Deep Kernel Inversion: Rapid and Accurate Molecular Interaction Prediction for Drug Design

Abstract

Computational drug design offers the opportunity to dramatically accelerate novel therapeutics for untreated diseases. Designing compounds with optimal efficacy and specificity, however, requires understanding and optimizing immense numbers of molecular interactions. While advances in predicting one-to-one molecular interactions continue, there has been limited progress in scaling one-to-many or many-to-many molecular interaction models. In this paper, we introduce a deep learning framework that embeds molecules into a high-dimensional vector space, which we have named Deep Kernel Inversion. In this framework, the dot product between vectors accurately predicts molecular interactions. This approach reduces the complexity of predicting an entire molecular interaction network from O(n2) to O(n), enabling new molecular design tasks previously inaccessible to computational approaches. In the case of human protein-protein interactions (PPI), we demonstrate a 100,000 fold decrease in the computation required to map the full human PPI network. We also demonstrate best-in-class performance across multiple molecular interaction tasks with this approach. This work offers a new way forward in scaling accurate molecular interaction predictions with applications in mapping biological pathways, target discovery, drug design, and therapeutic development.