Machine learning-based prediction of response to PARP inhibition across cancer types

PARP inhibitors (PARPi) are FDA approved for the treatment of BRCA1/2 deficient breast and ovarian cancer, but a growing body of pre-clinical evidence suggests the drug class holds therapeutic potential in other cancer types, independent of BRCA1/2 status. Large-scale pharmacogenomic datasets offer the opportunity to develop predictors of response to PARPi’s in many cancer types, expanding their potential clinical applicability. Response to the PARPi olaparib was used to identify a multi-gene PARPi response signature in a large in vitro dataset including multiple cancer types, such as breast, ovarian, pancreatic, lung cancer, osteosarcoma and Ewing sarcoma, using machine learning approaches. The signature was validated on multiple independent in vitro datasets, also testing for response to another PARPi, rucaparib, as well as two clinical datasets using the cisplatin response as a surrogate for PARPi response. Finally, integrative pharmacogenomic analysis was performed to identify drugs which may be effective in PARPi resistant tumors. A PARPi response signature was defined as the 50 most differentially transcribed genes between PARPi resistant and sensitive cell lines from several different cancer types. Cross validated predictors generated with LASSO logistic regression using the PARPi signature genes accurately predicted PARPi response in a training set of olaparib treated cell lines (80-89%), an independent olaparib treated in vitro dataset (66-77%), and an independent rucaparib treated in vitro dataset (80-87%). The PARPi signature also significantly predicted in vitro breast cancer response to olaparib in another separate experimental dataset. The signature also predicted clinical response to cisplatin and survival in human ovarian cancer and osteosarcoma datasets. Robust transcriptional differences between PARPi sensitive and resistant tumors accurately predict PARPi response in vitro and cisplatin response in vivo for multiple tumor types with or without known BRCA1/2 deficiency. These signatures may prove useful for predicting response in patients treated with PARP inhibitors.