6 min readLiganx team

Boltz-2 and co-folding: when to use it instead of docking

Co-folding models predict the protein-ligand complex from sequence and SMILES. Here is where Boltz-2 helps, where classical docking still wins, and how to combine them.

Classical docking starts from a receptor structure you already have and searches for where a ligand fits. Co-folding flips the premise: give it a protein sequence and a ligand SMILES, and a single neural network predicts the bound complex directly. Boltz-2 is the model that pushed this from a curiosity to something you should actually consider in a workflow. The question is when it earns its place and when you are better off docking.

What co-folding actually does

Co-folding descends from AlphaFold3, which extended structure prediction beyond single proteins to complexes that include small molecules, nucleic acids, and ions. Instead of treating the receptor as a fixed rigid body and sampling ligand poses against it, the model folds the protein and places the ligand in one joint prediction. That means the binding-site side chains and even backbone can rearrange around the ligand, which a rigid-receptor dock cannot do.

Boltz-1, released in late 2024, was the first openly licensed reproduction of the AlphaFold3 approach. Boltz-2, released in mid-2025, added the piece medicinal chemists actually care about: a binding affinity head. It predicts not just the pose but an estimate of how tightly the ligand binds, from sequence and SMILES alone.

Why people are excited

  • No receptor structure required. If your target has no crystal structure and homology models are poor, co-folding generates a complex from sequence. Docking has nothing to dock into without a structure.
  • Affinity that approaches FEP, far faster. The Boltz-2 authors report affinity correlations approaching free-energy perturbation on several benchmarks while running roughly three orders of magnitude faster, on the order of seconds per complex on a single GPU. FEP can take hours to days per edge.
  • Induced fit for free. Because the protein is folded jointly with the ligand, side-chain and loop rearrangements that a rigid dock would miss can appear in the predicted pose.

Where classical docking still wins

Co-folding is not a universal replacement, and treating it as one is the most common way people get burned.

  • Tiny, well-defined changes. If you have a good crystal structure and you are ranking a congeneric series differing by a methyl here and a halogen there, docking (or FEP) against that structure is usually more reliable than re-folding the whole complex for each analog.
  • Explicit, inspectable scoring. A docking score decomposes into interpretable terms — hydrogen bonds, hydrophobic contacts, clashes. A co-folding affinity is a single learned number with less mechanistic transparency, so it is harder to debug a surprising result.
  • Confidence that can mislead. Co-folding models can return a clean-looking pose with high internal confidence that is nonetheless wrong, especially for ligands or targets unlike the training data. The pose looks plausible, which makes the error dangerous. Pose validation (PoseBusters-style checks) is not optional.
  • Mutation deltas. When the whole point is the difference a single point mutation makes, a learned model may smooth over exactly the effect you are trying to measure. Docking the mutant and wild-type structures separately keeps that signal explicit.

How to combine them

The pragmatic workflow uses co-folding to get a structure and an affinity prior when you have nothing else, then hands the predicted complex to docking and explicit scoring for the work that needs an auditable, decomposable answer. Use co-folding to triage a large or structurally novel set; use docking to reason carefully about the survivors. And whatever generates the pose, run it through pose validation before you trust it.

Try it yourself

The fastest way to build intuition is to run the same ligand both ways and compare. Open Studio and dock a candidate against a target with a known crystal structure, then look at how a co-folding prediction places the same ligand. Where they agree, you can be more confident; where they diverge, you have learned where to look harder. Molecular docking and co-folding answer slightly different questions, and seeing both side by side is the clearest way to understand which to trust for a given problem.

Liganx is molecular docking online: free, browser-based, and set up so you can move between a docking run and a structure-prediction view without leaving the page. If you want to try molecular docking and compare it against a co-folded pose without a local install, that is the fastest path.

Primary sources