TF-GateNet: An Interpretable and Biologically Guided Framework for Primary-Metastatic State Prediction from Somatic Genomic Alterations.

Metastasis remains a major cause of cancer mortality, making reliable primary-metastatic state prediction from somatic genomic alterations clinically important yet technically difficult. We present TF-GateNet, a biologically constrained neural network that combines TF-aware feature integration based on TRRUST and DoRothEA TF-gene regulatory priors with sample-specific dynamic gating on a Reactome-defined hierarchical sparse backbone. The model was evaluated on multi-center prostate and breast-cancer cohorts using mutation and copy-number features across 10 repeated runs on a fixed 80/10/10 split, together with independent prostate external validation, and was compared with biologically informed neural-network baselines (P-NET, BKGNet-Pathway, and BKGNet-Protein), a dense feed-forward neural network (FNN), and conventional machine-learning baselines (LR, SVM, RF, DT, and XGBoost). On prostate, TF-GateNet achieved the best internal performance (AUROC 0.954 ± 0.005; AUPRC 0.925 ± 0.007) and the best combined external performance (AUROC 0.952 ± 0.009; AUPRC 0.898 ± 0.018). On breast, TF-GateNet achieved the strongest internal ranking performance, reaching AUROC 0.893 ± 0.004 and AUPRC 0.835 ± 0.006. Ablation analysis indicated that TF-aware integration accounted for the larger prostate gain, whereas within the TF-GateNet family on breast, both TF-aware integration and dynamic gating contributed positively. Interpretability analysis further supported a cross-level route from TF-related genomic perturbation cues to genes, pathways, and phenotype-associated predictions. These results position TF-GateNet as a biologically grounded and interpretable framework for primary-metastatic state prediction, with the strongest overall evidence in prostate cancer and favorable internal evidence in breast cancer.