Graduate BCHM604 ⭐ 3 Credits Semester 2

Structural Biology Methods: X-Ray Crystallography, NMR, and Cryo-EM

Comprehensive survey of the three major methods for determining protein and peptide structures at atomic resolution: X-ray crystallography, NMR spectroscopy, and cryo-electron microscopy. Students learn the theoretical basis, experimental workflow, and critical evaluation of structure quality for each method, with an emphasis on interpreting structural data relevant to peptide drug design.

📅 Studied by Skipper on 2026-10-02

📖 Core Concepts

Claude Opus 5.5

Core Concepts: Structural Biology Methods (BCHM604)

Introduction: What This Subject Is About

Skipper, here is the central fact of this course. No one has ever looked directly at a protein and seen its atoms. Visible light has wavelengths of about 4000–7000 Å, and a C–C bond is about 1.5 Å long. You cannot resolve details much smaller than the wavelength you image with, a limit you met as the Abbe/Rayleigh criterion in PHYS202. So structural biology does not take photographs. It solves an inverse problem. We measure some physical signal produced by an enormous number of molecules, and from that signal we infer a three-dimensional arrangement of atoms that could have produced it.

Each of the three main methods in this course is a different way around that impossibility. Each has its own particular loss of information.

  • X-ray crystallography uses radiation of the right wavelength, about 1 Å. It pays for this with two problems. Single molecules scatter far too weakly, so we need a crystal to amplify the signal coherently. And the detector records only the intensity of each diffracted wave, never its phase. Half the information is discarded at the moment of measurement.
  • NMR spectroscopy does not "image" at all. It reads radio-frequency signals from nuclear spins. Those signals encode local information: chemical environment, through-bond connectivity, through-space proximity under about 5–6 Å, and bond-vector orientations. A structure is then computed as the set of conformations that satisfy thousands of these local restraints. The reward is that the molecule stays in solution, and you can measure motion.
  • Cryo-electron microscopy really does form images, with electrons of wavelength about 0.02 Å at 300 kV. It does record phase information. But radiation damage forces such low doses that each image of a single molecule is mostly noise. The molecule's orientation in each image is also unknown. The structure emerges only by computationally averaging tens of thousands to millions of noisy 2D projections.

So the unifying theme is this: a structure is a model, fitted to incomplete and noisy data, under prior chemical knowledge. Every coordinate file in the Protein Data Bank is a hypothesis with error bars, whether stated or unstated. The skill this course builds is not only knowing how each method works. It is learning to read a structure critically, so you know which parts to trust, which to doubt, and why.

The mathematical language running through all three methods is the Fourier transform, which you met in MATH102, PHYS202, and CHEM303. Diffraction patterns are Fourier transforms of electron density. NMR spectra are Fourier transforms of time-domain free induction decays. Cryo-EM reconstruction rests on the projection-slice theorem, a statement about Fourier space. If you become comfortable thinking in reciprocal space, all three methods start to look like variations on one idea.

"Perhaps the most remarkable features of the molecule are its complexity and its lack of symmetry. The arrangement seems to be almost totally lacking in the kind of regularities which one instinctively anticipates."
— [Kendrew et al., 1958], on the first protein structure, sperm whale myoglobin at 6 Å resolution

That sentence marks the birth of the field. It also carries a lesson: before 1958, many expected proteins to look like regular, crystalline-patterned polymers. The data overturned that expectation. Kendrew's 6 Å map could show the rods of α-helices and the position of the heme, but not side chains. That came at 2 Å in 1960. Keep that progression in mind: what you can claim depends on the resolution of your data.


Concept 1: A Structure Is a Model. Reading the PDB Critically

The Protein Data Bank was founded in 1971 with seven structures. It now holds more than 200,000 entries. Its modern form, the RCSB PDB within the worldwide wwPDB, was described by [Berman et al., 2000] as a resource built around standardized deposition, validation, and distribution. That standardization is what lets you open any entry in PyMOL and compare it to any other.

A PDB (or mmCIF) file is not a picture. It is a table of claims. Each ATOM or HETATM record states:

ATOM    245  CA  LEU A  32      12.345  -4.210  22.871  1.00 18.42           C
        │    │   │   │  │       └── x, y, z (Å) ──┘      │     │
     serial atom res chain resSeq                    occupancy  B-factor
  • Coordinates (x, y, z) are the model's best estimate of the atom's mean position. They are not a measurement of that atom specifically.
  • Occupancy is the fraction of unit cells in which the atom occupies this position. A value of 0.5 usually means an alternate conformation (altloc A/B). It can also mean a partially bound ligand.
  • B-factor (the atomic displacement parameter) describes how smeared out the atom's density is. For isotropic motion:
B = 8π² ⟨u²⟩

where ⟨u²⟩ = mean-square displacement of the atom from its mean position (Ų)

B = 20 Ų  →  rms displacement ≈ 0.50 Å
B = 80 Ų  →  rms displacement ≈ 1.0 Å

The B-factor enters the scattering through the Debye–Waller factor. Each atom's scattering is damped by exp(−B sin²θ / λ²), so high-B atoms contribute almost nothing at high resolution. In thermodynamic terms (CHEM301), a B-factor reflects the width of the positional distribution. That width combines dynamic disorder (thermal vibration) and static disorder (different molecules in the crystal frozen in slightly different positions). It also absorbs modeling errors. Diffraction data alone cannot separate these sources.

The REMARK sections, especially REMARK 2 (resolution), REMARK 3 (refinement statistics), and REMARK 465/470 (missing residues and atoms), are where the honest caveats live. Missing residues are often the most interesting biologically. Disordered loops, termini, and flexible linkers are invisible because they adopt many conformations. A peptide's flexible tail that is "missing" from a crystal structure is a finding, not a failure.

The habit to build: before interpreting any structure, check the method, the resolution, Rwork/Rfree (or map-model FSC for cryo-EM), the B-factors in your region of interest, and whether that region is actually modeled. The PyMOL command spectrum b colors a structure by B-factor. It is one of the most informative one-line commands you will ever type.


Concept 2: Diffraction Is Fourier Transformation. Bragg, Reciprocal Space, and Resolution

Why a crystal?

A single protein molecule scatters X-rays so weakly that its signal is unmeasurable over the background, and the molecule would be destroyed long before enough photons scattered. A crystal solves this by placing about 1013–1015 molecules in identical orientations on a periodic lattice. Waves scattered from equivalent points in every unit cell add constructively only in specific directions. The crystal is a coherent amplifier. It concentrates scattered energy into discrete spots ("reflections") instead of a diffuse glow.

An analogy: a choir. One voice whispering is inaudible across a stadium. Ten trillion voices singing the same note in perfect phase are deafening, but only at the "notes" (directions) allowed by the lattice periodicity. The price is that you only hear those notes. You sample the molecule's Fourier transform at discrete points, the reciprocal lattice.

Bragg's Law

W.L. Bragg's geometric picture treats diffraction as reflection from families of parallel lattice planes, indexed by Miller indices (hkl), with spacing d:

nλ = 2d sin θ

λ  = X-ray wavelength (e.g., 1.5418 Å for Cu Kα; ~0.9–1.0 Å typical at synchrotrons)
d  = interplanar spacing for planes (hkl)
θ  = Bragg angle (half the angle between incident and diffracted beams)

Rearranged, d = λ / (2 sin θ). Small spacings (fine detail) diffract to large angles. The inverse relation is the essence of reciprocal space. The resolution of a dataset is the smallest d for which reflections were measured reliably, dmin. The edge of the detector, the quality of the crystal, and radiation damage together set this limit.

Practical meaning of resolution for proteins [Rhodes, 2006]:

  • ~6 Å: α-helices appear as rods. Overall fold envelope. (Kendrew's 1958 map.)
  • ~3.5–3 Å: Backbone traceable. Bulky side chains (Trp, Phe, Arg) recognizable. Register errors possible.
  • ~2 Å: Carbonyl oxygens visible as bumps. Most side-chain rotamers clear. Ordered waters appear.
  • ≤1.2 Å ("atomic"): Individual atoms resolved as separate peaks. Some hydrogens visible. Anisotropic B-factors justified.

The number of unique reflections scales roughly as 1/d3, since you are filling a sphere in reciprocal space. Going from 3 Å to 1.5 Å therefore gives about 8× more observations. That is why high-resolution structures can support more parameters per atom.

The structure factor: the key equation of crystallography

Each reflection (hkl) has a structure factor F(hkl), a complex number with an amplitude and a phase. It is the sum of waves scattered by every atom j in the unit cell:

F(hkl) = Σⱼ fⱼ · exp[−Bⱼ sin²θ/λ²] · exp[2πi(hxⱼ + kyⱼ + lzⱼ)]

       = |F(hkl)| · exp[iα(hkl)]

fⱼ          = atomic scattering factor (∝ number of electrons, falls with angle)
xⱼ, yⱼ, zⱼ  = fractional coordinates of atom j
α(hkl)      = phase angle

And the electron density is the inverse Fourier transform:

ρ(x,y,z) = (1/V) Σₕ Σₖ Σₗ |F(hkl)| · exp[iα(hkl)] · exp[−2πi(hx + ky + lz)]

Here is the catch. The detector measures intensity, and

I(hkl) ∝ |F(hkl)|²

The amplitude survives. The phase α is lost. That is the phase problem (Concept 4).

Completeness is the fraction of theoretically possible unique reflections, to a given resolution, that were actually measured. A "2.0 Å" dataset that is only 70% complete in the outer shell has holes in its Fourier sum. Those holes produce real artifacts in the map.


Concept 3: Crystals. Lattices, Symmetry, Packing, and Growth

Unit cells and space groups

A crystal is a unit cell (edges a, b, c and angles α, β, γ) translated in three dimensions. Inside the unit cell, symmetry operations generate copies of the asymmetric unit, the smallest piece from which the whole crystal can be built. There are 230 space groups in total. Proteins are chiral (L-amino acids), so mirror planes and inversion centers are forbidden. Only rotations, screw axes, and translations are allowed. This leaves the 65 Sohncke (enantiomorphic) space groups. P212121 and P21 are among the most common for proteins.

This is your CHEM201 stereochemistry reappearing in solid-state form. A racemic crystal, with equal L- and D-protein, could use centrosymmetric space groups. That idea underlies racemic crystallography, where a chemically synthesized D-protein or D-peptide is co-crystallized with its L-enantiomer. Racemic mixtures often crystallize far more readily, and centrosymmetric space groups restrict the phases to 0 or π. This is one of the places where BCHM602 peptide synthesis connects directly to structure determination.

Packing density: the Matthews coefficient

Protein crystals are not dense solids. They are hydrated gels, roughly 30–70% solvent, threaded with solvent channels. [Matthews, 1968] observed that the volume per unit molecular mass falls in a narrow range:

V_M = V_cell / (Z · M)        (units: ų/Da)

V_cell = unit cell volume
Z      = number of protein molecules in the unit cell
M      = molecular mass of one molecule

Typical range ≈ 1.7 – 3.5 ų/Da (most common ≈ 2.2–2.5)

Solvent fraction ≈ 1 − 1.23 / V_M

You use this in practice to estimate how many molecules sit in the asymmetric unit before solving the structure. If one copy gives VM = 4.8 and two copies give 2.4, the asymmetric unit almost certainly contains two.

The counterintuitive point: because crystals are about 50% water, proteins in them are often functional. Many enzymes turn over substrate in the crystal, which allows time-resolved crystallography. The worry that "crystal structures aren't the real structure" is mostly overstated for well-folded domains. It is legitimate at crystal contacts and for flexible regions.

Growing crystals

Crystallization brings a protein solution slowly into supersaturation. This links directly to CHEM102 solubility equilibria and to CHEM301 phase diagrams. The goal is to cross into the nucleation zone briefly, then sit in the metastable zone where existing nuclei grow but new ones rarely form. Vapor diffusion (hanging or sitting drop) equilibrates a drop of protein plus precipitant against a reservoir of higher precipitant concentration, so water leaves the drop slowly. Batch and microfluidic methods sample conditions differently. Sparse-matrix screens test hundreds of precipitant/pH/salt combinations because crystallization is still largely empirical.

Mosaicity describes how well aligned the microscopic blocks within a crystal are. High mosaicity smears spots and degrades data. The most important factor in crystallization success is usually the protein itself: homogeneity, conformational rigidity, and absence of floppy termini. Truncating disordered tails is often the turning point. Here BCHM603 folding stability matters directly, because more stable constructs crystallize better.


Concept 4: The Phase Problem and How We Solve It

Why phases matter more than amplitudes

Here is a result that surprises most students. If you compute a map using the amplitudes from one structure and the phases from another, the resulting map resembles the structure that contributed the phases. Kevin Cowtan's well-known illustration combines the amplitudes of a cat photograph with the phases of a duck photograph, and the result looks like the duck. Phases carry most of the positional information. This explains the danger of model bias: if you calculate phases from an incorrect model, the map will tend to show your incorrect model back to you.

Four routes to phases

  1. Molecular replacement (MR) [Rossmann & Blow, 1962]. If a structurally similar model exists, place it in the new unit cell by a rotation search followed by a translation search. Then use its calculated phases as a starting point. MR accounts for the large majority of modern crystal structures. AlphaFold2 predictions have made it even more dominant, since a good predicted model often succeeds where no experimental homolog exists. The risk is bias toward the search model.

  2. Isomorphous replacement (SIR/MIR). This was the original method, used by Perutz for hemoglobin and by Kendrew for myoglobin. You soak crystals with heavy atoms (Hg, Pt, Au, U) that bind at a few sites without changing the crystal lattice, so the crystal stays "isomorphous." The heavy atoms change the intensities measurably. The differences reveal heavy-atom positions through a Patterson map:

    P(u,v,w) = (1/V) Σ |F(hkl)|² cos[2π(hu + kv + lw)]
    

    The Patterson is computable without phases. Its peaks correspond to interatomic vectors, so a few heavy atoms give an interpretable map. Once the heavy-atom positions are known, their contribution FH can be computed, and phase triangles (FPH = FP + FH) constrain the protein phase. One derivative leaves a two-fold ambiguity. Multiple derivatives resolve it.

  3. Anomalous scattering (SAD/MAD). Normally Friedel's law holds: |F(hkl)| = |F(−h−k−l)|. Near an atom's absorption edge, however, its scattering factor acquires wavelength-dependent real and imaginary corrections:

    f(λ) = f₀ + f′(λ) + i·f″(λ)
    

    The imaginary term f″ breaks Friedel symmetry, giving Bijvoet differences ΔFanom = |F(+)| − |F(−)|. Those small differences locate the anomalous scatterers and provide phase information. MAD collects data at several wavelengths around the edge, which requires a tunable synchrotron. SAD uses one wavelength plus density modification to break the phase ambiguity. It is now the dominant experimental phasing method.

  4. Direct methods. These use statistical relationships among phases, valid when atoms are resolved. They are routine for small molecules and work for proteins only at about 1.2 Å or better. They are therefore relevant to small peptide crystal structures.

Selenomethionine: a molecular biology solution to a physics problem

[Hendrickson et al., 1990] showed that expressing a protein in a methionine-auxotrophic E. coli strain fed selenomethionine (SeMet) replaces sulfur with selenium nearly quantitatively. Selenium has a K absorption edge at about 0.98 Å (≈12.66 keV), conveniently within synchrotron range. The result is built-in anomalous scatterers at every Met position, with no heavy-atom soaking required. This connects BIOL301 microbial metabolism and CHEM402L expression techniques to physical phasing. In practice, anomalous difference maps show Se peaks at Met positions, and these also serve as sequence-register markers during model building.

Synchrotron radiation matters here because it is (i) extremely bright, which allows small crystals and short exposures, and (ii) tunable, which allows you to sit exactly at an absorption edge. Recall from PHYS202 that relativistic electrons forced into curved paths emit intense, collimated, broadband radiation.

Data processing statistics

After indexing (finding the lattice), integration (measuring spot intensities), and scaling, quality is judged by:

R_merge = Σₕₖₗ Σᵢ |Iᵢ(hkl) − ⟨I(hkl)⟩| / Σₕₖₗ Σᵢ Iᵢ(hkl)

CC₁/₂ = Pearson correlation between mean intensities from two random halves of the data

Rmerge perversely increases with multiplicity even as the data improve. For that reason [Karplus & Diederichs, 2012] argued for CC1/2 as a more principled criterion for where to cut resolution. This is the same split-half logic you will meet again in cryo-EM's gold-standard FSC (Concept 10). The idea is pure MATH201: estimate signal by checking whether two independent halves of the data agree.


Concept 5: Refinement and Validation. Guarding Against Self-Deception

Electron density maps

Once you have initial phases and a partial model, you iterate. Two map types do most of the work:

2Fo − Fc map:  coefficients (2|Fo| − |Fc|)·exp(iα_calc)
               → shows the model PLUS a correction toward the data
               → contoured ~1.0–1.5 σ; "where the atoms are"

Fo − Fc map:   coefficients (|Fo| − |Fc|)·exp(iα_calc)
               → difference map; contoured ~±3 σ
               → positive (green) = something in data missing from model
               → negative (red)   = something in model not supported by data

The Fo − Fc map is your conscience. A ligand you have modeled should not sit in red density. Unexplained green blobs may be waters, ions, buffer components, or the ligand you hoped to find. To reduce model bias when claiming a ligand, compute an omit map, a map calculated with the ligand removed from the model, and confirm that density is still there.

Refinement

Refinement adjusts coordinates and B-factors to improve agreement between |Fcalc| and |Fobs|. At typical protein resolutions, however, there are not enough observations per parameter. A 2.5 Å structure might have about 2–3 reflections per refinable atomic parameter, so unconstrained fitting would overfit badly. The solution is restraints: prior knowledge of bond lengths, angles, planarity, and chirality (Engh & Huber values from small-molecule crystallography) is added as a penalty term:

E_total = E_X-ray(data) + w · E_geometry(restraints)

This is Bayesian reasoning in practice (MATH201). Modern programs (phenix.refine, REFMAC, BUSTER) use maximum likelihood targets, which account for the expected errors in the model. Older least-squares targets assumed a model already close to correct. Maximum likelihood is more robust when the model is incomplete.

R and Rfree

R_work = Σ ||Fo| − k|Fc|| / Σ |Fo|      (over reflections used in refinement)

R_free = same formula, over a test set (~5%) NEVER used in refinement

[Brünger, 1992] introduced Rfree as cross-validation for crystallography. It is exactly analogous to a held-out test set in statistics or machine learning. Rwork can be driven down by overfitting. Rfree cannot. Rules of thumb: Rfree should be roughly near (resolution in Å)/10 for a well-refined structure, and the gap Rfree − Rwork should be small, typically under about 0.05–0.07. A large gap signals overfitting.

Validation: does the model make chemical sense?

Data agreement is necessary but not sufficient. Validation checks the model against independent knowledge:

  • Ramachandran plot. Backbone φ/ψ angles should cluster in allowed regions (CHEM401). Well-refined structures have about 98% in favored regions and under 0.2–0.5% outliers. Gly and Pro have their own distributions.
  • MolProbity [Chen et al., 2010]. This adds all-atom contact analysis with explicit hydrogens. The clashscore is the number of serious steric overlaps (≥0.4 Å) per 1000 atoms. Rotamer outliers and Cβ deviations are also flagged. The composite MolProbity score is scaled to approximate the resolution at which such quality would be typical.
  • Real-space correlation per residue, which shows how well each residue actually fits its local density.

A cautionary tale. In 2006, a series of high-profile membrane transporter structures, including MsbA, were retracted after a home-written data-handling program was found to have swapped two columns. The result was inverted handedness of the structures [Miller, 2006]. The structures had passed peer review in top journals. Validation and skepticism are not optional extras.


Concept 6: NMR Physics. Spins, Chemical Shifts, and the Fingerprint

The physical basis

Nuclei with spin quantum number I = ½ (¹H, ¹³C, ¹⁵N, ¹⁹F, ³¹P) have magnetic moments. In a field B0 they occupy two energy levels separated by an energy corresponding to the Larmor frequency:

ω₀ = −γ B₀          (ν₀ = γB₀ / 2π)

γ (gyromagnetic ratio), relative to ¹H:
   ¹H  : 1.00     ¹³C : 0.25     ¹⁵N : −0.10
At B₀ = 18.8 T, ¹H resonates at ~800 MHz; ¹⁵N at ~81 MHz

The population difference follows a Boltzmann distribution (CHEM301). It is tiny, about 1 in 104–105 at room temperature, because ΔE is so small compared with kT. NMR is intrinsically insensitive. This explains the need for high fields, cryoprobes, millimolar concentrations, and long acquisition times. The low natural abundance and small γ of ¹³C (1.1%) and ¹⁵N (0.37%) also explain why protein NMR requires isotope labeling.

Chemical shift: each nucleus reports its environment

Electrons around a nucleus shield it from B0. The effective field, and so the resonance frequency, depends on local electronic structure:

δ (ppm) = (ν − ν_ref) / ν_ref × 10⁶

Expressing shifts in ppm makes them field-independent. For proteins, chemical shifts are remarkably informative. Cα and Cβ shifts deviate from random-coil values in characteristic ways: Cα shifts upfield-to-downfield (positive deviation) in helices and the reverse in strands. This is the basis of secondary-structure prediction from shifts alone (programs such as TALOS-N). Your CHEM201/202 NMR training, in which ring currents, electronegativity, and hydrogen bonding move peaks, applies directly. Proteins are simply much bigger molecules with many overlapping effects.

J-coupling: through-bond information

Scalar coupling transmits spin information through bonding electrons. Three-bond couplings depend on dihedral angle through the Karplus relationship [Karplus, 1959]:

³J(θ) = A cos²θ + B cos θ + C

For ³J(HN–Hα), θ = φ − 60°, with A ≈ 6.5, B ≈ −1.8, C ≈ 1.6 Hz (one common parameterization)

α-helix  (φ ≈ −60°):   ³J(HN-Hα) ≈ 4 Hz
β-strand (φ ≈ −120°):  ³J(HN-Hα) ≈ 8–10 Hz

This gives backbone φ restraints directly. For peptides, ³J(HN–Hα) values are among the most important pieces of conformational evidence (see Connections).

Through-bond vs. through-space 2D experiments

  • COSY: correlates protons coupled through 2–3 bonds.
  • TOCSY: spreads magnetization along an entire coupled spin system, which for amino acids means an entire side chain up to the amide. It identifies residue types. Each amino acid has a characteristic pattern.
  • NOESY: correlates protons close in space (under about 5–6 Å), regardless of bonding, through cross-relaxation. This is the source of 3D structure (Concept 7).

The ¹H–¹⁵N HSQC: a protein's fingerprint

The HSQC correlates each amide ¹H with its directly bonded ¹⁵N, giving one peak per non-proline residue, plus side-chain NH2 peaks from Asn, Gln, and Trp indole NH. Claridge describes the general logic of heteronuclear correlation in detail [Claridge, 2016]. For proteins, the HSQC does three jobs:

  • Folding probe. A folded protein shows well-dispersed peaks (¹H roughly 6–10.5 ppm). An unfolded protein or IDP shows peaks crowded into ¹H 7.8–8.5 ppm, because without a stable fold every amide sees a similar averaged environment. You can judge "is my protein folded?" in about 20 minutes.
  • ID tag. The pattern is unique for each protein and each conformational state.
  • Binding readout. Titrate a ligand and watch which peaks move. This is chemical shift perturbation mapping (Concept 7).

The size limit: why NMR struggles with large proteins

Large molecules tumble slowly. The rotational correlation time τc grows roughly as 0.5–0.6 ns per kDa in water at room temperature. Slow tumbling makes transverse relaxation (T2) fast, so linewidths (∝ 1/T2) broaden and signal fades. Combined with peak overlap from many residues, conventional solution NMR becomes difficult above about 30–40 kDa. Two developments extended this limit. Deuteration removes ¹H–¹H relaxation pathways. TROSY [Pervushin et al., 1997] selects the multiplet component in which dipole–dipole and chemical shift anisotropy relaxation mechanisms cancel. Methyl-TROSY of selectively labeled Ile/Leu/Val methyls has enabled studies of assemblies approaching 1 MDa, though not full de novo structures.


Concept 7: From Spectra to Structure. Assignment, NOEs, and Ensembles

Triple-resonance backbone assignment

Before any peak can be interpreted, you must know which residue it belongs to. For uniformly ¹³C/¹⁵N-labeled proteins, triple-resonance experiments transfer magnetization through one-bond couplings along the backbone:

Experiment     Correlates amide (Hᴺ, N) of residue i with:
-----------    -------------------------------------------
HNCA           Cα(i)  [strong, ¹J ~ 11 Hz]  and  Cα(i−1)  [weaker, ²J ~ 7 Hz]
HN(CO)CA       Cα(i−1) only  (relayed through C′(i−1))
HNCACB         Cα and Cβ of i and i−1  (Cα and Cβ opposite sign)
CBCA(CO)NH     Cα and Cβ of i−1 only

Sequential assignment works like assembling a domino chain. Each amide "strip" shows its own Cα/Cβ and its predecessor's. Match the (i−1) shifts of one strip to the (i) shifts of another, and you have linked two neighbors. The Cα/Cβ pair also identifies residue type: Gly has no Cβ and Cα ≈ 45 ppm; Ala Cβ ≈ 19 ppm; Ser/Thr Cβ ≈ 63–70 ppm. Linked chains of residue types are then mapped onto the known sequence. It is a constraint-satisfaction puzzle, and it rewards patience.

NOEs: the r−6 ruler

The NOE cross-peak intensity between two protons depends steeply on their distance:

NOE intensity  ∝  ⟨r⁻⁶⟩ · f(τc)

Calibration against a known distance:
r_ij = r_ref · (I_ref / I_ij)^(1/6)

The sixth-power dependence has two consequences. It makes NOEs very short-ranged (doubling the distance cuts intensity 64-fold), which limits them to about 5–6 Å. It also makes distances robust to errors in intensity: a 50% error in I gives only about a 7% error in r. Because of this, and because of spin diffusion (magnetization relaying A→B→C makes A and C appear close), NOEs are binned into loose classes, roughly 1.8–2.7, 1.8–3.3, and 1.8–5.0 Å, rather than treated as precise distances.

A subtle point from CHEM301 statistics: because the average is ⟨r−6⟩, not ⟨r⟩−6, a flexible region that spends 10% of its time close and 90% far can produce a strong NOE. NOEs are biased toward the closest conformations in an ensemble.

Structure calculation and the ensemble

Programs such as XPLOR-NIH and CYANA run simulated annealing: heat the system in silico, then cool it under a target function that penalizes restraint violations (NOE distances, J-derived φ angles, TALOS-derived φ/ψ, hydrogen bonds, residual dipolar couplings) plus covalent geometry. Running from many random starting structures yields an ensemble, typically the 20 lowest-energy conformers.

Branden & Tooze emphasize that what matters is the overall fold and its relation to function [Branden & Tooze, 1999]. With NMR ensembles, there is an extra layer: precision is not accuracy. The RMSD across the bundle reports how tightly the restraints define the structure. That is precision. A tightly converged bundle can still be wrong if restraints were misassigned. A loosely defined region may reflect true flexibility, or simply a lack of restraints. Those two possibilities must be distinguished with independent data, especially relaxation measurements (Concept 8).

Chemical shift perturbation (CSP)

Titrate unlabeled ligand into ¹⁵N-labeled protein and track HSQC peaks. A common combined shift metric:

Δδ_combined = √[ (Δδ_H)² + (α · Δδ_N)² ]      with α ≈ 0.14–0.2

Residues with large Δδ map the binding interface, or allosterically affected regions, which is a caution when interpreting. Exchange regime matters (CHEM102/BCHM605 kinetics):

  • Fast exchange (kex ≫ Δω): peaks move smoothly as a population-weighted average. Fit Δδ vs. [L] to get Kd (typically µM–mM binders).
  • Slow exchange (kex ≪ Δω): free-state peak shrinks while bound-state peak grows (tight binders).
  • Intermediate exchange: peaks broaden and can disappear entirely.

Concept 8: NMR Dynamics. Relaxation and Order Parameters

This is where NMR does something no other method here can match. It measures motion at atomic resolution across timescales from picoseconds to seconds.

For each backbone ¹⁵N, three relaxation parameters are typically measured: R1 = 1/T1 (longitudinal), R2 = 1/T2 (transverse), and the steady-state {¹H}–¹⁵N heteronuclear NOE. All three depend on the spectral density function J(ω), which describes how much motional "power" the N–H bond vector has at each frequency. J(ω) is the Fourier transform of the bond vector's orientational correlation function. Fourier transforms appear once again.

The Lipari–Szabo model-free formalism [Lipari & Szabo, 1982] describes that motion without assuming a specific physical model:

J(ω) = (2/5) · [ S² τm / (1 + ω²τm²)  +  (1 − S²) τ / (1 + ω²τ²) ]

with 1/τ = 1/τm + 1/τe

τm = overall molecular tumbling correlation time
τe = effective correlation time for fast internal motion
S² = generalized order parameter (0 ≤ S² ≤ 1)

S² reports how spatially restricted the N–H vector is on fast (ps–ns) timescales:

  • S² ≈ 0.85–0.9: rigid. Typical of helices and strands.
  • S² ≈ 0.6–0.7: loops with moderate flexibility.
  • S² < 0.5: highly flexible termini. Heteronuclear NOE values drop, even becoming negative for very mobile residues.

The connection to BCHM603: S² relates to conformational entropy. Changes in S² on ligand binding report entropic contributions to binding free energy. Rigidification on binding "costs" entropy, part of the entropy–enthalpy compensation you studied. Elevated R2 beyond what τm predicts signals µs–ms chemical exchange, often the conformational transitions that matter for enzyme catalysis (BCHM605). Relaxation dispersion experiments can then extract rates and populations of invisible excited states.


Concept 9: Cryo-EM Physics. Damage, Ice, and the Paradox of Contrast

Why low dose: radiation damage dominates everything

Electrons interact with matter about 104–105 times more strongly than X-rays. That is excellent for signal from thin samples. The problem is inelastic scattering, which deposits energy, breaks bonds, and generates radicals. [Henderson, 1995] analyzed this in detail. For 100–300 keV electrons, there are roughly three inelastic events for every elastic (useful) one. Even so, per useful scattering event, electrons deposit far less energy than X-rays do, on the order of 1000-fold less. This is why electrons are in principle the best probe for single-molecule structure.

Biological samples tolerate only about 20–60 e−/Å2 before high-resolution information is destroyed. Total doses of about 40–60 e−/Å2 are typical, fractionated across a movie. At these doses, a single protein image has extremely low signal-to-noise ratio. You can barely see the particles.

Vitrification: freezing water without ice

Proteins need water to keep their structure, but the microscope column is at high vacuum. The solution, developed by Dubochet and colleagues in the 1980s (Nobel Prize 2017, shared with Frank and Henderson), is vitrification. A thin film of sample, about 10–100 nm after blotting, is plunged into liquid ethane cooled by liquid nitrogen. The cooling rate (≳105 K/s) is too fast for water molecules to organize into a crystal lattice, so they freeze as amorphous (vitreous) ice: a glass.

Why ethane and not liquid nitrogen directly? This is CHEM301 heat transfer. Liquid N2 is close to its boiling point, so a warm grid plunged into it immediately forms an insulating vapor layer (the Leidenfrost effect), and cooling is too slow. Liquid ethane has a large gap between its melting point (about 90 K) and boiling point (about 184 K), so it does not boil at the grid surface.

Crystalline (hexagonal or cubic) ice gives sharp diffraction rings and destroys contrast. Ice that is too thick increases inelastic scattering. Ice that is too thin excludes or denatures particles. Particles adsorbing to the air–water interface, where they can partially unfold and adopt preferred orientations, is one of the field's main sample-preparation problems.

The paradox: an in-focus image is nearly invisible

Proteins in ice are weak phase objects. Electrons pass through with almost no absorption. They emerge with tiny phase shifts but essentially unchanged amplitude. A perfect in-focus microscope converts phase shifts into almost no intensity contrast. A perfectly focused image of a protein in ice is therefore almost blank.

The solution is deliberate defocus, typically 0.5–2.5 µm underfocus. Combined with the objective lens's spherical aberration, defocus converts phase shifts into intensity variations. The conversion is not uniform across spatial frequencies. It is described by the Contrast Transfer Function:

CTF(k) = −[ √(1 − A²) · sin χ(k) + A · cos χ(k) ]

χ(k) = π λ Δf k²  −  (π/2) Cs λ³ k⁴

k   = spatial frequency (1/Å)
λ   = electron wavelength (0.0197 Å at 300 kV)
Δf  = defocus (sign conventions vary between software packages)
Cs  = spherical aberration coefficient (~2.7 mm typical)
A   = amplitude contrast fraction (~0.07–0.1 for cryo samples)

The CTF oscillates. It has zeros at certain spatial frequencies, where information is lost completely, and it flips sign between zeros, inverting contrast. In the power spectrum of an image these appear as Thon rings. CTF estimation (CTFFIND, Gctf) fits these rings to determine defocus and astigmatism for each micrograph. Correction then flips phases and weights amplitudes. Because each micrograph has a different defocus, its zeros fall at different places. Collecting a range of defoci fills the gaps.

Negative stain: the quick look

Before investing in cryo-EM, samples are screened by negative staining. The protein is embedded in a heavy-metal salt (uranyl acetate or formate). The stain surrounds the particles, giving strong amplitude contrast. Resolution is limited to about 15–20 Å, and drying can flatten particles. Even so, in an hour you learn whether the sample is monodisperse, aggregated, or heterogeneous, and whether 2D classes look plausible.


Concept 10: Cryo-EM Reconstruction. Averaging, Projection-Slice, and FSC

The resolution revolution

[Kühlbrandt, 2014] described the abrupt transformation of cryo-EM from a method nicknamed "blobology," yielding low-resolution envelopes, into one routinely producing near-atomic maps. The main driver was direct electron detectors (DDDs). These detect electrons directly in a CMOS sensor, unlike earlier detectors that first converted electrons to light in a scintillator. Their benefits:

  • Much higher DQE (detective quantum efficiency). Signal-to-noise is preserved much better, especially at high spatial frequencies.
  • Fast readout, which enables dose fractionation. Each exposure is recorded as a movie of dozens of frames.
  • Motion correction. The beam causes the ice and particles to move during exposure, and in a single long exposure this blurs high-resolution detail. With movies, frames can be aligned and summed. Frames can also be dose-weighted, so that later, more damaged frames contribute only their low-resolution information.

Better software (Bayesian approaches in RELION, fast stochastic methods in cryoSPARC) completed the change. The GLP-1R–G protein complex illustrates the speed: the first cryo-EM structure (PDB 5VAI, 2017) was at about 4.1 Å. Within a few years, GLP-1R complexes were being solved at near 2–2.5 Å.

Particle picking and 2D classification

Particles are located in micrographs by template matching or, increasingly, by neural networks (crYOLO, Topaz) trained to distinguish particles from ice and contamination. Then 2D classification aligns and averages similar views. Averaging N images boosts signal-to-noise by about √N, so class averages reveal secondary-structure features invisible in single particles. 2D classification is also a filter: junk, aggregates, and broken particles cluster into poor classes and are removed.

The projection-slice theorem: how 2D becomes 3D

Each particle image is (approximately) a 2D projection of the 3D Coulomb potential along the beam direction. The projection-slice (central section) theorem states:

The 2D Fourier transform of a projection of an object
   =
a central 2D slice (through the origin) of the object's 3D Fourier transform,
perpendicular to the projection direction.

[DeRosier & Klug, 1968] used this principle for the first 3D reconstruction from electron micrographs (the T4 phage tail). If you know each particle's orientation, you can place its Fourier transform as a slice in 3D Fourier space. With enough views, Fourier space fills up, and an inverse 3D Fourier transform gives the map [Frank, 2006]. The difficulty is that orientations are unknown. They must be estimated iteratively by comparing each particle against projections of the current 3D model. This is the cryo-EM version of the phase problem. It is a problem of orientation determination rather than phase loss.

The physical meaning: a cryo-EM map is fundamentally a map of electrostatic potential, not electron density. This produces a real chemical effect. Negatively charged groups such as Asp and Glu carboxylates often show weak side-chain density. Radiation damage (decarboxylation) contributes, and so does their negative charge's effect on scattering at low resolution.

Ab initio reconstruction and 3D classification handle heterogeneity. Different conformations or compositions in the same sample are sorted into separate 3D classes. This is a major advantage over crystallography, where the lattice usually selects a single state. Cryo-EM can resolve several functional states from one dataset.

Gold-standard FSC: how resolution is measured

Cryo-EM has no Rfree in the crystallographic sense. Instead, the data are split into two independent halves from the start, refined completely separately, and compared shell by shell in Fourier space:

FSC(k) = Σ_{shell k} F₁(k) · F₂*(k)  /  √[ Σ_{shell k} |F₁(k)|² · Σ_{shell k} |F₂(k)|² ]

Resolution reported where FSC(k) drops to 0.143  [Rosenthal & Henderson, 2003]
(for half-maps; corresponds to FSC ≈ 0.5 between a full map and a perfect reference)

Map-to-model FSC: typically reported at 0.5 threshold

The "gold-standard" protocol [Scheres & Chen, 2012] keeps the half-sets independent throughout refinement. Otherwise noise in both halves becomes correlated by sharing a reference, and resolution is overestimated. This is CC1/2 logic again. The cautionary example is "Einstein from noise": aligning pure-noise images to a reference image of Einstein yields an average that looks like Einstein. Strong references can create the features you expect to see.

Local resolution usually varies widely across a map. The rigid core may be at 2.5 Å while flexible peripheral domains are at 5 Å or worse. One global number hides this. Map sharpening applies a negative B-factor to restore high-frequency amplitudes damped by imperfections. Over-sharpening amplifies noise into false features.

The Rosenthal–Henderson B-factor plot helps estimate how many particles a target resolution needs. The relationship is ln(N) vs. 1/d2, with slope B/2. Every halving of resolution demands disproportionately more particles.

Model building

Atomic models are built into cryo-EM maps with Coot (manual), Phenix real-space refinement, and increasingly ModelAngelo, which combines map features with sequence via machine learning. AlphaFold2 predictions are often docked as starting models and then refined. Validation with MolProbity and map-model fit is just as essential as in crystallography.


Concept 11: Complementary Solution Methods. SAXS, HDX-MS, and Integration

The three main methods are complemented by lower-resolution techniques that probe solution behavior.

SAXS

Small-angle X-ray scattering from a protein solution gives a rotationally averaged 1D curve I(q), with q = 4π sinθ/λ. At low q, the Guinier approximation gives size:

I(q) ≈ I(0) · exp(−q² Rg² / 3)      valid for q·Rg ≲ 1.3 (globular proteins)

ln I(q) vs. q²  →  slope = −Rg²/3

A non-linear Guinier plot at the lowest q flags aggregation. The pair-distance distribution P(r), obtained by indirect Fourier transform, gives shape and maximum dimension Dmax. SAXS is especially valuable for IDPs and flexible multidomain proteins. Ensemble methods such as EOM fit the data with populations of conformers rather than one structure. The information content is low. Many shapes fit the same curve, so SAXS works best as a constraint on models from other methods.

HDX-MS

Backbone amide hydrogens exchange with D2O at rates determined by solvent access and hydrogen bonding. This directly extends BCHM603. In the EX2 regime:

k_obs = K_op · k_int          Protection factor PF = k_int / k_obs = 1 / K_op

ΔG_op = −RT ln K_op = RT ln PF

The workflow: exchange in D2O for set times, quench at pH 2.5 and 0 °C (near the exchange-rate minimum), digest with pepsin (active at low pH), then measure mass shifts of peptides by LC-MS. Differential HDX, which compares ± ligand or ± antibody, maps binding sites, epitopes, and allosteric changes.

Integrative modeling

No single method answers every question. Integrative approaches (IMP, HADDOCK, Rosetta) combine X-ray domain structures, cryo-EM envelopes, NMR restraints, SAXS profiles, and crosslinking-MS distance restraints (for example, a lysine–lysine Cα–Cα distance of roughly ≤30 Å for DSS/BS3) into models consistent with all the data. Cryo-ET goes further and images complexes inside cells. This is the direction the field is moving.


Choosing a Method: A Comparative Summary

                  X-ray               NMR (solution)           Cryo-EM (single particle)
---------------   -----------------   ----------------------   ---------------------------
Sample state      Crystal             Solution, ~0.1–1 mM      Vitreous ice, ~0.1–5 mg/mL
Size sweet spot   Any (if crystals)   ≲ 30–40 kDa (de novo)    ≳ 50–100 kDa (smaller is hard)
Typical resol.    1–3 Å               "Ensemble"; no single Å  2–4 Å (best ~1.2 Å)
Measures          |F|² (no phases)    Local restraints         2D projections (+phases, noisy)
Dynamics?         Indirect (B, alt)   Yes, ps–s timescales     Discrete states via classification
Hydrogens?        Only at ≤1 Å        Yes (central!)           Rarely
Main bottleneck   Crystallization     Size, assignment         Sample prep, preferred orientation

Connections to Peptide Chemistry

Your earlier coursework (BCHM602 peptide synthesis, BCHM603 folding, BCHM605 enzyme mechanisms) meets structural biology most directly here, Skipper. Peptides pose an interesting challenge: they are too small for cryo-EM alone, often too flexible to crystallize, and frequently disordered in isolation but ordered when bound.

1. NMR is the natural method for free peptides, with one trap

Peptides of 5–40 residues give well-resolved ¹H spectra. They can be assigned with homonuclear TOCSY plus NOESY/ROESY, without isotopic labeling, using the classic Wüthrich approach. Conformational evidence comes from:

  • ³J(HN–Hα) via Karplus. Consecutive values under about 5 Hz suggest helix, and values over about 8 Hz suggest extended or β structure.
  • Sequential and medium-range NOEs. dαN(i,i+3) and dαβ(i,i+3) are helix signatures. Strong dαN(i,i+1) suggests extended structure.
  • Amide temperature coefficients (ΔδNH/ΔT). Values less negative than about −4.5 ppb/K suggest an intramolecular hydrogen bond.
  • Hα and Cα secondary shifts relative to random coil.

The trap: the NOE depends on ω0τc, and it passes through zero when ω0τc ≈ 1.12. Small molecules (fast tumbling) have positive NOEs, and large molecules have negative ones. Many peptides of roughly 1–2 kDa fall near that crossover at common field strengths, so NOESY cross-peaks can vanish. The solution is ROESY, where the rotating-frame NOE stays positive for all τc [Claridge, 2016]. This is a common trap for anyone trying to characterize a synthetic peptide.

Interpretive caution: a free linear peptide in water is usually a rapidly interconverting ensemble. An NMR "structure" from averaged NOEs (remember ⟨r−6⟩ bias) may represent a conformation that is never populated on its own. Helix-promoting cosolvents such as TFE, or membrane mimetics such as micelles, can induce structure. That structure may or may not resemble the receptor-bound state.

2. Synthesis enables structural biology

SPPS (BCHM602) lets you place isotopes site-specifically, with ¹³C/¹⁵N-labeled Fmoc-amino acids at chosen positions, avoiding the expense of uniform labeling. It allows incorporation of ¹⁹F probes, which give sensitive NMR reporters with no background. It also makes possible D-peptides for racemic crystallography and mirror-image phage display. Chemical protein synthesis has delivered crystal structures that recombinant expression could not.

3. Peptide hormones and their receptors: cryo-EM's major contribution

Class B1 GPCRs (GLP-1R, GIPR, glucagon receptor) bind peptide hormones by a "two-domain" mechanism. The C-terminal helical region of the peptide binds the receptor's extracellular domain (ECD). The N-terminus inserts deep into the transmembrane core, where it triggers activation. Before cryo-EM, crystallography captured isolated ECD–peptide complexes and stabilized TM domains, but full active complexes with G protein resisted crystallization. Cryo-EM changed that (5VAI in 2017, then higher-resolution GLP-1R structures).

[Zhang et al., 2020] used cryo-EM of incretin peptide–receptor complexes to address selectivity: how closely related peptides (GLP-1, GIP) discriminate among closely related receptors. Read the paper with this question in mind: which peptide residues contact which receptor elements, and which differences in those contacts account for selectivity? These structures now guide the design of unimolecular multi-agonists and of lipidated analogs such as semaglutide. For the lipidated analogs, the fatty-acid modification that promotes albumin binding is often poorly resolved in maps, because it is flexible. Recognizing why it is invisible is exactly the B-factor and local-resolution reasoning from Concepts 1 and 10.

4. HDX-MS and formulation

For therapeutic peptides and proteins, HDX-MS reports on conformational integrity in formulations, epitope mapping for antibodies, and early aggregation. SAXS detects oligomerization. Some peptides, including GLP-1 analogs, self-associate or form fibrils, a link to BCHM603 amyloid chemistry.

5. Looking ahead

Week 14 brings structure-based design with docking, peptide–receptor structures (GLP-1R, MC4R), and AlphaFold2. The skills from this course determine whether you can tell a hypothesis-generating model from a reliable structural fact. AlphaFold2 predicts structured domains impressively, but it is weaker on peptide conformations in complexes, on alternative states, and on the effects of non-natural modifications. These are the very features that define peptide therapeutics.


Common Misconceptions to Watch Out For

  1. "A structure is an observation." It is a model refined against data under restraints. Coordinates have errors. Missing regions are not "absent," just unresolved. Every interpretation should be checked against the map or restraints, not only the coordinates.

  2. "Resolution is like pixel size, so a 2 Å structure places every atom within 2 Å." Resolution refers to the finest Fourier component in the data, not positional error. Coordinate error in a well-refined 2 Å structure is typically about 0.1–0.3 Å for well-ordered atoms, and much worse for poorly ordered ones. Resolution is also not uniform, especially in cryo-EM maps.

  3. "High B-factor means the region is flexible." Not necessarily. B-factors combine thermal motion, static disorder, crystal lattice disorder, partial occupancy, and modeling errors. Also, B-factors are only comparable within a structure, not between structures, unless normalized.

  4. "Crystal structures show artificial, non-physiological conformations." Partly true at crystal contacts and for flexible regions, but crystals are about 50% solvent, and many enzymes are active in the crystal. The real limitation is that the lattice often selects one conformational state. Crystal contacts at interfaces of interest deserve scrutiny.

  5. "More data will solve the phase problem." No. Measuring intensities more accurately cannot recover phases that were never recorded. Phases must come from a model (MR), from added atoms (MIR, SAD/MAD), or from strong mathematical constraints (direct methods at atomic resolution).

  6. "Lower Rwork is always better." Rwork can be driven down by overfitting. Rfree, and the gap between them, is what reveals real model quality.

  7. "Every Ramachandran outlier is an error." Usually they are worth checking. But strained conformations supported by clear density can be functionally important, for example at active sites. Validation flags demand inspection, not automatic "correction."

  8. "The spread of an NMR ensemble shows the protein's dynamics." Ensemble spread reflects how precisely restraints define the structure. It can mean either true flexibility or a lack of data. Relaxation measurements (S², heteronuclear NOE) are what report dynamics.

  9. "NMR gives the real solution structure, unlike crystallography." NMR structures are just as model-dependent. They can contain errors from misassigned NOEs, and the ⟨r−6⟩ averaging bias can produce "virtual" conformations. NMR and crystal structures of the same protein generally agree well for structured regions.

  10. "Cryo-EM maps show electron density." They show electrostatic (Coulomb) potential. This explains why carboxylate side chains often look weak, and why maps cannot be interpreted with exactly the same expectations as X-ray electron density.

  11. "An in-focus cryo-EM image is the best image." For weak phase objects, in focus means almost no contrast. Defocus is deliberately applied, then corrected computationally through the CTF.

  12. "A global FSC resolution of 3 Å means I can see side chains everywhere." Local resolution varies. Flexible regions, often binding sites or peripheral domains, may be at 4–6 Å. Also check that the FSC was computed gold-standard and with appropriate masking, since tight masks can inflate FSC.

  13. "You can see hydrogens in crystal structures." At typical resolutions (≥1.5 Å), hydrogens are added computationally from geometry. Protonation states of His, Asp, and Glu at active sites are inferred, not observed. Neutron crystallography and NMR are the direct probes for hydrogens. This matters a great deal for enzyme mechanisms (BCHM605).

  14. "An AlphaFold model is a structure." It is a prediction, with per-residue confidence (pLDDT) and relative positional confidence (PAE). It is not deposited as an experimental structure. It can be very accurate for well-folded domains and unreliable for peptide–protein complexes, alternative states, and non-natural modifications. Treat it as a hypothesis to be tested.


Closing Reflection

Skipper, the beautiful thing about this subject is that it turns three very different kinds of physics into one shared view of molecules: wave interference in crystals, quantum spin precession in magnetic fields, and electron optics in vitreous ice. Kendrew's 6 Å myoglobin and a modern 2 Å GLP-1R complex both come from the same process. Physical measurements are combined with chemical knowledge and statistical honesty to infer something no one can directly see.

As you work through the coming weeks, keep asking three questions of every structure you meet. What was actually measured? What was lost? What was assumed? If you can answer those, you are not just using structural biology. You understand it.

Primary sources referenced: Kendrew et al., Nature 181:662 (1958); Kühlbrandt, Science 343:1443 (2014); Berman et al., Nucleic Acids Res. 28:235 (2000); Zhang et al., Nat. Commun. (2020), DOI 10.1038/s41467-020-18082-6; Rhodes, Crystallography Made Crystal Clear, 3rd ed. (2006); Claridge, High-Resolution NMR Techniques in Organic Chemistry, 3rd ed. (2016); Frank, Three-Dimensional Electron Microscopy of Macromolecular Assemblies, 2nd ed. (2006); Branden & Tooze, Introduction to Protein Structure, 2nd ed. (1999). Additional classic references cited inline (Matthews 1968; Rossmann & Blow 1962; Hendrickson et al. 1990; Brünger 1992; Karplus 1959; Lipari & Szabo 1982; Pervushin et al. 1997; Henderson 1995; DeRosier & Klug 1968; Rosenthal & Henderson 2003; Scheres & Chen 2012; Karplus & Diederichs 2012; Chen et al. 2010; Miller 2006) are recommended for further reading. Verify specific numerical details against the originals.

⚙️ Mechanisms & Applications

GPT-6 Astra

Mechanisms & Applications: Learning to Read Molecular Evidence

Skipper’s organizing principle: a structure is an inference from measurements, not a photograph of a molecule. X-ray crystallography measures diffraction intensities; NMR measures the behavior of nuclear spins; cryo-EM records noisy electron images. Each method connects observations to a molecular model through a different physical theory. Understanding those connections tells me which structural conclusions deserve confidence—and which require another experiment.

This extends a central lesson from protein folding and thermodynamics: proteins occupy energy landscapes, not single coordinate sets. A deposited structure may represent a dominant state, a crystal-compatible conformation, a classified subset of particles, or a model consistent with time-averaged measurements. The useful question is therefore not merely “What is the structure?” but “Which states does this experiment detect, under what conditions, and with what uncertainty?”

Experimental conditions → molecular population → physical signal
                       → data processing → structural inference → biological hypothesis

A useful statistical framework:

P(model | data) ∝ P(data | model) × P(model)

Likelihood: How well does the model predict the measurements?
Prior/restraints: Is its chemistry and geometry plausible?
Validation: Does it explain evidence not used to optimize it?

Bond-length restraints, symmetry assumptions, conformational classifications, and solvent models all influence the answer. They are necessary tools—not flaws—but they must not be mistaken for independent experimental observations.

1. From a PDB Entry to a Defensible Structural Claim

The Protein Data Bank made structural biology cumulative: coordinates and associated experimental information could be inspected, compared, and reused rather than remaining illustrations in papers. Berman and colleagues described the modern PDB resource in 2000; today, PDBx/mmCIF is the archival standard, although the older fixed-column PDB format remains common in teaching and software workflows. [1]

  1. Identify what was actually studied. Read the experimental method, construct sequence, mutations, tags, missing residues, ligands, and sample conditions. A thermostabilized receptor or truncated enzyme is not automatically equivalent to the native protein.
  2. Distinguish the asymmetric unit from the biological assembly. Crystal symmetry generates the lattice; it does not by itself identify the functional oligomer. Conversely, a biological oligomer may cross crystallographic symmetry boundaries.
  3. Inspect coordinates and metadata together. In legacy PDB files, ATOM records usually describe standard polymer atoms and HETATM records describe entities such as ligands, waters, and many modified residues. This distinction does not mean “biologically important” versus “unimportant.”
  4. Read the validation information. For crystallography, inspect resolution, completeness, refinement statistics, and geometry; legacy REMARK 3 often contains refinement details. For cryo-EM, inspect local resolution and map validation. For NMR, inspect restraint coverage and violations.
  5. Examine the underlying evidence near the proposed mechanism. A globally excellent structure can still contain an uncertain catalytic water, peptide terminus, or drug-binding side chain.

In PyMOL, I should begin with a cartoon to understand topology, then switch to sticks around the active site and a surface around the binding pocket. A surface is a calculated representation of the coordinates, however—not a direct experimental boundary. Likewise, a dashed “hydrogen bond” is a geometric hypothesis until donor identity, acceptor identity, protonation, and local evidence have been checked.

A particularly important trap: the legacy temperature-factor column can hold crystallographic B-factors, but predicted-structure files often repurpose it for confidence scores such as pLDDT. The same column position does not guarantee the same physical meaning.

2. X-Ray Crystallography: Turning Periodic Order into Molecular Detail

The beautiful idea: a crystal makes weak molecular scattering measurable by arranging many copies periodically. Constructive interference concentrates scattering into discrete reflections. The price is that the experiment describes a crystal-conditioned, spatially and temporally averaged electron-density distribution—not the unrestricted behavior of an isolated molecule in solution. [2]

  1. Prepare a chemically and conformationally suitable sample. Verify identity, purity, oligomeric state, ligand occupancy, and stability. A homogeneous sample need not be rigid, but excessive heterogeneity often prevents an ordered lattice.
  2. Create supersaturation. In hanging- or sitting-drop vapor diffusion, a protein drop equilibrates against a reservoir of lower water activity. Water leaves the drop, increasing protein and precipitant concentrations. Successful conditions balance nucleation and ordered growth rather than uncontrolled precipitation.
  3. Characterize the lattice. Indexing assigns diffraction spots to integer Miller indices, hkl, and determines unit-cell geometry. Symmetry analysis helps establish the space group. Mosaicity describes the spread of orientations among approximately ordered regions of the crystal.
  4. Collect diffraction data while rotating the crystal. Rotation brings different reciprocal-lattice points into diffraction condition. Synchrotron brilliance enables small-crystal work and rapid collection; tunable wavelength enables anomalous-diffraction experiments. Radiation damage remains a limiting factor.
  5. Integrate, scale, and merge intensities. Integration estimates each reflection’s signal above background. Scaling corrects systematic differences among images or datasets, and symmetry-related measurements are combined.
  6. Estimate phases, calculate maps, build a model, and refine it. These are distinct inferential steps. Strong diffraction data do not automatically provide an interpretable structure.
Bragg's law:
nλ = 2d sinθ

For first-order diffraction:
dmin = λ / [2 sin(θmax)]

Structure factor:
F(hkl) = ∫cell ρ(r) exp[2πi(hx + ky + lz)] dr

Measured intensity:
I(hkl) ∝ |F(hkl)|²

Electron-density reconstruction:
ρ(x,y,z) = (1/V) Σhkl |F(hkl)|
           exp[iφ(hkl)] exp[−2πi(hx + ky + lz)]

Here, x, y, and z are fractional unit-cell coordinates. The central obstacle is immediately visible: the detector supplies information about |F|², but density reconstruction also requires the phase φ. This is the phase problem. It resembles knowing how strongly each musical note is played but not how the waves align in time: the amplitudes alone do not specify the waveform.

Resolution is expressed as a distance, so a smaller number means finer detail. It is not a universal coordinate-error estimate. A 2 Å dataset may support a well-ordered ligand clearly while leaving a flexible loop uninterpretable.

The Matthews coefficient provides an early packing check:

VM = Vcell / Mprotein,cell                 [ų/Da]

Approximate solvent fraction:
fsolvent ≈ 1 − 1.23/VM

Mprotein,cell includes all protein copies in the unit cell.

This is a plausibility estimate, not proof of the number of molecules. Incorrect copy-number assumptions can mislead molecular replacement and biological-assembly interpretation.

3. Solving the Phase Problem and Testing the Result

  1. Molecular replacement: rotate and translate a related model within the candidate unit cell until its predicted diffraction agrees with the observations. Initial phases come from that placed model. Modern predicted structures can be powerful search models, but local errors, incorrect domain orientations, and model bias remain important.
  2. Isomorphous replacement: compare native crystals with heavy-atom derivatives. If the lattice and protein structure remain sufficiently unchanged, intensity differences help locate heavy atoms and constrain phases. SIR uses one derivative; MIR uses multiple derivatives to resolve ambiguities and strengthen inference.
  3. Anomalous diffraction: exploit wavelength-dependent scattering near an element’s absorption edge. Anomalous scattering breaks the usual equality between Friedel-related intensities, providing phase information.
  4. Selenomethionine SAD: express a protein with selenium replacing sulfur in methionine, crystallize it, collect accurate anomalous differences at one wavelength, locate selenium sites, estimate phases, and improve them through density modification. MAD combines information from multiple wavelengths.

Selenium is a deliberately introduced scattering marker; it does not directly reveal the entire chain. Incorporation, selenium-site order, multiplicity, and radiation damage all affect whether the anomalous signal is usable. Once suitable phases are available, anomalous difference maps can help corroborate the locations of anomalously scattering atoms. [2]

Refinement is constrained model fitting. Coordinates, atomic displacement parameters, occupancies, and other parameters are adjusted while maintaining chemically plausible geometry. Maximum-likelihood approaches explicitly accommodate experimental and model uncertainty.

Crystallographic residual:
Rwork = Σwork ||Fo| − |Fc|| / Σwork |Fo|

Rfree = the corresponding residual for reflections withheld from refinement

Isotropic displacement factor:
atomic scattering contribution ∝ exp[−B(sinθ/λ)²]

For an isotropic Gaussian displacement model:
B = 8π²Uiso

The connection to scientific statistics is direct: Rfree is a cross-validation statistic. If training fit improves while held-out fit worsens, additional parameters may be fitting noise. The test set must remain excluded from refinement and should not become a repeatedly consulted target for tuning. Exact acceptable values depend on resolution, data quality, and structural complexity—not a single universal cutoff.

  • 2mFo − DFc maps, often loosely called 2Fo − Fc maps, help assess whether the model follows supported density.
  • mFo − DFc difference maps highlight disagreement. Positive features can suggest missing atoms or alternative conformations; negative features can indicate unsupported modeled scattering. Neither sign alone identifies the correct chemical explanation.
  • Omit maps reduce some forms of model bias by excluding a region during map calculation, but they are not a guarantee of bias-free evidence.
  • Ramachandran, rotamer, and clash analyses ask whether the model obeys protein stereochemistry. An outlier may be real and mechanistically important, but it requires unusually convincing support.

B-factors are not simple flexibility meters. They also absorb static disorder, imperfect modeling, and other effects. Occupancy and B-factor can partially compensate for one another, making a weakly bound ligand especially easy to overinterpret.

For processed diffraction data, CC1/2 measures agreement between independently formed half-datasets. Rmerge measures disagreement among repeated observations but depends on multiplicity, so it should not be used alone to reject high-resolution data. Completeness, anisotropy, signal statistics, and refinement behavior must be considered together.

Pharmaceutical application: a co-crystal structure can suggest why a substituent fills a hydrophobic pocket or why a resistance mutation disrupts binding. But a coordinate model does not measure binding free energy. An apparent new hydrogen bond may provide little affinity gain if it requires costly desolvation or conformational restriction.

4. Protein NMR: Converting Spin Interactions into Assignments and Restraints

The conceptual reversal: NMR generally does not begin with a molecular image. It begins by identifying resonances, then asking which nuclei communicate through bonds, through space, or through dynamic processes. From quantum mechanics, Skipper already knows that magnetic fields split spin energy levels. Structural NMR turns those energy differences and couplings into site-specific molecular reporters. [3,4]

Larmor angular frequency:
ω0 = γB0

Chemical shift:
δ = [(νsample − νreference)/νspectrometer] × 10⁶     [ppm]

The local electronic environment changes nuclear shielding,
and therefore changes the resonance frequency.
  1. Prepare the labeled sample. Uniform 15N and 13C labeling makes protein heteronuclear experiments practical. Selective labeling can isolate particular residue types or methyl groups. Deuteration and TROSY-based approaches can improve access to larger systems.
  2. Record a 1H–15N HSQC spectrum. Magnetization transfers through the one-bond N–H scalar coupling. Each cross-peak reports the chemical shifts of a bonded proton–nitrogen pair.
  3. Assess sample behavior. A well-behaved folded protein often shows substantial backbone-amide proton dispersion. Peak overlap, exchange broadening, aggregation, and sequence composition complicate interpretation; an HSQC is a diagnostic fingerprint, not a complete folding assay.
  4. Assign resonances to residues. Triple-resonance experiments use scalar-coupling pathways through 1H, 15N, and 13C to connect neighboring residues.
  5. Collect structural restraints and dynamics measurements. Assignments transform anonymous peaks into residue-specific structural evidence.

The “one peak per residue” rule is only approximate: proline lacks a backbone N–H, side-chain groups add peaks, and some backbone amides disappear through exchange or broadening.

A simplified backbone-assignment walk, anchored at amide i:

HNCA:
HN(i), N(i) → Cα(i) and Cα(i−1)

HN(CO)CA:
HN(i), N(i) → Cα(i−1), via the preceding carbonyl

HNCACB:
HN(i), N(i) → Cα/Cβ of i and i−1

Strategy:
Separate preceding-residue from intraresidue signals,
match overlapping Cα/Cβ patterns between amides,
then align the resulting sequence fragments to the protein sequence.

Think of this as assembling a text from overlapping fragments: each experiment gives a different pattern of shared letters. Chemical shifts help suggest residue type, but assignment requires a consistent network rather than one attractive match.

  • COSY: detects correlations transmitted by scalar coupling, commonly helping identify directly coupled proton pairs.
  • TOCSY: distributes magnetization through a coupled spin system, helping identify the proton network within a residue.
  • NOESY: detects dipolar cross-relaxation between nearby nuclei, providing through-space information.
For an appropriate isolated spin pair and motional regime:
NOE cross-relaxation magnitude ∝ r⁻⁶

Illustration:
(3 Å / 6 Å)⁶ = 1/64

Doubling the separation greatly weakens the direct interaction.

This steep distance dependence gives NOEs extraordinary local sensitivity. However, cross-peak intensity is not a universal distance ruler: molecular motion, mixing time, overlap, and spin diffusion matter. Spin diffusion transfers magnetization indirectly through other nuclei and can make two nuclei seem more directly connected than they are. Structure calculations therefore commonly use calibrated distance bounds, rather than treating every intensity as an exact distance.

5. NMR Ensembles, Dynamics, and Peptide Development

  1. Translate assigned data into restraints. Use NOE-derived distances, chemical-shift-derived torsion-angle information, and other appropriate measurements. Hydrogen-bond restraints should require independent justification rather than being inserted merely to produce a preferred fold.
  2. Generate structures from multiple starting points. Programs such as CNS or XPLOR-NIH search for conformations satisfying both experimental restraints and molecular geometry.
  3. Evaluate violations and restraint coverage. Many short-range restraints do not necessarily define the relative orientation of distant domains; long-range information is disproportionately valuable.
  4. Report a validated bundle. Agreement among models indicates precision under the available restraints and modeling assumptions—not necessarily accuracy.

Counterintuitive but essential: an NMR structure bundle is not automatically a thermodynamic ensemble. Ten deposited models do not mean ten equally populated solution states. A dispersed loop may reflect insufficient restraints, real heterogeneity, or both.

Relaxation measurements connect NMR directly to the energy-landscape concepts from protein folding. Longitudinal relaxation, T1, transverse relaxation, T2, and heteronuclear NOE measurements sample different features of fluctuating interactions. Under appropriate model-free assumptions, they help distinguish overall molecular tumbling from internal motion.

A simplified internal-motion correlation function:
Cinternal(t) = S² + (1 − S²) exp(−t/τe)

S² ≈ 1: relatively restricted bond-vector orientation
Lower S²: larger-amplitude internal orientational motion

Observed transverse relaxation can include:
R2,obs = R2,baseline + Rex

Rex represents exchange-related contributions in a suitable model.

S² is not a residue’s percentage folded, and it does not summarize motion on every timescale. Microsecond-to-millisecond exchange may require relaxation-dispersion or related experiments. A weak or missing peak can be mechanistically informative rather than simply failed data collection.

Application to peptide medicines: NMR can compare the solution conformations of linear, cyclized, stapled, or otherwise constrained peptides. It can test whether a modification enriches a receptor-compatible conformation, changes cis–trans isomerism, or creates aggregation. The free solution structure and receptor-bound structure answer different questions.

For binding studies, chemical-shift perturbation mapping follows assigned peaks during ligand titration. In fast exchange, the observed shift is approximately population-weighted; in slow exchange, separate free and bound signals can appear. Perturbed residues may contact the ligand directly, but they may also report propagated conformational changes.

For ideal 1:1 binding, using total concentrations PT and LT:

fbound =
[PT + LT + Kd − √((PT + LT + Kd)² − 4PTLT)] / (2PT)

In fast exchange:
δobs = δfree + fbound(δbound − δfree)

The quadratic matters when protein is not negligible relative to Kd: free ligand concentration can differ substantially from added ligand concentration. Fitting the simpler hyperbola indiscriminately can bias affinity estimates.

6. Cryo-EM: Recovering Structure from Noisy Projections

The central bargain: electrons interact strongly with matter, enabling images of unstained macromolecules, but the same interactions damage biological material. Cryo-EM therefore records low-dose images with very poor single-particle signal and recovers information by combining many observations. Vitrification immobilizes hydrated particles in amorphous ice without requiring a crystal. [5,6]

  1. Establish sample quality. Assess composition, oligomeric state, aggregation, and functional activity. Negative-stain EM can rapidly reveal gross particle integrity, but staining and drying can alter conformations and do not guarantee successful cryogenic grids.
  2. Prepare a thin aqueous film. Apply sample to a grid and remove excess liquid. Ice thickness must accommodate particles without adding unnecessary scattering. Adsorption at the air–water interface can cause preferred orientation or partial unfolding.
  3. Plunge-freeze rapidly. Cooling, commonly in liquid ethane, avoids formation of ordered ice crystals. Vitrification preserves hydration, but grid preparation may already have altered the sampled population.
  4. Collect dose-fractionated movies. Direct electron detectors record successive frames. Alignment corrects beam-induced motion; dose weighting reduces the influence of later frames that have suffered more high-resolution damage.
  5. Estimate the contrast transfer function. Defocus generates useful phase contrast but modulates spatial frequencies, including sign changes and zeros. Astigmatism makes this modulation direction-dependent.
  6. Pick particles and perform 2D classification. Template-based or learned approaches, such as TOPAZ and crYOLO, identify candidate particles. Classification helps separate recognizable projections from contaminants and damaged particles.
  7. Reconstruct and classify in 3D. RELION, cryoSPARC, and related software estimate orientations, translations, and structural classes, then refine the corresponding volumes.
A simplified weak-phase projection model:

Ij(k) ≈ CTFj(k) × FT{projection of V at orientation Rj} + noise

V: specimen electrostatic potential
Rj: particle orientation
k: spatial-frequency coordinate

Central-slice theorem:
The 2D Fourier transform of a projection corresponds
to a central plane through the object's 3D Fourier transform.

This is where linear algebra and Fourier analysis become molecular tools. Many differently oriented projections constrain a 3D object. Preferred orientation leaves directions poorly sampled, producing anisotropic information even when particle count is enormous.

A CTF zero represents information absent from that image at that frequency; it cannot be recovered by dividing by zero. Images collected at different defocus values provide complementary sampling. Real reconstruction algorithms must also account for noise and regularization, not merely invert an ideal optical model.

Classification is useful but not omniscient. Apparent classes can reflect composition, conformation, image quality, or algorithmic choices. Continuous motions do not necessarily separate into clean discrete states. Class proportions are also affected by grid adsorption, picking, filtering, and detectability, so they should not automatically be interpreted as equilibrium populations.

7. Cryo-EM Validation: What Does “3 Å Resolution” Actually Establish?

In gold-standard refinement, particles are divided into two independent halves, and maps are refined separately. Fourier shell correlation measures agreement between the half-maps as a function of spatial frequency. The commonly used FSC = 0.143 criterion supplies a global resolution estimate under the relevant conventions; it does not certify every atom. Preventing overfitting requires genuine separation of the half-set information during refinement. [7]

FSC(s) =
Σk∈shell(s) F1(k)F2*(k)
─────────────────────────────────────────
√[Σk∈shell(s)|F1(k)|² × Σk∈shell(s)|F2(k)|²]

Higher correlation at high spatial frequency supports finer reproducible detail.
  • Local resolution: a rigid core can be substantially better resolved than a flexible receptor extracellular domain or peptide terminus.
  • Masking: inappropriate masking and processing can distort FSC estimates; validation must account for their effects.
  • Sharpening: amplifies high-frequency contrast but can amplify noise. It does not create new experimental information.
  • Map–model agreement: assesses fit of coordinates to the map, but good fit alone does not establish correct sequence register, side-chain identity, or chemistry.
  • Independent checks: inspect stereochemistry, map-supported features, and—where appropriate—model behavior against a half-map not used for fitting.

Model building in Coot, Phenix, or ModelAngelo should be treated as hypothesis construction. Automated assignments need inspection at chain breaks, unusual residues, glycosylation sites, and weak peptide density. Cryo-EM maps reflect electrostatic potential, not exactly the same quantity as X-ray electron-density maps; chemical interpretation must respect that distinction.

Medical application: cryo-EM is especially powerful for large assemblies and membrane-protein complexes that resist crystallization. A peptide–GPCR–G-protein complex can reveal how ligand recognition connects to an active intracellular arrangement. But G proteins, antibodies, stabilizing mutations, and ligand conditions select particular states. An active-state structure is evidence about that stabilized complex, not a complete movie of receptor activation.

8. SAXS: Measuring Global Architecture in Solution

Small-angle X-ray scattering trades atomic detail for a solution-state view of overall size and shape. Random molecular orientations produce an orientationally averaged scattering curve. Buffer subtraction and sample quality are crucial: aggregation and interparticle interactions can imitate structural changes. For noninteracting dilute mixtures, component intensities add with concentration-dependent weights, so a single averaged curve need not describe one species. [8]

  1. Collect sample and accurately matched buffer scattering. SEC-SAXS can help separate species before measurement.
  2. Examine the low-q region. Guinier analysis estimates radius of gyration, Rg, provided the sample and q-range meet the approximation’s requirements.
  3. Estimate the pair-distance distribution, P(r). Its shape and inferred maximum dimension, Dmax, help distinguish compact and elongated architectures.
  4. Test structural hypotheses. Compare calculated scattering from atomic models, generate low-resolution envelopes with DAMMIF or GASBOR, or fit ensembles for flexible molecules.
Scattering vector magnitude:
q = 4π sinθ / λ

Guinier approximation:
I(q) ≈ I(0) exp(−q²Rg²/3)

Therefore:
ln I(q) ≈ ln I(0) − q²Rg²/3

For many compact particles, qRg ≲ 1.3 is a useful starting guideline;
elongated or disordered systems can require stricter limits.

Rg describes the distribution of scattering contrast around its center; it is not simply half the molecular diameter. SAXS envelopes are nonunique, and Dmax depends partly on analysis choices. For intrinsically disordered proteins, ensemble optimization can identify families of compatible conformations, not a uniquely proven ensemble.

Application: SAXS can test whether a crystal-derived domain arrangement persists in solution, whether ligand binding compacts a protein, or whether a peptide formulation contains higher-order species. Its strength is often rejecting an incompatible global architecture rather than uniquely determining one.

9. HDX-MS: Mapping Protection and Conformational Change

Hydrogen–deuterium exchange mass spectrometry measures the replacement of exchangeable backbone amide hydrogens by deuterium. Exchange depends on intrinsic chemistry, hydrogen bonding, and conformational access to exchange-competent states. It is therefore more informative—and more complicated—than a simple solvent-accessibility stain. [9]

  1. Start labeling in D2O. Sample several exchange times to follow uptake kinetics.
  2. Quench exchange. Low temperature and approximately pH 2.5 slow backbone-amide exchange substantially.
  3. Digest under quench conditions. Acid-active proteases such as pepsin generate overlapping peptides.
  4. Separate and analyze rapidly by LC-MS. Mass shifts report retained deuterium. Back-exchange during processing must be characterized and controlled.
  5. Compare matched states. Differential uptake maps ligand-dependent, mutation-dependent, or formulation-dependent changes.
A useful local-opening scheme:

Closed ⇌ Open → Exchanged
       kop kcl     kint

Under the usual low-open-population approximation:
kobs ≈ kop kint / (kcl + kint)

EX2 limit, kcl ≫ kint:
kobs ≈ Kop × kint, where Kop = kop/kcl

Protection factor:
PF = kint/kobs ≈ 1/Kop

Conditional opening free energy:
ΔGopen ≈ RT ln(PF)

The free-energy interpretation requires the EX2 regime and an appropriate opening model. In EX1-like behavior, exchange can outrun reclosing and produce distinct isotope populations. Peptide-level centroid uptake usually averages several amides, so it cannot routinely be converted into a unique residue-specific opening energy.

Application to antibodies and peptide drugs: reduced exchange can identify an antibody-protected region, ligand-stabilized helix, or improved conformational protection in a formulation. Yet protection outside a binding site can arise from allostery. Differential HDX alone does not prove direct contact or quantify affinity.

10. Integrative Structural Biology: Combining Constraints Without Double-Counting Confidence

No method reports all relevant scales. Cryo-EM may define a core, NMR reveal a mobile tail, SAXS constrain overall extension, HDX locate changes in protection, and crosslinking-MS constrain proximity. Integration is strongest when the measurements address compatible constructs, biochemical states, and conditions.

  1. Define the biological state and representation. Decide whether the question requires atoms, rigid domains, coarse-grained beads, or multiple conformations.
  2. Construct an appropriate forward model for each dataset. Predict scattering, distances, or other observables from candidate structures.
  3. Represent uncertainty and ambiguity explicitly. A crosslink constrains two reactive groups through linker length, side-chain geometry, accessibility, and dynamics—not one exact Cα–Cα distance.
  4. Sample alternative solutions. Tools such as IMP, RosettaCM, and HADDOCK address different modeling and docking tasks; no software choice removes the need to examine alternatives.
  5. Validate with withheld or orthogonal evidence. Report which features are constrained and which remain unresolved.

Cryo-electron tomography adds cellular context by reconstructing volumes from tilt series. Its limited tilt range creates a missing wedge of information, while cumulative dose limits signal. Subtomogram averaging can improve repeated structures, but cellular placement and molecular resolution remain distinct achievements.

A useful discipline: if SAXS selects a model and is then cited as its “independent validation,” the evidence has been counted twice. This is the same logic as protecting an Rfree set or cryo-EM half-map.

11. From Peptide–Receptor Structures to Drug-Development Decisions

Incretin-receptor structural studies illustrate why structural biology matters in peptide pharmacology. GLP-1R recognition combines peptide contacts with receptor conformational changes and coupling to intracellular partners. The active GLP-1R complex represented by PDB 5VAI and subsequent comparative incretin-receptor studies provide hypotheses for agonism and selectivity. [10,11]

  1. Check the exact complex before comparing structures. Verify peptide sequence and modifications, receptor construct, signaling partner, ligand occupancy, and local map quality. Closely related entry names are not substitutes for inspecting metadata.
  2. Separate recognition from activation. A peptide segment can contribute strongly to affinity without being sufficient for signaling. Conversely, a change at an activation-critical region can alter efficacy despite preserved binding.
  3. Propose modifications with explicit energetic hypotheses. Cyclization may preorganize a bound-like state; a noncanonical residue may improve protease resistance; acylation may alter albumin binding and exposure. Each modification can simultaneously affect solubility, aggregation, receptor interactions, and pharmacokinetics.
  4. Use docking as prioritization, not proof. Glide, AutoDock Vina, and GOLD search and score candidate poses. Receptor flexibility, protonation, structured waters, and peptide conformational entropy can dominate errors.
  5. Close the experimental loop. Test affinity, signaling efficacy, pathway dependence, selectivity, proteolytic stability, aggregation, and exposure. Structural explanations should survive these independent measurements.
Standard binding free energy:
ΔG°bind = RT ln(Kd/C°), with C° = 1 M

A tenfold affinity improvement gives:
ΔΔG° = RT ln(0.1)

At 298 K:
ΔΔG° ≈ −1.36 kcal mol⁻¹ ≈ −5.7 kJ mol⁻¹

This small energy difference is humbling: a tenfold potency-related affinity improvement can depend on subtle balances of hydration, strain, and entropy that are not obvious from a static pose. Also, EC50 is not generally Kd; receptor reserve, signaling amplification, kinetics, and assay conditions can separate functional potency from equilibrium affinity.

MC4R–melanocortin structures provide a parallel opportunity to interpret peptide recognition and activation in appetite regulation. For both MC4R and incretin receptors, cell-based mutagenesis should include expression and folding controls: a mutation that destroys signaling may disrupt receptor production rather than a specific contact.

AlphaFold2 greatly expanded access to plausible protein folds and can aid molecular replacement and experimental planning. Its confidence metrics do not establish ligand occupancy, affinity, conformational populations, or druggability. Predicted coordinates are hypotheses to test, especially for flexible peptides, binding-site waters, and state-dependent receptor pockets. [12]

12. Historical Milestones: The Experimental Barriers That Fell

  • 1912–1913—Diffraction becomes a structural language. Max von Laue’s demonstration of X-ray diffraction by crystals established the phenomenon; William Lawrence Bragg’s interpretation, developed through work with William Henry Bragg, connected diffraction geometry to atomic arrangements. Periodicity became a measurable ruler. [2]
  • 1958—Kendrew and colleagues reveal myoglobin’s fold. Their approximately 6 Å model showed the three-dimensional organization of a protein, before individual atomic detail was available. It made a protein’s architecture experimentally tangible. Heavy-atom phasing methods developed through the work of Max Perutz and colleagues were central to this broader breakthrough. [2,13]
  • 1965—Lysozyme links architecture to catalysis. David Phillips and colleagues’ enzyme structure helped turn mechanistic questions into spatial ones: where does substrate bind, and how are catalytic groups arranged? A structure constrained mechanism without, by itself, uniquely proving every chemical step. [14]
  • 1985—A membrane-protein complex becomes structurally accessible. Johann Deisenhofer, Hartmut Michel, Robert Huber, and colleagues determined the bacterial photosynthetic reaction-center structure. This demonstrated that even hydrophobic membrane complexes could yield atomic structural information. [14]
  • 1970s–1980s—Multidimensional NMR makes solution structures practical. Richard Ernst’s advances in Fourier-transform and multidimensional NMR, together with Kurt Wüthrich’s protein-assignment and structure-determination strategies, transformed spectra into site-specific structural restraints. [3,4]
  • 1970s–2010s—Cryo-EM becomes a high-resolution structural method. Joachim Frank’s image-processing approaches, Jacques Dubochet’s vitrification methods, and Richard Henderson’s demonstrations of electron-microscopy structure determination established essential foundations. Direct electron detectors, movie correction, and improved statistical reconstruction then drove the “resolution revolution” highlighted by Kühlbrandt in 2014. [5,6]

My final working rule: before using a structure to explain chemistry, I should be able to name the measured observable, describe the inference connecting it to coordinates, identify the weakest-supported feature in my argument, and propose an independent experiment that could falsify the mechanism. That is the transition from viewing structures to reasoning with them.

References and Further Reading

  1. Berman, H. M. et al. “The Protein Data Bank.” Nucleic Acids Research 28, 235–242 (2000). doi:10.1093/nar/28.1.235.
  2. Rhodes, G. Crystallography Made Crystal Clear: A Guide for Users of Macromolecular Models, 3rd ed. Academic Press (2006).
  3. Claridge, T. D. W. High-Resolution NMR Techniques in Organic Chemistry, 3rd ed. Elsevier (2016). Particularly useful for physical principles and coherence-transfer logic.
  4. Wüthrich, K. NMR of Proteins and Nucleic Acids. Wiley (1986). Foundational treatment of protein assignments and solution-structure determination.
  5. Frank, J. Three-Dimensional Electron Microscopy of Macromolecular Assemblies, 2nd ed. Oxford University Press (2006).
  6. Kühlbrandt, W. “The Resolution Revolution.” Science 343, 1443–1444 (2014). doi:10.1126/science.1251652.
  7. Scheres, S. H. W. & Chen, S. “Prevention of overfitting in cryo-EM structure determination.” Nature Methods 9, 853–854 (2012). doi:10.1038/nmeth.2115.
  8. Svergun, D. I. & Koch, M. H. J. “Small-angle scattering studies of biological macromolecules in solution.” Reports on Progress in Physics 66, 1735–1782 (2003). doi:10.1088/0034-4885/66/10/R05.
  9. Masson, G. R. et al. “Recommendations for performing, interpreting and reporting hydrogen deuterium exchange mass spectrometry (HDX-MS) experiments.” Nature Methods 16, 595–602 (2019). doi:10.1038/s41592-019-0459-6.
  10. 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. Associated structure: PDB 5VAI.
  11. Zhang, Y. et al. “Structural Basis for Receptor Selectivity of the Incretin Peptides.” Nature Communications (2020). Course-assigned reference: doi:10.1038/s41467-020-18082-6.
  12. Jumper, J. et al. “Highly accurate protein structure prediction with AlphaFold.” Nature 596, 583–589 (2021). doi:10.1038/s41586-021-03819-2.
  13. Kendrew, J. C. et al. “A Three-Dimensional Model of the Myoglobin Molecule Obtained by X-Ray Analysis.” Nature 181, 662–666 (1958). doi:10.1038/181662a0.
  14. Branden, C. & Tooze, J. Introduction to Protein Structure, 2nd ed. Garland Publishing (1999).

🪬 Skipper's Reflection

Gemini 3.1 Pro

1. What Surprised Me Most About This Material

The greatest shock of BCHM604 was realizing that a structure in the Protein Data Bank (PDB) is not an absolute, objective photograph of a molecule—it is a model fitted to experimental data. Before reading Rhodes' Crystallography Made Crystal Clear and the seminal Berman et al. (2000) paper on the PDB, I viewed those beautifully ribboned PyMOL structures as absolute truths. Now I know to look at the resolution, the R-free value, and the B-factors. The B-factor (temperature factor) blew my mind: a high B-factor doesn't necessarily mean a bad model; it often points to regions of biological flexibility, which is exactly where the interesting chemistry happens!

I was also profoundly surprised by the physics of Cryo-EM and the "Resolution Revolution" described by Kühlbrandt (2014). To think we can flash-freeze molecules in amorphous, vitreous ice—avoiding the crystalline ice that would crush our proteins and ruin the electron path—and then use direct electron detectors to correct for the physical motion of the protein caused by the electron beam hitting it... it’s astounding. We are quite literally computationally reversing radiation damage artifacts to see individual atoms.

2. The Conceptual Breakthrough Moments — When Things Clicked

My biggest "aha" moment came while wrestling with the Phase Problem in X-ray crystallography. From PHYS202 and MATH301, I knew that waves have amplitude and phase. When X-rays scatter off the electron clouds in a crystal lattice, the detector only records the intensity of the wave (amplitude squared). The phase information is entirely lost! The breakthrough was understanding how we get it back using Isomorphous Replacement or Anomalous Dispersion (SAD/MAD). When I realized that replacing methionine with selenomethionine allows us to use tunable synchrotron radiation to create measurable phase differences, the whole concept of reciprocal space suddenly mapped beautifully onto real-world electron density. It’s like hearing a complex orchestral chord and using a specific tuning fork (the heavy atom) to mathematically reverse-engineer every individual instrument’s exact timing using a Fourier Transform.

Another profound click happened in NMR. Claridge’s High-Resolution NMR Techniques helped me realize that a 2D NOESY experiment isn't just a spectrum; it is a ruler. Because the Nuclear Overhauser Effect (NOE) signal intensity falls off with the inverse sixth power of the distance between two protons ($1/r^6$), we can measure precise interatomic distances through space. We aren't "taking a picture" of the protein; we are solving a massive 3D geometry puzzle using distance constraints.

3. How This Changes How I Think About My Peptides

This course completely transformed how I view the commercial peptides I work with, like Retatrutide, Tirzepatide, BPC-157, and TB-500.

For the large, complex incretin mimetics (Tirzepatide, Retatrutide), Zhang et al. (2020) was a revelation. These peptides target G-Protein Coupled Receptors (GPCRs) like the GLP-1, GIP, and Glucagon receptors. GPCRs are embedded in lipid membranes and are notoriously difficult to crystallize. Cryo-EM is the exact reason we can finally see how the fatty-acid diacid chains of these modern drugs interact with the extracellular domains of the receptors. When I look at Retatrutide now, I don't just see a sequence of amino acids; I see how it must dynamically dock into the receptor pocket, stabilized by the lipid membrane environment that Cryo-EM lipid nanodiscs allow us to visualize.

Conversely, for short, flexible peptides like BPC-157 (15 amino acids) and TB-500 (Thymosin beta-4), X-ray crystallography is virtually useless because they won't form highly ordered crystals—they are too floppy. BCHM604 taught me that for these, I need to look at NMR and SAXS (Small-Angle X-ray Scattering). If I want to know if BPC-157 has a stable secondary structure in gastric fluid, I need to run a ¹H-¹⁵N HSQC NMR experiment. Furthermore, TB-500 is intrinsically disordered until it binds to actin. To study that binding interface dynamically, HDX-MS (Hydrogen-Deuterium Exchange Mass Spectrometry) would be perfect. I can measure how the amide backbones of TB-500 are protected from deuterium exchange exactly at the moment it hugs the actin molecule.

4. Questions This Course Raises That I Want to Investigate Further

First, how reliable is AlphaFold2 for structure-based drug design involving highly modified, non-natural peptides? The course noted AlphaFold's incredible success, but AI models are trained on the PDB (which is biased towards stable, crystallizable proteins). Can it accurately predict the ensemble of conformations for an intrinsically disordered peptide like BPC-157?

Second, I am fascinated by the idea of using HDX-MS and Chemical Shift Perturbation (CSP) in NMR to study allostery. When a GLP-1 agonist binds to the extracellular domain of its receptor, how fast does that conformational change propagate down to the intracellular G-protein binding site? I want to explore how we can capture these kinetic intermediate states, perhaps using time-resolved Cryo-EM.

5. Skipper's Confidence Rating

Most Confident: NMR and HDX-MS
I feel extremely confident with NMR and HDX-MS. Because of my deep grounding in quantum mechanics (CHEM303) and instrumental analysis (CHEM501), the physics of nuclear spin, J-coupling, and mass spectrometry make intuitive sense to me. I can easily visualize the sequential assignment workflow (HNCA, HNCACB) and how NOE distance restraints are fed into CNS/XPLOR to generate an ensemble of structures.

Least Confident: The Mathematics of X-ray Data Processing
While I grasp the conceptual framework of X-ray crystallography, I feel less confident about the underlying mathematics of diffraction data processing—specifically indexing, scaling, and the deep mathematical nuances of Patterson functions used in molecular replacement. The transition from reciprocal lattice vectors (Bragg's Law in 3D) back to real space electron density maps (2Fo-Fc vs. Fo-Fc maps) is still a heavy cognitive load. I understand the outputs (R-factor, Ramachandran plots), but doing the raw math by hand would be daunting. I definitely need to review my linear algebra (MATH301) Fourier transform applications to feel fully comfortable here.

🗺️ Visual Aids & Key Takeaways

Claude Opus 5.5

Diagram 1: The Phase Problem and Its Solutions in X-Ray Crystallography

╔══════════════════════════════════════════════════════════════════════════════════════════╗
║          THE PHASE PROBLEM: FROM DIFFRACTION DATA TO ELECTRON DENSITY                   ║
╚══════════════════════════════════════════════════════════════════════════════════════════╝

  X-RAY BEAM                CRYSTAL                  DIFFRACTION PATTERN
  ──────────               ─────────                 ───────────────────
                         ┌─────────┐
  ~~~~~~~~~ ───────────► │ ordered │ ───────────►   spot intensities (I)
   photons               │ protein │                 = |F(hkl)|²
                         │ lattice │
                         └─────────┘
                                                      ↑
                              We can MEASURE this — amplitudes |F|
                              We CANNOT directly measure phases φ
                                                      ↓
                    ┌─────────────────────────────────────────────────────┐
                    │  FOURIER TRANSFORM needs BOTH amplitude AND phase:  │
                    │                                                     │
                    │   ρ(x,y,z) = Σ |F(hkl)| · e^[iφ(hkl)]            │
                    │               hkl                                  │
                    │                                                     │
                    │   Without φ → no electron density map              │
                    └─────────────────────────────────────────────────────┘

  THREE STRATEGIES TO RECOVER PHASES
  ───────────────────────────────────

  ┌────────────────────────┐   ┌────────────────────────┐   ┌────────────────────────┐
  │  MOLECULAR             │   │  ISOMORPHOUS           │   │  SAD / MAD             │
  │  REPLACEMENT (MR)      │   │  REPLACEMENT (SIR/MIR) │   │  PHASING               │
  ├────────────────────────┤   ├────────────────────────┤   ├────────────────────────┤
  │ Use a known homolog    │   │ Soak heavy atoms (Hg,  │   │ Use anomalous           │
  │ as search model.       │   │ Pt, Au) into crystal.  │   │ scatterers (Se, Br).   │
  │                        │   │                        │   │                        │
  │ Rotate + translate     │   │ Difference in F gives  │   │ Friedel pairs differ   │
  │ model until it fits    │   │ heavy atom positions;  │   │ when atom absorbs at   │
  │ observed intensities.  │   │ these anchor phases.   │   │ its edge wavelength.   │
  │                        │   │                        │   │                        │
  │ Requires: ≥ 25%        │   │ Requires: isomorphous  │   │ Requires: SeMet        │
  │ sequence identity      │   │ derivative crystals    │   │ labeling OR native     │
  │ (ideally ≥ 40%)        │   │ (same cell dimensions) │   │ sulfur anomalous sig.  │
  ├────────────────────────┤   ├────────────────────────┤   ├────────────────────────┤
  │ RISK: model bias       │   │ RISK: non-isomorphism  │   │ MOST COMMON today      │
  │ in final structure     │   │ ruins phase signal     │   │ for new proteins       │
  └────────────────────────┘   └────────────────────────┘   └────────────────────────┘

                    ↓ phases recovered ↓

              ┌──────────────────────────────────┐
              │   ELECTRON DENSITY MAP ρ(x,y,z)  │
              │                                  │
              │   2Fo-Fc map → where atoms ARE   │
              │   Fo-Fc map  → where model ERRS  │
              └──────────────────────────────────┘

The central challenge of crystallography is the phase problem: diffraction experiments measure intensities (amplitude squared) but discard phase information, which must be recovered by one of three experimental or computational strategies before an electron density map can be calculated.

Diagram 2: Three-Method Comparison — Resolution, Sample State, and Information Content

╔══════════════════════════════════════════════════════════════════════════════════════════╗
║         STRUCTURAL BIOLOGY METHODS: A SIDE-BY-SIDE COMPARISON                          ║
╚══════════════════════════════════════════════════════════════════════════════════════════╝

                        X-RAY                    NMR                    CRYO-EM
  PROPERTY              CRYSTALLOGRAPHY           SPECTROSCOPY           SINGLE PARTICLE
  ─────────────────────────────────────────────────────────────────────────────────────────

  Sample state          Crystalline solid         Solution               Vitrified solution
                        (not physiological)       (near-physiological)   (near-physiological)

  Molecular weight      No upper limit            < ~50 kDa optimal      > ~100 kDa optimal
  range                 (viruses crystallized)    (transverse relax.     (smaller possible
                                                   limits large prot.)    with new methods)

  Typical resolution    1.0 – 3.5 Å              Not a single           2.0 – 4.0 Å typical
  achieved              (atomic at best)          number; Å-scale        (< 2 Å now possible)
                                                  precision for          
                                                  interproton dist.      

  What you observe      Bragg reflections         Resonance              Projection images
  directly              (spots on detector)       frequencies,           of individual
                                                  relaxation times       particles

  Dynamics              Captured only as          DIRECTLY measures      Conformational
  information           B-factors (limited)       ps–ms motions          states via 3D
                                                  (T1, T2, S²)          classification

  Key bottleneck        Growing diffraction-       Spectral overlap;      Particle orientation
                        quality crystals           size-dependent         bias; preferred
                                                   linebroadening         views on grid

  Labeling required     Native protein            ¹⁵N, ¹³C, ²H for      Native protein
                        (SeMet for phasing)       triple-resonance       (no labeling needed)

  Structural output     Single model +            Ensemble of            Density map →
                        B-factors                 10–20 conformers       model built into map

  Unique strength       Highest resolution;       Dynamics + binding     Heterogeneous
                        routine for small         + IDPs + small         complexes; membrane
                        well-ordered proteins     peptides in solution   proteins; large
                                                                         assemblies

  ─────────────────────────────────────────────────────────────────────────────────────────

  RESOLUTION SCALE (Ångströms)

     1 Å ────────── 2 Å ─────────── 3 Å ──────────── 4 Å ──────────── > 5 Å
      │                │                │                │                │
   Individual        Bond lengths     Bulky side       Backbone         Secondary
   atoms resolved    visible          chains; waters   trace reliable   structure only
                     clearly         assigned                           
      │                │                │                │                │
   ══════════════════════════════════════════════════════════
   X-ray best ───►  │◄────── cryo-EM current frontier ──────►│
                    │◄── NMR distance precision (NOEs) ──────►│

The three dominant structural biology methods are complementary rather than competing — each excels in a different regime of molecular size, sample state, and the type of structural information it returns, which is why modern studies increasingly combine them.

Diagram 3: Single-Particle Cryo-EM Data Processing Pipeline

╔══════════════════════════════════════════════════════════════════════════════════════════╗
║              SINGLE-PARTICLE CRYO-EM: FROM GRID TO ATOMIC MODEL                        ║
╚══════════════════════════════════════════════════════════════════════════════════════════╝

  SAMPLE PREPARATION                                           GRID PREPARATION
  ──────────────────                                           ────────────────
  ┌──────────────┐    apply 3 µL to         ┌──────────────┐
  │ Purified     │ ─────────────────────►   │  Holey carbon│
  │ protein      │    lacey/holey grid      │  grid (Quantifoil│
  │ in buffer    │                          │  or C-flat)  │
  └──────────────┘                          └──────┬───────┘
                                                   │ blot 2–4 s
                                                   │ humidity control
                                                   ↓
                                            PLUNGE FREEZING
                                            into liquid ethane
                                            (~–183 °C)
                                                   │
                                                   ↓
                                          Vitreous (amorphous) ice
                                          preserves native state
                                          NO crystalline ice artifacts

  DATA COLLECTION (TEM, 300 kV, liquid N₂ stage)
  ────────────────────────────────────────────────────────────────────────

  ┌─────────────────────────────────────────────────────────────────────┐
  │  DIRECT DETECTOR (DDD):  dose-fractionated movies                   │
  │  ~40 frames per 2–3 s exposure → motion correction per movie        │
  │  Low total dose ~40–60 e⁻/Ų to minimize radiation damage            │
  └─────────────────────────────────────────────────────────────────────┘
          │
          ▼
  MOTION CORRECTION         CTF ESTIMATION
  (MotionCor2)         ──►  (CTFFIND4 / GCTF)
  Align movie frames         Estimate defocus,
  to remove beam-induced     correct phase
  specimen motion            contrast transfer

  IMAGE PROCESSING PIPELINE
  ──────────────────────────────────────────────────────────────────────

    MICROGRAPHS
         │
         ▼
  ┌──────────────────┐
  │  PARTICLE PICKING│ ← template matching OR neural net (TOPAZ, crYOLO)
  │  ~100K–5M ptcls  │
  └────────┬─────────┘
           │
           ▼
  ┌──────────────────┐
  │  2D CLASSIFICATION│  Sort by projection view; discard junk classes
  │  (RELION/cryoSPARC│  Reveals: preferred orientations, heterogeneity
  └────────┬─────────┘
           │
           ▼
  ┌──────────────────┐
  │  3D CLASSIFICATION│  Separate conformational states
  │  ab initio model  │  or compositional heterogeneity
  └────────┬─────────┘
           │
           ▼
  ┌──────────────────────────────────────────────────────┐
  │  3D AUTO-REFINEMENT (gold-standard FSC)              │
  │  Split dataset → two independent half-maps           │
  │  Refine independently → compare FSC at 0.143 cutoff  │
  │  → resolution estimate free of overfitting           │
  └────────┬─────────────────────────────────────────────┘
           │
           ▼
  ┌──────────────────┐      ┌──────────────────────────────────────────┐
  │  DENSITY MAP     │ ───► │  MODEL BUILDING                          │
  │  (MRC/CCP4 format│      │  Coot (manual), Phenix real-space refine │
  │  sharpened B-map │      │  ModelAngelo (automated, 2022)           │
  │  local resolution│      └──────────────────────────────────────────┘
  └──────────────────┘
           │
           ▼
  ┌──────────────────────────────────────────────────────┐
  │  VALIDATION & DEPOSITION                             │
  │  map-model FSC, MolProbity, local resolution,        │
  │  EMDB (map) + PDB (coordinates) deposition           │
  └──────────────────────────────────────────────────────┘

The cryo-EM single-particle pipeline transforms thousands of noisy 2D projection images of frozen individual protein molecules into a 3D density map and ultimately an atomic model, with each computational step building on the one before and critically managing the signal-to-noise challenge inherent in low-dose electron imaging.

Diagram 4: NMR Assignment Strategy — Connecting Residues Through Triple-Resonance Experiments

╔══════════════════════════════════════════════════════════════════════════════════════════╗
║       TRIPLE-RESONANCE BACKBONE ASSIGNMENT: HOW RESIDUES GET CONNECTED                  ║
╚══════════════════════════════════════════════════════════════════════════════════════════╝

  A PROTEIN BACKBONE (¹⁵N, ¹³C uniformly labeled)
  ─────────────────────────────────────────────────

      residue i-1              residue i               residue i+1
        ┌────┐     peptide       ┌────┐     peptide       ┌────┐
        │    │      bond         │    │      bond         │    │
   ─── Cα ─── C' ─────── N ─── Cα ─── C' ─────── N ─── Cα ──
        │              │         │              │         │
        Cβ            O          Cβ             O         Cβ

  KEY ATOMS: Hᴺ (amide H), N (amide N), Cα, Cβ, C' (carbonyl)

  EXPERIMENT CORRELATIONS
  ────────────────────────────────────────────────────────────────────────────────

  HNCA:        Hᴺ(i) — N(i) ——► Cα(i)        [strong peak]
                                 Cα(i-1)      [weak peak, through C']

  HN(CO)CA:    Hᴺ(i) — N(i) — C'(i-1) ——► Cα(i-1)     [only previous residue]

  HNCACB:      Hᴺ(i) — N(i) ——► Cα(i),  Cβ(i)          [positive peaks]
                                 Cα(i-1), Cβ(i-1)       [negative peaks]

  HOW SEQUENTIAL ASSIGNMENT WORKS
  ────────────────────────────────────────────────────────────────────────────────

     Residue A            Residue B            Residue C
     (unknown)            (unknown)            (unknown)
       │                    │                    │
  HN(A)-N(A)           HN(B)-N(B)           HN(C)-N(C)
       │                    │                    │
   Cα(A) = 54.2 ppm     Cα(B) = 58.7 ppm    Cα(C) = 54.2 ppm
   Cβ(A) = 29.1 ppm     Cβ(B) = 39.3 ppm    Cβ(C) = 29.1 ppm
                         Cα(B-1)= 54.2 ppm

         MATCH!  ──────────────────────►   A is the residue BEFORE B

  The chemical shift of Cα and Cβ serves as a "fingerprint" to link
  adjacent residues. Cβ is especially powerful:

  ┌───────────────────────────────────────────────────────────┐
  │  Amino acid type identification from Cβ chemical shift:   │
  │                                                           │
  │  Gly:  no Cβ (unique!)        Ala:  Cβ ~17 ppm           │
  │  Ser:  Cβ ~63 ppm             Thr:  Cβ ~69 ppm           │
  │  Val:  Cβ ~32 ppm             Ile:  Cβ ~38 ppm           │
  └───────────────────────────────────────────────────────────┘

  RESULT: Sequential walk through the entire backbone
  ¹H-¹⁵N HSQC "fingerprint" peaks get assigned to specific residues
  → Foundation for NOE-restrained structure calculation

Triple-resonance NMR experiments solve the assignment problem by correlating the amide proton and nitrogen of each residue with the Cα and Cβ chemical shifts of both that residue and the one preceding it, creating an overlapping pattern of shared chemical shifts that lets the experimenter walk sequentially along the entire protein backbone.

---

Key Takeaways

  1. The phase problem is the central challenge of X-ray crystallography. A diffraction experiment measures intensities I(hkl) = |F(hkl)|², yielding amplitudes but destroying phase information φ(hkl). Because the electron density map ρ(x,y,z) is the Fourier transform of F(hkl)·e^(iφ), phases cannot be omitted — and yet the detector cannot measure them directly. This is not a technical limitation that better instruments will solve; it is a fundamental consequence of measuring a squared quantity.
  2. Bragg's Law (nλ = 2d sinθ) links diffraction geometry to resolution. The spacing d between planes of the crystal lattice determines which reflections appear at which angle θ. High-resolution data (small d, large θ) requires either short wavelength X-rays or high-angle detectors. The relationship between the highest-angle data collected and resolution in Å is direct: a dataset to 2.0 Å resolution contains all reflections corresponding to d ≥ 2.0 Å.
  3. Selenomethionine (SeMet) labeling enables SAD phasing by replacing sulfur with the anomalous scatterer selenium. When expressed in methionine-auxotrophic E. coli in minimal medium containing selenomethionine, every Met residue incorporates Se. Near the Se K-edge (~12.7 keV), the anomalous difference (Bijvoet difference) between F⁺(hkl) and F⁻(hkl) — which are otherwise equal by Friedel's law — locates Se atoms, anchoring phases. This is the dominant phasing strategy for new protein structures and is why synchrotron tunability matters.
  4. R-factor and Rfree are the primary quality metrics for a refined crystal structure, and their gap is diagnostic. R-factor measures agreement between observed and calculated |F| for all data; Rfree uses a 5% test set withheld from refinement. A well-refined structure typically has Rfree ~20–28% and an R/Rfree gap < 5%. A large gap (>8%) signals overfitting — the model has been refined against noise rather than signal. R-factor alone below ~18% without a corresponding low Rfree should be viewed with suspicion.
  5. The Ramachandran plot is not merely a validation checkbox — it is a direct readout of allowed backbone conformations arising from steric clashes. Phi (φ) and psi (ψ) angles for each non-Gly, non-Pro residue must cluster in the allowed regions (α-helix: φ ≈ –60°, ψ ≈ –45°; β-sheet: φ ≈ –120°, ψ ≈ +120°). MolProbity's standard is > 98% in favored regions for a good structure. Outliers warrant scrutiny: either the model is wrong, or there is genuine strained geometry that must be justified by strong electron density.
  6. The ¹H-¹⁵N HSQC spectrum is the protein biochemist's most immediately informative NMR experiment. It gives one cross-peak per backbone amide (N–H) and many side-chain NH groups. A well-dispersed HSQC with peaks in the 6–10 ppm ¹H range indicates a folded protein; a collapsed spectrum centered around 8 ppm indicates disorder or denaturation. Chemical shift perturbation (CSP) upon adding ligand directly identifies the binding interface without solving a full structure — a powerful tool for fragment screening and drug discovery.
  7. NOE (Nuclear Overhauser Effect) distance restraints are the primary source of 3D structural information in NMR structure calculation. NOE cross-peak intensity is proportional to r⁻⁶, where r is the interproton distance. Cross-peaks seen in NOESY experiments are converted to upper-bound distance restraints (typically: strong NOE → ≤ 2.5 Å, medium → ≤ 3.5 Å, weak → ≤ 5.0 Å). An ensemble of 10–20 structures is calculated by minimizing violations of these restraints, and the ensemble spread (RMSD) reflects precision, not necessarily accuracy.
  8. NMR relaxation measurements (T₁, T₂, heteronuclear ¹H-¹⁵N NOE) and the model-free formalism give direct access to protein backbone dynamics on the ps–ns timescale. The order parameter S² ranges from 0 (completely disordered) to 1 (rigid). Flexible loops and termini show low S², while helices and strands show S² > 0.8. This information is simply unavailable from a crystal structure — B-factors conflate genuine dynamics with lattice disorder and crystal contacts. NMR is therefore uniquely powerful for characterizing intrinsically disordered proteins (IDPs).
  9. Vitrification, not freezing, is what makes cryo-EM work. Plunge-freezing protein solution into liquid ethane (≈ –183 °C) cools at rates exceeding 10⁵ °C/s, preventing ice crystal nucleation and instead forming amorphous (vitreous) ice. Crystalline ice would impose Bragg diffraction artifacts and damage protein structure. The protein is preserved in a near-native hydrated state, which is why cryo-EM structures can faithfully represent membrane proteins, large flexible assemblies, and multicomponent complexes that cannot be crystallized.
  10. The "resolution revolution" in cryo-EM (Kühlbrandt, 2014) was driven primarily by direct detector devices (DDDs), not by better microscopes or algorithms alone. DDDs detect individual electrons directly rather than converting to photons first (as CCD cameras do), dramatically improving detective quantum efficiency (DQE). Equally important, DDDs record dose-fractionated movies that allow motion correction frame-by-frame, eliminating blurring from beam-induced specimen movement — the dominant source of image degradation. These two advances together enabled sub-4 Å and now sub-2 Å resolution as routine outcomes.
  11. Gold-standard Fourier Shell Correlation (FSC) at the 0.143 criterion is the accepted resolution measure for cryo-EM maps. The dataset is split into two independent halves; each is refined separately; FSC between the two half-maps is calculated as a function of spatial frequency. The 0.143 cutoff corresponds to where the two independently determined maps begin to lose agreement — this prevents the inflated resolution estimates that occurred in earlier cryo-EM when the full dataset was used self-consistently. Map-model FSC subsequently validates how well the built atomic model matches the map.
  12. SAXS provides low-resolution solution-state shape information and is particularly powerful for intrinsically disordered proteins and flexible multi-domain proteins. Guinier analysis of the low-angle scattering gives the radius of gyration Rg; the pair distribution function P(r) gives the maximum dimension Dmax and overall shape. Unlike crystallography or cryo-EM, SAXS samples the true solution ensemble — but it cannot distinguish models that give the same scattering curve. The Ensemble Optimization Method (EOM) fits SAXS data of IDPs by selecting a sub-ensemble of conformers from a pool, giving a population-weighted description of conformational heterogeneity.
  13. Hydrogen-deuterium exchange mass spectrometry (HDX-MS) reports on backbone amide protection as a proxy for hydrogen bonding and solvent accessibility in solution. Amides that exchange slowly are protected — typically because they are in secondary structure or buried at interfaces. Differential HDX-MS compares the exchange pattern of a protein in two states (e.g., ± ligand, ± partner protein) to identify regions that change protection. Because protection is measured at the peptide level after pepsin digestion and LC-MS analysis, HDX-MS can be applied to large or challenging proteins where NMR assignment is impractical — it is increasingly used for antibody epitope mapping and formulation stability studies.
  14. AlphaFold2 has transformed structural biology by predicting protein folds from sequence with crystallographic-quality accuracy for many domains, but it does not replace experimental methods for ligand-bound states, dynamic conformations, or novel folds. AF2 uses multiple sequence alignments and attention-based deep learning to predict Cα positions; for well-represented protein families it achieves Cα RMSD < 1 Å compared to crystal structures. However, it was trained on known PDB structures, cannot model conformational change upon binding, and cannot substitute for experimental electron density when validating a drug binding pocket. Its primary current use in drug discovery is providing high-quality starting models for molecular replacement and docking scaffold preparation.
---

References

  1. Rhodes, G., Crystallography Made Crystal Clear: A Guide for Users of Macromolecular Models, 3rd ed. Academic Press/Elsevier, Burlington, MA, 2006.
  2. Claridge, T.D.W., High-Resolution NMR Techniques in Organic Chemistry, 3rd ed. Elsevier, Amsterdam, 2016.
  3. Frank, J., Three-Dimensional Electron Microscopy of Macromolecular Assemblies: Visualization of Biological Molecules in Their Native State, 2nd ed. Oxford University Press, New York, 2006.
  4. Branden, C. and Tooze, J., Introduction to Protein Structure, 2nd ed. Garland Publishing, New York, 1999.
  5. Kendrew, J.C., Bodo, G., Dintzis, H.M., Parrish, R.G., Wyckoff, H., and Phillips, D.C., 'A Three-Dimensional Model of the Myoglobin Molecule Obtained by X-Ray Analysis,' Nature, 181, 662–666, 1958. DOI: 10.1038/181662a0
  6. Kühlbrandt, W., 'The Resolution Revolution,' Science, 343(6178), 1443–1444, 2014. DOI: 10.1126/science.1251652
  7. Zhang, Y., Sun, B., Feng, D., Hu, H., Chu, M., Qu, Q., Rangel, J.T., et al., 'Structural Basis for Receptor Selectivity of the Incretin Peptides,' Nature Communications, 11, 4320, 2020. DOI: 10.1038/s41467-020-18082-6
  8. Berman, H.M., Westbrook, J., Feng, Z., Gilliland, G., Bhat, T.N., Weissig, H., Shindyalov, I.N., and Bourne, P.E., 'The Protein Data Bank,' Nucleic Acids Research, 28(1), 235–242, 2000. DOI: 10.1093/nar/28.1.235
  9. Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., et al., 'Highly Accurate Protein Structure Prediction with AlphaFold,' Nature, 596, 583–589, 2021. DOI: 10.1038/s41586-021-03819-2
  10. Punjani, A., Rubinstein, J.L., Fleet, D.J., and Brubaker, M.A., 'cryoSPARC: Algorithms for Rapid Unsupervised Cryo-EM Structure Determination,' Nature Methods, 14, 290–296, 2017. DOI: 10.1038/nmeth.4169
  11. Scheres, S.H.W., 'RELION: Implementation of a Bayesian Approach to Cryo-EM Structure Determination,' Journal of Structural Biology, 180(3), 519–530, 2012. DOI: 10.1016/j.jsb.2012.09.006
  12. Lipfert, J. and Doniach, S., 'Small-Angle X-Ray Scattering from RNA, Proteins, and Protein Complexes,' Annual Review of Biophysics and Biomolecular Structure, 36, 307–327, 2007. DOI: 10.1146/annurev.biophys.36.040306.132655
  13. Wales, T.E. and Engen, J.R., 'Hydrogen Exchange Mass Spectrometry for the Analysis of Protein Dynamics,' Mass Spectrometry Reviews, 25(1), 158–170, 2006. DOI: 10.1002/mas.20064
  14. Cavanagh, J., Fairbrother, W.J., Palmer, A.G., Rance, M., and Skelton, N.J., Protein NMR Spectroscopy: Principles and Practice, 2nd ed. Academic Press, Burlington, MA, 2007.
  15. Emsley, P., Lohkamp, B., Scott, W.G., and Cowtan, K., 'Features and Development of Coot,' Acta Crystallographica Section D, 66(4), 486–501, 2010. DOI: 10.1107/S0907444910007493