A multi-task domain-adapted model to predict chemotherapy response from mutations in recurrently altered cancer genes

Next generation sequencing (NGS) of tumours is increasingly utilised in oncological practice, however only a minority of patients harbour oncogenic driver mutations benefiting from targeted therapy. Development of a drug response prediction (DRP) model based on available genomic data is important for the ‘untargetable’ majority of cases. Prior DRP models typically rely on whole transcriptome and whole exome sequencing (WES), which is often unavailable in clinical practice. We therefore aim to develop a DRP model towards repurposing of standard chemotherapy, requiring only information available in clinical grade NGS (cNGS) panels of recurrently mutated genes in cancer. Such an approach is challenging due to the sparsity of data in a restricted gene set and limited availability of patient samples with documented drug response. We first show that an existing DRP performs equally well with whole exome data and a cNGS subset comprising ∼300 genes. We then develop Drug IDentifier (DruID), a DRP model specific for restricted gene sets, using a novel transfer learning-based approach combining variant annotations, domain-invariant representation learning and multi-task learning. Evaluation of DruID on pan-cancer data (TCGA) showed significant improvements over state-of-the-art response prediction methods. Validation on two real world - colorectal and ovarian cancer - clinical datasets showed robust response classification performance, suggesting DruID to be a significant step towards a clinically applicable DRP tool.