---
id: APP-PROT-001
title: "Protein Structure Prediction (AlphaFold)"
alternative_names:
  - "AlphaFold"
  - "Computational protein structure prediction"
  - "The protein folding problem (computational)"
category: "Applications → Structural Biology"
volume: VII
chapter: 37
entry_type: "Standard Article"
level: 3
status: "Draft"
last_reviewed: 2026-10-15
version: "0.1.0"
prerequisites:
  - "APP-SCI-001"
related_entries:
  - "TECH-DL-001"
  - "TECH-DIFF-001"
  - "TECH-LLM-001"
  - "APP-HEALTH-001"
  - "SAFE-XRISK-001"
---

# Protein Structure Prediction (AlphaFold)

## 1. Concise Definition
**Protein structure prediction** is the computation of a protein's three-dimensional structure from its amino-acid sequence. **AlphaFold** is a family of deep-learning systems from Google DeepMind (with later work at Isomorphic Labs) that, since its second version in 2020, predicts single-protein structures with accuracy that a community blind test assessed as rivalling experimental structures in most cases, and whose later version extends to complexes with nucleic acids, small molecules and ions.

## 2. Overview & Significance
Experimental determination of a structure (X-ray crystallography, cryo-electron microscopy, NMR) takes months to years of effort per protein (Jumper et al., 2021). Sequences are known for far more proteins than have experimental structures. A fast, accurate predictor therefore changes which structural questions can be asked. This article is the detailed case study of the most thoroughly validated AI-for-science result; the broader field is covered in APP-SCI-001, which is not repeated here. Evidence is sorted into three classes:

| Class | Meaning | Examples below |
| :--- | :--- | :--- |
| **Deployed product / resource** | Openly available tools and databases in routine research use | AlphaFold Protein Structure Database; open-source and server tools |
| **Validated result** | Blind-test or independent experimental comparison | CASP14 assessment; comparison with crystallographic maps |
| **Research demonstration** | Published capability claims awaiting broad independent evaluation | AlphaFold 3 complexes; protein language-model predictions; generative design |

## 3. Historical Development
The prediction of structure from sequence was framed as a "grand challenge" for about half a century after Anfinsen's work on sequence-structure relations. Progress was tracked by CASP (Critical Assessment of protein Structure Prediction), a biennial community experiment with blind testing and independent assessment, first run in 1994 (Baek et al., 2021, for context; Kryshtafovych et al., 2021). In CASP14 (2020), the assessors reported that deep-learning methods from one group consistently delivered computed structures rivaling experimental ones in accuracy, a result they described as a solution to the classical protein-folding problem at least for single proteins (Kryshtafovych et al., 2021). The method was published in 2021 (Jumper et al., 2021). The same year, a second group published RoseTTAFold, a three-track network that approached AlphaFold's accuracy (Baek et al., 2021), and DeepMind with EMBL-EBI released predictions for almost the entire human proteome (Tunyasuvunakool et al., 2021) and opened the AlphaFold Protein Structure Database (Varadi et al., 2022). A protein language model approach, ESM-2, followed in 2023 (Lin et al., 2023). AlphaFold 3 was described in 2024 (Abramson et al., 2024). In 2024 the Nobel Prize in Chemistry was awarded half to David Baker for computational protein design and half to Demis Hassabis and John Jumper for protein structure prediction (Nobel Prize Outreach, 2024; reported by news sources, citation to the Nobel announcement page to be confirmed).

## 4. Technical / Conceptual Explanation
**Inputs and representation.** AlphaFold 2 takes the target sequence, a multiple sequence alignment (MSA) of related sequences and optionally template structures. Jumper et al. describe a redesigned neural network that incorporates physical and biological knowledge about protein structure, using multi-sequence alignments, in the design of the deep-learning algorithm (Jumper et al., 2021).

**Outputs and confidence.** The system returns atomic coordinates together with per-residue and pairwise confidence estimates and predicted aligned errors (Varadi et al., 2022). Per-residue confidence is reported on a 0-100 scale (pLDDT). The confidence outputs are part of the model, not an afterthought, and are needed to interpret results: low-confidence regions are often disordered.

**Accuracy metrics.** CASP compares a predicted structure with the experimental one after superposition, using measures such as the global distance test, the fraction of residues within given distance cut-offs of their true positions. Because the experimental structure can itself be uncertain, assessment includes comparison with experimental density.

**AlphaFold 3.** The third version uses a substantially updated diffusion-based architecture to predict the joint structure of complexes of proteins, nucleic acids, small molecules, ions and modified residues (Abramson et al., 2024). The paper reports far greater accuracy for protein-ligand interactions than state-of-the-art docking tools, much higher accuracy for protein-nucleic acid interactions than nucleic-acid-specific predictors, and substantially higher antibody-antigen accuracy than AlphaFold-Multimer v2.3. For diffusion models see TECH-DIFF-001.

**Language-model approach.** ESM-2 scaled protein language models up to 15 billion parameters and inferred atomic-level structure directly from a single sequence, giving an order-of-magnitude speed-up and enabling structure prediction for more than 617 million metagenomic sequences, over 225 million of them with high confidence (Lin et al., 2023). Accuracy relative to MSA-based AlphaFold depends on the protein class and was not compared here.

## 5. Types, Components & Classifications
- **Single-chain structure prediction:** AlphaFold 2, RoseTTAFold, ESM-2 (and related folding models).
- **Complex prediction:** protein-protein (AlphaFold-Multimer), and protein with nucleic acids, ligands and ions (AlphaFold 3).
- **Databases:** AlphaFold Protein Structure Database (predictions), distinct from the experimental Protein Data Bank.
- **Downstream design tools:** inverse folding (ProteinMPNN) and generative backbone design (RFdiffusion) that use structure-prediction ideas to design new proteins (Dauparas et al., 2022; Watson et al., 2023).

## 6. Applications & Concrete Examples

### 6.1 Deployed resources
- **AlphaFold Protein Structure Database.** Created by Google DeepMind and EMBL-EBI. The initial release in 2021 contained over 360,000 predicted structures across 21 model-organism proteomes (Varadi et al., 2022). As of the database's website in 2026, it states open access to over 200 million predictions under a CC-BY-4.0 licence (EMBL-EBI, 2026; value read through a search summary, [VERIFICATION NEEDED: confirm on the site and note date]). Current counts belong in the Living Reference collections.
- **Human proteome.** Predictions for 98.5% of human proteins covered 58% of residues with a confident prediction, 36% of all residues at very high confidence, compared with 17% of residues covered by experimental structures at the time (Tunyasuvunakool et al., 2021).

### 6.2 Validated results
- **CASP14 blind test.** Independent assessors reported that the deep-learning predictions from one group rivalled experimental structures in accuracy, and that the models were already helpful for problematic crystal structures (Kryshtafovych et al., 2021). The AlphaFold paper reports accuracy competitive with experimental structures in a majority of cases and greatly outperforming other methods (Jumper et al., 2021).
- **Comparison with crystallographic maps.** Terwilliger et al. (2024) compared predictions directly with experimental electron-density maps. In many cases predictions matched closely; in others even very high-confidence predictions differed on a global scale (distortion, domain orientation) and locally (backbone and side-chain conformations). The authors recommend treating predictions as "exceptionally useful hypotheses" and verifying details experimentally, especially interactions that were not part of the prediction.

### 6.3 Research demonstrations
- **AlphaFold 3 complexes** (Abramson et al., 2024). Evidence is the developers' benchmark comparisons; at the time of drafting, broad independent evaluation of each claim was not verified [VERIFICATION NEEDED: independent benchmarks of AlphaFold 3; and the terms under which model weights are available].
- **Generative design.** ProteinMPNN reported sequence recovery of 52.4% on native backbones versus 32.9% for Rosetta, with experimental validation of previously failed designs (Dauparas et al., 2022). RFdiffusion applied diffusion to de novo protein design (Watson et al., 2023). These are design methods (see APP-SCI-001), not structure predictors.

## 7. Advantages, Limitations & Unresolved Questions
**Advantages.** Speed (a prediction replaces months of experiment as a first model), open databases, and confidence estimates that flag unreliable regions.

**Limitations.**
- **Predictions are models.** High-confidence predictions can still deviate from experiment in domain arrangement and local conformation, and the method does not account for ligands, covalent modifications or environmental factors in its single-chain form (Terwilliger et al., 2024).
- **Single static structures.** The headline problem solved at CASP14 was prediction of single-protein structures (Kryshtafovych et al., 2021); the dynamic ensembles that determine function are a separate scientific question.
- **Disorder and low-confidence regions.** Confidence outputs flag regions that are likely disordered (Tunyasuvunakool et al., 2021).
- **From structure to drug.** A predicted structure does not by itself yield a drug. No source verified in this draft shows an approved medicine whose discovery depended on AlphaFold predictions; the claim that AlphaFold accelerates drug discovery is Preliminary and should be read against the long timescale of drug development (see APP-HEALTH-001 for clinical evidence standards).
- **Dual use.** Better structure prediction and design tools may lower barriers to harmful biological design; this is discussed in SAFE-XRISK-001 and is not evidenced by the sources used here.

**Unresolved questions.** Accuracy for novel folds, multi-chain assemblies, conformational change and effects of mutations; how to benchmark complex prediction independently; how structure-prediction tools combine with experimental pipelines.

## 8. Current State of Knowledge
As of 2026-10-15: single-protein structure prediction at near-experimental accuracy for most targets is Established as a CASP-validated result and by independent comparison with experimental maps, with documented exceptions (Well Supported to Established, depending on the claim). The practice of treating predictions as hypotheses to verify is the recommendation of crystallographers who tested them (Terwilliger et al., 2024). AlphaFold 3's complex-prediction advantages are Preliminary pending independent evaluation. The 2024 Nobel Prize recognised both structure prediction and computational protein design.

## 9. Future Trajectory & Strategic Outlook
Near-term (0-3 years): broader complex and ensemble prediction, integration with experimental data, larger databases of predicted complexes (EMBL-EBI has announced plans in 2026 to add millions of predicted complexes, per search summary; unverified). Medium-term (3-10 years): routine use in drug and enzyme discovery pipelines; structure-aware generative design. Long-term claims, such as a general ability to design arbitrary functional proteins, are speculative.

| Field | Content |
| :--- | :--- |
| **Forecast ID** | `FCST-PROT-001` |
| **Forecast Statement** | A new molecular entity approved by the US FDA or European Medicines Agency will have peer-reviewed discovery literature stating that machine-learning-predicted protein structures (AlphaFold family or equivalent) guided lead identification or optimisation. |
| **Operational Definition** | An approval of a new molecular entity (not a biologic made before 2020) for which at least one peer-reviewed discovery paper explicitly credits predicted structures as guiding design decisions. |
| **Time Horizon** | By 2030-12-31 |
| **Assumptions** | Drugs in current pipelines that used predicted structures complete trials; discovery papers are published. |
| **Confidence Level** | Medium (about 45%); editorial judgement. |
| **Alternative Scenarios** | Approvals occur but attribution is ambiguous; trial attrition delays approvals past 2030. |
| **Review Schedule** | Annual review each October; final evaluation January 2031. |
| **Status** | `Active` |

## 10. Related Entries & Knowledge Graph
- **Foundational Concepts**:
  - [[APP-SCI-001] AI in Scientific Discovery](../vol07_applications/APP-SCI-001_ai_in_science.md): The broader field in which this case study sits.
  - [[TECH-DL-001] Deep Learning](../vol03_technical/TECH-DL-001_deep_learning.md): Training methods.
- **Related Methods & Technologies**:
  - [[TECH-DIFF-001] Diffusion Models](../vol03_technical/TECH-DIFF-001_diffusion_models.md): Architecture used in AlphaFold 3 and RFdiffusion.
  - [[TECH-LLM-001] Large Language Models](../vol03_technical/TECH-LLM-001_large_language_models.md): Background for protein language models.
  - [[APP-HEALTH-001] AI in Healthcare and Medicine](../vol07_applications/APP-HEALTH-001_ai_in_healthcare_and_medicine.md): Clinical evidence standards for downstream use.
- **Advanced Implications & Future Trajectories**:
  - [[SAFE-XRISK-001] Existential Risk from AI](../vol08_safety_governance/SAFE-XRISK-001_existential_risk.md): Biological misuse discussion.

## 11. References & Foundational Literature
1. Abramson, J., Adler, J., Dunger, J., et al. (2024). *Accurate structure prediction of biomolecular interactions with AlphaFold 3*. Nature, 630(8016), 493-500. https://doi.org/10.1038/s41586-024-07487-w. [Describes AlphaFold 3 and its benchmark claims for complexes.]
2. Baek, M., DiMaio, F., Anishchenko, I., et al. (2021). *Accurate prediction of protein structures and interactions using a three-track neural network*. Science, 373(6557), 871-876. https://doi.org/10.1126/science.abj8754. [RoseTTAFold, approaching AlphaFold accuracy.]
3. Dauparas, J., Anishchenko, I., Bennett, N., et al. (2022). *Robust deep learning-based protein sequence design using ProteinMPNN*. Science, 378(6615), 49-56. https://doi.org/10.1126/science.add2187. [Sequence design with 52.4% native sequence recovery versus 32.9% for Rosetta.]
4. EMBL-EBI. (2026). *AlphaFold Protein Structure Database*. https://alphafold.ebi.ac.uk/. [Official database site; the figure of over 200 million predictions was read via a search summary, to be confirmed.]
5. Jumper, J., Evans, R., Pritzel, A., et al. (2021). *Highly accurate protein structure prediction with AlphaFold*. Nature, 596(7873), 583-589. https://doi.org/10.1038/s41586-021-03819-2. [AlphaFold 2 and its CASP14 performance.]
6. Kryshtafovych, A., Schwede, T., Topf, M., Fidelis, K., & Moult, J. (2021). *Critical assessment of methods of protein structure prediction (CASP)-Round XIV*. Proteins, 89(12), 1607-1617. https://doi.org/10.1002/prot.26237. [Community assessment of CASP14 results.]
7. Lin, Z., Akin, H., Rao, R., et al. (2023). *Evolutionary-scale prediction of atomic-level protein structure with a language model*. Science, 379(6637), 1123-1130. https://doi.org/10.1126/science.ade2574. [ESM-2 and the metagenomic atlas.]
8. Nobel Prize Outreach. (2024). *The Nobel Prize in Chemistry 2024*. https://www.nobelprize.org/prizes/chemistry/2024/summary/. [Award to Baker, Hassabis and Jumper; facts taken from news reports of the announcement, official page not retrieved. VERIFICATION NEEDED: confirm on nobelprize.org.]
9. Terwilliger, T. C., Liebschner, D., Croll, T. I., et al. (2024). *AlphaFold predictions are valuable hypotheses and accelerate but do not replace experimental structure determination*. Nature Methods, 21(1), 110-116. https://doi.org/10.1038/s41592-023-02087-4. [Comparison of predictions with experimental crystallographic maps.]
10. Tunyasuvunakool, K., Adler, J., Wu, Z., et al. (2021). *Highly accurate protein structure prediction for the human proteome*. Nature, 596(7873), 590-596. https://doi.org/10.1038/s41586-021-03828-1. [Human proteome predictions and their coverage and confidence.]
11. Varadi, M., Anyango, S., Deshpande, M., et al. (2022). *AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space with high-accuracy models*. Nucleic Acids Research, 50(D1), D439-D444. https://doi.org/10.1093/nar/gkab1061. [Description of the database and its initial release.]
12. Watson, J. L., Juergens, D., Bennett, N. R., et al. (2023). *De novo design of protein structure and function with RFdiffusion*. Nature, 620(7976), 1089-1100. https://doi.org/10.1038/s41586-023-06415-8. [Diffusion-based protein design.]

## 12. Evidential Classification Matrix
| Section / Claim | Evidence Category | Supporting Source | Confidence / Status |
| :--- | :--- | :--- | :--- |
| Near-experimental single-protein accuracy at CASP14 | Established | Kryshtafovych et al. (2021); Jumper et al. (2021) | High |
| Predictions can deviate from experiment despite high confidence | Well Supported | Terwilliger et al. (2024) | High |
| Human proteome coverage figures | Well Supported | Tunyasuvunakool et al. (2021) | High (2021 figures) |
| Database exceeds 200 million predictions | Well Supported | EMBL-EBI (2026) | Moderate; verify date |
| AlphaFold 3 outperforms docking and specialised predictors | Preliminary | Abramson et al. (2024) | Moderate; developer benchmarks |
| ESM-2 single-sequence structure prediction at scale | Preliminary | Lin et al. (2023) | Moderate |
| AlphaFold accelerates drug discovery to approval | Hypothetical / Forecast | `FCST-PROT-001` | Medium |
| 2024 Nobel Prize award | Established | Nobel Prize Outreach (2024) | Moderate; official page to confirm |

## 13. Revision History
- **v0.1.0** (2026-10-15): Initial draft. Citations checked against arXiv/Crossref/publisher pages where marked; not independently fact-checked.