AcornHRD algorithm construction
AcornHRD was based on the results of sequencing depth ratio as input data to estimate LCNA events on the genome. For those samples of low sequencing depth, a fitness window size seems particularly important. To address the question of optimal window size, we adopted seven different window sizes (40kb, 100kb, 300kb, 500kb, 800kb, 1Mb and 1.4Mb) to estimate the number of LCNA based on the samples from cohort I. For each sample, the mode of the count of LCNA in different window sizes was calculated. There were up to all samples in every window size. The statistical results showed that the 100kb window covered the most 31 (75.6%) samples (Figure 2A). Besides, the 500kb-window size had a close value (73.2%) regarding the number of samples. To validate the stability of the 100kb-window size, breast cancer samples from the TCGA cohort were utilized to identify LCNA events. As expected, the number of samples covered by the 100kb-window size still remained the maximum number (Figure 2B). Therefore, the 100kb-window size had good stability and was defined as the optimal window size for the following analysis.
Mutations in the BRCA are strongly associated with HRD positive [23, 24]. To verify AcornHRD sensitivity, the patients with BRCA mutation from cohort II were constructed into a test panel with a 100kb-window size and 50kb-step size. For a 95% confidence detection rate, the score of 10 was defined as the cut-off threshold value. Of 55 tumors, 53 (96.4%) were determined to be HRD positive with a score greater than or equal to 10 (Supplementary Table 3). The result of 95% sensitivity demonstrated a good degree of credibility.
Comparing HRD status between different algorithms
For a more comprehensive evaluation of the algorithm, ShallowHRD software [25] was joined into the later comparative analysis. Six standard samples were sequenced with whole genomic DNA as described in the method section. The sensitivity and specificity of AcornHRD are both 100%, while those of the ShallowHRD are 40% and 100%, respectively (Table 1). Subsequently, the clinical cohort including 15 BRCA-positive and 35 BRCA-negative patients from the clinical cohort, was used to detect HRD status. The HRD results of the clinical cohort showed that AcornHRD sensitivity is far superior to ShallowHRD, but its specificity is not as good (Supplementary Table 4 and Supplementary Table 5).
Correlation between BRCA mutations and HRD status
Variation located in the BRCA as a prominent hallmark had been studied in many cancers [26-30]. As previous studies have reported, BRCA mutations are the most critical factors in patients with breast cancer to assess risk [31, 32]. HRD, another tumor biomarker, is being used in guiding therapy in more and more studies [33-36]. BRCA mutations are known to be strongly associated with HRD; therefore, we applied BRCA mutation status to compare the accuracy of the two HRD assessment methods, TCGA cohort (2 samples without somatic mutation information were filtered out) was analyzed using AcornHRD and ShallowHRD software [25], respectively. Firstly, the mutations of BRCA genes were confirmed in tumor sequencing reads by in-house calling variation pipeline (more details presented in Methods Part). Of the 120 patients, 20 harbored BRCA mutations. Subsequently, the HRD status of 120 tumor tissue samples was identified by two algorithms. The results of AcornHRD (Table 2) and ShallowHRD (Table 3) both show that BRCA status is significantly correlated with HRD status (P=0.00838 and P=0.00284, respectively). However, the positive agreement rate of AcornHRD is higher than the ShallowHRD algorithm, which is 70% (14/20) and 60% (12/20), respectively (Table 2 and Table 3). In summary, AcornHRD is more stable in the application performance of three different cohort of WGS data, which is superior to the published algorithm.
The HRD in clinical cohort
High HRD score was also significantly correlated with BRCA mutation (Table 4) in the 50 clinical cohort. Among the 50 patients who received anthracycline-based neoadjuvant therapy, 28 had high HRD scores, and 22 had low HRD scores. The breast cancer samples selected for clinical study were all HER-2 negative, including 24 TNBC samples and 26 ER and/or PR positive samples. High HRD score was significantly correlated with TNBC and high Ki-67 expression (Table 4). HRD score high trend to ER negative and PR negative (Table 4).
Correlation between HRD status and NAC efficacy
In this study, HRD positive includes either a high HRD score or a BRCA mutation, whereas HRD negative includes a low HRD score and no BRCA mutation. Of the 50 patients, 30 were identified as HRD positive. Moreover, pCR and residual tumor burden (RCB 0/1) are both important indicators for tumor efficacy evaluation, of which pCR (RCB 0) is the main evaluation indicator.
Regarding the HRD status of all patients (n = 50), patients with HRD positive were more likely to respond to standard NAC containing anthracyclines than HRD negative patients, with a pCR (RCB 0) outcome (OR = 9.5, 95% CI 1.11 - 81.5, p = 0.04) (Table 5). Similar results were observed for the endpoint RCB of 0/1. In the entire cohort of 50 patients, patients with HRD positive were more likely to achieve RCB 0/1 compared to non-deficient patients (OR = 10.29, 95% CI 2.02 - 52.36, p = 0.005) (Table 6). This applied to a cohort of 35 patients lacking germline BRCA mutations; patients with HRD trended toward an RCB 0/1 response compared with HRD negative patients (OR = 6.0, 95% CI 1.00 - 35.91, p = 0.05) (Table 6).