An Active Learning Framework Improves Tumor Variant Interpretation

For precision medicine to reach its full potential for treatment of cancer and other diseases, protein variant effect prediction tools are needed to characterize variants of unknown significance (VUS) in a patient's genome with respect to their likelihood to influence treatment response and outcomes. However, the performance of most variant prediction tools is limited by the difficulty of acquiring sufficient training and validation data. To overcome these limitations, we applied an iterative active learning approach starting from available biochemical, evolutionary, and functional annotations. With active learning, VUS that are most challenging to classify by an initial machine learning model are functionally evaluated and then reincorporated with the phenotype information in subsequent iterations of algorithm training. The potential of active learning to improve variant interpretation was first demonstrated by applying it to synthetic and deep mutational scanning datasets for four cancer-relevant proteins. The utility of the approach to guide interpretation and functional validation of tumor VUS was then probed on the nucleotide excision repair (NER) protein xeroderma pigmentosum complementation group A (XPA), a potential biomarker for cancer therapy sensitivity. A quantitative high-throughput cell-based NER activity assay was used to validate XPA VUS selected by the active learning strategy. In all cases, active learning yielded a significant improvement in variant effect predictions over traditional learning. These analyses suggest that active learning is well suited to significantly improve interpretation of VUS and cancer patient genomes. Significance: A novel machine learning approach predicts the impact of tumor mutations on cellular phenotypes, overcomes limited training data, minimizes costly functional validation, and advances efforts to implement cancer precision medicine.