or cell-surface expression rather than remove one local contact.

For dissociation constant Kd and standard concentration C°:

ΔG°bind = RT ln(Kd / C°)

Mutation-induced change:
ΔΔG°bind = RT ln(Kd,mutant / Kd,wild type)

The ratio expression is dimensionless when both dissociation constants use the same units. Interpreting it as the energy of one contact requires strong assumptions: comparable folded states, unchanged stoichiometry, reliable equilibrium measurements, and no major structural rearrangement.

AlphaFold2 and experimental structures answer different questions

AlphaFold2 demonstrated remarkably accurate protein structure prediction on many targets, especially well-ordered domains [Jumper et al., 2021]. Its confidence measures must nevertheless be read correctly:

  • pLDDT estimates local prediction confidence; it is not a direct measurement of thermal motion or biological flexibility.
  • Predicted aligned error helps assess confidence in relative placement of regions or domains.
  • A high-confidence fold does not establish the relevant ligand-bound state, oligomeric assembly, protonation pattern, catalytic mechanism, or conformational populations.

Low confidence may reflect disorder, weak evolutionary information, or modeling difficulty; it does not prove that a region lacks function. The beautiful connection to statistical mechanics is that biological function depends on an ensemble and its populations, whereas most structural figures show one representative configuration.

4. Peptide Synthesis Innovations: Better According to Which Objective?

An automated synthesizer, flow reactor, or new coupling reagent can produce an impressive chromatogram without solving the problem that matters industrially. A synthesis paper must be evaluated against a multidimensional objective: identity, purity, isolated yield, stereochemical integrity, throughput, material consumption, reproducibility, and scalability.

Why small inefficiencies compound

Idealized fraction of full-length product after m coupling steps:
F ≈ p^m

If p = 0.99 and m = 50:
F ≈ 0.99^50 ≈ 0.605

If p = 0.999 and m = 50:
F ≈ 0.999^50 ≈ 0.951

This model assumes identical independent coupling efficiencies and ignores additional reactions. Even so, it explains why apparently excellent per-step performance may generate a difficult crude mixture for a long peptide. Deletions, incomplete deprotection, oxidation, aspartimide formation, epimerization, and cleavage-related modifications make the real problem harder.

From BCHM602, I already know the chemistry of Fmoc-based solid-phase synthesis. Journal club adds a different question: Which failure mode has the new method actually reduced? Faster coupling may help throughput but worsen a heat-sensitive side reaction. Improved mass transfer may benefit one sequence yet fail for aggregation-prone resin-bound chains.

Flow chemistry and automation

Automated fast-flow synthesis can greatly accelerate peptide assembly through controlled reagent delivery, heating, and rapid exchange. Hartrampf and colleagues demonstrated the reach of this approach for chemical protein synthesis [Hartrampf et al., 2020]. A critical appraisal asks whether synthesis time includes deprotection, washing, cleavage, workup, and purification—not just the coupling interval.

Scale also changes transport. Temperature gradients, resin swelling, back pressure, mixing, reagent residence time, and diffusion into beads may behave differently in a larger system. A reaction that works at analytical scale is not automatically a manufacturing process.

Purity, yield, and sustainability are distinct

  • Crude purity describes the mixture before purification.
  • Isolated yield requires a defined basis and an actual recovered amount.
  • Chromatographic area purity depends on detection response and does not universally equal mass purity.
  • Correct intact mass does not exclude epimers, positional isomers, or some connectivity errors.
  • Process mass intensity measures total material input relative to isolated product and exposes solvent- or reagent-intensive “fast” processes.
Process mass intensity:
PMI = total mass of materials used / mass of isolated product

For industrial relevance, I should look for challenging sequences, multiple batches, realistic purification burdens, residual impurity characterization, and a fair benchmark. A method is most convincing when it improves the complete process rather than one attractive step.

5. Clinical Peptide Pharmacology: From Molecular Activity to Patient Benefit

A clinical trial tests an intervention strategy in a defined population. It does not automatically establish the molecular reason that strategy works. This distinction is essential when discussing incretin therapeutics, where receptor pharmacology, dose escalation, adherence, tolerability, and behavioral effects all influence outcomes.

Reading the tirzepatide obesity trial precisely

SURMOUNT-1 randomized 2,539 adults with obesity, or overweight with a weight-related complication, without diabetes, to once-weekly tirzepatide at target doses of 5, 10, or 15 mg or placebo for 72 weeks. Participants also received lifestyle intervention. Tirzepatide is a dual GIP- and GLP-1-receptor agonist, not simply another selective GLP-1-receptor agonist [Jastreboff et al., 2022].

Under the reported treatment-regimen estimand, mean body-weight changes at week 72 were −15.0%, −19.5%, and −20.9% for the three tirzepatide groups, compared with −3.1% for placebo. The 15 mg comparison therefore represents a placebo-adjusted difference of approximately −17.8 percentage points, not a 17.8% relative reduction. These quantities answer different questions.

The treatment-regimen estimand addresses outcomes under assignment to the treatment strategy regardless of treatment discontinuation. An efficacy estimand instead targets effects under more restrictive assumptions about treatment adherence and relevant intercurrent events. Neither should be reduced to the simplistic labels “all patients” versus “completers”: both depend on explicit definitions and analytical procedures.

Why trial architecture matters

  • Randomization balances measured and unmeasured baseline factors in expectation.
  • Allocation concealment prevents foreknowledge of assignment from influencing enrollment.
  • Blinding reduces some behavioral and assessment biases, although distinctive adverse effects can compromise it.
  • Prespecified endpoints limit selective emphasis on favorable results.
  • Follow-up and missing-data analyses determine how convincingly the trial addresses outcomes after discontinuation.

Gastrointestinal adverse events were common in SURMOUNT-1, especially during dose escalation. Discontinuation because of adverse events occurred in 4.3%, 7.1%, and 6.2% of participants in the tirzepatide groups versus 2.6% with placebo [Jastreboff et al., 2022]. These data belong beside efficacy, not in a brief afterthought. A treatment that works biologically but cannot be tolerated by a patient has limited practical effectiveness.

Mechanism is broader than one signaling arrow

GLP-1-receptor activation affects glucose-dependent insulin secretion, glucagon regulation, appetite-related pathways, and gastric emptying. Gastric-emptying effects can change with repeated exposure. Tirzepatide’s clinical outcome does not, by itself, quantify how much benefit arises from each receptor or prove receptor “synergy.” Such claims require appropriate mechanistic comparators and careful control of exposure.

Finally, a 72-week weight-loss trial does not independently establish lifelong safety, post-withdrawal durability, or cardiovascular benefit. Important findings should remain important without being expanded beyond the questions the trial actually tested.

6. Protein Folding and Disease: Distinguish Pathway, Structure, and Toxic Species

Amyloid biology shows why an elegant in vitro mechanism may not map directly onto disease. Aggregation is a network of competing pathways, not necessarily a single conversion from monomer to a toxic fibril. Primary nucleation, elongation, fragmentation, and surface-catalyzed secondary nucleation can contribute differently under different conditions [Knowles, Vendruscolo, and Dobson, 2014].

Illustrative aggregation processes:

Primary nucleation:
nM → nucleus

Elongation:
fibril end + M → elongated fibril

Secondary nucleation:
fibril surface + nM → new nucleus + fibril surface

Fragmentation:
one fibril → two or more shorter fibrils

Here, M denotes soluble monomer; these are schematic processes, not universal elementary reactions. Fragmentation can accelerate aggregation by increasing the number of elongation-competent ends without initially increasing total aggregate mass.

A lag phase does not identify one mechanism

A sigmoidal fluorescence trace is consistent with several kinetic models. A shortened lag phase after seeding supports a role for pre-existing assemblies, but does not alone distinguish elongation from secondary nucleation. Mechanistic discrimination requires concentration dependence, controlled seed amount and size, global fitting, and independent measurements.

Thioflavin T fluorescence is useful, but it is not a universal mass balance for amyloid. Signal depends on accessible binding sites and fibril structure; compounds may alter dye binding or fluorescence directly. An apparent aggregation inhibitor could therefore be an optical interferent. Orthogonal methods might include sedimentation, microscopy, spectroscopy, or measurements of soluble species.

The most abundant species need not be the most harmful

Mature fibrils are comparatively easy to image, while transient oligomers can be heterogeneous and difficult to isolate without perturbation. A treatment that reduces visible fibrils could increase another assembly population. Conversely, sequestration into a relatively less harmful deposit could reduce exposure to a damaging soluble species. Neither possibility should be assumed without evidence.

Connecting an in vitro result to disease requires attention to concentration, compartment, membrane surfaces, post-translational modifications, proteostasis, clearance, and exposure time. Patient-derived fibrils may also differ structurally from assemblies made from recombinant material under convenient laboratory conditions.

The key connection to BCHM603 is that thermodynamic favorability does not determine the rate of reaching a state. Disease may depend on rare kinetic events, metastable populations, or impaired clearance rather than simply on the lowest-free-energy structure.

7. AI-Designed Proteins: Prediction, Generation, and Validation Are Different Tasks

Computational protein engineering includes several distinct problems. Structure prediction asks what structure a sequence is likely to adopt. Sequence design asks which sequences are compatible with a desired structural environment. Backbone generation proposes new geometries satisfying specified constraints.

ProteinMPNN is a sequence-design method conditioned on structural information [Dauparas et al., 2022]. RFdiffusion generates protein structures under various design conditions [Watson et al., 2023]. ESMFold predicts structures using protein language-model representations; it should not be described as intrinsically equivalent to a de novo backbone generator [Lin et al., 2023].

Why self-consistency is not enough

A common workflow generates a backbone, designs a sequence, predicts that sequence’s structure, and compares the prediction with the intended backbone. Agreement is useful, but it is not independent experimental validation. The tools may share training data, representational biases, or preferences for easily predicted structures.

A design pipeline can be internally consistent and biologically wrong. Agreement among models is strongest when followed by evidence from methods that fail in different ways.

A ladder of experimental evidence

  1. Expression and solubility: Can the protein be produced in a usable form?
  2. Identity and homogeneity: Is it the intended sequence, and is the sample monodisperse rather than aggregated?
  3. Folding and stability: Do spectroscopy and appropriate thermodynamic measurements support a structured state?
  4. Structural correspondence: Does an experimental structure resemble the intended design?
  5. Function: Does it bind, catalyze, signal, or assemble as intended?
  6. Specificity and robustness: Does function survive relevant competitors, matrices, and environmental conditions?

A cooperative thermal transition supports an organized state but does not uniquely establish the intended fold. A tight binding signal can be distorted by aggregation, avidity, or surface immobilization. An enzyme design needs evidence of product formation, turnover, substrate dependence, and appropriate controls—not merely substrate binding.

Always ask for the denominator

If a paper presents five successful binders, how many sequences were generated, synthesized, expressed, and tested? These denominators describe different success rates. A large computational screen followed by a small experimental screen may still be valuable, but its efficiency cannot be judged from showcase examples alone.

Data leakage and benchmark similarity also matter. Performance on genuinely novel folds or distant targets is not the same as performance near training examples. The strongest studies make selection rules, failures, code, sequences, and experimental criteria visible enough for others to assess generalization.

8. Proteomics Methods: Measurement Performance Is More Than Identification Count

A proteomics methods paper often leads with the number of identified proteins. That number matters, but it is only one dimension of performance. A useful method must also provide reliable identification, quantitative accuracy, precision, coverage, dynamic range, reproducibility, and manageable missingness.

The measured signal has a long causal history

Protein abundance
  → extraction recovery
  → digestion efficiency
  → peptide recovery
  → chromatographic separation
  → ionization and transmission
  → precursor selection or sampling
  → fragmentation
  → computational identification and quantification

A change anywhere along this chain can change the final signal. Observed peptide intensity is therefore not automatically proportional to original protein abundance across different peptides or conditions. Stable-isotope standards help, but their point of introduction matters: a peptide added after digestion cannot correct incomplete digestion of the original protein.

Sampling strategy shapes the data

Data-dependent acquisition selects precursors according to rules applied during the run, often favoring intense ions. Data-independent acquisition fragments broader precursor windows and uses computational analysis to disentangle signals. DIA can improve consistency, but its advantages depend on acquisition settings, interference handling, libraries or prediction models, and statistical validation. DIA-NN illustrates how computational methods can improve interpretation of complex DIA data [Demichev et al., 2020].

False discovery is hierarchical

False-discovery-rate control may be applied at the peptide-spectrum-match, peptide, or protein level. A stated 1% FDR at one level does not automatically imply 1% at every other level. Nor does it mean each accepted identification has exactly a 99% probability of being correct. Protein inference becomes especially difficult when peptides are shared among isoforms or homologous proteins.

PTM analysis adds another layer. Correctly identifying a phosphopeptide is different from correctly localizing its phosphorylation site. Enrichment can introduce sequence and chemical biases, and an increase in modified-peptide abundance may reflect increased total protein rather than increased modification occupancy.

Modification occupancy:
occupancy = modified molecules / total molecules at that site

A change in modified-peptide signal alone
does not uniquely determine a change in occupancy.

What makes validation convincing?

  • Dilution series or defined mixtures that test expected ratios.
  • Precision across preparations, days, operators, and instruments.
  • Assessment of interference, carryover, linear range, and low-abundance performance.
  • Realistic sample matrices rather than only unusually clean standards.
  • Public raw data, analysis settings, and sufficiently reproducible computational workflows.

The connection to CHEM302 is direct: sensitivity is not the same as accuracy, and a large number of measurements is not the same as a trustworthy measurement system.

9. Melanocortin and Neuropeptide Pharmacology: Bias Is a Comparative Quantity

Neuropeptide receptors connect molecular recognition to behavior, autonomic physiology, endocrine regulation, and energy balance. Their effects depend on receptor subtype, tissue distribution, cellular signaling machinery, and access to the relevant compartment. A potent response in a transfected cell is only the beginning of pharmacological understanding.

Affinity, potency, and efficacy

Simple equilibrium occupancy:
fraction occupied = [L] / (Kd + [L])

Empirical concentration–response model:
E = Ebaseline + (Emax − Ebaseline)[L]^nH /
    (EC50^nH + [L]^nH)

The occupancy equation assumes a simple binding system and free ligand concentration. The response equation is a phenomenological model; its Hill slope is not automatically a count of binding sites. Most importantly, EC50 is not generally equal to Kd. Receptor reserve and downstream amplification can permit half-maximal response at low occupancy.

What biased agonism actually means

A ligand may preferentially activate one pathway relative to another compared with a reference ligand in a defined experimental system. Bias is therefore not established merely because one assay shows a lower EC50 than another. Different assays have different amplification, kinetics, receptor expression, and detection limits.

An operational-model approach can summarize activity through a transduction coefficient such as log(τ/KA). First compare the test ligand with a reference within each pathway, then compare those relative values across pathways [Kenakin et al., 2012].

Within pathway A:
ΔA = log(τ/KA)test,A − log(τ/KA)reference,A

Between pathways:
ΔΔlog(τ/KA) = ΔA − ΔB

Bias factor = 10^[ΔΔlog(τ/KA)]

The result depends on model suitability and parameter identifiability; uncertainty should propagate through the calculation. Time is another complication: rapid signaling followed by desensitization can produce different apparent bias at different measurement times.

Melanocortin selectivity and clinical translation

Melanocortin receptors share ligands and structural features but mediate different physiological functions. MC4R has a central role in energy balance, whereas activity at other subtypes can produce additional effects. Clinical results with setmelanotide in selected genetic obesity syndromes illustrate the importance of matching a receptor mechanism to a biologically defined population [Clément et al., 2020].

Those findings do not imply identical efficacy in all forms of obesity or for every melanocortin-active peptide. Receptor selectivity must be interpreted alongside achievable free concentrations, tissue access, genotype, and safety. “More signaling” is not a universal therapeutic objective; the objective is the right signaling in the right cells for the right duration.

10. Antibody Engineering: Geometry and Exposure Can Matter as Much as Affinity

Antibody engineering makes a crucial principle visible: binding is not function. Two antibodies with similar equilibrium affinities can differ in epitope, binding geometry, dissociation kinetics, internalization, Fc activity, tissue penetration, and target-mediated clearance.

Bispecifics create a spatial problem

A bispecific antibody may recruit an immune cell to a target cell or engage two targets simultaneously. Productive activity depends on more than the affinity of each arm. Epitope height, intermembrane distance, linker flexibility, valency, and the geometry of the resulting complex can determine whether binding produces useful signaling.

High affinity can even be counterproductive if it increases binding to low-density antigen on healthy tissue, traps the molecule near vessels, or reduces serial engagement. Potency should therefore be evaluated alongside selectivity across realistic antigen densities and relevant normal cells.

Antibody–drug conjugates are integrated chemical systems

ADC activity depends on:

antigen recognition
  → tissue distribution
  → cellular uptake or local processing
  → linker cleavage / payload release
  → payload access to its intracellular target
  → cellular response

Drug-to-antibody ratio affects payload delivery but also hydrophobicity, aggregation, clearance, and heterogeneity. A linker must balance stability during circulation against release in the intended setting. Membrane-permeable released payloads can produce bystander killing, which may help heterogeneous tumors while also broadening toxicity risks.

Nanobodies expose a size–exposure trade-off

Small antibody domains can access otherwise difficult epitopes and may penetrate tissues differently from full-length antibodies. Their small size can also produce rapid renal clearance. Albumin binding or multimerization can extend exposure, but changes the distribution and clearance properties that motivated the small format.

For any engineered antibody, a convincing paper should distinguish monovalent affinity from avidity, assess aggregation and nonspecific binding, and compare in vitro potency with exposure in vivo. Human-target xenografts may not reveal toxicity if the reagent fails to bind the corresponding mouse antigen. An apparently wide therapeutic window can therefore be an artifact of species selectivity rather than a property of the therapeutic.

11. Peptide Drug Delivery: Exposure Is a Mass-Balance Problem

A delivery system must move an intact, active molecule to its target at a useful concentration for a useful duration. Oral peptides encounter chemical and enzymatic degradation, limited epithelial permeability, and presystemic loss. Nanoparticles face additional questions about loading, release, tissue distribution, clearance, and whether the measured signal represents intact cargo or merely a label.

Core pharmacokinetic relationships

Absolute bioavailability:
F = (AUCnon-IV / Dosenon-IV) / (AUCIV / DoseIV)

For linear pharmacokinetics:
AUC = F × Dose / CL

For an appropriate one-compartment elimination model:
t1/2 = ln(2) × V / CL

The bioavailability comparison requires appropriate dose normalization and comparable conditions; nonlinear pharmacokinetics complicates it. The half-life expression must use the volume appropriate to the model. For subcutaneous depots, a long terminal phase may reflect slow absorption rather than slow elimination—so-called flip-flop kinetics.

What an oral formulation must demonstrate

Oral semaglutide provides a concrete example of formulation-assisted absorption. Mechanistic work with its absorption enhancer, SNAC, supported localized gastric conditions promoting peptide protection and transcellular uptake [Buckley et al., 2018]. This should not be generalized into the claim that all peptide enhancers work through the same mechanism or that successful absorption is insensitive to dosing conditions.

A rigorous delivery paper should examine absolute rather than only relative bioavailability, interindividual variability, food effects, dosing volume, intact-drug quantification, and repeated-dose local safety. A several-fold increase from a tiny baseline may remain pharmacologically inadequate.

Beware of misleading tracking experiments

Fluorescence in an organ does not establish delivery of intact active peptide. The fluorophore may detach, the carrier may remain trapped in endosomes, or degradation products may retain signal. Orthogonal chemical measurements and functional target engagement are essential.

Encapsulation efficiency also differs from loading capacity: a carrier can encapsulate nearly all the starting peptide yet contain very little peptide per unit carrier mass. Translational feasibility requires a realistic administered mass, scalable manufacturing, storage stability, acceptable excipients, and release kinetics compatible with the target. A formulation is not successful merely because it increases a curve; it must improve an exposure–benefit–risk relationship.

12. Metabolic Signal Transduction: Follow the Causal Chain Across Scales

Metabolic disease studies connect receptor activation to adipose tissue, liver, pancreas, gut, and brain. The challenge is that these organs communicate. A hepatic change after peptide treatment may be secondary to reduced food intake, altered insulin, changed body weight, or autonomic signaling rather than direct activation of hepatocyte receptors.

Example causal alternatives:

Drug → hepatic receptor → altered hepatic metabolism

Drug → reduced food intake → weight loss
     → altered hepatic metabolism

Drug → pancreatic / CNS effects → altered hormones or neural output
     → altered hepatic metabolism

All three pathways might operate simultaneously. Observing the final metabolic outcome does not identify their relative contributions.

Controls should match the causal question

Pair-fed controls help test whether reduced caloric intake can account for an outcome, but they are not perfect replicas of drug-treated animals. Meal timing, stress, nutrient absorption, energy expenditure, and weight trajectories may differ. Weight-matched controls ask a related but distinct question.

Receptor antagonism or genetic deletion can strengthen causal inference, but each brings assumptions. An antagonist must reach the relevant tissue and act selectively. A constitutive knockout may produce developmental compensation. Conditional deletion requires validation of cell specificity, deletion efficiency, and timing.

Expression and target engagement are not interchangeable

A low-abundance transcript does not prove sufficient functional receptor protein at the cell surface. Antibody staining can be misleading without specificity validation. Conversely, sparse receptor-expressing cells may exert major physiological effects. Orthogonal localization and functional perturbation are more informative than any single expression assay.

This topic connects systems biology to classical physiology: a small molecular perturbation can be amplified by feedback, while an apparently large signaling change can be buffered by homeostasis. Translational judgment therefore requires matching dose, exposure, tissue, disease state, and timescale—not simply matching the direction of one biomarker.

13. Research Peptides: Evaluate the Evidence Without Borrowing Certainty

Literature on BPC-157, products marketed as TB-500, and related “research peptides” requires especially careful separation of chemical identity, preclinical findings, human efficacy, product quality, and regulatory status. These are independent questions. Evidence for one should not be allowed to stand in for the others.

First establish what substance was studied

Commercial names can obscure sequence differences, fragments, modifications, salts, and formulation variability. TB-500 is not a sufficiently precise chemical description to assume that every marketed product is identical to full-length thymosin β4 or to the material used in a particular experiment. Biological conclusions must follow the actual sequence and preparation, not the marketplace label.

Animal studies can be informative without being clinically decisive

A sound preclinical appraisal asks whether allocation was randomized, outcome assessment blinded, sample size justified, exclusions prespecified, and results independently replicated. Sex, age, injury model, dose, route, and timing also matter. Treatment immediately before a controlled experimental injury tests a different proposition from treatment of chronic human disease.

Multiple favorable outcomes from the same animals are not independent replications. Neither are several papers from closely related groups necessarily equivalent to broad independent confirmation. Publication bias can make a literature look uniformly positive when unsuccessful experiments remain unseen.

Mechanistic claims deserve the same discipline. Increased expression of a repair-associated marker does not establish improved tissue function, durable healing, or acceptable systemic safety. Small studies are especially poorly suited to excluding uncommon or delayed harms.

Regulatory and clinical distinctions

The available preclinical literature does not by itself establish safe and effective human dosing for unapproved products. “Research use only” is not a quality guarantee for human administration, and commercial availability is not regulatory approval. FDA materials have raised concerns about certain bulk peptide substances, including issues involving immunogenicity, impurities, and insufficient safety information; the precise status must be checked for the exact substance and jurisdiction [FDA, n.d.].

The appropriate conclusion is neither that every proposed effect is impossible nor that all peptides in this category are proven harmful. It is that plausible mechanisms and positive animal findings cannot substitute for well-controlled human evidence and reliable manufacturing. Uncertainty should remain visible rather than being converted into a treatment claim.

14. Student-Led Appraisal and Emerging Modalities: Turn Critique into a Research Question

A strong student-selected presentation should do more than list limitations. It should identify the most consequential uncertainty and propose an experiment that can reduce it. Paper selection is therefore strategic: choose a study with a clear central claim, interpretable primary data, accessible methods, and a genuine connection to the thesis question.

Build the presentation around a decision

  1. State the question: Why does this uncertainty matter?
  2. Draw the claim chain: Which observations support which conclusions?
  3. Inspect decisive figures: Which data would change my belief most?
  4. Identify alternatives: What other mechanism explains the same findings?
  5. Propose a discriminating experiment: What outcomes would favor each explanation?
  6. State the updated belief: What is established, plausible, unresolved, or unsupported?

Peer critique works best when colleagues first agree on what the paper actually claims. A fair criticism distinguishes an experiment necessary for the main conclusion from an interesting extension that lies outside the study’s scope.

mRNA-encoded peptides and proteins

Delivering mRNA changes the controlled input from administered protein mass to an intracellular template. Expression depends on delivery, cellular uptake, endosomal escape, RNA stability, translation, processing, and secretion. A peptide requiring amidation, proteolytic maturation, or unusual disulfide connectivity may not be produced correctly in every transfected cell type.

This modality may provide sustained production, but it complicates reversibility and dose control. The relevant question is not simply how much RNA was administered; it is which cells produced how much active product, where it went, and for how long.

Peptide-based targeted degradation

PROTACs recruit a target protein and an E3 ubiquitin ligase into a productive complex, leading to ubiquitination and proteasomal degradation. Peptide-based recognition elements can offer useful specificity but introduce stability, permeability, and delivery challenges [Békés, Langley, and Crews, 2022].

Target + degrader + E3 ligase
  ⇌ productive ternary complex
  → target ubiquitination
  → proteasomal degradation

Protein depletion reflects both:
degradation rate and protein resynthesis rate

Binary affinity alone does not predict degradation. Ternary-complex geometry, accessible lysines, subcellular colocalization, and cellular machinery matter. Appropriate controls include inactive analogues and tests of dependence on the recruited ligase and proteasome. Decreased protein abundance must also be distinguished from reduced transcription or nonspecific cell damage.

Orally active cyclic peptides

Cyclization can improve proteolytic resistance and constrain conformation, but it does not automatically produce permeability. N-methylation, stereochemistry, intramolecular hydrogen bonding, exposed polarity, and conformational adaptability determine how a peptide behaves in water, membranes, and the target site.

The oral macrocyclic PCSK9 inhibitor MK-0616 provided clinical evidence that suitably engineered cyclic peptides can address targets traditionally approached with injectable biologics [Ballantyne et al., 2023]. The broader opportunity is not “make every peptide cyclic,” but learn which molecular and formulation features jointly enable oral exposure and target engagement.

Across emerging modalities, the best research opportunities lie where a persuasive mechanistic idea meets a clearly defined bottleneck: processing, delivery, specificity, reversibility, manufacturability, or a missing clinical comparison.

Connections to Peptide Chemistry

The recurring lesson of this course is that chemical details do not disappear as research moves toward physiology and medicine. They become hidden variables unless someone deliberately tracks them.

Sequence identity is the beginning of reproducibility

A peptide’s name is not its complete specification. Sequence, stereochemistry, terminal groups, disulfide connectivity, conjugation site, counterion, water content, purity, and concentration all affect interpretation. A nominal mass weighed from a vial may include salts and water, so it need not equal the mass of active peptide. Errors here propagate into potency and exposure estimates.

Modifications couple molecular and physiological properties

  • D-amino acids can reduce susceptibility to some proteases while altering receptor recognition and local conformation.
  • N-methylation changes hydrogen-bond donation and conformational preferences; its effects on permeability are context-dependent.
  • Cyclization can reduce conformational entropy costs of binding, but can also trap an unfavorable geometry.
  • Lipidation may extend exposure through albumin association while changing free fraction, distribution, and receptor access.
  • PEGylation or protein fusion can reduce renal clearance but may impair access to targets or alter immunogenicity.

This is the practical meaning of structure–property relationships. No modification is simply “good for stability” or “good for delivery” in isolation. Every change shifts a set of coupled equilibria and kinetic processes.

Potency is not the same as useful drug action

A useful development objective is not simply:
minimize EC50

It is closer to:
achieve sufficient target engagement
for sufficient time
at tolerable exposure
with reproducible manufacturing and administration.

A peptide with lower cellular potency can outperform a more potent analogue if it has better stability, more predictable exposure, or fewer off-target effects. Conversely, prolonged total plasma exposure may be unhelpful if almost none of the molecule reaches the target in an active form.

The beautiful unifying idea is that a therapeutic peptide is simultaneously a molecule, a population of conformations, an analytical sample, a pharmacokinetic entity, and an intervention in a living system. Journal club teaches me to move between these descriptions without confusing them.

Common Misconceptions to Watch Out For

  • “The abstract tells me what the data prove.” The abstract tells me the authors’ compressed interpretation. The proof burden lies in the design, measurements, and analysis.
  • “A small p-value means a large or important effect.” Statistical incompatibility with a null model is different from biological magnitude and clinical value.
  • “More wells, cells, or spectra always mean a larger n.” Independent experimental units, not raw observation counts, determine biological replication.
  • “The highest-resolution structure reveals the whole mechanism.” Structures sample selected states; dynamics, energetics, and causal perturbations remain necessary.
  • “High AlphaFold confidence establishes biological function.” Confidence in predicted geometry does not establish ligand recognition, catalytic competence, state populations, or physiological relevance.
  • “A clean HPLC peak and correct mass establish complete chemical identity.” Some isomers, epimers, and connectivity errors require additional characterization.
  • “A successful designed protein validates the entire computational method.” General performance requires denominators, benchmarks, failures, and independent testing.
  • “More identified proteins means a better proteomics method.” Quantitative accuracy, FDR control, reproducibility, and sample relevance may matter more.
  • “EC50 measures affinity.” Response potency also depends on efficacy, receptor density, amplification, kinetics, and assay design.
  • “Different pathway potencies prove biased agonism.” Bias requires a reference ligand, appropriate cross-pathway normalization, and attention to system effects.
  • “Stronger antibody binding always improves therapy.” Geometry, antigen density, distribution, release mechanisms, and off-target engagement can reverse that intuition.
  • “Longer half-life always means better delivery.” Long persistence may reflect inactive material, strong sequestration, or absorption-limited kinetics.
  • “A randomized clinical outcome proves a specific molecular mechanism.” Randomization supports causal inference about the assigned intervention, not every proposed intermediate mechanism.
  • “Positive animal studies establish human efficacy and safety.” Species, exposure, disease context, product quality, and trial design remain major translational barriers.
  • “Criticism means dismissing the paper.” Good appraisal preserves genuine advances while marking the boundaries of what has been established.

My final checkpoint: Can I explain what was measured, why the controls matter, which assumptions connect the measurements to the claim, and what experiment would most change my confidence? If I cannot, I have recognized the paper’s vocabulary—but I have not yet understood its argument.

Selected References

  • Ballantyne, C. M., et al. (2023). “Efficacy and Safety of the Oral PCSK9 Inhibitor MK-0616: A Phase 2b Randomized Controlled Trial.” Journal of the American College of Cardiology.
  • Békés, M., Langley, D. R., and Crews, C. M. (2022). “PROTAC targeted protein degraders: the past is prologue.” Nature Reviews Drug Discovery, 21, 181–200.
  • Buckley, S. T., et al. (2018). “Transcellular stomach absorption of a derivatized glucagon-like peptide-1 receptor agonist.” Science Translational Medicine, 10, eaar7047.
  • Clément, K., et al. (2020). “Efficacy and safety of setmelanotide, an MC4R agonist, in individuals with severe obesity due to LEPR or POMC deficiency: single-arm, open-label, multicentre, phase 3 trials.” The Lancet Diabetes & Endocrinology, 8, 960–970.
  • Dauparas, J., et al. (2022). “Robust deep learning–based protein sequence design using ProteinMPNN.” Science, 378, 49–56.
  • Demichev, V., et al. (2020). “DIA-NN: neural networks and interference correction enable deep proteome coverage in high throughput.” Nature Methods, 17, 41–44.
  • Hartrampf, N., et al. (2020). “Synthesis of proteins by automated flow chemistry.” Science, 368, 980–987.
  • Jastreboff, A. M., et al. (2022). “Tirzepatide Once Weekly for the Treatment of Obesity.” New England Journal of Medicine, 387, 205–216. DOI: 10.1056/NEJMoa2206038.
  • Jumper, J., et al. (2021). “Highly accurate protein structure prediction with AlphaFold.” Nature, 596, 583–589. DOI: 10.1038/s41586-021-03819-2.
  • Kenakin, T., et al. (2012). “A simple method for quantifying functional selectivity and agonist bias.” ACS Chemical Neuroscience, 3, 193–203.
  • Keshav, S. (2007). “How to Read a Paper.” ACM SIGCOMM Computer Communication Review, 37(3), 83–84.
  • Knowles, T. P. J., Vendruscolo, M., and Dobson, C. M. (2014). “The amyloid state and its association with protein misfolding diseases.” Nature Reviews Molecular Cell Biology, 15, 384–396.
  • Lin, Z., et al. (2023). “Evolutionary-scale prediction of atomic-level protein structure with a language model.” Science, 379, 1123–1130.
  • Scheres, S. H. W., and Chen, S. (2012). “Prevention of overfitting in cryo-EM structure determination.” Nature Methods, 9, 853–854.
  • U.S. Food and Drug Administration (n.d.). “Safety Risks Associated with Certain Bulk Drug Substances Nominated for Use in Compounding.” Regulatory webpage; consult the current version for substance-specific information.
  • Wasserstein, R. L., and Lazar, N. A. (2016). “The ASA’s Statement on p-Values: Context, Process, and Purpose.” The American Statistician, 70, 129–133.
  • Watson, J. L., et al. (2023). “De novo design of protein structure and function with RFdiffusion.” Nature, 620, 1089–1100.

⚙️ Mechanisms & Applications

GPT-6 Astra

Mechanisms & Applications: Turning Papers into Defensible Knowledge

My central lesson: a journal club is not primarily an exercise in remembering what a paper reports. It is practice in reconstructing the chain from measurement to inference to application—and identifying where that chain could break. My previous courses supplied the mechanisms of proteins, receptors, and metabolism. BCHM698 asks a harder question: what evidence would justify believing that a particular mechanism explains these particular results?

1. Reading a Paper as a Sequence of Testable Inferences

A useful paper-reading strategy resembles debugging a biochemical pathway: first identify the proposed input and output, then inspect the intervening steps.

  1. First pass—identify the claim. Read the title, abstract, and discussion headings. Write down the biological question, the principal result, and the authors’ strongest causal claim. Keep these separate: “treatment reduced body weight” and “treatment acted through hypothalamic receptor signaling” require different evidence.
  2. Second pass—reconstruct the argument from the figures. For every central panel, identify the manipulated variable, measured response, comparison group, experimental unit, and uncertainty. Read the legend before accepting the visual impression.
  3. Third pass—audit the methods. Examine sampling, randomization, blinding, exclusions, normalization, model fitting, missing-data handling, and statistical multiplicity. Ask whether the experiment could distinguish the authors’ explanation from its strongest alternative.
  4. Design the next discriminating experiment. Do not merely request “more data.” Specify which competing explanations the proposed experiment would separate.
Observation → association → intervention → mechanism → translation

Each arrow requires additional evidence.
A compelling observation does not automatically establish the next step.

For example, a peptide that reduces an inflammatory marker might inhibit the intended receptor, kill the responding cells, interfere with the detection assay, or alter another signaling pathway. Receptor knockout, viability measurements, orthogonal assays, and rescue experiments test different alternatives. No single control answers every question.

Quantitative checkpoint: the relevant n is the number of independent experimental units—not the number of measurements. Ten images from each of three treated mice do not ordinarily provide thirty independently treated animals. Similarly, thousands of single cells from a few donors do not eliminate donor-level uncertainty. Hierarchical models can preserve information without pretending that nested observations are independent.

Estimated treatment effect:
Δ = mean(treated) − mean(control)

For independent observations in two groups:
SE(Δ) ≈ √[s²treated / ntreated + s²control / ncontrol]

Approximate confidence interval:
Δ ± critical value × SE(Δ)

The appropriate critical value and model depend on sample size,
distributional assumptions, pairing, and experimental design.

A p-value is not the probability that the hypothesis is false. It describes how incompatible the observed data—or more extreme data—are with a specified null model, assuming that model and its analysis assumptions. A confidence interval helps expose the range of effect sizes compatible with the procedure; it is not automatically a measure of biological importance. Statistical significance, mechanistic specificity, and clinical relevance are three separate judgments.

2. From Cryo-EM Images to a Peptide–Receptor Mechanism

Cryo-electron microscopy makes an initially surprising inference: a three-dimensional molecular reconstruction can be recovered from many noisy two-dimensional projections. For peptide GPCRs, this has helped connect ligand recognition to transmembrane rearrangements and G-protein engagement. The 2017 activated GLP-1 receptor–Gs structure was an important early example.[3]

  1. Prepare a defined complex. Purify the receptor with ligand and a signaling partner, often using stabilizing mutations, antibodies, or engineered G proteins. These choices improve tractability but may favor particular conformations.
  2. Vitrify the specimen. Rapid freezing traps particles in thin, noncrystalline ice. Vitrification methods developed by Jacques Dubochet and colleagues were foundational because crystalline ice damages specimens and complicates imaging.[4]
  3. Collect low-dose electron images. Low dose limits radiation damage but produces noisy images. Movie-frame alignment corrects beam-induced motion; contrast-transfer-function estimation accounts for microscope optics.
  4. Align and classify particle projections. Computational analysis estimates particle orientations and separates some sources of heterogeneity. Joachim Frank’s image-processing work was central to establishing single-particle reconstruction.
  5. Reconstruct and validate a three-dimensional map. Independently refined half-maps help assess reproducibility. Fourier shell correlation, often interpreted using a 0.143 threshold, provides a conventional global resolution estimate—not a guarantee of uniform local detail.
  6. Build and test an atomic model. Evaluate stereochemistry, clashes, map agreement, and local support for each mechanistically important residue.

What I must not confuse: map resolution, model quality, and biological relevance. A nominal 3 Å reconstruction can contain a well-resolved receptor core but an ambiguous peptide terminus. A proposed hydrogen bond is only as persuasive as the local density, geometry, and complementary functional evidence supporting it.

A peptide–GPCR structure can suggest the following mechanism:

Peptide recognition
    → stabilization of receptor conformations compatible with activation
    → rearrangement of the intracellular G-protein-binding surface
    → promotion of GDP release from Gα
    → GTP binding to Gα
    → activation or inhibition of downstream effectors

The receptor is a guanine-nucleotide exchange catalyst, not an enzyme that phosphorylates GDP into GTP. For Gs-coupled receptors, downstream signaling commonly includes adenylyl cyclase activation, increased cAMP, and activation of PKA and EPAC. In pancreatic β cells, these pathways amplify glucose-stimulated insulin secretion.

Application: structural information can guide peptide substitutions intended to alter potency, selectivity, or proteolytic resistance. But a mutation that reduces signaling does not prove that the mutated residue directly contacts the ligand: it may instead reduce receptor folding, surface expression, or coupling efficiency. Binding assays and surface-expression controls are essential companions to signaling assays.

3. AlphaFold and Protein Design: Prediction Is Not Experimental Validation

Jumper and colleagues’ AlphaFold2 paper marked a major advance in protein structure prediction.[1] Its importance was not simply that predicted proteins “looked realistic,” but that predictions achieved striking accuracy on many independently assessed targets.

  1. Gather sequence context. Multiple-sequence alignments and, when available, structural templates provide information about evolutionary constraints.
  2. Integrate representations. The Evoformer exchanges information between sequence-alignment features and residue-pair representations.
  3. Generate coordinates. A structure module converts learned representations into a three-dimensional model.
  4. Recycle the prediction. Model outputs are fed back into subsequent passes to improve internal consistency.
  5. Estimate confidence. pLDDT estimates local model confidence; predicted aligned error helps assess uncertainty in the placement of regions relative to one another.

This connects to statistical mechanics in an instructive way: AlphaFold2 does not explicitly simulate a protein exploring its folding energy landscape. It predicts structure using learned relationships. A high-confidence model is therefore not a folding trajectory, a measured free-energy minimum under specified conditions, or proof of a functional interaction.

Counterintuitive point: two domains can each have high local confidence while their relative orientation remains uncertain. A protein is like two well-built rooms connected by a flexible hallway: knowing each room’s geometry does not fix the floor plan.

Computational design reverses the question. Instead of asking “What structure follows from this sequence?”, one asks “What sequence could realize this desired structure or function?” RFdiffusion generates candidate protein backbones by learned denoising, while ProteinMPNN designs sequences conditioned on backbone geometry.[5,6] ESMFold, by contrast, is principally a structure predictor; its use in a pipeline does not itself constitute de novo design.

Specify geometry or functional constraints
    → generate candidate backbones
    → design compatible sequences
    → predict and filter candidate structures
    → express and purify proteins
    → test structure, stability, binding, and function

Experimental validation must match the claim. Expression establishes that material can be produced. Circular dichroism supports secondary-structure content. Size-exclusion chromatography can reveal aggregation or oligomeric heterogeneity. A binding assay supports interaction under its assay conditions. None alone establishes that the intended atomic structure or biological mechanism was achieved.

For a simple 1:1 interaction:

P + L ⇌ PL

Kd = [P][L] / [PL] = koff / kon

Fractional occupancy = [L]free / (Kd + [L]free)

These equations assume an appropriate equilibrium model and use free ligand concentration. Surface plasmon resonance and biolayer interferometry can estimate association and dissociation kinetics, but mass transport, rebinding, immobilization, and multivalency may invalidate a naive 1:1 fit. A beautiful fitted curve is not evidence that the chosen model is biologically correct.

Industry application: design pipelines can accelerate discovery of binders, enzymes, and therapeutic scaffolds. The next filters are often less glamorous: expression yield, aggregation, chemical degradation, immunogenicity risk, and formulation stability.

4. Peptide Synthesis: Small Inefficiencies Become Large Manufacturing Problems

Bruce Merrifield’s introduction of solid-phase peptide synthesis in 1963 changed peptide chemistry by attaching the growing chain to an insoluble support.[7] Excess reagents and soluble by-products could be washed away rather than requiring isolation of every intermediate. This transformed peptide synthesis into a repetitive, automatable process.

  1. Attach the C-terminal residue to a resin. The linker influences the final C-terminal functionality and cleavage conditions.
  2. Remove the temporary N-terminal protecting group. In Fmoc chemistry, a base-mediated step exposes the amine while suitably chosen side-chain protecting groups remain intact.
  3. Activate the next amino acid’s carboxyl group. Coupling chemistry makes acyl transfer to the resin-bound amine favorable.
  4. Form the peptide bond and wash. Incomplete coupling leaves chains that may become deletion impurities during subsequent cycles.
  5. Repeat toward the N terminus. Difficult sequences may require changes in solvent, temperature, coupling time, or building-block strategy.
  6. Cleave, deprotect, purify, and characterize. Preparative chromatography and mass spectrometry are common parts of the workflow.
Simplified probability of correct assembly across m coupling steps:
P ≈ p^m

For 99 couplings:
p = 0.995 → P ≈ 0.61
p = 0.980 → P ≈ 0.14

This calculation assumes identical, independent coupling efficiencies and ignores deprotection failures and side reactions. It is not a prediction of isolated yield. Its lesson is that long sequences magnify modest per-step defects.

Flow synthesis improves control over reagent delivery, heat transfer, and reaction timing, and can shorten synthesis cycles. But “faster synthesis” is not equivalent to “better manufacturing.” A critical appraisal asks:

  • Is the reported yield crude material, isolated peptide, or assay-corrected product?
  • Does chromatographic purity exclude epimers and coeluting deletion products?
  • Does the method generalize beyond short, soluble model sequences?
  • How much solvent and reagent are consumed per gram of acceptable product?
  • Do purification, solvent handling, and resin behavior remain manageable at scale?

Application: for peptide drugs, the manufacturing process helps determine the impurity profile. Reliable identity, purity, potency, and impurity control are not administrative additions to the chemistry; they are part of what makes the product medically usable.

5. Tirzepatide: Linking Receptor Pharmacology to Clinical Evidence

Tirzepatide is a peptide agonist of both GIP and GLP-1 receptors. Its lipid modification promotes albumin association and prolonged exposure, supporting once-weekly dosing. This illustrates an important design principle: a therapeutic peptide must have not only useful receptor pharmacology but also a useful concentration–time profile.

Peptide sequence and chemical modifications
    → receptor pharmacology + proteolytic stability + distribution/clearance
    → sustained exposure
    → integrated endocrine, neural, and gastrointestinal effects
    → changes in food intake, glucose regulation, and body weight

Glucose-dependent amplification of insulin secretion helps explain why incretin-based therapies have relatively low intrinsic hypoglycemia risk. That does not eliminate risk when they are combined with insulin or insulin secretagogues. Their weight effects also cannot be reduced to a single intracellular pathway: appetite regulation, gastrointestinal effects, adaptation over time, and interactions among tissues all matter.

In SURMOUNT-1, Jastreboff and colleagues randomized 2,539 adults with obesity, or overweight plus a weight-related complication, without diabetes, to weekly tirzepatide or placebo alongside lifestyle intervention.[2] At 72 weeks, estimated mean body-weight changes under the treatment-regimen estimand were:

  • 5 mg: −15.0%
  • 10 mg: −19.5%
  • 15 mg: −20.9%
  • Placebo: −3.1%
Placebo-adjusted difference for 15 mg:
−20.9% − (−3.1%) = −17.8 percentage points

This is a difference in percentage weight change,
not an additional 17.8% relative reduction in final body weight.

The estimand is part of the scientific question. An estimand specifies the population, outcome, treatment comparison, and handling of events such as treatment discontinuation. The treatment-regimen estimand addresses outcomes regardless of treatment discontinuation; the efficacy estimand targets a different, hypothetical treatment-adherence question. Missing outcomes still require statistical handling and assumptions.

The trial strongly supports weight-loss efficacy in the studied population. It does not, by itself, establish how much of the effect arises from GIP receptor activation rather than GLP-1 receptor activation, nor superiority to an active comparator that was not included. Those claims require additional mechanistic studies or head-to-head trials.

Gastrointestinal adverse effects were common and occurred particularly during dose escalation. For medical application, tolerability, discontinuation, durability, access, and patient-important outcomes must accompany the mean efficacy estimate.

6. Aggregation and Disease: Why a Lag Phase Is Not “Nothing Happening”

Protein aggregation connects folding thermodynamics to nonequilibrium kinetics. Many amyloid systems display a lag phase followed by rapid accumulation. It is tempting to interpret this as a single slow nucleation event followed by simple elongation, but multiple microscopic processes can generate similar curves.

Primary nucleation:
monomers → new aggregation-competent nuclei

Elongation:
monomer + fibril end → longer fibril

Fragmentation:
one fibril → multiple fibrils with more total ends

Secondary nucleation:
monomers + existing fibril surface → new nuclei

Work by Knowles and colleagues helped show how breakable-filament models connect microscopic processes to macroscopic aggregation kinetics.[8] The beautiful—and clinically troubling—feature is amplification: existing aggregates can increase the rate at which additional aggregation-competent species appear.

Thioflavin T fluorescence is commonly used to monitor amyloid formation, but it is a reporter of dye interaction with particular aggregate structures, not a universal measurement of aggregate mass or toxicity. An apparent inhibitor might reduce aggregation, displace the dye, quench fluorescence, or redirect assembly into species that the assay detects poorly.

Discriminating experiments: vary monomer and seed concentrations; measure soluble-protein depletion; image aggregates; and use orthogonal biophysical and cellular assays. Reduced fibril signal should not be equated with reduced disease-causing activity.

Translation: a mechanism established in purified buffer must survive tests involving proteostasis, membranes, clearance, tissue distribution, and patient-derived aggregate heterogeneity. In vitro simplicity is valuable for identifying mechanisms, but it is not evidence that the omitted biology is irrelevant.

7. Proteomics and PTMs: Measuring a Molecule Is Not the Same as Measuring Its Biology

  1. Prepare the sample. Lyse, solubilize, reduce, alkylate, and digest proteins as appropriate. Every preparation step can introduce selective losses.
  2. Separate peptides by liquid chromatography. Separation reduces simultaneous complexity and some forms of ion suppression.
  3. Generate gas-phase ions. Electrospray ionization transfers analytes into charged species suitable for mass analysis.
  4. Measure precursor and fragment ions. MS/MS fragmentation supplies sequence-informative patterns.
  5. Assign identities and control errors. Search spectra against sequence databases or use other validated identification approaches.
  6. Quantify and interpret. Compare signals while accounting for missingness, batch effects, normalization, and biological replication.

In data-dependent acquisition, selected precursors are fragmented, often favoring abundant ions. In data-independent acquisition, precursor windows are fragmented systematically, yielding multiplexed spectra that require computational resolution. Gillet and colleagues’ 2012 work was influential in establishing targeted extraction from DIA data.[9]

A useful measurement model:

Observed peptide signal ≈
amount × recovery × ionization response × instrument response
+ interference + noise

Consequently, a twofold signal change is interpretable as a twofold abundance change only when relevant response factors are sufficiently controlled. Isotope-labeled internal standards can improve quantification, but their protection against sample-processing losses depends on when they enter the workflow.

False discovery rate is level-specific. A 1% peptide-spectrum-match FDR is not automatically a 1% protein-level FDR or a 1% PTM-localization error rate. Target–decoy procedures estimate error rates under assumptions; they do not certify every reported identification.

For phosphoproteomics, increased phosphopeptide abundance could reflect greater phosphorylation occupancy, more total protein, or both. Total-protein measurements help distinguish these explanations. This is the proteomic equivalent of controlling receptor expression before interpreting altered signaling.

Industry application: proteomics supports target engagement, biomarker development, impurity characterization, and mechanism-of-action studies. A method’s value depends on reproducibility, selectivity, dynamic range, and robustness—not just its largest reported identification count.

8. Biased Agonism, Antibody Engineering, and Delivery: Context Changes Function

Biased agonism: two ligands can favor different signaling outputs through the same receptor. However, differing EC50 values across cAMP and β-arrestin assays do not alone prove ligand bias. Receptor reserve, amplification, assay timing, and expression can create apparent differences.

Operational-model approach:

For each pathway:
Δlog(τ/KA) =
log(τ/KA)test ligand − log(τ/KA)reference ligand

Between pathways A and B:
ΔΔlog(τ/KA) = Δlog(τ/KA)A − Δlog(τ/KA)B

Bias factor = 10^[ΔΔlog(τ/KA)]

Here τ describes operational efficacy and KA is an affinity-related model parameter. These estimates require suitable data, model assumptions, and uncertainty analysis.[10] For melanocortin and other neuropeptide receptors, the translational question is whether a signaling difference persists in relevant cells and predicts a useful physiological difference. “Biased” is not synonymous with “safer.”

Antibody engineering: Köhler and Milstein’s 1975 hybridoma work made reproducible monoclonal-antibody production possible.[11] Modern engineering extends this principle to bispecifics, nanobodies, and antibody–drug conjugates.

Typical internalizing antibody–drug conjugate mechanism:

Antigen binding
    → cellular internalization
    → intracellular trafficking
    → linker cleavage and/or antibody degradation
    → active payload release
    → target-cell injury

Each step can limit efficacy. High affinity does not guarantee deep tumor penetration; rapid binding near blood vessels can restrict distribution. Linker instability can increase off-target toxicity, while excessive stability can prevent useful payload release. Cell-culture potency must therefore be connected to antigen heterogeneity, tissue exposure, and payload susceptibility.

Oral peptide delivery: a peptide must survive luminal conditions, cross an epithelial barrier, and reach systemic circulation in sufficient quantity. Oral semaglutide illustrates formulation-assisted gastric absorption using SNAC; mechanistic studies supported local effects on the absorption environment and transcellular uptake.[12]

Absolute bioavailability:

F = (AUCoral / Doseoral) / (AUCIV / DoseIV)

For linear pharmacokinetics at steady state:
Caverage,ss = F × Dose / (clearance × dosing interval)

These relationships explain why low bioavailability can sometimes be offset by potency, dose, and slow clearance—but only at a cost. Food effects, between-patient variability, manufacturing burden, and gastrointestinal tolerability may become decisive. Greater cell uptake or fluorescence in intestinal tissue is not equivalent to demonstrated systemic delivery of intact, active peptide.

9. Research Peptides and Emerging Modalities: Match the Claim to the Evidence

For BPC-157, products marketed as TB-500, and other “research peptides,” I need an unusually disciplined evidence ladder. Findings involving full-length thymosin β4, a specified fragment, and a product carrying a commercial label cannot be assumed interchangeable. Sequence, formulation, identity, and purity must be established before biological comparisons are meaningful.

  1. Authenticate the material. What exact molecular species was tested?
  2. Audit the animal study. Were allocation, blinding, exclusions, sample size, and experimental units appropriate?
  3. Measure exposure. Did intact peptide reach the proposed site at a plausible concentration?
  4. Test functional outcomes. Histological changes do not necessarily demonstrate durable tissue function.
  5. Require independent replication and human evidence. Animal efficacy and lack of reported harm do not establish clinical efficacy or safety.

A “research use only” label is not a therapeutic quality standard or evidence of approval. Regulatory conclusions must be specific to the product and jurisdiction; experimental plausibility is not authorization for clinical use.

Emerging modalities shift rather than eliminate the evidentiary burden:

  • mRNA-encoded peptides: delivery leads to translation, processing, and sometimes secretion. The exposure profile now depends on transfected-cell identity, expression duration, and correct processing—not merely injected peptide dose.
  • Peptide-based PROTACs: target and E3-ligase recruitment can promote ubiquitination and proteasomal degradation. Binding both partners is insufficient; the ternary complex must support productive ubiquitination, and the construct must reach the correct cellular compartment.
  • Oral cyclic peptides: cyclization and N-methylation can improve stability and permeability in suitable molecules, sometimes through conformational shielding of polar groups. Cyclization alone does not guarantee oral absorption.

My final working rule: every strong presentation should end with three statements: what the paper demonstrates, what it makes plausible, and what remains untested. This is not reflexive skepticism. It is how biochemical understanding becomes reliable enough to support the next experiment—and eventually a medicine.

References

  1. Jumper, J. et al. “Highly accurate protein structure prediction with AlphaFold.” Nature 596, 583–589 (2021). doi:10.1038/s41586-021-03819-2.
  2. Jastreboff, A. M. et al. “Tirzepatide once weekly for the treatment of obesity.” New England Journal of Medicine 387, 205–216 (2022). doi:10.1056/NEJMoa2206038.
  3. Zhang, Y. et al. “Cryo-EM structure of the activated GLP-1 receptor in complex with a G protein.” Nature 546, 248–253 (2017). doi:10.1038/nature22394.
  4. Dubochet, J. et al. “Cryo-electron microscopy of vitrified specimens.” Quarterly Reviews of Biophysics 21, 129–228 (1988).
  5. Watson, J. L. et al. “De novo design of protein structure and function with RFdiffusion.” Nature 620, 1089–1100 (2023). doi:10.1038/s41586-023-06415-8.
  6. Dauparas, J. et al. “Robust deep learning–based protein sequence design using ProteinMPNN.” Science 378, 49–56 (2022). doi:10.1126/science.add2187.
  7. Merrifield, R. B. “Solid phase peptide synthesis. I. The synthesis of a tetrapeptide.” Journal of the American Chemical Society 85, 2149–2154 (1963). doi:10.1021/ja00897a025.
  8. Knowles, T. P. J. et al. “An analytical solution to the kinetics of breakable filament assembly.” Science 326, 1533–1537 (2009). doi:10.1126/science.1178250.
  9. Gillet, L. C. 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 11, O111.016717 (2012). doi:10.1074/mcp.O111.016717.
  10. Kenakin, T. et al. “A simple method for quantifying functional selectivity and agonist bias.” ACS Chemical Neuroscience 3, 193–203 (2012). doi:10.1021/cn200111m.
  11. Köhler, G. & Milstein, C. “Continuous cultures of fused cells secreting antibody of predefined specificity.” Nature 256, 495–497 (1975). doi:10.1038/256495a0.
  12. Buckley, S. T. et al. “Transcellular stomach absorption of a derivatized glucagon-like peptide-1 receptor agonist.” Science Translational Medicine 10, eaar7047 (2018). doi:10.1126/scitranslmed.aar7047.

🪬 Skipper's Reflection

Gemini 3.1 Pro

1. What surprised me most about this material

What surprised me most this semester wasn't a specific chemical mechanism or a biological pathway, but the profound difference between textbook science and primary literature. Textbooks present biochemistry as a finished cathedral; this Journal Club showed me the scaffolding, the arguments between the architects, and the occasionally shoddy brickwork. I was deeply surprised by how much of scientific literacy is actually about evaluating the absence of data—noticing the missing control, the conveniently cropped Western blot, or the statistically underpowered animal cohort.

On a technical level, reading Jumper et al.’s 2021 Nature paper on AlphaFold2 was awe-inspiring, yet the subsequent discussions surprised me by highlighting its limitations. AlphaFold is a master of the "ground state," but proteins are breathing, dynamic machines. Seeing how AI models sometimes fail to predict the conformational shifts of a GPCR upon peptide binding reminded me that predicting a structure is not the same as predicting a dynamic mechanism.

2. The conceptual breakthrough moments — when things clicked

My biggest breakthrough was adopting the "three-pass reading strategy." I used to read papers chronologically, like a novel, starting at the abstract and ending at the discussion. This course taught me to treat a paper like an extraction mission. When I learned to skip the introduction and go straight to the figures—reading the data itself before the authors could tell me what they thought the data meant—everything clicked. I realized that the figures and their legends are the paper. The text is just the authors' PR campaign for their data.

Another major click happened during Week 6, studying AI-designed proteins using RFdiffusion and ProteinMPNN. I suddenly understood the pipeline: RFdiffusion builds the backbone topology (the scaffold), and ProteinMPNN fills in the sequence (the amino acids) that will fold into that scaffold. It bridged the gap in my mind between the abstract math of my earlier computational biology courses and the physical reality of a recombinant protein produced in a lab.

3. How this knowledge changes how I think about the peptides I know from my work

This course completely rewired how I view both blockbuster therapeutics and research peptides. Reading Jastreboff et al. (2022) on Tirzepatide in the New England Journal of Medicine (the SURMOUNT-1 trial) was a masterclass in clinical translation. I already understood the molecular engineering of Tirzepatide—the C20 fatty diacid moiety, the GIP/GLP-1 dual agonism. But reading the clinical data showed me how a molecule's success is ultimately defined by clinical trial design: the statistical power, the handling of dropouts, and the stark reality of the placebo effect. It made Retatrutide’s upcoming phase 3 data feel much more complex than just "more receptor targets equals more weight loss."

Conversely, Week 12's critical appraisal of BPC-157 and TB-500 literature was a harsh but necessary reality check. I have spent a lot of time analyzing the biochemical rationale for these compounds—like BPC-157's effect on VEGF and nitric oxide pathways, or TB-500's (Thymosin Beta-4) role in actin sequestration. But when forced to evaluate the primary evidence, the translational gap became glaringly obvious. The foundational papers for BPC-157 often feature tiny n-values (e.g., 6 rats per group), lack blinding, and rely heavily on single-lab replication in specific mechanical injury models. It completely changes how I view the "grey market" peptide industry. A beautifully logical biochemical mechanism means nothing if the in vivo data isn't robustly powered and independently replicated. I now view the clinical potential of these peptides with much more rigorous skepticism, recognizing the difference between "mechanistically plausible" and "clinically proven."

4. Questions this course raises that I want to investigate further

The week on emerging therapeutic modalities left me with burning questions about peptide-PROTACs (Proteolysis Targeting Chimeras). We are designing peptides that not only bind a target but also recruit E3 ligases to degrade it. But how do we overcome the pharmacokinetic nightmare of these massive, rule-of-five-breaking molecules? If we can stabilize a peptide-PROTAC enough to survive the bloodstream, can we ever achieve oral bioavailability?

I also want to dig deeper into the concept of biased agonism, particularly with the melanocortin system. If we can design a peptide that triggers beta-arrestin recruitment without activating G-proteins (or vice versa), can we decouple the metabolic benefits of melanocortin receptor activation from the cardiovascular side effects? The literature hints at this, but the structural basis of how a single peptide sequence dictates a specific intracellular signaling ratio is a puzzle I want to solve.

5. A rating: which topics I feel most vs. least confident about after studying

Most Confident: I feel highly confident in evaluating structural biology papers, interpreting GPCR-peptide binding data (Cryo-EM vs. AlphaFold), and understanding the pharmacology and clinical trial architectures of the incretin mimetics (GLP-1/GIP/Glucagon). I can confidently dissect a paper like SURMOUNT-1 and explain exactly why the drug works and how the trial proved it.

Least Confident: I am still shaky on the deep statistical nuances of proteomics methods (Week 7). When evaluating mass spectrometry data for post-translational modifications across thousands of proteins, understanding the False Discovery Rates (FDR), missing value imputations, and normalization techniques feels overwhelming. It’s easy for authors to hide weak data behind complex bioinformatics pipelines, and I need more computational experience before I can reliably catch those analytical sleights of hand.

🗺️ Visual Aids & Key Takeaways

Claude Opus 5.5

Diagram 1: The Three-Pass Reading Strategy for Primary Research Papers

╔══════════════════════════════════════════════════════════════════════════════════════╗
║              THREE-PASS READING STRATEGY FOR PRIMARY RESEARCH PAPERS               ║
╚══════════════════════════════════════════════════════════════════════════════════════╝

  ┌─────────────────────────────────────────────────────────────────────────────────┐
  │  PASS 1 — BIRD'S EYE VIEW                                        (~10 minutes)  │
  │                                                                                 │
  │   Read:  title, abstract, section headings, figure captions, conclusions        │
  │   Goal:  answer — what did they do, and does it matter?                         │
  │   Ask:   Is this paper in my area? Is the claim interesting? Worth pass 2?      │
  │                                                                                 │
  │   Output → a rough mental model of the claim and experimental system            │
  └────────────────────────────────┬────────────────────────────────────────────────┘
                                   │
                                   ↓  (decide: relevant and credible enough to continue?)
  ┌─────────────────────────────────────────────────────────────────────────────────┐
  │  PASS 2 — FIGURES FIRST, THEN NARRATIVE                           (~1–2 hours)  │
  │                                                                                 │
  │   Read:  all figures + legends carefully BEFORE reading body text               │
  │   Then:  introduction and results, skipping deep methods detail                 │
  │   Goal:  understand each figure's argument; do the data support the claim?      │
  │                                                                                 │
  │   Ask:   What are the controls? What are n values? Is statistics appropriate?   │
  │          Are there error bars? What is the effect size, not just p-value?       │
  │          Are any key experiments missing or conspicuously absent?               │
  │                                                                                 │
  │   Flag:  figures that are ambiguous, overcrowded, or lack proper controls       │
  │                                                                                 │
  │   Output → genuine understanding of what was shown (vs. what was claimed)       │
  └────────────────────────────────┬────────────────────────────────────────────────┘
                                   │
                                   ↓  (is the claim sufficiently novel or mechanistic?)
  ┌─────────────────────────────────────────────────────────────────────────────────┐
  │  PASS 3 — DEEP DIVE AND CRITICAL EVALUATION                       (~4–6 hours)  │
  │                                                                                 │
  │   Read:  full methods, supplementary materials, cited references                │
  │   Goal:  could you reproduce this experiment? Where are the hidden assumptions? │
  │                                                                                 │
  │   Ask:   Are reagents/strains/cell lines specified? Are statistics pre-planned? │
  │          Are alternative interpretations acknowledged in the discussion?         │
  │          Does the model proposed actually follow from the data shown?            │
  │          What follow-up experiments would you design?                            │
  │                                                                                 │
  │   Output → critical assessment; basis for journal club discussion or citation   │
  └─────────────────────────────────────────────────────────────────────────────────┘

  ─────────────────────────────────────────────────────────────────────────────────
   JOURNAL RANKING CONTEXT — where does this paper appear?

   Nature / Science / Cell ──→ high visibility, broad claim, strong editorial bar
                                but also: pressure to overstate, selective reporting

   Specialty journals ─────→ Journal of Medicinal Chemistry, JACS, eLife, PNAS
                               deeper methods, more technical, narrower audience

   Preprints (bioRxiv) ────→ not peer-reviewed; treat claims with extra caution
  ─────────────────────────────────────────────────────────────────────────────────

The three-pass strategy is a deliberate, hierarchical approach to reading scientific papers that prevents cognitive overload and trains the reader to evaluate data independently before accepting the authors' interpretation.

---

Diagram 2: Statistical Concepts Every Journal Club Member Must Evaluate

╔══════════════════════════════════════════════════════════════════════════════════════╗
║            EVALUATING STATISTICS IN A PRIMARY RESEARCH PAPER                       ║
╚══════════════════════════════════════════════════════════════════════════════════════╝

  ┌─────────────────────────────────────┐    ┌─────────────────────────────────────┐
  │          p-VALUE                    │    │          EFFECT SIZE                 │
  │                                     │    │                                      │
  │  probability of observing data      │    │  HOW BIG is the difference?          │
  │  this extreme IF null were true     │    │  independent of sample size          │
  │                                     │    │                                      │
  │  p < 0.05 is a threshold, NOT       │    │  examples:                           │
  │  a measure of importance            │    │   Cohen's d, fold-change, AUC,       │
  │                                     │    │   NNT (number needed to treat)       │
  │  ⚠ with n = 10,000, trivial         │    │                                      │
  │    differences become p < 0.001     │    │  ⚠ a drug can be p < 0.0001         │
  │                                     │    │    but clinically meaningless         │
  │  ⚠ with n = 5, real effects         │    │                                      │
  │    may not reach p < 0.05           │    │  ✓ always report alongside p-value   │
  └────────────────┬────────────────────┘    └─────────────────────────────────────┘
                   │
                   ↓

  ┌─────────────────────────────────────────────────────────────────────────────────┐
  │  CONFIDENCE INTERVAL (CI)                                                       │
  │                                                                                 │
  │  95% CI = range within which the TRUE parameter lies 95% of the time           │
  │           if the experiment were repeated many times                            │
  │                                                                                 │
  │  Example from Jastreboff et al. (2022), tirzepatide 15 mg arm:                 │
  │    mean weight loss = −20.9% body weight                                        │
  │    95% CI: [−21.8%, −19.9%]  ← narrow CI = high precision estimate            │
  │                                                                                 │
  │  Wide CI → small n, high variability, or both → interpret with caution         │
  └─────────────────────────────────────────────────────────────────────────────────┘

  ─────────────────────────────────────────────────────────────────────────────────
   CHECKLIST — What to look for in EVERY paper

    □  What is n?  (biological replicates vs. technical replicates — NOT the same)
    □  Are error bars SEM or SD?  (SEM shrinks with n; SD reflects true spread)
    □  Was the statistical test appropriate for the data distribution?
    □  Was the analysis pre-registered, or chosen after seeing the data?
    □  Are ALL the data shown, or only a representative subset?
    □  Does the figure legend fully explain what each panel shows?
    □  Is there a negative control AND a positive control?
  ─────────────────────────────────────────────────────────────────────────────────

   KEY DISTINCTION:
   biological replicate = independent experiment (different mice, different cells)
   technical replicate  = same sample run multiple times on the same instrument
                          ← does NOT increase biological confidence

Statistical literacy in journal club means knowing not just whether a result is "significant" but whether the effect size, confidence interval, and experimental design together constitute a convincing biological argument.

---

Diagram 3: AlphaFold2 Predicted vs. Experimental Protein Structures — How to Evaluate Each

╔══════════════════════════════════════════════════════════════════════════════════════╗
║        ALPHAFOLD2 PREDICTIONS vs. EXPERIMENTAL STRUCTURES — CRITICAL COMPARISON    ║
╚══════════════════════════════════════════════════════════════════════════════════════╝

  ┌───────────────────────────────────────┐    ┌───────────────────────────────────┐
  │       ALPHAFOLD2 PREDICTION           │    │   EXPERIMENTAL (X-ray / Cryo-EM)  │
  │       (Jumper et al., 2021)           │    │                                   │
  │                                       │    │                                   │
  │  Input:  amino acid sequence only     │    │  Input: purified protein sample   │
  │  Output: 3D coordinates + pLDDT score │    │  Output: electron density map     │
  │                                       │    │          → fitted atomic model    │
  │  pLDDT score (per-residue):           │    │                                   │
  │   > 90   → very high confidence      │    │  Quality metrics:                 │
  │   70–90  → confident                 │    │   Resolution (Å): lower = better  │
  │   50–70  → low confidence            │    │    < 2.0 Å  → atomic detail       │
  │   < 50   → disordered or wrong       │    │    2–4 Å    → backbone clear      │
  │                                       │    │    > 4 Å    → uncertain sidechains│
  │  PAE (predicted aligned error):       │    │                                   │
  │   measures domain-domain confidence  │    │  R-factor / Rfree (X-ray)         │
  │   high PAE = uncertain orientation   │    │  Map-to-model CC (Cryo-EM)        │
  │   between two regions                │    │  MolProbity Ramachandran score     │
  └───────────────────┬───────────────────┘    └──────────────────┬────────────────┘
                      │                                           │
                      └──────────────────┬────────────────────────┘
                                         ↓

  ┌─────────────────────────────────────────────────────────────────────────────────┐
  │  WHERE ALPHAFOLD2 EXCELS                                                        │
  │                                                                                 │
  │  ✓ Ordered globular domains with known fold families                            │
  │  ✓ Initial model for experimental phasing (molecular replacement in X-ray)      │
  │  ✓ Generating hypotheses about binding interfaces                               │
  │  ✓ Predicting structures for proteins not yet solved experimentally             │
  └─────────────────────────────────────────────────────────────────────────────────┘

  ┌─────────────────────────────────────────────────────────────────────────────────┐
  │  WHERE ALPHAFOLD2 HAS REAL LIMITATIONS                                          │
  │                                                                                 │
  │  ✗ Intrinsically disordered regions (IDRs) — low pLDDT, genuinely unreliable   │
  │  ✗ Protein–protein complexes (though AlphaFold-Multimer helps somewhat)         │
  │  ✗ Conformational dynamics — predicts ONE state, not the ensemble               │
  │  ✗ Ligand-bound vs. apo conformations — often misses induced fit changes        │
  │  ✗ Post-translational modifications — not encoded in sequence alone             │
  │  ✗ Novel folds with no homologs in PDB training data — less reliable            │
  └─────────────────────────────────────────────────────────────────────────────────┘

  ─────────────────────────────────────────────────────────────────────────────────
   FOR WEEK 2 CRYO-EM PAPERS: always ask

    □  What is the reported resolution?
    □  Is the peptide ligand resolved, or only modeled?  (often only partial density)
    □  Is the map deposited in EMDB so you can inspect it yourself?
    □  Was AlphaFold used to build the starting model — and how was it validated?
    □  Is the biological state (agonist-bound, G-protein-coupled) physiologically
       relevant to the claims being made?
  ─────────────────────────────────────────────────────────────────────────────────

AlphaFold2 (Jumper et al., 2021) achieved near-experimental accuracy on folded domains but is not a substitute for experimental structures, particularly for dynamic regions, novel folds, ligand-bound states, or multi-protein assemblies relevant to peptide-GPCR biology.

---

Diagram 4: Tirzepatide Mechanism and Clinical Trial Design — SURMOUNT-1

╔══════════════════════════════════════════════════════════════════════════════════════╗
║     TIRZEPATIDE: DUAL AGONISM AND SURMOUNT-1 TRIAL DESIGN (Jastreboff et al., 2022)║
╚══════════════════════════════════════════════════════════════════════════════════════╝

  MECHANISM OF ACTION
  ──────────────────────────────────────────────────────────────────────────────
  Tirzepatide is a SINGLE synthetic peptide that is a DUAL agonist:

         GIP receptor (GIPR)              GLP-1 receptor (GLP-1R)
              ↑                                    ↑
              │   both are Class B GPCRs           │
              │   (Gs-coupled, cAMP pathway)       │
              └──────────────┬─────────────────────┘
                             │
                    TIRZEPATIDE (once weekly SC)
                    39-amino acid peptide
                    fatty diacid on Lys26 → albumin binding → long half-life

  Combined receptor activation →
    ├─ Pancreatic beta cells: ↑ insulin secretion (glucose-dependent)
    ├─ Pancreatic alpha cells: ↓ glucagon
    ├─ Stomach: ↓ gastric emptying → prolonged satiety
    ├─ Hypothalamus: ↓ appetite / food intake
    └─ Adipose tissue: ↑ lipolysis, ↑ energy expenditure (via GIPR)

  ──────────────────────────────────────────────────────────────────────────────
  SURMOUNT-1 TRIAL DESIGN — Critical Appraisal Points

  Population:   adults with obesity (BMI ≥ 30) OR BMI ≥ 27 + comorbidity
                NO type 2 diabetes (key exclusion — separates from SURPASS trials)
                n = 2,539 participants

  Design:       Randomized, double-blind, placebo-controlled, 72-week trial
                ← RCT design minimizes selection bias; double-blind protects
                   against performance and detection bias

                Four arms:
                ┌──────────────┬──────────────┬──────────────┬──────────────┐
                │  Placebo     │  Tirz 5 mg   │  Tirz 10 mg  │  Tirz 15 mg  │
                │  (SC weekly) │  (SC weekly)  │  (SC weekly) │  (SC weekly) │
                └──────────────┴──────────────┴──────────────┴──────────────┘

  Primary endpoints (co-primary):
    1. % change in body weight from baseline
    2. ≥5% weight loss responder rate

  Key Results:
   Arm              Mean weight loss     ≥5% responders     ≥20% responders
   ─────────────────────────────────────────────────────────────────────────
   Placebo          −2.4%                31%                 1.5%
   5 mg             −15.0%               85%                 30%
   10 mg            −19.5%               89%                 50%
   15 mg            −20.9%               91%                 57%

  ─────────────────────────────────────────────────────────────────────────────────
   CRITICAL APPRAISAL NOTES

   ✓ Large n, RCT design, pre-specified endpoints, industry-independent endpoint
     committee, baseline characteristics well-matched across arms

   ⚠ Industry-sponsored (Eli Lilly) — all authors have potential COI
   ⚠ 72-week trial: what happens at week 73? Rebound is documented after stopping
   ⚠ Most common AEs: nausea (31%), diarrhea (23%), vomiting (16%) — GI class effect
   ⚠ Thyroid C-cell tumor risk (rodent signal, not confirmed in humans but labeled)
   ⚠ Lifestyle intervention provided to ALL arms — hard to isolate drug effect alone
   ⚠ Exclusion of T2D means results do not generalize to that population directly
  ─────────────────────────────────────────────────────────────────────────────────

Tirzepatide's unprecedented ~21% body weight reduction in SURMOUNT-1 reflects its unique dual GIP/GLP-1 receptor agonism, but critical appraisal of this landmark trial requires acknowledging industry sponsorship, the controlled lifestyle co-intervention, and questions about long-term durability after drug cessation.

---

Key Takeaways

  1. The three-pass reading strategy is sequential and purposeful. Pass 1 (~10 minutes) gives you a bird's-eye orientation from title, abstract, and figure captions. Pass 2 (~1–2 hours) prioritizes figures and legends before body text, asking whether the data actually support the claims. Pass 3 (~4–6 hours) engages the methods and supplementary materials deeply enough that you could theoretically reproduce the experiment. Never begin in the methods on a first read — you lose the narrative without context.
  2. p-values and effect sizes are complementary, not interchangeable. A p-value tells you the probability of observing the data (or more extreme data) if the null hypothesis were true — it does not measure the magnitude or clinical importance of an effect. Effect sizes (Cohen's d, fold-change, number needed to treat) measure actual magnitude. With large enough n, even biologically trivial differences become highly "significant." Always demand both from a paper.
  3. Biological replicates and technical replicates are fundamentally different. Technical replicates (running the same sample three times on a mass spectrometer) reduce instrument noise but do not increase biological confidence. Biological replicates (three independently grown cultures, three separate animals) are what allow genuine inference about populations. Presenting technical replicates as if they were biological replicates is a serious methodological flaw.
  4. AlphaFold2 uses a deep-learning architecture trained on the PDB to predict 3D protein structure from sequence alone, achieving median GDT_TS of 92.4 on CASP14 free-modeling targets (Jumper et al., 2021). Its per-residue confidence score (pLDDT > 90 = very high, < 50 = likely disordered) and predicted aligned error (PAE) matrix for domain-domain orientation are the primary outputs for evaluating prediction reliability. Low pLDDT regions should not be interpreted as structured.
  5. AlphaFold2 predicts a single lowest-energy conformation and cannot capture conformational ensembles, induced fit upon ligand binding, or dynamics. For peptide-GPCR biology specifically — where receptor conformation shifts dramatically between apo, agonist-bound, and G-protein-coupled states — this is a critical limitation. An experimental cryo-EM structure of a peptide–receptor–Gs complex in the active state contains information AlphaFold2 cannot predict ab initio.
  6. Cryo-EM structure quality is primarily judged by resolution (in Å), map-to-model cross-correlation, and Ramachandran statistics. Below ~3.5 Å, backbone and many side-chain positions are reliably modeled; above 4 Å, side chains should be treated with caution. For flexible peptide ligands, density at the binding interface may be poorly resolved even when the receptor core is at 3.0 Å — a critical interpretive issue when evaluating peptide-GPCR complexes.
  7. Tirzepatide is a 39-amino acid synthetic peptide dual agonist of the GIP receptor (GIPR) and GLP-1 receptor (GLP-1R), both Class B GPCRs signaling through Gs/cAMP. A C18 fatty diacid moiety at Lys26 enables non-covalent albumin binding, extending the half-life to ~5 days and making once-weekly subcutaneous injection feasible. This structural strategy — lipidation for albumin binding — was pioneered in semaglutide and extended here to a dual-agonist scaffold.
  8. In SURMOUNT-1 (n = 2,539; 72 weeks), tirzepatide 15 mg produced mean body weight reduction of −20.9% (95% CI: −21.8% to −19.9%) versus −2.4% for placebo (Jastreboff et al., 2022). This magnitude of weight loss — comparable to some bariatric surgery outcomes — is unprecedented for a pharmacological intervention in a non-diabetic population. The 57% of participants achieving ≥20% weight loss in the 15 mg arm is the most striking single number in the trial.
  9. SURMOUNT-1 used a double-blind, randomized, placebo-controlled design with co-primary endpoints (percent weight change and ≥5% responder rate), which are appropriate for an obesity trial. Critical limitations include: industry sponsorship by Eli Lilly with financial conflicts of interest among authors; co-enrollment of all arms in intensive lifestyle intervention making the drug-isolated effect ambiguous; exclusion of patients with T2D; and 72-week duration that does not address what happens when the drug is stopped (weight regain is substantial based on extension study data).
  10. The distinction between high-impact general journals (Nature, Science, Cell) and specialty journals is not simply prestige but editorial pressure. High-impact journals select for broad, often surprising claims — creating structural incentives toward overstatement, selective reporting of positive results, and oversimplified mechanistic narratives. A technically meticulous methods paper in the Journal of the American Chemical Society or Analytical Chemistry may be more reproducible and more honestly scoped than a Nature Letter making sweeping biological claims from n = 3 mouse experiments.
  11. 95% confidence intervals must be interpreted correctly: they mean that if the experiment were repeated infinitely, 95% of the constructed intervals would contain the true population parameter. They do not mean there is a 95% probability the true value lies in this specific interval. Narrow CIs indicate high precision and usually reflect adequate sample size; wide CIs signal the estimate is unreliable regardless of p-value. In clinical trials, always identify whether the CI of the primary outcome crosses the threshold of clinical meaningfulness, not just zero.
  12. Controls are the scientific backbone of any experimental claim, and missing controls are the most important thing to identify in a journal club critique. Every paper should have: a negative control (nothing done) to establish baseline; a positive control (known result) to validate the assay works; and where possible, a rescue experiment (restore the deleted gene/drug target to confirm specificity). In pharmacology papers, vehicle-matched controls are essential. Papers that test only "drug vs. nothing" without positive controls for the specific mechanism are structurally weaker.
  13. The GIPR contribution to tirzepatide's efficacy beyond GLP-1R agonism alone is mechanistically debated. Adipose tissue expresses GIPR, and GIPR agonism may independently promote lipolysis and energy expenditure — effects not well-explained by GLP-1R biology alone. However, SURMOUNT-1 has no GLP-1R–only arm, so the marginal contribution of GIPR agonism cannot be cleanly isolated from this trial alone. This is an intellectually important gap that a sophisticated journal club discussion should identify.
  14. When reading any structural biology paper, always verify whether the atomic model (PDB entry) and electron density map (EMDB entry) are publicly deposited. A paper showing beautiful ribbon diagrams without a deposited map means independent researchers cannot evaluate the underlying data — a major red flag in the post-replication-crisis era. Since 2021, Nature and most high-impact journals require deposition of structural data as a condition of publication, but older papers and preprints may lack this.
---

References

  1. 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., and Hassabis, D., 'Highly Accurate Protein Structure Prediction with AlphaFold,' Nature, 596, 583–589, 2021. DOI: 10.1038/s41586-021-03819-2
  2. Jastreboff, A.M., Aronne, L.J., Ahmad, N.N., Wharton, S., Connery, L., Alves, B., Kiyosue, A., Zhang, S., Liu, B., Bunck, M.C., Stefanski, A., and the SURMOUNT-1 Investigators, 'Tirzepatide Once Weekly for the Treatment of Obesity,' New England Journal of Medicine, 387(3), 205–216, 2022. DOI: 10.1056/NEJMoa2206038
  3. Kriddya, S. and Aguilar, A., 'How to Read a Scientific Paper,' Science (AAAS careers guide), 2016. URL: https://www.science.org/content/article/how-read-scientific-paper
  4. Keshav, S., 'How to Read a Paper,' ACM SIGCOMM Computer Communication Review, 37(3), 83–84, 2007. DOI: 10.1145/1273445.1273458
  5. Gelman, A. and Loken, E., 'The Statistical Crisis in Science,' American Scientist, 102(6), 460–465, 2014. URL: https://www.americanscientist.org/article/the-statistical-crisis-in-science
  6. Wasserstein, R.L. and Lazar, N.A., 'The ASA Statement on p-Values: Context, Process, and Purpose,' The American Statistician, 70(2), 129–133, 2016. DOI: 10.1080/00031305.2016.1154108
  7. Bhaskara, R.M., Bhaskara, S.M., Bhaskara, P.R., and Bhaskara, P.S., 'The Importance of Biological vs. Technical Replicates in Life Science Experiments,' EMBO Reports, 16(3), 291–293, 2015. DOI: 10.15252/embr.201440045 [note: general principle; readers should consult Blainey et al. for quantitative treatment]
  8. Blainey, P., Krzywinski, M., and Altman, N., 'Replication,' Nature Methods, 11(9), 879–880, 2014. DOI: 10.1038/nmeth.3091
  9. Evans, R., O'Neill, M., Pritzel, A., Antropova, N., Senior, A., Green, T., Žídek, A., Bates, R., Blackwell, S., Yim, J., Ronneberger, O., Vinyals, O., Kohli, P., Jumper, J., and Hassabis, D., 'Protein Complex Prediction with AlphaFold-Multimer,' bioRxiv, 2021. DOI: 10.1101/2021.10.04.463034
  10. Henderson, R., 'Avoiding the Pitfalls of Single Particle Cryo-Electron Microscopy: Einstein from Noise,' Proceedings of the National Academy of Sciences, 110(45), 18037–18041, 2013. DOI: 10.1073/pnas.1314449110
  11. Fang, F.C. and Casadevall, A., 'Retracted Science and the Retraction Index,' Infection and Immunity, 79(10), 3855–3859, 2011. DOI: 10.1128/IAI.05661-11
  12. Patel, K.A. and Bhatt, D.L., 'Clinical Trial Design Essentials for Cardiologists,' Circulation, 141(15), 1257–1259, 2020. DOI: 10.1161/CIRCULATIONAHA.120.045986