| Summary: | Abstract Machine learning and artificial intelligence (AI) are actively applied in drug discovery, such as virtual screening, wherein appropriate molecular representation is critical. Conventional compound representations have limited use because they cannot encode the 3D spatial arrangement of atoms. An atom pair map (APM) represents a compound using a numerical matrix that encodes the physicochemical properties of all atom pairs and interatomic distances. In this way, APM inherently captures the 3D shape of a compound, whereas other conventional representations do not, such as fingerprints, SMILES, and molecular graphs. In this study, we performed a step-by-step evaluation of (i) how well APMs encode common molecular characteristics shared among ligands for target or phenotypic screening hits and (ii) how our APM-based attention model (APNet) compares with other conventional and advanced models. We demonstrated that APM and APNet consistently outperformed other representations and related models across various benchmarks.
|
|---|
