FitScore: A Fast Machine Learning-Based Score for 3D Virtual Screening Enrichment

Abstract Enhancing virtual screening enrichment has become an urgent problem in computational chemistry, driven by increasingly large databases of commercially available compounds, without a commensurate drop in in vitro screening costs. Docking these large databases is possible with cloud-scale computing. However, rapid docking necessitates compromises in scoring, often leading to poor enrichment and an abundance of false positives in docking results. This work describes a new scoring function composed of two parts – a knowledge-based component that predicts the probability of a particular atom type being in a particular receptor environment; and a tunable weight matrix that converts the probability predictions into a dimensionless score suitable for virtual screening enrichment. This score, the FitScore, represents the compatibility between the ligand and the binding and is capable of a high degree of enrichment across standardized docking test sets.