AlphaFold 3 applied to ligands
- Google DeepMind and Isomorphic Labs pushed AlphaFold 3 beyond protein folding in May 2024, adding direct prediction of protein complexes with small-molecule ligands. - The important detail is the caveat: AlphaFold Server supports ligands, but with limits, and newer benchmarks still show weak spots in binding affinity and novel pockets. - So the bottleneck is shifting from raw structure prediction toward proprietary data, assay design, and fast lab validation.
Protein structure AI is no longer just about folding a chain into a shape. The newer game is whether the model can place a drug-like molecule into the right pocket, in the right pose, on the right protein state. That is the part that matters for drug discovery — and it is exactly where AlphaFold 3 moved the conversation. The change started with the May 8, 2024 AlphaFold 3 release from Google DeepMind and Isomorphic Labs, but the bigger story now is what happened after people tried to use it on ligands. ### What changed with ligands? AlphaFold 2 was great at proteins, but it was not built to jointly model proteins with DNA, RNA, ions, and small molecules in one system. AlphaFold 3 was. DeepMind’s paper and server docs both make that explicit — the model can predict complexes containing proteins plus ligands, not just isolated folded proteins. That sounds like a small extension, but basically it turns a structure predictor into something much closer to a docking-and-complex modeler. (nature.com) ### Why does that matter so much? Because a drug does not work by admiring a protein’s fold from across the room. It works by binding — in a specific orientation, with specific contacts, often in a pocket that changes shape as the ligand arrives. If your model can only say “here is the protein,” you still need another stack of tools to guess the ligand pose. If your model can reason over the whole assembly, you cut out a lot of that glue code. (nature.com) ### Is this just docking now? Not quite. The catch is that AlphaFold 3 is better thought of as a structure generator for biomolecular complexes, not a full replacement for medicinal chemistry workflows. The official GitHub notes that the public server has a more limited set of ligands and covalent modifications than the full code path. And several later assessments say the model still struggles when the protein has to move a lot, when the pocket is unfamiliar, or when you need reliable affinity ranking rather than a plausible pose. (nature.com) ### Where does it still break? The recurring failure mode is state. Many proteins are not one rigid object — they are more like a lock that changes shape while the key is entering. Benchmarks published after the AlphaFold 3 launch flagged trouble with induced fit, ternary complexes, and GPCR systems where the model leans toward one conformation even when the ligand implies another. That means a nice-looking complex can still be the wrong biological answer. (github.com) ### So why are people still excited? Because even imperfect ligand-aware prediction is a huge step up from the old split workflow. You can now generate candidate complexes much faster, especially for targets without solved structures. That changes the economics of early screening. The expensive part becomes less “can we make a 3D guess at all?” and more “which guesses deserve scarce assay time?” ### Why does data suddenly matter more? (nature.com) Turns out better models expose a different bottleneck — training and validation data. A 2025 Nature news piece highlighted that drug firms are building their own AlphaFold-like systems partly because public structural data are thin where pharma cares most, especially protein–ligand complexes. Isomorphic Labs’ February 2026 IsoDDE report makes the same point from the other side: AlphaFold 3 advanced the field, but generalizing to novel chemical space, finding new pockets, and predicting affinity still need better systems and better data. (nature.com) ### What does this do to virtual screening? It reframes it. Virtual screening used to be bottlenecked by weak structures and brittle docking. Now the front end is stronger, but the back end matters more — curated ligand datasets, proprietary complex structures, orthogonal scoring, and quick wet-lab loops. In other words, AlphaFold 3 did not end validation. It made validation the center of gravity. (pubmed.ncbi.nlm.nih.gov) ### Bottom line? AlphaFold 3 made ligand modeling real enough to change workflows, but not reliable enough to let chemistry teams skip experiments. The breakthrough is not “AI solved docking.” It is that the bottleneck moved — from getting a structure at all to proving which predicted interactions are actually worth betting on. (nature.com) (storage.googleapis.com)
