LoopPlay

Blind flexible docking asks a model to do two hard things at once: find where a small molecule binds on a protein it has never seen, and predict how both partners change shape on binding. Regression-based docking networks answer in a fraction of a second, far faster than sampling-based diffusion methods or classical search software, but the multi-task versions of these networks have a persistent asymmetry — they predict the bound protein pocket noticeably better than they predict the bound ligand. Ligands and pockets have different structural statistics, and training one shared objective over both lets the easier task dominate.

LoopPlay, developed at the Centre for Artificial Intelligence and Robotics of the Chinese Academy of Sciences in Hong Kong, reframes that asymmetry as a game. The Docking Game is a two-player formulation in which a ligand player predicts the pocket location and the bound ligand pose, while a protein player predicts the bound pocket conformation. The two are trained by Loop Self-Play, a nested procedure: in the outer loop the players swap predicted structures so each conditions on the other's current best guess, and in the inner loop each player feeds its own predictions back into itself to refine them iteratively. A shared distance-map term couples the two payoff functions, and the authors prove the procedure converges to a Nash equilibrium.

The contribution is the training algorithm rather than a new network. LoopPlay builds on the FABind family of equivariant docking layers and starts from FABFlex's pretrained weights, producing an improved checkpoint within that established lineage rather than a new architecture.

#Key Features

• Game-theoretic training: Ligand and pocket prediction are modeled as separate players with their own loss functions, coupled through a shared distance map, so neither task's gradient swamps the other.
• Two-level loop self-play: Outer-loop pose exchange lets each player adapt to the other's structural predictions; inner-loop self-refinement lets a player iterate on its own output, converting a single-shot regression into a short refinement trajectory.
• Blind and flexible: No binding pocket is supplied at inference, and both the ligand and the pocket are allowed to move from their unbound conformations to predicted bound states.
• Fast fixed-weight inference: A complex is docked in 0.32 seconds on average, roughly 319 times faster than the sampling-based flexible-docking method DynamicBind at 102.12 seconds.
• Proved convergence: The paper provides a theoretical convergence guarantee for the alternating training scheme, addressing the instability that alternating optimization often introduces.

#Technical Details

Each player is built from stacked FABind layers, an E(3)-equivariant graph neural network operating on a heterogeneous protein-ligand graph, with protein residue features drawn from ESM-2. The system has three modules — a one-layer pocket-prediction module and five-layer ligand-docking and pocket-docking modules, with hidden sizes of 128, 512, and 512. Inference starts from unbound inputs: an apo ligand conformer generated with RDKit and an apo protein structure predicted by AlphaFold 2, with the ligand initially placed at the protein's center. Training uses PDBBind v2020, filtered to 12,807 complexes, with 734 pre-2019 complexes held out for validation and 303 post-2019 complexes for testing. On the full test set LoopPlay reaches a mean ligand RMSD of 4.90 Å and a 41.91% success rate at the standard 2 Å threshold, against 5.44 Å and 40.59% for FABFlex, the previous best regression-based method. On the stricter subset of 114 complexes whose protein receptors do not appear in training at all, it records 6.59 Å mean RMSD and a 32.46% success rate, the highest of the compared methods.

#Applications

Fast blind docking is most useful where throughput dominates: virtual screening across large compound libraries, hit triage before more expensive physics-based rescoring, and target-agnostic pose prediction when no cocrystal structure or annotated pocket exists. Sub-second inference with fixed weights means a screen can be run over many targets without per-target refitting, and because the pipeline consumes RDKit conformers and predicted apo structures rather than experimental holo structures, it applies to targets with no solved complex.

#Impact

LoopPlay's broader claim is methodological — that self-play, familiar from game-playing agents, transfers to structural biology when a prediction task decomposes into interacting sub-tasks with asymmetric difficulty. The measured gain over FABFlex is real but incremental, roughly 10% in mean RMSD overall and 8% on unseen receptors, and it is uneven across the distribution: LoopPlay is better at the median and upper quartile of RMSD while FABFlex is marginally better at the 25th percentile. Evaluation is confined to PDBBind with no experimental validation, and the model inherits its lineage's dependence on the quality of RDKit conformers and predicted apo structures. The work is an arXiv preprint that has not been peer reviewed, and no code repository or trained checkpoint has been released, so the results have not been independently reproduced.