ImmuneNet: Composition-Aware Quantification of Adaptive Lymphocytes in High-Grade Serous Ovarian Cancer

Bulk RNA-seq mixes signals from tumor, stromal, and immune cells, which obscures the adaptive lymphocyte readouts needed for prognosis and treatment design. In high-grade serous ovarian cancer, recovering T and B cell fractions is particularly difficult due to marker leakage from malignant programs, cohort and platform mismatch between references and targets, and miscalibrated predictions. We present ImmuneNet, a context-aware estimator of T and B fractions from bulk RNA-seq. The pipeline builds a tumor-matched single-cell reference, curates a compact marker panel by removing genes with measured malignant leakage, and trains on pseudo-bulk mixtures that cover the composition simplex. The model uses branched, gene-wise attention to aggregate lineage evidence and a parallel ridge-regularized linear component on the same standardized panel; a fixed blend and post hoc calibration (temperature scaling and isotonic regression) yields well-calibrated fraction estimates. Evaluation emphasizes composition-aware criteria alongside standard correlations. On held-out mixtures, ImmuneNet attains Spearman 0.967 (T) and 0.991 (B), RMSE 0.049 (T) and 0.032 (B), and the lowest total variation distance 0.047 over (T,B,O), outperforming strong linear and PLS baselines on joint composition error while remaining robust and data efficient. On TCGA-OV, ImmuneNet correlates strongly with tumor-matched T and B module proxies. These design choices produce a practical and reliable deconvolution approach tailored to ovarian cancer and suitable for downstream analyses on bulk cohorts.