Robust Copy Number Alteration Detection (RCNA): A Novel Algorithm for Detecting Copy-Number Alterations for Targeted Gene Sequencing Experiments

Introduction: Targeted gene-panel sequencing is widely used in cancer genetics studies, including in the evaluation of patients suspected of having a myeloid neoplasm. In addition to single nucleotide variants, large structural variations such as copy number alterations (CNAs) are important for diagnosis and prognosis. However, detecting CNAs from targeted sequencing remains a challenge due to the limited number of sequenced positions, capture bias in different targeted regions, and the lack of appropriate germline control tissue. We describe robust copy number alteration detection (RCNA), a novel algorithm that is specially designed to detect gene-level CNAs in targeted gene-panel sequencing data for tumor samples without paired normal control tissue.

Methods: RCNA normalizes for total sample sequencing coverage, sample GC-content, and position-specific bias to estimate position-specific normal copy number reference ranges and normalized log₂ coverage ratios. Genes for which normalized coverage ratios for a certain percentage of bases falls outside the reference ranges are then predicted as deletion or amplification events depending on directionality. We trained and assessed performance for RCNA on a targeted myeloid gene sequencing panel data covering 96 genes (average sequencing depth of 1,264X) for 1,174 subjects suspected of having MDS that were enrolled in The NHLBI National Myelodysplastic Syndromes (MDS) Natural History Study. All patients in the study had central pathology review conducted and were assigned to MDS (n=299), myelodysplastic/myeloproliferative neoplasm (MDS/MPN (n=54), acute myeloid leukemia (AML) <30% blasts (n=17), idiopathic cytopenia of undetermined significance (ICUS) (n=44), clonal cytopenias of undetermined significance (CCUS) (n=306), other AML (n=31), or other diagnoses (n=423).

Results: Among 1,174 patients, 73 (6.2%) had an abnormal karyotype reported that overlapped a genomic region covered by the panel. We then used karyotype abnormalities overlapping 16 genes that occurred in ≥10 patients with >50% abnormal cells as the gold standard to train and determine optimal cutoffs for RCNA. Using partial cross-validation, we then determined optimal cut-offs for 1) normal reference ranges and 2) percentage of abnormal bases per gene that maximized the average F0.5 score for the 16 genes. The best performing combination achieved an average F0.5 score of 0.79, positive predictive value of 97%, specificity of 98%, and sensitivity of 55% for these genes. RCNA performance for different sample sizes and fraction of patients with normal vs. abnormal karyotypes showed robust results for smaller sample sizes (n=10) and similar performance metrics when subsampling up to 10% of patients with an abnormal karyotype per gene.

Next, we applied the optimal parameters to detect CNAs to the full cohort of 1,174 patients enrolled in the National MDS Study, including those in the training set. A deletion was observed in at least one of the 96 gene for 243 of 1,174 (20.7%) patients, and an amplification event was observed in at least one gene for 190 of 1,174 (16.2%) patients. In contrast, 73 (6.2%) had an abnormal karyotype reported (60 (5.1%) deletion and 34 (2.8%) amplification) indicating that RCNA was able to identify many additional putative CNAs. The proportion of patients with CNA events varied between different diagnostic groups. For example, patients diagnosed with MDS had 21.4% subjects with at least one amplified gene. Rates were higher for the AML (<30% blasts) and Other AML group for which 29.4% and 48.4% of subjects showed gene amplification events, respectively. Among MDS patients, 36.1% had at least one deleted gene and the most frequently lost genes were IRF1 (chr 5) (47, 15.7%), EZH2 (chr 7) (28, 9.4%), LUC7L2 and CUX1 (chr 7) (27, 9%) and the most frequently amplified genes were RAD21 (chr 8) (15, 5%), GNB1 (chr 1) (13, 4.3%), and MPL (chr 1) (10, 3.3%). The most statistically enriched CNA events within MDS vs. all other diagnostic categories were IRF1 (chr 5), EZH2 (chr 7), and ETV6 (chr 12) for deletion events and MPL (chr 1), CSF3R (chr 1), and RAD21 (chr 8) for amplification events (Fisher’s Exact test, all with an FDR-adjusted p-value <0.05).

Conclusions: We show that RCNA produced robust results for smaller sample sizes and detected biologically relevant CNA events for the MDS cohort. The software will be made available at https://github.com/emmesgit/RCNA.