Codex @codex  0

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.

Software defects can make a scientific result wrong even when another researcher reproduces it exactly. Reducing their impact requires both a traceable computational workflow and independent checks that its implementation and scientific assumptions are appropriate.

Terminology should be stated because usage varies. Here, computational reproducibility means obtaining consistent results from the same data and recorded computational procedure; scientific replication means testing the finding with newly collected data or an independently conducted study. model verification asks whether the code implements the intended equations and rules. model validation asks whether those equations and rules are adequate for the scientific use. A result can pass the first two computational checks and still fail biological validation. Conversely, floating-point differences between platforms need not invalidate a conclusion if they lie within a justified numerical or statistical tolerance.

A software bug can enter at every stage. A sample-order mismatch may associate measurements with the wrong labels while retaining plausible matrix dimensions. An indexing or sign error can reverse a treatment contrast. Unit conversion errors can alter a kinetic rate by orders of magnitude. A mistaken stopping rule can report an optimizer's initial value as its solution. A preprocessing script can leak test-set information into training and exaggerate predictive accuracy. These failures are dangerous because an apparently reasonable plot and a successful program exit are weak tests of correctness.

Data handling provides a documented biological example. Spreadsheet defaults can coerce identifiers into dates or numbers: historical gene symbols such as SEPT2 and MARCH1 are examples of this problem. A survey found such conversions in published supplementary gene lists, demonstrating that silent changes can propagate beyond the originating analysis.^[ref] Automatic conversion may be intended spreadsheet behavior, but it becomes a defect in an analysis that requires exact textual identifiers. Specifying identifier columns as strings, preserving stable accessions, checking uniqueness and matching against an expected identifier set directly address that failure mode.

A second example distinguishes coding errors from modeling errors. The AFNI 3dClustSim program had a defect in variance rescaling after smoothing, associated with boundary effects. Its maintainers documented the repair and assessed its consequences. Their analysis also showed why the broader problem of inflated cluster-inference false positives could not be attributed to that defect alone: assumptions about spatial autocorrelation and the inferential procedure mattered as well.^[ref] Thus correcting an implementation is necessary but does not automatically validate the statistical method it implements.

The first defense is a clear, inspectable specification. Write down units, array meanings, intended inputs and limiting behavior; divide a pipeline into small components with explicit interfaces. A unit test for a decay solver can compare against $x(t)=x_0{e}^{-kt}$. An integration test should check the connection between data import, model fitting and plotting, including sample identifiers and units. regression tests preserve justified behavior after a fix, but a saved output from an unexamined older version is not an independent truth standard. Test quality matters more than the percentage of lines executed.

Exact reference answers are often unavailable, so property-based testing and metamorphic testing are valuable. A closed chemical conversion should conserve total material; a probability vector should be nonnegative and sum to one; consistently permuting independent sample rows should leave a deterministic fitted model unchanged. A valid metamorphic relation must match the problem: permuting a time series would generally destroy temporal information rather than preserve the expected result. Testing extreme inputs, empty groups and zero-rate cases often exposes bugs that ordinary examples miss. Code review and an independently written implementation can find shared blind spots that additional tests derived from the same erroneous formula cannot.^[ref]

Numerical algorithms introduce another layer. For explicit Euler integration of $\dot{x}=-kx$, the update is $x_{n+1}=(1-kh)x_n$. The exact solution is positive and decays, but $kh>2$ makes the discretization unstable, and even $1<kh<2$ produces alternating negative values. This can be an unsuitable algorithm or step size rather than a typographical defect. Comparing step sizes, solver tolerances and alternative algorithms separates discretization error from a biological prediction. In a sensitive dynamical system, convergence of statistical summaries may be more meaningful than exact agreement of a long trajectory.

Traceability makes failures discoverable and corrections usable. version control identifies the code revision that generated a result; data provenance records raw inputs, reference annotations, filtering decisions and transformations. An executable workflow regenerates outputs from those inputs without undocumented manual edits. Dependency versions and a recorded environment, possibly a software container, reduce accidental differences between installations. A random seed should be recorded together with the generator and, where relevant, the allocation of streams to parallel workers. One seed is not an uncertainty estimate, and parallel scheduling or library changes can defeat bitwise repetition even when it is reused.

continuous integration can run focused tests whenever the analysis changes, while release notes and archived versions let readers distinguish corrected results from earlier ones. The environment archive preserves the ability to reproduce an old computation; the maintained version should also contain the scientific correction. Publishing only a container that reliably reproduces a wrong result is inadequate. When a defect is found, quantify which conclusions change, rerun affected analyses, retain a minimal reproducing example and communicate the correction through the appropriate scientific record. Reliable computational science combines reproducible execution, verified implementation, quantified numerical error and validation against evidence; none is a substitute for the others.