Evidence checked October 31, 2026. A cited October 2026 preprint is identified as preliminary; version-specific statements distinguish AlphaFold2 from AlphaFold3. Worked scenarios are hypothetical.
1. The image is a proposal about matter
A protein structure prediction can look finished before the scientific question is finished. Its ribbon has a shape, its atoms have coordinates, and a molecular viewer allows the user to rotate it as if it were an object recovered from the world. That visual completeness is useful. It also hides a distinction between producing coordinates and establishing what those coordinates explain. A model can be an excellent starting point without settling the behavior of the molecule in a cell.
The distinction remains active research rather than an objection left over from an earlier generation of prediction. A preprint posted October 6, 2026, by Jonathan Feldman and Jeffrey Skolnick investigated how structural information emerges within AlphaFold3, using interventions on internal representations and inputs. The authors argue that information in multiple sequence alignments can act as a structural shortcut. The preprint is preliminary, not a settled explanation of every prediction or evidence that the system is generally unreliable. It makes a narrower point relevant to readers: an accurate output and a complete account of how that output is obtained are different achievements.
For researchers using predicted structures, there is a further distinction. Even a correct structural arrangement does not automatically answer how strongly a molecule binds, how rapidly an enzyme works, which state dominates in a particular environment, or whether changing a residue will alter a person's health. Those are related questions, but they ask for different evidence.
The useful response is neither to dismiss prediction nor to treat it as a universal measurement. It is to identify which uncertainty the prediction reduces and which uncertainty the next experiment must still address.
2. Local confidence is not a verdict on the whole protein
A protein model can contain regions with very different levels of support. One domain may be predicted confidently while a connecting segment is uncertain. If the viewer displays only a single average confidence number, that variation can disappear. A high average can coexist with uncertainty precisely where the user's biological question is located.
EMBL-EBI's AlphaFold2 training material distinguishes local confidence from confidence in the relative placement of regions. AlphaFold2 reports pLDDT and predicted aligned error, or PAE, for these purposes. The confidence-score guide recommends interpreting predictions critically using both kinds of information. The scores are tools for locating uncertainty, not a replacement for deciding which structural feature matters.
Consider a hypothetical protein with a well-predicted central domain and an uncertain terminal segment. A researcher investigating the central fold may have a useful structural hypothesis. Another researcher proposing that the terminal segment blocks a binding site has a weaker foundation, even though both are looking at the same file. The validity of a conclusion depends on the region and relation being interpreted, not merely on the model's overall reputation.
This is why a figure should preserve confidence information when it supports a structural argument. A uniformly colored ribbon can make unsupported details look as solid as supported ones. A well-designed illustration can instead show which part of the proposed mechanism rests on high-confidence structure and which part remains speculative. That is not visual clutter. It is the information needed to assess the claim.
3. Two good domains can have an uncertain relationship
Imagine two accurately drawn pieces of machinery placed on a table. Knowing the shape of each piece does not establish how they connect or move relative to each other. The same distinction appears in multidomain proteins. Local structure can be reliable while the arrangement between regions remains uncertain. A compelling overall silhouette may therefore be less secure than its components.
DeepMind's explanation of AlphaFold's confidence measures describes PAE as an expected positional error after alignment on another part of the structure. It specifically notes that high inter-domain PAE should discourage interpretation of relative domain positions. The developer's account also cautions that predicted coordinates remain predictions even when they are often highly accurate.
A hypothetical mechanism might depend on two domains touching. If the individual domains have strong local confidence but their relative placement is uncertain, the contact is not established by the image. The model may still help design a test of whether such a contact occurs. The correct scientific statement would describe a proposed interaction and the evidence needed to discriminate it from other arrangements.
This matters for comparisons between models as well. Two runs may agree on each domain and disagree on the arrangement between them. Reporting that the predictions are similar because most atoms align within each domain could conceal the biologically important difference. The comparison should match the question: local folds for local-fold questions, domain arrangement for domain-arrangement questions. A structural metric is informative only after the object of comparison has been defined.
4. Confidence is not a probability of biological truth
A score on a zero-to-one-hundred scale invites a familiar interpretation: ninety might sound like a ninety percent chance that the proposed biological story is correct. That is not what a local structural confidence score means. It concerns an estimated aspect of structural accuracy under the model's definitions. A mechanism assembled from that structure can contain additional assumptions that the score does not evaluate.
For example, a hypothetical enzyme model might confidently place a loop near a pocket. A researcher could then propose that the loop determines substrate selectivity. The prediction supports a structural hypothesis about the loop, not the entire causal explanation. The selectivity claim requires evidence about substrate interactions, competing conformations, and functional consequences. Multiplying the visual plausibility of those ideas does not convert the original confidence score into a probability for the full mechanism.
Calibration also depends on the evaluated setting. A score can be informative across a benchmark while still being wrong in particular cases. The presence of exceptions does not make the score useless; the absence of guaranteed correctness makes independent evidence important when the consequences of error are substantial. A useful confidence measure helps allocate attention rather than eliminating the need for judgment.
This distinction should shape headlines. Saying that a model is confident about a structural region is more precise than saying AI has confirmed a biological function. The latter phrase skips several inferential steps. The former can still describe a major advance: researchers may now have a sharply specified hypothesis where they previously had little structural guidance. Precision about that advance preserves its value without promising knowledge the calculation did not produce.
5. Version changes alter the meaning of the output
The name AlphaFold covers systems with different capabilities and output conventions. An explanation accurate for one version can mislead when transferred to another without qualification. This is particularly important when proteins are modeled alongside other molecular components rather than in a protein-only setting.
EMBL-EBI's AlphaFold3 guidance describes pLDDT at the atom level and PAE for pairs of tokens, reflecting the system's treatment of heterogeneous molecular complexes. It also explains that the inclusion of relevant non-polymer context can affect confidence. The AlphaFold3 quality guide supports a practical reporting requirement: identify the model version, inputs, and relevant score definitions before interpreting the result.
A hypothetical comparison between an older single-protein prediction and a newer complex prediction may show a different arrangement. That difference could reflect the changed molecular context, the model architecture, or other settings. It should not automatically be described as the newer system discovering motion or correcting a previously proven error. The runs may be answering different questions.
Reproducibility requires preserving those distinctions. A downloadable coordinate file without its input composition and generation details may be insufficient to reconstruct the claim. If a ligand, ion, partner chain, or sequence modification is central to the interpretation, it belongs in the description of the analysis. Readers should not have to infer the scientific setup from a screenshot. A structure prediction is the output of a specified computational experiment, and its provenance is part of the evidence.
6. A static structure cannot by itself measure motion
Proteins participate in processes that unfold over time. A single arrangement of atoms can illuminate those processes, but it does not directly provide a distribution of conformations, the rate of transitions, or the conditions that shift one state into another. A model's apparent openness or closure is a structural proposal, not a movie of the molecule's behavior.
EMBL-EBI's discussion of AlphaFold3 limitations explicitly distinguishes static prediction from dynamic behavior in solution. It notes that generating multiple predictions does not guarantee recovery of additional relevant states. The limitations guide also gives examples of state-selection difficulties. Those are version-specific cautions, not a claim that every protein or every prediction has the same limitation in practice.
Consider a hypothetical transporter whose activity involves more than one arrangement. A plausible model of one arrangement can help identify a candidate pathway or contact. It cannot alone establish how often that arrangement occurs or whether it is the rate-limiting state. A second model with a different shape does not automatically supply the missing kinetic information. The two outputs might reflect model uncertainty rather than an experimentally meaningful transition.
The right next question is what observation could distinguish the proposed states or measure their relevance. Different experimental approaches answer different aspects of that question. The key is to avoid treating computational diversity as a physical ensemble without evidence. A collection of coordinate files may be useful for exploration; assigning those files biological frequencies requires an additional argument.
7. Disorder is not simply a failed picture
A low-confidence region can be frustrating because it appears to be the part the model could not solve. But the interpretation is not always that a single stable structure exists and the program failed to find it. In some cases, the region may not have a single well-defined arrangement under the relevant conditions. In others, the model may lack sufficient information. These possibilities should not be collapsed into one explanation.
For analysis, the distinction matters because a poorly supported coordinate trace should not be treated as a precise object. Its apparent distance from another region may be arbitrary. Its shape may not justify a claim about a pocket, obstacle, or interaction. The presence of atoms in a file does not guarantee that the file contains useful spatial information for every segment.
A hypothetical study might map several disease-associated variants onto a low-confidence tail and infer that they cluster in one three-dimensional site. If the tail's arrangement is uncertain, that spatial cluster may be an artifact of the chosen model. The variants' positions in the sequence remain real, but the proposed spatial relationship needs a stronger basis.
An uncertainty-aware figure can show the segment without inviting overinterpretation. It might separate confidently modeled regions from uncertain ones and explain what can still be learned from sequence location or other evidence. The goal is not to erase difficult regions from biology. It is to keep a weakly supported shape from becoming an apparently precise explanation. Sometimes the most informative structural conclusion is that the available prediction does not resolve the relationship the researcher wants to discuss.
8. Correct overall shape can leave the decisive detail wrong
Many biological arguments depend on small spatial differences. An overall fold may be correct while the orientation of a side chain, a local loop, or a contact geometry remains unsuitable for the proposed mechanism. Global similarity is therefore not a guarantee of atomic precision at the site of interest.
EMBL-EBI's discussion of AlphaFold2 accuracy makes this distinction explicitly: high-confidence regions can be very useful for experimental design, while features requiring atomic precision may exceed what a particular prediction supports. The accuracy guide also warns against extending statistics from evaluated regions to disordered or low-confidence segments. The relevant question is not whether the entire model is broadly good, but whether the detail used in the argument is adequately supported.
Suppose a hypothetical explanation of catalysis depends on one atom being oriented toward another. A small local error could change that explanation without substantially changing the protein's global shape. Conversely, a model might be imperfect in remote regions yet provide a useful hypothesis about the central fold. Accuracy is not a single property that applies equally to every inference.
This suggests a practical standard for structural claims: inspect the evidence at the scale of the claim. A statement about domain membership needs different precision from a statement about a specific chemical interaction. A strong manuscript makes that scale visible. It does not cite a good global alignment as if it independently validates every atomic contact drawn in the figure.
9. Experimental structures also require interpretation
Avoiding overconfidence in predictions should not lead to the opposite simplification that every experimental structure is a perfect, context-free photograph. Experimental structural models are inferred from measurements and require assessment of their agreement with those measurements and with physical constraints. Their strengths and uncertainties depend on the method and the particular data.
RCSB PDB's guide distinguishes validation based on experimental data from assessment of computed models. It discusses agreement with structure-determination data and stereochemical plausibility as separate aspects of quality. The RCSB quality guide supports a more demanding comparison than prediction versus unquestionable truth: examine the evidence supporting each model, especially in the region relevant to the question.
A hypothetical disagreement between a prediction and an experimental model could have several explanations. The prediction could be wrong. The experimental region could be poorly resolved. The preparations could represent different molecular contexts or states. Determining which explanation is most plausible requires looking beyond two overlaid ribbons. It may involve the underlying data, preparation conditions, confidence measures, and independent observations.
This does not erase the distinction between measured evidence and a prediction. It clarifies why the distinction matters. Experimental data can directly constrain a structural model in ways a sequence-based prediction cannot. A computational model can supply a hypothesis that helps interpret those data. The strongest conclusion comes from understanding how the pieces of evidence relate, rather than declaring one image authoritative because of its label and dismissing the other before inspecting it.
10. A benchmark result has a population
A statement that a prediction method performs well is incomplete without knowing the set of structures on which it was evaluated. Performance can vary with the kinds of molecules, the availability of related information, the definition of similarity, and the regions included in scoring. A benchmark establishes evidence within an evaluation design; it does not automatically describe every future target.
A 2023 study by Terwilliger and colleagues compared AlphaFold predictions with experimental structural evidence and found both substantial accuracy and meaningful differences. The authors emphasized the value of predictions as hypotheses that accelerate experimental work. The primary study is useful precisely because it examines agreement in detail rather than treating a high confidence score as universal correctness. Its findings concern the structures and AlphaFold generation studied, not every later model or every molecular complex.
For a hypothetical new benchmark, suppose a method performs exceptionally well on compact proteins similar to examples already represented in structural databases. That result may be important, but it leaves uncertainty about unusual complexes or targets with different information available. A second evaluation designed around those cases could reveal a different performance profile. Neither result should be made to stand in for the other.
The editorial requirement is to preserve the benchmark's denominator and scope. Which targets were attempted, which were excluded, and which regions were scored? Were failures included? Was the method selected or tuned using the same examples? These questions do not diminish a strong result. They establish where the result can responsibly guide new research.
11. A plausible binding pose is not a binding measurement
When a predicted small molecule fits into a protein pocket, the picture can seem to answer the drug-discovery question directly. But a pose describes an arrangement. Binding strength concerns the relationship between molecular states under specified conditions. The visual compatibility of one arrangement is not itself a measurement of that relationship.
A hypothetical compound might be placed neatly in a pocket and still fail to bind detectably in an experiment. Another might bind in a different arrangement from the predicted one. A third might bind but have little effect on the biological process of interest. These possibilities show why pose, binding, and functional effect need separate terms. They are connected steps in an investigation, not synonyms.
A confidence score for a predicted interface also needs careful handling. It can help prioritize a structural hypothesis, but it should not be relabeled as an affinity or a probability of therapeutic success. Those quantities require different definitions and evidence. A model may be confident about how supplied components could be arranged without establishing that the interaction occurs meaningfully in the biological setting being discussed.
The practical value of the prediction is still considerable. It can identify candidate contacts, suggest discriminating measurements, or help interpret an observed interaction. The important question is what the next experiment would show if the proposed pose were right and what it would show if a plausible alternative were right. A model becomes more useful when it generates a testable distinction rather than merely a convincing illustration.
12. Function requires more than location
A protein's shape can suggest what it might do, especially when interpreted alongside other information. Yet structural resemblance alone does not settle every functional claim. A pocket may resemble a site associated with a known activity without establishing that the protein performs that activity under physiological conditions. The missing evidence may concern substrate availability, cellular location, partner molecules, regulation, or a direct functional measurement.
Consider a hypothetical protein with a predicted fold similar to a known enzyme. The resemblance can motivate a functional hypothesis. It does not establish the reaction rate, the preferred substrate, or the conditions in which the activity matters. Those are separate questions that may require different experiments. A useful prediction narrows the search rather than replacing the search with a label.
The same caution applies when interpreting a contact network. Drawing a chain of interacting residues can provide a mechanistic proposal. It does not by itself establish the direction or magnitude of causal influence through that chain. A structural route is not automatically a demonstrated signaling route. The word mechanism should be reserved for an explanation supported at the relevant causal level, or explicitly presented as a hypothesis.
For readers, a simple distinction helps: what was observed, what was predicted, and what was inferred? A paper may contain all three legitimately. Trouble arises when the categories merge in the summary. Preserving them allows the reader to appreciate the computational contribution and understand why functional experiments remain necessary. The experiment is not an old-fashioned ritual after the AI answer; it addresses a question the structural prediction may never have been designed to answer.
13. A variant placed on a structure is not a clinical interpretation
Mapping a genetic variant onto a predicted structure can make its location easier to understand. It can reveal whether the altered residue lies in a proposed domain, near a candidate interface, or in an uncertain region. That information can support research. It does not by itself establish whether the variant causes disease or how a person carrying it will be affected.
A hypothetical substitution near a modeled pocket might plausibly alter a contact. Several uncertainties remain: whether the local geometry is correct, whether the substitution changes the relevant molecular behavior, whether that change matters in cells, and whether it explains the observed clinical association. A visually persuasive chain of reasoning is still a chain with multiple links to test.
The opposite inference is equally risky. A variant outside a confidently modeled region is not automatically harmless. The model may not represent the relevant interaction, state, or regulatory role. Structural absence is not biological absence. Nor does a prediction that preserves the overall fold guarantee preservation of function.
This article offers no interpretation of any individual's genetic result. The broader scientific point is that structural models can contribute one line of evidence within a larger evaluation. Their contribution should be stated at the level actually supported: a hypothesis about molecular consequences, a prioritization for functional work, or a context for interpreting other measurements. Turning a structural illustration into a clinical conclusion skips the validation that gives such conclusions their meaning. The fact that the coordinates are precise does not make the downstream health inference equally precise.
14. Repeated predictions do not automatically become independent evidence
Running a model several times can reveal sensitivity to settings or stochastic variation. Agreement across runs may be reassuring in a limited sense. But repeated outputs from the same underlying method and closely related inputs are not independent confirmations in the way that different kinds of experimental evidence can be. Shared assumptions can produce shared errors.
A hypothetical researcher might generate fifty predictions and find that most show the same interface. That observation establishes consistency under the chosen procedure. It does not establish that the interface exists in the cell. If the method favors a particular arrangement because of its learned patterns or inputs, repeated runs can reproduce that preference faithfully.
The analysis should therefore describe what variation was explored and why. Did the runs test different molecular contexts, alternative sequence information, or merely different random seeds? Were the settings chosen before inspecting the preferred outcome? Were unsuccessful or inconvenient outputs included in the report? Those details determine what the agreement actually means.
A stronger argument may combine computational consistency with evidence that constrains the same relationship from another direction. The value lies in the different failure modes of the evidence sources. If two observations would be unlikely to agree under a specific alternative explanation, their combination can be informative. Merely increasing the number of similar coordinate files does not necessarily achieve that. Repetition can measure stability of the computational procedure; it should not be silently promoted into independent verification of the biological claim.
15. The input sequence is an experimental choice
A prediction begins with a representation of the molecule. That representation may differ from the form studied experimentally or present in a particular biological context. Sequence boundaries, included partners, modifications, and other components can matter to interpretation. A result can be computationally reproducible and still answer the wrong molecular question if the input does not match the intended target.
Imagine a hypothetical comparison between a full-length prediction and an experiment using a shortened construct. Differences might arise from the removed region, the conditions of the experiment, or an error in either model. Treating the two as identical objects can obscure the explanation. The comparison should state precisely what molecular material each structure represents.
This becomes especially important when figures combine information from multiple sources. A contact drawn from one model, a functional annotation from another sequence version, and an experimental measurement from a different construct may not align as neatly as the illustration suggests. Residue numbering and sequence identity are mundane checks with substantial consequences for the argument.
A good research record preserves the exact input and links it to the reported output. It also distinguishes deliberate simplifications from accidental mismatches. Simplification is often necessary and useful; the problem is forgetting what was simplified when making the final claim. A structural prediction can be interpreted only in relation to the molecule that was actually modeled. The name of a protein is not always a sufficiently precise description of that molecule.
16. A membrane, a partner, and the surrounding environment
A protein does not occupy an abstract empty coordinate system in biology. Its surroundings can constrain which arrangements are possible and which interactions are relevant. Depending on the question, that context may include a membrane, other macromolecules, chemical conditions, or modifications. A model that omits part of the context can still be useful, but its scope should remain visible.
A hypothetical prediction might place a domain in a position that appears geometrically reasonable in isolation but conflicts with a proposed membrane orientation. That conflict is a reason to investigate the model and assumptions, not to ignore the surrounding biology because the ribbon looks confident. Alternatively, the conflict may reveal that the initial biological assumption needs revision. Either way, the structure and context must be evaluated together.
Context also affects how one interprets an apparent binding site. A pocket accessible in an isolated model might be blocked in a larger assembly. A contact absent from one model might form only in the presence of a partner. These examples do not establish that any particular prediction is wrong; they show why a structural claim needs a specified biological setting.
The practical question is what the model includes and what the explanation assumes. If the explanation relies on a missing component, the gap should be stated. Researchers can then decide whether to model additional context, seek experimental constraints, or limit the claim. The important habit is to treat the environment as part of the scientific problem rather than scenery added after a structure has been produced.
17. Prediction can make experiments more discriminating
The value of a structural hypothesis is not limited to being confirmed. It can help design an experiment whose possible outcomes distinguish explanations. A useful model identifies what would be different if the proposed arrangement were correct. That is a stronger contribution than providing a decorative figure after the experimental story has already been written.
Suppose a hypothetical study proposes two possible interfaces between proteins. Each arrangement implies a different set of nearby regions. Researchers can seek observations that discriminate between those arrangements rather than collecting a generic measurement that both would explain. The prediction helps turn a broad question about interaction into a more precise question about geometry.
Care is still needed when interpreting perturbations. If changing a residue disrupts activity, the result may support a proposed contact, but it could also reflect another consequence of the change. A persuasive design considers alternatives such as altered stability, expression, or assembly. The structure guides the experiment; it does not eliminate the need for controls that separate competing explanations.
This is where computational and experimental work can become genuinely complementary. The model focuses attention on a tractable hypothesis. The experiment supplies information the model did not directly measure. Unexpected outcomes can then improve the explanation rather than being treated as inconveniences. A failed prediction can be scientifically productive if the test was designed clearly enough to reveal why the original account was incomplete. The aim is not to preserve the first attractive structure at all costs, but to improve understanding of the molecular system.
18. A structural claim needs a traceable record
A coordinate file is necessary for many structural analyses, but it is not a complete research record. The interpretation may depend on the software version, input composition, sequence sources, selected outputs, confidence measures, and subsequent processing. If those details are missing, another researcher may be able to view the same image without being able to reconstruct the analysis that produced the claim.
Selection is especially important. A paper may show the highest-ranked model, a representative model, or a model chosen because it supports a proposed mechanism. These are different selection rules. The rule should be stated, along with the relevant alternatives. Otherwise, readers cannot tell whether the displayed arrangement is typical of the calculation or one favorable output among many.
Subsequent editing also needs clarity. Removing uncertain regions, adding components from another structure, or changing the display can improve readability. But a composite figure should not be mistaken for one unmodified prediction. The record should distinguish what the model produced from what the authors added for explanation.
Traceability is not bureaucratic decoration. It allows errors to be located at the correct level. A disagreement may come from a different input, a changed model, a selection rule, or a biological assumption rather than from a fundamental contradiction. Preserving the chain makes correction easier and comparison fairer. In a fast-moving field, the ability to reconstruct a claim can matter as much as the ability to generate another plausible structure quickly.
19. The drug-discovery story has more than one endpoint
Structural prediction is often discussed in the language of drug discovery, where a compelling molecular image can be mistaken for proximity to a treatment. But identifying a plausible target interaction, demonstrating a biochemical effect, establishing cellular activity, and showing benefit in people are different endpoints. Progress at one stage can be important without establishing success at the next.
A hypothetical discovery program might use a predicted pocket to prioritize compounds for testing. If the prioritization improves the yield of useful experimental candidates, that is a measurable contribution. It does not require claiming that the prediction itself discovered a medicine. Conversely, a striking predicted interaction that never survives experimental testing may be an informative hypothesis but not a demonstrated development success.
Evaluating such programs requires an appropriate comparator. How would candidates have been selected otherwise? Were the test conditions comparable? Were unsuccessful predictions included? Did the method improve a decision that mattered, or merely produce many possibilities? Without these questions, a sequence of anecdotes can create an exaggerated impression of effectiveness.
The same discipline protects against exaggerated criticism. Failure of a candidate later in development does not prove that its structural prediction was wrong; the failure may concern a different property. Attribution should match the stage of the claim. A structural method should receive credit for the uncertainty it genuinely reduces and scrutiny for the errors relevant to that task. Neither universal praise nor universal blame helps identify where the technology is scientifically useful.
20. The alignment can decide what looks similar
Comparing structures requires choices about which atoms are matched and how the objects are aligned. Those choices can change the visual impression of agreement. A comparison focused on one domain can show excellent overlap while leaving another domain displaced. A whole-protein alignment can distribute the disagreement across both regions. Neither display is inherently dishonest, but each answers a different geometrical question.
Consider a hypothetical protein with two rigid domains connected by a flexible region. Two models preserve both domain shapes but rotate one domain relative to the other. Aligning each domain separately would show close local agreement. Aligning the entire protein would show a larger discrepancy. Reporting only one similarity number without the alignment definition could make either the agreement or the difference sound more general than it is.
Excluding uncertain residues raises a related issue. It may be sensible when evaluating the reliable core, provided the exclusion is explicit. It becomes misleading if the resulting accuracy is then attributed to the entire molecule. The fraction evaluated and the reason for exclusions belong beside the score. A beautiful overlay of a selected core should not quietly stand in for successful prediction of every region.
For a mechanistic argument, the best comparison is the one that preserves the feature at issue. If the claim concerns domain motion, aligning each domain independently removes the very difference that needs examining. If it concerns local fold conservation, a global displacement may distract from a genuinely conserved feature. Structural comparison therefore requires an analytical plan, not merely a graphics command. The image should expose the relevant relationship rather than optimize the appearance of agreement.
21. The experiment stays open for a reason
A good protein structure prediction changes the starting position of research. It can offer a detailed hypothesis where there was once little spatial information. It can help organize existing observations and propose new ones. Those are substantial achievements. They do not require treating every coordinate as established fact or every downstream biological story as a consequence of the model's confidence score.
The practical reading strategy is to separate four questions. What structure was predicted? Which parts and relationships are supported by the confidence measures? What biological claim is being made from that structure? What evidence tests the additional steps between geometry and function? The questions are connected, but answering the first does not automatically answer the other three.
The sources checked for this article span developer explanations, database guidance, an experimental comparison study, and a preliminary October 2026 investigation of model internals. They address different levels of the problem. Their shared relevance is not that prediction has failed. It is that interpreting a prediction remains a scientific activity, requiring attention to scale, context, and uncertainty.
The most convincing future demonstrations will show exactly how a structural prediction improved an investigation: a better hypothesis, a more discriminating experiment, a resolved ambiguity, or a useful decision tested against alternatives. The experiment remains open because the world contains information the model has not directly measured. Keeping that distinction clear is not a retreat from the power of AI in biology. It is how that power becomes reliable knowledge rather than a persuasive picture.