1214P - A novel machine learning based method to detect homozygous deletion of homologous recombination repair (HRR) genes in prostate cancer

Background

HRR gene alterations are enriched in prostate cancer (PC) and are associated with sensitivity to PARP inhibitors. Homozygous deletion (HD), which accounts for up to 15% of HRR gene alterations in PC, is an important type of pathogenic alteration. However, the reliable detection of HD is technically challenging. Here we introduced a novel machine learning based method for HD detection in formalin-fixed paraffin-embedded (FFPE) PC tissues.

Methods

HD/non-HD simulation events (∼30000) of varying tumor cellularity and fragment size were constructed using the sequencing data of matched tumor and wild type cell lines tested by AmoyDx HRD Complete NGS panel. Seven artificial features were calculated based on genomic segment depth and minor allele frequency of single nucleotide polymorphisms, tumor cellularity and tumor ploidy. The feature data were fed into the xgboost with 10-StratifiedKFold to generate HD recognition model (training :validation datasets=7:3). Analytical performance was assessed with reference cell lines: specificity with wildtype cell lines, sensitivity with gradient HD-positive cell lines (tumor cellularity: 20%, 30%, 40%, 50%) with varying DNA input (30ng, 50ng, 100ng), reproducibility of intra- and inter-runs, and anti-interference against hemoglobin, triglycerides, xylene, paclitaxel, and ethanol. Clinical performance was evaluated using FFPE PC tissues.

Results

The trained HD model achieved 99.8% accuracy in the in silico validation dataset. Analytical validation demonstrated: 100% specificity; 100% sensitivity at gene level with 30% tumor cellularity and at exon level with 40% tumor cellularity with 100ng DNA input; 100% reproducibility; strong anti-interference capability with all HD events detected and no false positives. 132 FFPE PC tissues showed high consistency with a second validated NGS assay (PPA 100%, NPA 98.4%, OPA 98.5%).

Conclusions

A machine learning based method was successfully developed and adopted in multiple NGS assays to detect HD of HRR genes in PC tissues. It provides a reliable approach to effectively identify PC patients with HRR HD.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Janssen.

Disclosure

J. Wang, P. Cheng, W. Xu, Z. Huang, X. Chen, S. Liu, Y. Guo, W. Shi, S. Yang: Financial Interests, Personal, Full or part-time Employment: Amoy Diagnostics. X. Ye, U. Singh, K. Bell, K. Urtishak, L. Luo, X. Lyu, L. Zhou: Financial Interests, Institutional, Full or part-time Employment: Janssen Research & Development.