5583 Background: Tumors with homologous recombination repair deficiencies (HRD+) are susceptible to PARP inhibitors (PARPi) which exploit a synthetic lethality by targeting an essential base excision repair (BER) pathway. Current biomarkers designed to predict efficacy of PARPi (BRCA1/2 mut , HRD+) are ineffective, particularly in advanced stages of therapy.There is an urgent unmet need for improved PARPi biomarkers that is particularly acute for patients who fail multiple lines of therapy. We sought to develop and validate a predictive signature for identifying olaparib-sensitive patients that uses real-world data (RWD) as input. Methods: We developed a machine learning (ML) model, called DrugBERT, that learns a relationship between 1) drug structure, 2) DNA alterations, 3) drug response and 4) binding affinity between a drug and its purported target. DrugBERT utilizes genetic information from clinical commercial NGS panels and a subset of clinical features, and outputs a predictive signature for a drug of interest, without need for additional training. Each prediction is accompanied by a Vulnerability Network (VN) which represents a predicted latent genetic sensitivity in tumors and provides biological interpretability to model outputs. Because our model uses RWD as input, we were able to retrospectively evaluate our signature in a RWD ovarian cancer cohort treated with the PARPi, olaparib. Samples were filtered to biopsies taken &lt; 2 years prior to olaparib treatment (N= 48 samples). Drug combinations were analyzed by processing each drug separately and stratifying patients based on sensitivity predictions to all drugs in the regimen. Results: Zephyr's signature outperformed existing PARPi biomarkers in a complex real-world olaparib-treated ovarian cancer cohort. OS and PFS were not significantly different when stratifying patients by BRCA1/2 mutations or HRD+. For Zephyr’s signature, both OS (p-value &lt; 5x10 -3 ; HR = 3.37, 95% CI: 1.27-5.46) and PFS (p-value &lt; 10 -3 ; HR = 2.04, 95% CI: 1.35-3.10), were statistically and clinically significantly prolonged. Characterization of VNs enriched in olaparib-predicted sensitive patient tumors indicated enhanced sensitivity to perturbing HR and cell cycle pathways. Conversely, olaparib-predicted non-sensitive ovarian tumors were characterized by VNs indicating perturbation sensitivity to cell motility and stress response pathways and displayed higher expression of DNA repair, drug resistance, and cancer stem cell pathways. Conclusions: We present a novel and enhanced approach for stratifying patients for PARPi therapy, retrospectively validated in a complex real-world ovarian cancer cohort treated with olaparib. Our interpretable model not only offers a biological hypothesis for predicted responses but is also readily applicable in clinical settings and compatible with standard multi-gene commercial NGS panels for immediate use.

Evaluation of a novel signature for PARP inhibitor sensitivity prediction using real-world data.