Abstract
Protein kinase inhibitors have ushered in a paradigm shift in targeted cancer therapy and precision oncology. Despite their clinical success, resistance frequently arises through adaptive kinome reprogramming, in which cancer cells dynamically rewire signaling pathways to bypass inhibition. This, combined with the vast complexity of chemical space and cellular heterogeneity, necessitates scalable, data-driven strategies to identify compounds with desirable polypharmacological profiles and to predict therapeutic response across diverse biological contexts.
Two conceptually aligned computational frameworks were developed to address these challenges. The first is a multi-task deep neural network trained on a harmonized, kinome-wide bioactivity dataset, enabling large-scale virtual screening, compound–target interaction profiling, and quantitative analysis of selectivity and polypharmacology. The second, DrugSSeq, is a multimodal deep learning framework that integrates compound molecular descriptors and gene expression profiles to predict drug sensitivity across diverse cancer types, supporting the prioritization of compounds in biologically responsive cellular contexts. To enable broad access to kinome-scale predictions, KNet, a web-based platform, was developed to deliver structure-informed predictions across the human kinome. When combined with molecular modeling tools such as AlphaFold and Schrödinger, these frameworks enable systematic exploration of kinase–ligand interactions across diverse scaffolds and both established and understudied targets.
In summary, this work presents cohesive and extensible computational frameworks that span molecular and phenotypic levels of prediction, advance methodologies for kinase-targeted drug discovery and precision oncology, and deliver scalable tools to enable compound prioritization,support polypharmacology profiling, and inform the development of next-generation cancer therapies.