Lab proof

Introduction

AlphaFold made protein prediction dramatically faster and more accessible, but it did not remove the need for experiments. Its models are best understood as highly informed hypotheses: often extremely useful, sometimes astonishingly accurate, yet still incomplete descriptions of molecules that exist inside messy, changing biological systems. Even DeepMind, EMBL-EBI and structural biologists routinely caution that AlphaFold outputs carry uncertainty and should be interpreted alongside confidence metrics and experimental evidence. [alphafold.ebi.ac.uk]alphafold.ebi.ac.ukAlphaFold Protein Structure Database - EMBL-EBIAll AlphaFold and AlphaMissense Data and other information provided on this site contain p… [ebi.ac.uk]alphafold.ebi.ac.ukAlphaFold Protein Structure Database - EMBL-EBIAll AlphaFold and AlphaMissense Data and other information provided on this site contain p…

Lab proof illustration 1 That distinction matters because AI-generated structures can look more certain than they really are. A polished three-dimensional protein model encourages the intuition that scientists are seeing “the actual molecule”. In reality, many proteins shift shape, interact with partners, contain flexible regions or depend on chemical conditions that a static prediction cannot fully capture. Wet-lab methods such as cryo-electron microscopy, X-ray crystallography, biochemical assays and molecular biology experiments remain essential for checking whether a prediction corresponds to what living systems actually do.

For the broader idea of AI-driven scientific acceleration, this is an important lesson. AlphaFold shows how AI can massively lower the cost of generating scientific ideas and starting points. But faster hypothesis generation is not the same thing as automated scientific truth. The bottleneck often moves from producing candidate explanations to experimentally testing them.

What a predicted structure can and cannot prove

AlphaFold predicts the probable three-dimensional arrangement of a protein from its amino-acid sequence. In many cases, especially for stable single proteins, the results are close to experimentally measured structures. Independent comparisons against cryo-EM and crystallographic data found that many AlphaFold predictions align remarkably well with laboratory measurements. [ebi.ac.uk]ebi.ac.ukThese metrics can be used to identify regions of the predicted structure, and relative…Read more… [Nature]nature.comNatureHighly accurate protein structure prediction with AlphaFoldby J Jumper · 2021 · Cited by 49928 — Here we provide the first computat…

That success has practical value. Researchers can now:

identify likely active sites
design mutations to test
prioritise drug targets
infer broad evolutionary relationships
accelerate structural biology workflows [researchgate.net]researchgate.netAdvantages and Limitations of AlphaFold in Structural…1 Dec 2025 — This narrative review synthesizes applications reported in the 2022…

But a good geometric prediction does not automatically establish biological truth.

A predicted structure alone usually cannot prove:

that the protein actually folds that way inside cells
that the structure is biologically active
how the protein moves over time
which molecules it binds in living systems
whether the prediction changes under different conditions
whether a disease mutation truly alters function
how strongly two proteins interact
whether a proposed drug-binding pocket is real

This distinction matters because biology is not only about shape. It is also about dynamics, chemistry, timing and cellular context.

One reason wet-lab proof remains essential is that experimental biology tests consequences, not merely geometry. A researcher may use AlphaFold to predict that a mutation disrupts an enzyme pocket, but only experiments can establish whether enzyme activity actually falls, whether cells compensate through other pathways, or whether the effect matters in disease.

Structural biologists increasingly describe AlphaFold models as guides for experimentation rather than replacements for it. A 2024 analysis in Nature Methods explicitly framed AlphaFold predictions as “valuable hypotheses”, while noting that predictions may omit ligands, modifications and environmental effects visible in experiments. [Nature]nature.comNatureAlphaFold predictions are valuable hypotheses and…by TC Terwilliger · 2024 · Cited by 396 — Artificial intelligence-based protei…

In practice, many labs now work in a hybrid loop:

AI generates a structural hypothesis.
Scientists design experiments around the model.
Experimental data confirms, refines or contradicts the prediction.
The updated evidence feeds back into modelling.

That workflow is still a major acceleration compared with starting from nothing. But the experimental stage remains the point where claims become reliable science.

Hard cases: disorder, motion and protein assemblies

AlphaFold works best when proteins adopt a relatively stable structure. Biology, unfortunately for neat prediction systems, is full of molecules that do not behave that way.

Intrinsically disordered proteins

Many proteins or protein regions are intrinsically disordered, meaning they do not settle into one stable shape. Instead they fluctuate between multiple conformations or only fold when interacting with other molecules.

EMBL-EBI training materials explicitly note that AlphaFold cannot reliably predict “dynamic substructures” or proteins without a single fixed conformation. [ebi.ac.uk]ebi.ac.ukstrengths and limitations of alphafold25 Jan 2024 — AlphaFold2 can be used to identify intrinsically disordered regions. Naturally, the system cannot predict disordered or dyn…

This creates a subtle problem for non-specialists. AlphaFold still outputs a structure-like object even for disordered regions. Without understanding the confidence scores, a reader may mistake a low-confidence ribbon model for a real biological structure.

Researchers studying disordered proteins have warned that some medium-confidence outputs can even be misleading. In certain cases, low confidence does indicate genuine disorder, but not always in a simple or easily interpretable way. [ScienceDirect]sciencedirect.comScienceDirectAlphaFold and Implications for Intrinsically Disordered…by KM Ruff · 2021 · Cited by 748 — The C-terminal domain of p27…

That matters because intrinsically disordered proteins are deeply involved in signalling, regulation, cancer biology and neurodegenerative disease. Some of the most biologically important molecules are precisely the ones least suited to static prediction.

Proteins are not frozen objects

AlphaFold generally predicts one dominant structure. Real proteins often behave more like moving machines than static sculptures.

They bend, twist, open and close. They shift shape when binding drugs, DNA, ions or other proteins. Some proteins cycle through many conformations as part of their function.

Multiple reviews emphasise that AlphaFold predictions remain fundamentally static representations of systems that are dynamic in reality. PMC [OUP Academic]academic.oup.comOUP AcademicBeyond static structures: protein dynamic conformations…by X Cui · 2025 · Cited by 34 — This review outlines the fundament…

This limitation becomes especially important in drug discovery. A predicted binding pocket may exist only transiently, or disappear under physiological conditions. A protein might adopt one shape in isolation and another inside a crowded cellular environment.

Experimental methods capture at least part of this complexity. Cryo-EM can sometimes reveal multiple conformational states. Biochemical assays can measure activity changes. Molecular dynamics simulations attempt to model movement over time. Without those additional layers, a predicted structure can create false confidence about how a protein actually behaves.

Lab proof illustration 2

Large assemblies remain difficult

Cells depend heavily on molecular assemblies rather than isolated proteins. Ribosomes, ion channels, transcription complexes and signalling machinery often involve many interacting components.

AlphaFold-Multimer and AlphaFold 3 improved complex prediction substantially, including predictions involving nucleic acids and ligands. [Nature]nature.comStructural validation and assessment of AlphaFold2…by M van Breugel · 2022 · Cited by 67 — Another limitation of AF2 is that the predi… But large assemblies remain difficult, especially when interactions are weak, transient or highly dynamic.

Confidence scores for complexes require careful interpretation, and EMBL-EBI guidance warns that some global scores should be treated cautiously. [ebi.ac.uk]ebi.ac.ukHow have AlphaFold2's predictions of protein structure…AlphaFold2 structures fit well into experimental cryo-EM electron density maps…

Experimental work is still needed to establish:

whether the interaction occurs in cells
which assembly state dominates biologically
whether the orientation is correct
how flexible the interface is
whether additional molecules are required

In practice, many of the hardest and most medically important molecular systems remain precisely those where experiments matter most.

How confidence scores should guide experiments

One of AlphaFold’s most important innovations was not only structure prediction itself, but the inclusion of uncertainty estimates.

These scores help scientists distinguish between regions that are probably reliable and regions that may be speculative.

pLDDT: local confidence

AlphaFold’s best-known confidence metric is pLDDT, a residue-by-residue confidence score running from 0 to 100. Higher values generally indicate more reliable local geometry. [ebi.ac.uk]ebi.ac.ukHow accurate are AlphaFold 2 structure predictions?Overall, AlphaFold2 gets the vast majority of the side chains right, but is marginally…

Broadly speaking:

above 90: often highly reliable
70–90: generally good backbone prediction
50–70: caution required
below 50: likely unreliable or disordered

These are not absolute guarantees, but practical guides.

Importantly, low confidence is itself biologically informative. It may suggest flexible loops, disordered regions or conformational variability rather than simple algorithmic failure. [alphafold.ebi.ac.uk]alphafold.ebi.ac.ukFAQs - AlphaFold Protein Structure DatabaseFor regions that are intrinsically disordered or unstructured in isolation AlphaFold is expect…

Good experimental practice therefore treats confidence maps as navigation tools.

For example:

high-confidence regions may guide mutation experiments
uncertain interfaces may become targets for cryo-EM or crosslinking studies
low-confidence segments may suggest disorder experiments
questionable active sites may require biochemical validation before drug design

The confidence scores help researchers allocate scarce experimental time more intelligently.

Lab proof illustration 3

PAE: relative uncertainty

Another important metric is Predicted Aligned Error, or PAE, which estimates confidence in the relative positioning between different regions of the structure. [ebi.ac.uk]alphafold.ebi.ac.ukFAQs - AlphaFold Protein Structure DatabaseFor regions that are intrinsically disordered or unstructured in isolation AlphaFold is expect…

A protein may contain highly accurate local regions while still having uncertain domain orientations overall. That distinction matters enormously for understanding function.

For example, two domains connected by a flexible linker might each be individually correct while their relative arrangement is wrong. A casual viewer might see a polished complete model and assume the whole structure is trustworthy.

Specialists instead inspect the uncertainty maps first.

This is one reason AlphaFold has not eliminated structural biology expertise. In some ways it has increased the importance of interpretation. Scientists now need to judge where the model is likely strong, where it is weak, and which experiments are most informative next.

Why this still counts as scientific acceleration

The need for wet-lab proof does not diminish AlphaFold’s importance. In many fields, the expensive part of science is not only collecting data but deciding which hypotheses are worth testing.

AlphaFold dramatically expands the number of plausible hypotheses researchers can generate quickly.

That matters for the wider AI bloom argument because scientific progress often depends on reducing search costs. If AI systems help scientists navigate vast spaces of biological possibility more efficiently, the long-run effects on medicine and biotechnology could still be enormous even when experiments remain necessary.

The practical change is already visible:

structural guesses that once took years can appear in hours
small labs can begin projects previously limited to elite centres
experimentalists can prioritise more promising targets
failed directions may be discarded earlier
interdisciplinary researchers can access structural reasoning without becoming full-time structural biologists

But AlphaFold also illustrates a broader truth about AI and science. Advanced AI may compress some bottlenecks while leaving others stubbornly physical.

Biology ultimately happens in cells, tissues, organisms and ecosystems, not only in computation. Testing drugs still requires chemistry, manufacturing, animal models, clinical trials and regulation. Understanding disease still depends on measurement and intervention in the real world.

In that sense, AlphaFold is neither “AI solved biology” nor “AI hype”. It is a powerful example of a more realistic pattern: AI systems can massively increase humanity’s ability to generate useful scientific models, while experimental reality remains the final judge of which models are true.

Endnotes

Source: alphafold.ebi.ac.uk
Link: https://alphafold.ebi.ac.uk/download
Source snippet
AlphaFold Protein Structure Database - EMBL-EBIAll AlphaFold and AlphaMissense Data and other information provided on this site contain p...
Source: ebi.ac.uk
Link: https://www.ebi.ac.uk/training/online/courses/alphafold/inputs-and-outputs/evaluating-alphafolds-predicted-structures-using-confidence-scores/
Source snippet
These metrics can be used to identify regions of the predicted structure, and relative...Read more...
Source: ebi.ac.uk
Title: strengths and limitations of alphafold
Link: https://www.ebi.ac.uk/training/online/courses/alphafold/an-introductory-guide-to-its-strengths-and-limitations/strengths-and-limitations-of-alphafold/
Source snippet
25 Jan 2024 — AlphaFold2 can be used to identify intrinsically disordered regions. Naturally, the system cannot predict disordered or dyn...
Source: ebi.ac.uk
Link: https://www.ebi.ac.uk/training/online/courses/alphafold/validation-and-impact/how-have-alphafolds-predictions-of-protein-structure-been-validated/
Source snippet
How have AlphaFold2's predictions of protein structure...AlphaFold2 structures fit well into experimental cryo-EM electron density maps...
Source: nature.com
Link: https://www.nature.com/articles/s41586-021-03819-2
Source snippet
NatureHighly accurate protein structure prediction with AlphaFoldby J Jumper · 2021 · Cited by 49928 — Here we provide the first computat...
Source: ebi.ac.uk
Link: https://www.ebi.ac.uk/training/online/courses/alphafold/validation-and-impact/how-accurate-are-alphafold-structure-predictions/
Source snippet
How accurate are AlphaFold 2 structure predictions?Overall, AlphaFold2 gets the vast majority of the side chains right, but is marginally...
Source: nature.com
Link: https://www.nature.com/articles/s41592-023-02087-4
Source snippet
NatureAlphaFold predictions are valuable hypotheses and...by TC Terwilliger · 2024 · Cited by 396 — Artificial intelligence-based protei...
Source: alphafold.ebi.ac.uk
Link: https://alphafold.ebi.ac.uk/faq
Source snippet
FAQs - AlphaFold Protein Structure DatabaseFor regions that are intrinsically disordered or unstructured in isolation AlphaFold is expect...
Source: sciencedirect.com
Link: https://www.sciencedirect.com/science/article/pii/S0022283621004411
Source snippet
ScienceDirectAlphaFold and Implications for Intrinsically Disordered...by KM Ruff · 2021 · Cited by 748 — The C-terminal domain of p27...
Source: pmc.ncbi.nlm.nih.gov
Link: https://pmc.ncbi.nlm.nih.gov/articles/PMC9601767/
Source snippet
Indeed when...Read mor...
Source: pmc.ncbi.nlm.nih.gov
Link: https://pmc.ncbi.nlm.nih.gov/articles/PMC12956938/
Source snippet
AlphaFold's static predictions do not reflect the reality of conformational landscapes.Read more...
Source: academic.oup.com
Link: https://academic.oup.com/bib/article/26/4/bbaf340/8202937
Source snippet
OUP AcademicBeyond static structures: protein dynamic conformations...by X Cui · 2025 · Cited by 34 — This review outlines the fundament...
Source: nature.com
Link: https://www.nature.com/articles/s42003-022-03269-0
Source snippet
Structural validation and assessment of AlphaFold2...by M van Breugel · 2022 · Cited by 67 — Another limitation of AF2 is that the predi...
Source: nature.com
Link: https://www.nature.com/articles/s41586-024-07487-w
Source snippet
NatureAccurate structure prediction of biomolecular interactions...by J Abramson · 2024 · Cited by 14701 — Here we describe our AlphaFol...
Source: ebi.ac.uk
Title: confidence scores in alphafold multimer
Link: https://www.ebi.ac.uk/training/online/courses/alphafold/inputs-and-outputs/evaluating-alphafolds-predicted-structures-using-confidence-scores/confidence-scores-in-alphafold-multimer/
Source snippet
Confidence scores in AlphaFold-MultimerA TM score below 0.5 means the predicted structure is likely wrong: pTM score follows the same def...
Source: ebi.ac.uk
Link: https://www.ebi.ac.uk/training/online/courses/alphafold/inputs-and-outputs/evaluating-alphafolds-predicted-structures-using-confidence-scores/plddt-understanding-local-confidence/
Source snippet
It is scaled from 0 to 100, with higher scores indicating higher confidence and usually a more accurate...Read more...
Source: nature.com
Link: https://www.nature.com/articles/s41598-022-14382-9
Source snippet
AlphaFold2 models indicate that protein sequence...by HB Guo · 2022 · Cited by 312 — High pLDDT scores (e.g., > 80) indicate high confid...
Source: ebi.ac.uk
Link: https://www.ebi.ac.uk/training/online/courses/alphafold/inputs-and-outputs/evaluating-alphafolds-predicted-structures-using-confidence-scores/pae-a-measure-of-global-confidence-in-alphafold-predictions/
Source snippet
PAE: A measure of global confidence in AlphaFold2...Predicted aligned error (PAE) is a measure of how confident AlphaFold2 is in the rel...
Source: nature.com
Link: https://www.nature.com/articles/s42004-026-01899-7
Source snippet
Specifically, for large...Read more...
Source: nature.com
Link: https://www.nature.com/articles/s41467-025-56572-9
Source snippet
B Structural ensemble of TDP...Read more...
Source: nature.com
Link: https://www.nature.com/articles/s41467-023-44288-7
Source snippet
From interaction networks to interfaces, scanning...by H Bret · 2024 · Cited by 93 — We show that when using the full sequences of the p...
Source: alphafold.ebi.ac.uk
Title: ebi.ac.uk Alpha Fold Protein Structure Database
Link: https://alphafold.ebi.ac.uk/
Source snippet
It regularly achieves accuracy...Read more...
Source: ebi.ac.uk
Link: https://www.ebi.ac.uk/training/online/courses/alphafold/accessing-and-predicting-protein-structures-with-alphafold/accessing-predicted-protein-structures-in-the-alphafold-database/
Source snippet
Accessing predicted protein structures in the AlphaFold...Users can download many predicted protein structures from the AlphaFold Protei...
Source: ebi.ac.uk
Link: https://www.ebi.ac.uk/training/online/courses/alphafold/validation-and-impact/
Source snippet
Validation and impact | AlphaFoldBy the end of this section you will be able to: Explain how AlphaFold structure predictions were validat...
Source: ebi.ac.uk
Link: https://www.ebi.ac.uk/training/online/courses/alphafold
Source snippet
AlphaFold | EMBL-EBI TrainingThis tutorial is aimed at researchers who are interested in using AlphaFold to predict protein structures an...
Source: ebi.ac.uk
Link: https://www.ebi.ac.uk/training/online/courses/alphafold/alphafold-3-and-alphafold-server/how-to-assess-the-quality-of-alphafold-3-predictions/
Source snippet
How to assess the quality of AlphaFold 3 predictionsA value above 90 indicates high confidence; a value below 50 indicates the correspond...
Source: ebi.ac.uk
Link: https://www.ebi.ac.uk/training/online/courses/navigating-alphafold-database/understanding-the-structure-prediction-page/summary-and-model-confidence-tab/
Source snippet
This tab gives you the key information for a specific AlphaFold Database entry. At the top, you find related information to the protein e...
Source: embl.org
Link: https://www.embl.org/
Source snippet
European Molecular Biology Laboratory | EMBL.orgEMBL is Europe's life sciences laboratory. Research EMBL performs fundamental research in...
Source: embl.org
Title: alphafold potential impacts
Link: https://www.embl.org/news/science/alphafold-potential-impacts/
Source snippet
This metric predicts the...Read more...
Source: embl.org
Title: alphafold using open data and ai to discover the 3d protein universe
Link: https://www.embl.org/news/science/alphafold-using-open-data-and-ai-to-discover-the-3d-protein-universe/
Source snippet
Case study: AlphaFold uses open data and AI to discover...9 Feb 2023 — EMBL-EBI has also helped to integrate the AlphaFold 2 protein str...
Source: embl.org
Title: sameer velankar alphafold training
Link: https://www.embl.org/news/people-perspectives/sameer-velankar-alphafold-training/
Source snippet
The AI revolution hinges on accessible training12 Mar 2024 — EMBL-EBI Team Leader Sameer Velankar explains why AI tools such as AlphaFold...
Source: embl.org
Link: https://www.embl.org/news/science-technology/first-complexes-alphafold-database/
Source snippet
Millions of protein complexes added to AlphaFold...4 days ago — Of these, 1.7 million high-confidence homodimer predictions have been ad...
Source: embl.org
Title: compare experimental and predicted structures with pdbe kb
Link: https://www.embl.org/news/updates-from-data-resources/compare-experimental-and-predicted-structures-with-pdbe-kb/
Source snippet
Compare experimental and predicted structures with PDBe...24 Oct 2022 — This allows easy comparison of the experimental structures in th...
Source: sciencedirect.com
Link: https://www.sciencedirect.com/science/article/pii/S2001037025001734
Source snippet
Bridging prediction and reality: Comprehensive analysis of...by A Mazhibiyeva · 2025 · Cited by 11 — Here we present the first comprehen...
Source: sciencedirect.com
Link: https://www.sciencedirect.com/science/article/pii/S2001037025001722
Source snippet
Multimeric protein interaction and complex predictionby D Lu · 2025 · Cited by 8 — This article reviews the new methods in multimeric pro...
Source: academic.oup.com
Link: https://academic.oup.com/nar/article/50/D1/D439/6430488
Source snippet
Protein Structure Database: massively expanding...by M Varadi · 2022 · Cited by 8512 — The initial version of AlphaFold DB contains over...
Source: deepmind.google
Link: https://deepmind.google/science/alphafold/
Source snippet
AlphaFold — Google DeepMindAlphaFold has revealed millions of intricate 3D protein structures, and is helping scientists understand how l...

Additional References

Source: researchgate.net
Link: https://www.researchgate.net/publication/398214210_Advantages_and_Limitations_of_AlphaFold_in_Structural_Biology_Insights_from_Recent_Studies
Source snippet
Advantages and Limitations of AlphaFold in Structural...1 Dec 2025 — This narrative review synthesizes applications reported in the 2022...
Source: researchgate.net
Link: https://www.researchgate.net/publication/382971763_Evaluation_of_AlphaFold_3%27s_Protein-Protein_Complexes_for_Predicting_Binding_Free_Energy_Changes_upon_Mutation
Source snippet
Evaluation of AlphaFold 3's Protein-Protein Complexes for...It also discusses current limitations and offers perspectives on integrating...
Source: chemistry.unist.ac.kr
Link: https://chemistry.unist.ac.kr/new-study-tests-the-limits-of-alphafold2s-accuracy-in-protein-structure-prediction/
Source snippet
Study Tests the Limits of AlphaFold2's Accuracy in...Nov 27, 2023 — AlphaFold2 found to be accurate enough to predict changes in protein...
Source: youtube.com
Link: https://www.youtube.com/watch?v=UqeQfRDA8Yk
Source snippet
How to interpret AlphaFold structuresI'll be covering how alpha fold works on a very high level how to interpret predictions and some of...
Source: repository.cam.ac.uk
Link: https://www.repository.cam.ac.uk/server/api/core/bitstreams/326f82fb-13e6-4e56-b085-754ca9c69efc/content
Source snippet
predictions are valuable hypotheses and...by TC Terwilliger · 2024 · Cited by 377 — Both experimentally determined protein structures an...
Source: frontiersin.org
Link: https://www.frontiersin.org/journals/artificial-intelligence/articles/10.3389/frai.2026.1739303/full
Source snippet
This review first outlines AlphaFold's architectural evolution...Read more...
Source: mindwalkai.com
Link: https://www.mindwalkai.com/blog/explained-how-to-plot-the-prediction-quality-metrics-with-alphafold2
Source snippet
how to plot the prediction quality metrics with AlphaFold2Apr 14, 2026 — Regions with pLDDT between 50 and 70 are low confidence and shou...
Source: biorxiv.org
Title: 2022.09.14.507881v1.full text
Link: https://www.biorxiv.org/content/10.1101/2022.09.14.507881v1.full-text
Source snippet
Topological Links in Predicted Protein Complex Structures...16 Sept 2022 — When using AlphaFold-Multimer to predict protein–protein comp...
Source: pmc.ncbi.nlm.nih.gov
Link: https://pmc.ncbi.nlm.nih.gov/articles/PMC11956457/
Source snippet
power and pitfalls of AlphaFold2 for structure prediction...These studies provide compelling evidence that AF2 can be applied across pep...
Source: biorxiv.org
Link: https://www.biorxiv.org/content/10.64898/2025.12.05.691730v1.full-text
Source snippet
AlphaFold3 and Intrinsically Disordered Proteins10 Dec 2025 — For monomers, AF3's pLDDT scores reliably predict disorder (MCC: 0.693), ma...