AAAI Publications, The Thirtieth International Flairs Conference

Font Size: 
Virtual Screening Assisted by Siamese Neural Networks
Alan Diego dos Santos, Duncan Ruiz

Last modified: 2017-05-08


High-throughput virtual screening relies on scoring functions to evaluate binding affinity between ligand and receptor. Although useful for identification of new potential drugs, imperfections in these scoring functions can lead to incorrect classification of small molecules. In this context, non-parametric machine-learning approaches can identify implicit binding interactions that can be hard to model explicitly. We present an approach to distinguish between ligands and decoys using energy-based models with siamese neural networks. Taking as inputs 3D biochemical property grids from ligand and receptor, it is computed the compatibility between them. We show that this model outperforms other machine-learning approaches in a Fully Flexible Receptor model of InhA-NADH complex.


virtual screening;machine learning;scoring function;

Full Text: PDF