Predicting Oncogenic Mutants of Epidermal Growth Factor Receptor in Lung Cancer by Molecular Dynamics Simulation with Machine Learning

Abstract The epidermal growth factor receptor (EGFR) is a frequently mutated protein kinase found in patients suffering from non-small cell lung cancer (NSCLC). Although therapeutic drugs targeting EGFR are available, not all patients respond to these drugs, and it is not yet known whether some clinically detected mutations are actionable. Although the oncogenicity of some mutants is well established, many mutants found in patients are not yet classified or classified as Variants of Uncertain Significance. For newly discovered rare mutants, strong evidence supporting their classification into oncogenic or benign is often lacking. Under such situations, multiple pieces of evidence can still help classification, even if each individual piece of evidence is weaker. Computational predictions have been considered as supporting information by the Standard Operating Procedure (SOP) of the Clinical Genome Resource/Cancer Genomics Consortium/Variant Interpretation for Cancer Consortium (ClinGen/CGC/VICC). The SOP considers predictions from multiple computational programs as providing only one piece of supporting information to avoid double counting because these programs share common approaches. A computational approach that explicitly accounts for the conformational effects of mutations via molecular dynamics (MD) simulation could add another piece of complementary supporting information. We tested the hypothesis that features of structural dynamics observed from MD simulation could help distinguish oncogenic from benign mutants. To this end, we performed three independent MD simulations lasting 1 μs each for wild-type EGFR and 32 of its oncogenic or benign mutants. We found that random forest machine-learning models constructed using suitable structural features were able to distinguish oncogenic and benign mutations effectively. Although we did not supply features of structural motifs commonly used by scientists to distinguish active and inactive conformations, our machine-learning approach identified and utilized some features associated with these motifs for classifying EGFR mutants.