Comprehensive profiling of mutational signatures and machine learning and subtypes of homologous recombination deficiency.

• DOI: 10.1200/jco.2023.41.16_suppl.568
• Publication Date: 2023-06-04T14:02:11Z
• Authors: Joonoh Lim; Seongyeol Park; Ryul Kim; Baek-Lok Oh; Sangmoon Lee; Jeong Seok Lee; Ji In Ryu; Jai Min Ryu; Se-Kyung Lee; Byung-Joo Chae; Jeongmin Lee; Ji-Yeon Kim; Yeon Hee Park; Young Seok Ju
• Keywords: PALB2; Subtyping; CHEK2; Indel

568 Background: Recently, genomic features have proven effective at gene-agnostic detection of homologous recombination deficiency (HRD). However, it remains to be explored what extent we can exploit genome analysis assess HRD status. Here, show that a machine learning (ML) classifier based on mutational signatures enables robust and subtyping outperforms current state-of-the-art classifier. Methods: We whole-genome-sequenced (WGS) ~700 breast cancers identified pathogenic variants in HR-related genes ( BRCA1/2, PALB2, CHEK2, RAD51B, ATM, etc.). Mutational all variant types, single-base substitution (SBS), indel (ID), structural (SV) copy-number (CN), were included as features. Biallelic inactivation BRCA1/2 by germline somatic loss-of-heterozygosity (LOH) was regarded true HRD. trained four different classifiers with n-fold cross validation used averaged scores for prediction. In addition, performed cluster subgroup within the predicted cases unveil heterogeneity Results: total 88 (12%) carriers LOH, including 29 BRCA1 (4%), 28 BRCA2 4 PALB2 (0.6%), 3 RAD51B (0.4%). Among them, 70 (80%) classified positives. As expected, most (28/29) (28/28) HRD-positive. All (4/4) (3/3) also HRD-positive, confirming their importance HR. Inclusion biallelic mutants (10; 1.4%) reached statistical significance enrichment (Fisher’s; P = 0.045). 171 (24%) having least one 57 PTEN (33%), 40 CDK12 (23%), 17 (10%), ARID1A 15 BRIP1 (9%). 72 these (42%) Our 12 (7%) more than HRDetect (167 vs. 155; F1 0.99 0.95) using signature ID6, SV3 (or RS3), LST, SBS3. Further, correctly (33/34) (27/28) from each other (F1 0.97), RS3, RS1, SBS37. Features similar BRCA1, BRCA2, but our able distinguish combination CN11, ID11, CN8, SBS40 ~1). Total 112 (67%), 48 (29%), 7 (4%) BRCA1-like, BRCA2-like, BRCA1/2-like, among which 47 (28%) monoallelic HR-gene 27 (16%) wild types. results suggest identify ~40% (47/120) mutation profile alone. Conclusions: believe method developed here, ML detect classify HRD, will benefit larger WGS data patients who HRD-related therapeutics.