What Changed: AlphaFold 3 Is No Longer Just About Proteins
In May 2024, Google DeepMind and Isomorphic Labs released AlphaFold 3, a fundamentally different tool from its predecessor. AlphaFold 2 (released in 2021) made protein structure prediction practical for most researchers. AlphaFold 3 expands the scope entirely: it now models DNA, RNA, ligands, ions, and modified residues alongside proteins.
This is more than an incremental improvement. AlphaFold 2 had no realistic way to predict how a protein binds a drug molecule or how a DNA-binding transcription factor actually grips DNA. AlphaFold 3 does both, with documented accuracy gains. The shift from “protein structure prediction” to “biomolecular interaction prediction” changes which questions you can ask of your structures.
The underlying architecture also changed. AlphaFold 3 uses a diffusion-based model (similar to image diffusion models) rather than the transformer-based architecture of AF2. This diffusion approach treats coordinate generation as a generative process, which gives AF3 better flexibility to handle diverse molecular types.
For computational researchers and structural biologists, the question is immediate: Should I migrate my workflows? The answer depends on what you are trying to do.
AlphaFold 3 vs. AlphaFold 2: When to Use Each
Here is the practical decision matrix:
Use AlphaFold 3 if you need to:
- Model protein-ligand complexes (including drug binding)
- Predict nucleic acid-protein interactions
- Study RNA or DNA structure in isolation or bound to proteins
- Model protein complexes with higher accuracy than AF2
Stick with AlphaFold 2 if:
- You are predicting single-chain protein structures only
- You need the full source code and can run it locally on your infrastructure (AF3 code was restricted to non-commercial use until February 2025; source code was made publicly available in Feb 2025 with ongoing non-commercial restrictions)
- You rely on commercial use and cannot accept licensing restrictions
- Speed is critical for high-throughput screening (AF2 may still be faster on your hardware)
AlphaFold 2 remains production-ready and will not disappear. Its accuracy for protein monomers is solid, it is fully open source under the Apache 2 license, and it integrates into established pipelines. Use it when AF3 is unnecessary.
The Three Ways to Access AlphaFold 3
1. AlphaFold Server (Web Interface, Free, Limited)
The easiest entry point is AlphaFold Server, a free web portal launched alongside AF3. You submit a sequence or complex, and the server runs AF3 for you. No installation required.
Limits: Each user gets 30 predictions per day with up to 5,000 “tokens” per job (roughly equivalent to the total number of residues in your input). You cannot use custom ligands or drugs; the server accepts only a predefined list of common biological molecules. This limitation exists to avoid cannibalizing the drug discovery services of Isomorphic Labs.
2. AlphaFold 3 Open-Source Code (Non-Commercial)
In February 2025, Google DeepMind released the AF3 code and weights publicly on GitHub. It is free to download but restricted to non-commercial academic research.
Advantages: You run locally, no daily quota, full customization (custom ligands, modifications, batch processing). Disadvantages: High computational cost (GPU memory and runtime), licensing audit risk if your institution is unclear on commerciality, setup complexity.
3. Third-Party Implementations
Researchers have built open-source implementations. ESMFold (Meta) and RoseTTAFold-All-Atom (Baker Lab) are mature alternatives with fewer licensing strings attached.
AlphaFold 3 vs. Competitors: Accuracy, Speed, and Coverage
Here is how AF3 compares to ESMFold and RoseTTAFold-All-Atom:
| Metric | AlphaFold 3 | AlphaFold 2 | ESMFold | RoseTTAFold-All-Atom |
|---|---|---|---|---|
| Coverage | Proteins, DNA, RNA, ligands, ions, modifications | Proteins only | Proteins only | Proteins, DNA, RNA, ligands, ions |
| Protein-ligand accuracy | 50% improvement over physics-based tools | Limited | Not designed for ligands | Comparable to AF3 |
| Speed | Moderate (high GPU cost) | Moderate | 60x faster than AF2 | Slower than ESMFold |
| Protein monomer accuracy | Slightly better than AF2 | Baseline | Competitive with AF2 | Competitive with AF2 |
| Availability | Web server (quota) or code (non-commercial) | Open source (Apache 2) | Open source (MIT) | Open source (no restriction) |
| Licensing | Non-commercial use only | Fully open | Open source | Open source |
| Learning curve | Easy (web) or high (local) | Moderate | Low | Low |
AF3’s real win: ligand and nucleic acid modeling. ESMFold is fastest for protein-only tasks. RoseTTAFold-All-Atom is a robust open-source alternative if you need customization without licensing friction.
For most researchers, the AlphaFold Server web interface is the practical entry point. The 5,000-token limit covers most single-chain proteins and small complexes.
Key Improvements Under the Hood
According to the foundational paper by Abramson et al. in Nature, AF3 achieves:
- Minimum 50% improvement in accuracy for protein-ligand interactions compared to traditional physics-based docking
- First AI system to outperform physics-based tools on the PoseBusters benchmark for small-molecule docking
- Ability to model modified residues (phosphorylation, glycosylation, ubiquitination) and covalent modifications
The diffusion architecture allows AF3 to generate atomic coordinates in a learned way, making it more flexible than AF2’s constraint-based approach. This also means AF3 can reason about bond geometry and chemical constraints directly.
Where AlphaFold 3 Fails (and Where You Need Other Tools)
AlphaFold 3 is powerful but not universal. Know its limits:
Intrinsically Disordered Proteins (IDPs) and Disordered Regions: AF3 hallucinates structure in genuinely disordered regions. If your protein contains an intrinsically disordered domain or linker, AF3 may confidently predict a spurious fold. You will need other tools like IUPred or PONDR to annotate disorder regions first.
Large Complexes: AF3 struggles with multi-protein complexes beyond a few thousand residues. The 5,000-token limit on the server enforces this. Large assemblies (ribosomes, proteasomes, virus capsids) are beyond current scope.
Membrane Proteins: AF3 does not reliably predict the lipid bilayer or conformational changes induced by membrane embedding. If you study GPCRs or transporters, you will still rely on cryo-EM, MD simulations, or hybrid approaches.
Protein Dynamics: AF3 predicts a static snapshot, not an ensemble or conformational transitions. For allosteric mechanisms or kinetically important states, molecular dynamics or cryo-EM with sample heterogeneity remains essential.
Novel Ligands Not in Training Data: AlphaFold Server does not accept custom drug molecules. The open-source code can theoretically accept novel molecules, but AF3 is less reliable for out-of-distribution compounds with no training precedent.
Practical Workflow: When to Use AF3 in Your Research
-
Drug discovery lead identification: Use AF3 to model protein-drug poses before you synthesis. Compare AF3 predictions to experimental binding data or MD simulations. AF3 is fastest for initial triage.
-
Protein engineering: Design mutations to improve ligand binding. Predict the mutant complex with AF3, then validate experimentally.
-
RNA-protein interactions: Study transcription factors, helicases, spliceosomal proteins. AF3 can model RNA backbone and base contacts explicitly.
-
Cryo-EM model interpretation: Use AF3 to build initial models or interpret low-resolution density.
-
Genomics and variant effect prediction: Predict structures of variants to infer pathogenicity.
What AF3 is not: a replacement for molecular dynamics, cryo-EM, crystallography, or biochemical validation.
The Licensing Question: What You Need to Know
AlphaFold 3 open-source code is restricted to “non-commercial” use. This term is vague and creates friction:
- Academic researchers at universities: clear green light
- Researchers at nonprofits: generally okay, but check your licensing
- Industry researchers: not permitted without Isomorphic Labs licensing deal
- Startups: risky unless you have explicit permission
If you work in industry or an ambiguous setting, start with the AlphaFold Server (clearly non-commercial) or commit to RoseTTAFold-All-Atom and ESMFold, both of which carry fewer restrictions.
What This Means for Your Field
AlphaFold 3 will reshape structural biology workflows over the next 1-2 years. Drug discovery teams will integrate AF3 into scoring pipelines before wet-lab screening. Protein designers will use AF3 to validate interface predictions. Genomicists will incorporate AF3 predictions into variant pathogenicity classifiers.
The accessibility bottleneck (daily quota, licensing, GPU cost) will likely ease as competitors catch up and Isomorphic Labs releases commercial licenses. For now, the free web server is your entry point.
Verdict: Should You Switch?
If you work with protein-ligand or nucleic acid interactions, yes. The AlphaFold Server costs you nothing and your time to predict one complex is minutes. Try it on a real question in your project this week.
If you predict only monomeric protein structures, AlphaFold 2 is still efficient and fully open. No urgent need to migrate.
If you work in industry, be cautious of licensing. Use RoseTTAFold-All-Atom or negotiate a commercial AF3 license with Isomorphic Labs.
The gap between AF3 and older tools is wide enough that predictions are now part of the standard structural bioinformatics toolkit. Learn it.
Next Steps
Start with the free AlphaFold Server. Predict a known complex and compare the result to experimental structure. Check the confidence metrics (pLDDT for confidence, pAE for interface confidence).
Want to go deeper? Check out the official AlphaFold training materials for a walkthrough of AF3 architecture and when to use each tool.
Interested in open alternatives? Our recent bioinformatics tools roundup covers other recent structure prediction updates, including ESMFold and RoseTTAFold.
Sign up for our newsletter below to stay updated on structure prediction benchmarks, new releases, and practical workflows for your research.