Panel-Specific Optimization and Benchmarking of Copy-Number Callers for ERBB2 Amplification Detection Using the...

Background/Objectives: ERBB2 amplification is a clinically relevant breast-cancer biomarker, but CNV detection from targeted sequencing depends on assay design, normalization, tumor fraction, and decision rules. We evaluated panel-specific ERBB2 analysis configurations against a binary FISH reference, with emphasis on the distinction between development-set performance and internal validation. Methods: We retrospectively evaluated 126 FFPE breast-tumor specimens (59 FISH-positive and 67 FISH-negative; 123 recorded patient identifiers) sequenced using NanOnco Plus Panel v3.0 on DNBSEQ-G400. Seven tumor-only, panel-compatible CNV caller configurations were compared; no matched normal tissue was available, although panel-matched normal reference pools were used where required by a caller. PureCN was run as a standalone tumor-only workflow, and CNVkit was run independently. Original fixed-rule results were calculated on the development cohort. An additional common 5-fold, 20-repeat patient-grouped cross-validation procedure selected decision rules using training data only, maximizing sensitivity subject to training specificity ≥90%. Upstream caller outputs were held fixed. Technical no-calls were retained. Results: On the development cohort, the fixed CNVkit median-bin log2 rule yielded 76.3% sensitivity (Wilson 95% CI, 64.0-85.3%) and 91.0% specificity (81.8-95.8%), with F1 81.8% and κ = 0.679. Native standalone PureCN yielded 16.9% sensitivity and 95.5% specificity at absCN ≥ 14; PURPLE had three numerical no-calls. Under the common patient-grouped decision-calibration procedure, CNVkit yielded 67.8% sensitivity and 91.0% specificity, whereas native PureCN yielded 19.3% sensitivity and 91.2% specificity. Exploratory computational purity subgroups and descriptive profiles across 554 genes illustrated heterogeneity of the measured DNA copy-number signal; independent pathology cellularity was unavailable, while quantitative FISH fields were recovered only for 45 of 126 specimens and were not used to define the full-cohort benchmark. Conclusions: The results support assay-specific development of ERBB2 decision rules but do not establish an intrinsic ranking of CNV software or clinical validation of other panel genes. The original full-cohort metrics are development-set estimates, and the additional cross-validation is conditional on previously generated caller outputs. External validation with independently assessed tumor content and complete clinical HER2 adjudication remains necessary.