Hierarchical Neural Networks for Drug Response Prediction Using Targets, Transcriptomics, and Pharmacogenomic Biomarkers Download PDF

Precision oncology seeks to tailor anticancer therapies to individual patients by leveraging molecular profiles of tumors and, when available, patient-specific data; however, predictive models trained on large pharmacogenomic screens often struggle to translate clear in vitro associations into clinically meaningful response predictions. Many existing drug response models combine drug descriptors, gene expression levels, mutations, and copy-number alterations into a single flat representation, disregarding the biological sequence in which a drug engages molecular targets, perturbs cellular programs, and is influenced by the patient’s genomic context. To address this, this MDL article proposes a hierarchical neural network for predicting drug response in cancer, designed to integrate drug-target interactions, tumor transcriptomes, and pharmacogenomic biomarkers within a biologically ordered framework. The architecture first encodes drug-target and target-inhibition information in a lower-level target-engagement module, then merges this latent drug mechanism representation with transcriptomic context in a middle module, and finally modulates the resulting signal using mutation and copy-number biomarkers in an upper pharmacogenomic module to produce a personalized sensitivity estimate. Conceptually, this nested structure is expected to yield more biologically coherent predictions than flat multimodal approaches and could clarify whether predicted resistance arises primarily from inadequate target engagement, an unfavorable transcriptomic state, or genomic alterations affecting drug response. By respecting the hierarchy from drug mechanism to cellular state to patient genotype, such a biologically structured deep learning model offers a principled framework for interpretable precision oncology, potentially bridging preclinical pharmacogenomic screening and clinical decision support.