Predicting how a protein folds is one problem. Predicting how it grabs onto another molecule — a drug, a strand of DNA, another protein — is a harder, more commercially urgent one, because that’s the interaction most drugs actually depend on. Google DeepMind and Isomorphic Labs’ AlphaFold 3, released May 8, 2024, was built specifically to predict that binding. Two years on, the accuracy gains are real and measured. Whether “solves” is the right word is worth being honest about.
What AlphaFold 3 actually does
AlphaFold 3 predicts the structure of biomolecular complexes — not just a single folded protein, but how that protein sits alongside DNA, RNA, small-molecule drug candidates (ligands), ions, and chemically modified residues, all in one unified model. That’s the real leap from AlphaFold 2: earlier versions were built to solve protein folding in isolation, while AF3 models the interactions those folded proteins have with everything around them — which is closer to how drug discovery actually works.
The peer-reviewed paper was published in Nature (Vol. 630, 2024), with a later addendum refining some of the original claims.
The accuracy numbers, specifically
DeepMind’s own reported figure, on the industry-standard PoseBusters benchmark, is that AlphaFold 3 is 50% more accurate than the best traditional docking methods — and it gets there without needing a known binding pocket fed in beforehand, which prior tools typically required. On ligand-docking success specifically (a standard pocket-accuracy measure), AF3 substantially outperformed established tools like AutoDock Vina and Gold, and RoseTTAFold All-Atom, in statistically significant comparisons published in the Nature paper. Protein-protein and antibody-antigen interface prediction also improved meaningfully over AlphaFold-Multimer, AF3’s own predecessor for multi-molecule structures.
Where “solved” doesn’t hold up
This is the part coverage of AlphaFold 3 tends to skip, and it’s worth stating plainly: independent analysis has documented real, specific limitations.
- Stereochemistry errors persist. Roughly 4.4% of predictions still contain chirality violations — molecules mirrored incorrectly — even after the model’s own error-correction step.
- It hallucinates in disordered regions. A 2025 study focused specifically on AF3’s tendency to generate confident-looking but wrong structures for intrinsically disordered protein regions, a known hard problem in structural biology.
- It appears to memorize common ligands rather than fully generalizing. Accuracy on frequently-seen training ligands is over 10 percentage points higher than on novel or rare ones — a meaningful concern for drug discovery specifically, where the whole point is usually predicting binding for a molecule that hasn’t been seen before.
- It only predicts static structures. Real proteins move, flex, and shift between conformations; AlphaFold 3 doesn’t model that dynamic behavior, which limits how far its predictions can be trusted without follow-up wet-lab validation.
Put together, the accurate framing is: significantly more accurate than prior computational methods, not a replacement for experimental validation.
The real-world traction backs up “significant,” not “solved”
As of DeepMind’s own November 2025 “Five Years of Impact” recap, the AlphaFold Server has run over 8 million structure predictions for more than 3 million researchers in 190+ countries, and AlphaFold has been cited in 35,000+ papers. On the commercial side, Isomorphic Labs — DeepMind’s drug-discovery spinoff built on this technology — signed roughly $3 billion combined in milestone-based deals with Eli Lilly and Novartis back in January 2024, and raised a further $2.1 billion round in May 2026 led by Thrive Capital to push its next-generation “IsoDDE” drug-design engine toward the clinic. As of mid-2025, Isomorphic Labs said it was preparing to dose its first patients in oncology trials for AI-designed drug candidates — real progress, but trial-stage, not yet an approved drug on the market.
FAQ
Did AlphaFold 3 actually solve protein binding?
Not entirely. It substantially improved prediction accuracy over prior computational methods (DeepMind cites a 50% gain on the PoseBusters benchmark), but documented limitations — stereochemistry errors, hallucination in disordered regions, weaker performance on novel ligands, and no modeling of protein dynamics — mean it’s a major advance, not a complete solution.
What’s the difference between AlphaFold 2 and AlphaFold 3?
AlphaFold 2 predicted the 3D structure of a single protein. AlphaFold 3 predicts how proteins interact with DNA, RNA, other proteins, and drug-like molecules — modeling binding and complexes, not just individual folding.
Is AlphaFold 3 used in real drug discovery today?
Yes — Isomorphic Labs, DeepMind’s drug-discovery spinoff, has multi-billion-dollar partnerships with Eli Lilly and Novartis and, as of 2025, was preparing early-stage oncology trials for AI-assisted drug candidates.
Can I use AlphaFold 3 myself?
The AlphaFold Server is publicly accessible and has run millions of predictions for researchers worldwide, though usage is generally aimed at researchers rather than casual/non-scientific use.
Bottom line
AlphaFold 3 is a genuine, measured leap in predicting how molecules bind to each other — the “50% more accurate than the best docking tools” claim is real and independently reportable, not marketing fluff. But “solved” is doing more work than the science supports; researchers who work with the model directly are candid that it still gets stereochemistry wrong, memorizes familiar molecules more than it truly generalizes, and can’t model how proteins actually move. Best read as one of the biggest steps forward in structural biology in years — not the final word.
As of August 2026.
