AI vs Physics: Outperformance in Protein–Ligand Docking

AI vs Physics: Outperformance in Protein–Ligand Docking

PARTNERSHIP
Announcement

Summary

Full Text

We evaluate docking approaches on the PoseX benchmark and outline a hybrid workflow that combines AI-generated poses with brief physics-based minimization for optimal accuracy and throughput. Physics-based methods are robust but slow and often struggle with metal-coordinating sites or structured water. AI docking predicts plausible poses in seconds and captures interaction patterns that are hard to encode with hand-crafted scoring.

Overall Outperformance

Standardized benchmarks of experimentally solved complexes allow direct pose-accuracy comparisons. We used the PoseX dataset, reflecting real discovery challenges.

Figure 1. Percentage of correct predictions by protein–ligand docking methods in PoseX.

Pure AI engines such as ArtiDock and Uni-Mol yield a higher fraction of correct poses than classic force-field methods like AutoDock Vina or Glide. Accuracy improves further when ArtiDock’s pose is followed by a brief physics-based refinement (Vina or UFF), adding little compute. Co-folding models that dock while predicting the pocket are useful in some cases but currently trail both specialized AI docking and physics-based methods.

Performance in Difficult Areas

Sites with metal ions, cofactors, or structured water are the hardest; generic parameterizations can miss coordination geometry and directional hydrogen bonding, causing misplaced poses.

Figure 2. ArtiDock performance in pockets containing no heteroatoms, cofactors, ions, or bound water: median RMSD (Å), IDTI-PLI Median and PB-Valid (fraction of poses passing PoseBusters checks).

ArtiDock shows lower median RMSD and better recovery of true contacts across difficult-site categories. A short minimization may slightly increase RMSD due to subtle shifts but typically boosts chemical validity by restoring key interactions.

How It Works: ArtiDock Architecture

Figure 3. General scheme of ArtiDock architecture and inference pipeline.

ArtiDock encodes the ligand and pocket as compact graphs with essential chemical and spatial features. A lightweight network learns an inter-graph distance matrix that describes how the ligand should fit. A fast alignment then converts the predicted distances into a 3D pose, bypassing exhaustive search.

Conclusion

A hybrid AI–physics workflow delivers precise poses in seconds, scales to large libraries, and performs especially well in metal/cofactor/water-rich pockets, accelerating decision-making in drug discovery.