Machine learning identifies a core gene set predictive of acquired resistance to EGFR tyrosine kinase inhibitor.

PURPOSE: Acquired resistance (AR) to epidermal growth factor receptor tyrosine kinase inhibitors (EGFR-TKIs) is a major issue worldwide, for both patients and healthcare providers. However, precise prediction is currently infeasible due to the lack of an appropriate model. This study was conducted to develop and validate an individualized prediction model for automated detection of acquired EGFR-TKI resistance.
METHODS: Penalized regression was applied to construct a predictive model using publically available genomic cohorts of acquired EGFR-TKI resistance. To develop a model with enhanced generalizability, we merged multiple cohorts then updated the learning parameter via robust cross-study validation. Model performance was evaluated mainly using the area under the receiver operating characteristic curve.
RESULTS: Using a multi-study-derived machine learning method, we developed an extremely parsimonious model with generalized predictors (DDK3, CPS1, MOB3B, KRT6A), which has excellent prediction performance on blind cohorts for AR to EGFR-TKIs (gefitinib, erlotinib and afatinib) and monoclonal antibody against EGFR (cetuximab). In addition, our model also showed high performance for predicting intrinsic resistance (IR) to EGFR-TKIs from two large-scale pharmacogenomic resources, the Cancer Genome Project and the Cancer Cell Line Encyclopedia, suggesting that these general predictive features may work across AR and IR.
CONCLUSIONS: We successfully constructed a multi-study-derived prediction model for acquired EGFR-TKI resistance with excellent accuracy, generalizability and transferability.