Over 90% of cancer deaths are due to drug resistance caused by genetic variants. A large number of variants are variants of unknown significance meaning that they are not clinically validated biomarkers. Existing biomarkers capture only single variants in a single protein. In this study, a computational pipeline was developed that applies machine learning to the dihedral angles extracted from replica exchange molecular dynamics to predict drug resistance status of single or multiple variants in a specific protein. A network model is then applied to predict the cumulative effect on drug resistance of one or more variant protein. This method was applied to BRAF, PTEN, KRAS, and MEK1. The random forest machine learning model achieved high accuracy across proteins, ranging from 89.5% to 100%. For BRAF, the model reached 91.7% accuracy on 12 variants treated with dabrafenib and 100% accuracy on 11 variants treated with vemurafenib. For PTEN, the model achieved 94.7% accuracy on 38 variants, for KRAS it achieved 89.5% accuracy on 19 variants, and for MEK1 it achieved 100% accuracy on 8 variants. In addition, the pipeline generated predictions for variants of unknown significance, including 4 in BRAF with dabrafenib, 5 in BRAF with vemurafenib, 4 in PTEN, 5 in KRAS, and 3 in MEK1. The network model was able to capture the complex dynamics of the growth and proliferation pathways to predict the cumulative effect of multiple variant proteins. Further clinical studies are needed to validate the pipeline to gauge utility for providing drug resistance information to the oncologist.
Xie et al. (Sun,) studied this question.