DockFormer: Efficient Multi-Modal Receptor-Ligand Interaction Prediction using Pair Transformer

Abstract

Protein-small molecule interactions, or receptor-ligand interactions, are essential for understanding biological processes and advancing drug design. Despite advancements, existing prediction models of these interactions still lack capabilities and accuracy needed to replace traditional screening. In this paper, we introduce DockFormer, a method that leverages multi-modal learning to predict both the binding affinity and structure of these interactions. DockFormer employs fully flexible docking, where no part of the receptor remains rigid, by adapting the AlphaFold2 architecture. Instead of relying on protein sequences and Multiple Sequence Alignments, DockFormer uses predicted receptor structures as input. This modification enables the model to concentrate on ligand docking prediction rather than protein folding, while preserving full receptor flexibility. The stream-lined design also reduces the model size to just 8 layers, compared to AlphaFold2s 48 layers, greatly accelerating the inference process and making it more efficient for large-scale screening. When evaluated on affinity benchmarks such as CASF-2016, PLINDER, and the recently released CASP16 ligand screening benchmark, DockFormer performs comparably to or better than state-of-the-art methods, which typically rely on templates or bound structures as input. On structural benchmarks such as Posebusters and PLINDER, DockFormer demonstrated success rate of 20% and 15%, respectively.