Contrastive Alignment of Expression and Copy Number Highlights Dosage-Insensitive Genes in Cancer

Copy number variations (CNVs) are a hallmark of cancer genomes, yet the relationship between CNV and gene expression is not strictly deterministic. Some genes maintain stable expression despite copy number changes through regulatory compensation. Identifying these dosage-insensitive genes is challenging, requiring methods that distinguish true regulatory escape from technical noise in heterogeneous single-cell data. Here, we present a contrastive learning framework that learns a shared latent space aligning single-cell RNA-seq expression profiles with inferred CNV patterns. Our key innovation is hard negative mining: explicitly training on cell pairs with similar CNV but divergent expression patterns, which represent potential dosage insensitivity. By combining InfoNCE loss with hard negative triplet loss, we learn embeddings where expression-CNV distance quantifies regulatory concordance. We apply this framework to 10 lung adenocarcinoma patients (80k cells) from the GSE131907 atlas, classifying cancer cells as "concordant" (expression follows CNV) or "discordant" (expression escapes CNV). Differential expression analysis between these groups reveals two gene categories: escape genes upregulated in discordant cells despite CNV status, and compensation genes downregulated in discordant cells. Pooled analysis across 40,775 cancer cells identifies significant escape genes including VSIG4, FCGR1A, TREM2, and MARCO, as well as compensation genes such as MALAT1, CCL5, and CD8A. These genes represent candidate therapeutic targets and biomarker hypotheses for CNV-independent tumor behavior. Our approach provides a generalizable frame-work for discovering regulatory escape mechanisms in cancer using standard single-cell RNA-seq data.