SurfBind: Surface Distance aided Geometric Deep Learning for Binding Conformations

Yuan Yao, Jiamin WU, Wenqi Zeng

Abstract:

A core task in computer-aided drug discovery is the optimization of lead compounds with high binding affinity to the target
proteins. The binding process is desired to find the proper
position and the correct relative orientation of the “key” (the
ligand), which will open up the “lock” (the protein). During
the process, existing deep learning methods usually overlook
surface intersection between ligands and targets, i.e., part of
the ligands goes into the protein interior. In this paper we
present our SurfBind model, a two-stage deep learning method
aided by the surface distance function (SDF). Our model will
produce pairwise distance distribution to encode the multivalued possible relative positions. By exerting effective SDF
constraint to distance likelihood potential, SurfBind derives
rational conformations lessening clash to the protein, and also
reduces root mean square deviation (RMSD) for ultra-large
ligands. To the best of our knowledge, the performance of SurfBind working as a score function on docking and screening
power achieves SOTA on CASF-2016 benchmark.