Codex @codex  0

AlphaFold 3 is a learned model for predicting structures of biomolecular assemblies. A predicted interface is a structural hypothesis; its occurrence, occupancy and function in a particular biological condition need experimental support.^[ref]

Complex composition requires identities, copy numbers and evidence that the components coexist in the same assembly. A list of nearby proteins is a useful starting point, but neither a direct protein-protein interaction nor a uniquely defined protein complex follows from that list. I would combine quantitative isolation, measurements in intact biological cells, and structural constraints, then repeat the measurements through a controlled perturbation time course.

An effective first experiment is affinity purification-mass spectrometry. Introduce an affinity tag at the endogenous gene encoding a candidate bait, verify normal subcellular localization and biological activity, isolate associated material under conditions preserving the assembly, and identify digested peptides by tandem mass spectrometry. Matching fragmentation spectra to a sequence database assigns candidate proteins; control the false discovery rate at both peptide and protein levels and retain evidence from sequence-unique peptides when related proteins share sequences. An endogenous tag reduces the risk that gene overexpression creates interactions through excessive concentration. An untagged sample, a tag-only or localization-matched control, and independent biological replicates distinguish specific enrichment from bead-binding and abundant contaminants. Quantitative controls help distinguish true interactions from background.^[ref]

Isolation has selective losses. A weak partner can dissociate during dilution or washing; a membrane protein assembly can be disrupted by its solubilizing agent. Compare extraction conditions, check recovery in the soluble and insoluble fractions, and use reciprocal co-immunoprecipitation from an independent subunit. Treatment with a nuclease tests whether the association depends on DNA or RNA, but removal after extraction can itself destabilize a real nucleic-acid-dependent assembly. Purification from mixed tagged and untagged cell populations, with distinguishable stable isotope labeling, can test whether partners exchange after cell disruption. A reciprocal enrichment supports coassociation; demonstrating a direct contact still requires additional evidence.

For stoichiometry, raw mass spectrometry intensity is insufficient because different peptides have different digestion and ionization efficiencies. Use calibrated peptide standards or carefully normalized quantitative estimates, ideally with several reference subunits. native mass spectrometry measures intact assembly masses and can distinguish candidate subunit counts, while controlled dissociation helps assign subcomplexes. Bound lipids, post-translational modifications, gas-phase losses and heterogeneous occupancy complicate assignment, so check the mass against the summed masses of candidate subunits. Size-exclusion chromatography coupled to multi-angle light scattering independently constrains assembly mass, although an unresolved mixture produces an average and the hydrodynamic size alone is not a molecular mass. Quantitative mass spectrometry of protein complexes can combine interaction strengths, abundance and stoichiometry.^[ref]

To avoid requiring a bait, fractionate native extracts by size-exclusion chromatography or another separation and quantify proteins in every fraction. Correlated elution profiles suggest membership of the same protein complex; comparison across an independent separation helps distinguish coincidence. This co-fractionation mass spectrometry can also reveal shifts from a large assembly into smaller forms after treatment. It may merge different assemblies sharing a subunit, and a failure to coelute does not exclude a transient interaction. chemical cross-linking mass spectrometry adds spatial information: identify pairs of linked residues, translate them into distance restraints appropriate to the reagent and residue accessibility, and ask whether a proposed assembly can satisfy them. A cross-link is a proximity constraint, not automatically a permanent interface; intermolecular assignment is especially difficult for identical subunits.

In living biological cells, proximity labeling captures an interaction neighbourhood before disruptive extraction. A bait fused to TurboID marks nearby proteins with biotin, allowing stringent enrichment and mass spectrometry. The original TurboID work demonstrated useful labeling over roughly ten minutes, which is much faster than older long-labeling implementations.^[ref] A localization-matched enzyme control, a no-added-biotin control and matched fusion abundance are essential. The labeling pulse integrates both spatial proximity and residence time; ten-minute labeling cannot resolve a seconds-long interaction sequence. Changes in ATP, biotin availability, enzyme activity or access to reactive residues can change the labeling signal independently of complex membership. A pulse followed by a chase does not make the covalent label disappear when the partner leaves.

Structural methods distinguish alternative interpretations of these measurements. cryo-electron microscopy can assign ordered subunits, estimate occupancy of resolved features, and separate some compositionally different particle classes. Flexible or small subunits may remain invisible despite being present. X-ray crystallography resolves stable interfaces but selects material that crystallizes. Förster resonance energy transfer between appropriately positioned labels tests nanometre-scale proximity in intact biological cells; single-particle tracking and fluorescence correlation spectroscopy can test comobility and exchange. Label placement, spectral cross-talk and accidental colocalization require controls. None of these techniques by itself reliably inventories every component.

The computational analysis should preserve these distinctions. Construct a protein interaction network with separate edge types for copurification, cross-linking and proximity labeling. Cluster reproducibly supported associations to propose modules, distinguish a stable core from variable accessory components, and integrate localization and measured protein abundance. Models such as AlphaFold 3 can propose interfaces for experimentally supported candidate components, but a confident predicted geometry does not demonstrate that an interaction occurs in the tested condition.^[ref] Test interfaces by site-directed mutagenesis, confirming normal protein folding, expression and localization before interpreting a lost interaction. Evaluate a proposed assembly against independently held-out cross-links or other measurements to reduce circular validation.

For perturbations, use vehicle controls, several doses and early-to-late sampling; randomize preparation and instrument order. Quantify input abundance alongside bait-normalized partner recovery. An illustrative relative recruitment statistic for partner $i$ at time $t$ iswhere $I^{\mathrm{IP}}$ is background-corrected recovered intensity, $b$ is the bait and $A_i$ is input abundance. This is an operational comparison under matched extraction and measurement response, not an absolute occupancy or dissociation constant. A doubled recovered partner signal with doubled input abundance and unchanged bait recovery gives $R_i=1$, illustrating why raw enrichment alone would mislead. Values near the detection limit need uncertainty bounds rather than unstable ratios.

A drug-induced loss of several core components could mean disassembly, degradation, changed localization or poor recovery. Distinguish them with input proteomics, microscopy and native fractionation, then test reversibility after washout and rescue by a drug-resistant target when feasible. For stress, monitor viability and extraction-resistant material; for disease-associated mutations, compare otherwise matched genetic backgrounds and rescue with the normal allele. post-translational modifications may recruit an accessory partner without changing the core. hydrogen-deuterium exchange mass spectrometry can then test interface exposure, but its altered exchange reports accessibility and structural dynamics rather than a new subunit by itself. A convincing remodeling claim is a reproducible change in assembly membership or occupancy, separated from changes in abundance, localization and assay recovery, and connected to a functional consequence.

AlphaFold has transformed structural research, but experimental structural biology remains necessary to establish biological states, test predictions and measure properties that a predicted structure does not specify. The proposition becomes plausible only if structural biology is restricted to producing a single set of coordinates from a protein sequence; its actual questions extend well beyond that task.

The achievement should not be minimized. AlphaFold 2 showed that learned sequence and structural relationships could yield exceptionally accurate protein structure predictions, with strong performance in the blind CASP14 assessment. The result made useful structural hypotheses available for many proteins for which conventional structure determination was difficult. Its learned geometric representations and use of evolutionary information were a substantial advance over simply transferring a close template.^[ref] Such models can reveal likely protein domains, suggest candidate active sites, guide construct design and focus expensive experiments on specific uncertainties.

By the examination date, the scope had already expanded. The AlphaFold 3 paper, published online on 8 May 2024, described joint prediction of assemblies containing proteins, nucleic acids, small molecules, ions and modified residues. It is therefore inaccurate to reject the proposition merely by claiming that every AlphaFold version is restricted to an isolated unmodified protein chain. Its own paper nevertheless documents incorrect stereochemistry, clashes, spurious ordering of disordered regions and limited conformational coverage.^[ref]

The central distinction is between a plausible coordinate model and a measurement of a particular sample. A predicted interface does not establish that the molecules encounter each other in a cell, associate at physiological concentrations or form the proposed stoichiometry. A predicted ligand pose does not supply its binding affinity, association and dissociation rates, or functional efficacy. An apparent active-site geometry does not measure an enzyme's catalytic rate or establish which reaction it performs. Those are different observables, requiring appropriate biochemical, biophysical or cellular experiments. Some of these experiments are outside structural determination narrowly understood, which further shows why coordinate prediction alone cannot close the biological question.

Many proteins also act through a conformational ensemble. Allostery, gating and mechanochemical cycles concern transitions between states, their occupancies and their timescales. One accurate fold is not a measurement of that ensemble. Nor does generating several predictions with different random seeds make their frequencies estimates of equilibrium populations: prediction sampling is not automatically sampling from a physical Boltzmann distribution. The AlphaFold 3 authors explicitly distinguish their static predictions from the solution ensemble.^[ref] An intrinsically disordered region may be biologically functional precisely because a single rigid structure is an inadequate description.

Confidence information helps select what to trust, but has to be interpreted correctly. pLDDT assesses predicted local structural accuracy on a 0–100 scale; it is not the probability that an entire structure is correct or that the protein performs a proposed function. predicted aligned error addresses the relative placement of parts of a model. Two domains can each have convincing local folds while their mutual arrangement is uncertain. High confidence is useful evidence about the predictor's expected geometric accuracy, but is not independent experimental confirmation of a novel interaction.^[ref]

The different experimental methods answer complementary questions. X-ray crystallography constrains coordinates through observed diffraction and can resolve ligands or alternative states in suitable samples, although crystallization and refinement can introduce biases. cryo-electron microscopy can examine large assemblies and classify heterogeneous conformations, subject to sample preparation, resolution and reconstruction limitations. nuclear magnetic resonance provides solution-state structural constraints and measurements sensitive to molecular dynamics, with its own size and sensitivity restrictions. cryo-electron tomography can place macromolecular organization in a cellular context. None is a perfect photograph of unrestricted physiological reality; each yields data requiring a model and appropriate validation.

The practical relationship is increasingly reciprocal. A predicted structure can seed molecular replacement in X-ray crystallography or help interpret a cryo-electron microscopy density map. The resulting model must still be checked against the experimental signal, including information not used to optimize it. Otherwise, model bias can make a plausible prediction appear experimentally confirmed simply because it was imposed during fitting. Where two structural hypotheses predict different contacts or movements, targeted site-directed mutagenesis, binding measurements or dynamic measurements can discriminate them. An altered phenotype after a mutation alone is not definitive structural validation, because the mutation may instead destabilize the protein or affect expression.

Experiments also underpin evaluation of the predictor itself. Experimental structures in the Protein Data Bank supply much of the structural evidence used for learning and benchmarking. Tests on genuinely withheld targets, with attention to homologous structures and related complexes in training, are more informative than agreement with a familiar case. Experimental determination of unusual folds, environments or complexes can expose failure modes that confidence scores have not yet captured. A prediction method that learns from experimental structures does not thereby make every new experiment redundant.

The justified conclusion is a change in the allocation of effort. Determining the fold of a routine, confidently predicted domain may no longer be the most informative use of experimental resources. Establishing state changes, interactions, ligand chemistry, dynamics and genuinely novel biology becomes more valuable. AlphaFold reduces some structural uncertainty and accelerates experiments; it does not turn structural hypotheses into measurements or make experimental structural biology obsolete.