Physically valid protein-ligand co-folding for unseen chemotypes and pockets
Hard but with a visible path: in-distribution accuracy is already strong (AF3 baseline LDDT-PLI about 0.8 in CASP16), and open models plus company claims are improving generalisation quickly, so it is easier than affinity prediction (63) and comparable to hardware engineering problems such as CFET (52). Application value 60 as every structure-based pipeline uses these models; leverage 46 as it gates affinity prediction and generative design but docking baselines remain.
The problem
Do AI models that draw a protein with a drug understand the physics, or just recall what they have seen?
- 1Model predicts protein with drug
- 2Works near training examples
- 3Pocket destroyed, pose stays
- 4Accuracy drops on unfamiliar chemistry
- 5Need physics-aware, calibrated models
What co-folding is. A single neural network (AlphaFold 3, Boltz, Chai-1, Protenix, OpenFold3) takes protein sequence and ligand structure and outputs the complex, using diffusion to generate coordinates. It beats classical docking on in-distribution pose accuracy (in CASP16 a baseline AlphaFold 3 reached mean LDDT-PLI about 0.8 versus 0.69 for the best participants).
Built from
earlier problems → this oneThe problems this one is built from: progress or tools it needs (builds on), special cases that come first (generalizes), and weaker results on the way (implies). On the right, what builds on it.
- Start with
- Converged reference interaction energies for 100-atom noncovalent complexes: reconciling CCSD(T) and diffusion Monte CarloComputational Chemistry · D40
- One classical force field accurate for folded proteins, disordered proteins and weak protein–protein associationComputational Chemistry · D51.5
- Predicting fold-switching proteins and both of their folds from sequenceComputational Medicine · D58
- Step 2
- A transferable exchange–correlation functional free of delocalization and self-interaction errorComputational Chemistry · D72
- Boltzmann-weighted conformational ensembles of multi-state proteins from sequenceComputational Medicine · D64
- Step 3
Path to understanding
foundations → problemThe research entries this problem rests on, in the order to read them. Each links to its explainer.
- Start with
- Reverse-mode automatic differentiationAlgorithm
- Gravitational N-body simulationAlgorithm
- Step 2
- The transformer architectureAlgorithm
- Molecular dynamics for biomoleculesAlgorithm
- Step 3
- Step 4
- Molecular docking and binding free-energy calculationAlgorithm · host
- Then
- Physically valid protein-ligand co-folding for unseen chemotypes and pocketsThe problem
State of the art
- Physics tests. Masters, Mahmoud and Lill (Nature Communications 2025) tested AlphaFold 3, RoseTTAFold All-Atom, Chai-1 and Boltz-1 with binding-site mutagenesis. On CASF-2016 (n = 285), for AlphaFold 3 predictions that were initially correct (RMSD < 2 Å), more than half kept the same pose after the pocket was disrupted, and among high-confidence complexes (confidence > 80) 42-52% kept an unaltered pose depending on the disruption versus 78% without disruption. Funnel-metadynamics on CDK2 gave a bound probability of 0 for packed or inverted pockets, so the retained poses are physically unsupported.
- Training-set similarity. Skrinjar et al. ("Runs N' Poses", bioRxiv 2025) built 2,600 post-cut-off protein-ligand systems and found AlphaFold 3, Chai-1, Protenix and Boltz-1 degrade as similarity to the training set falls, including when the ligand is only moved slightly from a training position; pockets are often correct while ligand placement is not, and the effect is strongest for ligands seen in only one pocket. Their conclusion is that these models largely recall ligand poses.
- Validity. PoseBusters (Buttenschoen et al., 2024) showed that many deep-learning docking poses are physically invalid despite low RMSD; newer co-folding models are better on pose validity but not uniformly.
- Industry claims. Isomorphic Labs has described a closed drug-design engine (IsoDDE) and reports substantial gains over AlphaFold 3 on internal generalisation benchmarks. The specific figures circulating in secondary coverage could not be confirmed against a primary technical report for this entry and are therefore not quoted; the model and weights are not released, so no independent verification exists. Any such numbers are company claims, not measured results.
- CASP16. AlphaFold 3 baselines scored mean LDDT-PLI of about 0.8 against 0.69 for the best CASP participants, so pose prediction for in-distribution ligands is strong.
Realistic applications
Reliable co-folding on new chemistry would turn structure-based design from a retrospective into a prospective tool for novel targets and chemotypes, allow rapid pose generation to initialise free-energy calculations, and make learned structure models safe to use in generative design loops without constant physical re-validation.
What counts as solved
A refereed, independent benchmark (with code, time-split data and adversarial perturbation set released) in which a publicly available model meets all four criteria above, with the same evaluation protocol applied to AlphaFold 3 and at least two other models for comparison, reproduced by a group not involved in building the model.
How it gets verified
ReproductionA training run, experiment or computation rerun by someone else.
A publicly available model must meet all four criteria on an independent benchmark (code, time-split data and adversarial perturbation set released), with the same protocol applied to AlphaFold 3 and at least two other models, reproduced by a group not involved in building the model.
Claims under verification
+ add a claimNo claimed results on the Verification board for this problem yet.
Where the work is
Both modelling work (physics-aware training, validity constraints) and benchmark engineering are needed; new experimental structures of novel chemotypes provide test data.
Bottlenecks
ranked by votes- #1TheoryMemorisation versus physical learning
Models keep poses after the pocket is destroyed and degrade with training-set similarity; training objectives do not enforce physical response.
Community grades
Sign in to add your own grade.
Connected problems
Progress
- Feb 10, 2026IsoDDE described by Isomorphic Labs (company claim)source
Isomorphic Labs reports a closed co-folding and design engine with large gains over AlphaFold 3 on internal generalisation benchmarks. No model, weights or refereed report are available, and the circulating figures were not confirmed from a primary source.
- Oct 6, 2025Adversarial co-folding tests publishedsource
AlphaFold 3 and three other models keep poses after binding-site removal; metadynamics shows they are not bound.
- Feb 5, 2025Runs N' Poses (preprint)source
2,600 post-cut-off systems: accuracy falls with similarity to the training set.
- Aug 1, 2023PoseBusterssource
Introduces physical-validity checks and shows many deep-learning poses are implausible.