A convolutional attention model classifies copy number variants from whole exome sequencing

Copy number variants are important biomarkers in genetic disease and cancer, yet whole-exome CNV callers often rely on read-depth heuristics that capture limited positional or chromosomal context and generalize poorly across platforms. We present a dual-input convolutional neural network with attention that ingests normalized read depth, genomic coordinates, and chromosome identity. The model was pretrained on ECOLE-labeled 1000 Genomes data and fine-tuned on seven expert-annotated samples. On a held-out test set, the method achieved macro F1 = 0.83 and macro PR-AUC = 0.93. In additional contextual analyses reported in the Supplementary Material, CNN-Att exhibits a sensitivity-precision trade-off consistent with established WES CNV callers. Cross-platform evaluations on HiSeq 4000, NovaSeq 6000, MGISEQ 2000, and BGISEQ 500 yielded overall F1 up to 0.96. Fine-tuning increased deletion and duplication recall at the cost of a higher false positive rate, reflecting an explicit trade-off between sensitivity and precision. The architecture masks padded depth tokens and uses attention to highlight weak depth signals that are characteristic of small or noisy events. These results indicate strong sensitivity and robust performance across sequencing technologies, supporting use in clinical triage, multi-site genomics, and large-scale screening.