Stochasticity in threshold of homologous recombination deficiency (HRD) score.

Abstract

e17561 Background: HRD test from Myriad Inc. was approved by FDA as a biomarker for niraparib in ovarian cancer. Its threshold (as 42) was initially established by capturing 95% of tumors with BRCA1/2 mutation and BRCA1 promoter methylation in the training cohort. Here, we explored the stochasticity of such threshold, when the test noises themselves were concerned. Methods: We simulated HRD scores from figure 1 of Telli et al., 2016 by assigning the median from each histogram bin to the associated samples. The threshold was calculated, as described in the original paper, and the standard error (SE) and confidence interval (CI) for such threshold were estimated from bootstrapping the cohort. Furthermore, the entire cohort was down-sampled to determine the effect of sample size on the threshold. Results: The threshold of 42.5 was calculated from our simulated dataset, close to the original one, with SE of 1.62 and 95% CI of 40.75 - 47.5 for the bootstrap. As expected, SE steadily increased to 2.21, 2.84, 2.97 when the sample sizes decreased to 250, 230, and 210 respectively. When the dataset was further down-sampled to include only 200 BRCA deficient samples, SE increased to 3.76, and drastically, CI ranged between 36.5 and 51.2. Conclusions: We have showed there exists manifest stochasticity in the original threshold of 42 for the Myriad HRD test, which is likely close to platform noise in magnitude. The minimum size of BRCA deficient patients to obtain a stable threshold is 200, as shown in our study.