Abstract 1249: Development of a pan-cancer algorithm to predict homologous recombination deficiency and sensitivity to PARPi therapy

Abstract

Abstract Background: PARP inhibitors (PARPi) are approved for multiple indications with ongoing trials to explore broader utility. However, identifying the right patients for these therapies across multiple disease types remains a challenge. In ovarian cancer, genomic-scar based measures for homologous recombination deficiency (HRD) are approved diagnostics (genome-wide LOH [gLOH] and genomic instability score [GIS]); however, broader utility has not been established. Methods: A pan-cancer genomic profiling dataset (n = 202,472; Foundation Medicine, Cambridge, MA) was split 70:30 for training and validation of an HRD signature using an XGB machine learning model (mlHRD). A broad set of copy number (Macintyre 2018) and indel features (Alexandrov 2020) were used to identify signatures of HRD. gLOH (Coleman 2017) and GIS (Timms 2014) were calculated using copy number profiles. Biallelic alterations were predicted using a computational zygosity algorithm (Sun, 2018). The nationwide, de-identified Flatiron Health-Foundation Medicine ovarian and prostate clinico-genomic databases (FH-FMI CGDB) were utilized for outcomes analysis. The de-identified data originated from approximately 280 US cancer clinics (~800 sites of care). Time to therapy discontinuation (TTD) was estimated with Kaplan-Meier analysis. Hazard ratios were calculated using unadjusted Cox proportional Hazard models. Results: We developed an algorithm to predict HRD status using indel and copy number features (see methods). Across the pan-cancer dataset, the rate of mlHRD was 6.4% with the highest frequency in fallopian tube (30%), ovarian (30%), peritoneal (23%), breast (16%), and prostate cancers (15%). Sensitivity to detect biallelic BRCA1/2 alterations was high across tumors [ovary (93%), prostate (87%), breast (85%), pancreas (80%)]. Beyond BRCA1/2, mlHRD positivity was associated with biallelic alterations in RAD51D (OR = 24, p<1E-10), PALB2 (OR = 23, p<1E-10), BARD1 (OR = 23, p<1E-10), and RAD51C (OR = 19, p<1E-10). In the FH-FMI CGDB ovarian cancer cohort, 150 patients were treated with PARPi (mlHRD positive = 73; negative = 77); mlHRD positivity was associated with improved TTD (median 8.9 mo v 3.9 mo; HR = 0.49 [0.34-0.71], p < 0.001), with similar predictive power to gLOH >16% (HR = 0.55 [0.38-0.79], p = 0.001) and GIS >42 (HR = 0.59 [0.41-0.86], p = 0.006). For 62 patients with prostate cancer treated with PARPi (mlHRD positive = 27; negative = 35), mlHRD was associated with prolonged TTD on PARPi (median 6.8 mo v 3.4 mo; HR = 0.56 [0.30-1.03], p = 0.064), trending more predictive than gLOH >8.29% (Sokol 2020) and GIS >42 (HR = 0.64 [0.29-1.40] and 0.80 [0.37-1.73], respectively; p>0.05). Conclusion: These findings suggest that HRD is associated with genomic scarring beyond ovarian cancer. Additional retrospective and prospective analyses in clinical datasets are needed to explore the utility of this signature. Citation Format: Emmanuel Antonarakis, Jay Moore, Dexter Jin, Tim Chen, Justin Newberg, Zoe Fleischmann, Karthikeyan Murugesan, Garrett Frampton, David Fabrizio, Russell Madison, Ethan Sokol. Development of a pan-cancer algorithm to predict homologous recombination deficiency and sensitivity to PARPi therapy [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 1249.