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

Area of Science:

• Computational Biology
• Genomics
• Cancer Research

Background:

• Metastasis is a primary driver of cancer mortality.
• Predicting the primary-metastatic state from genomic alterations is crucial but challenging.
• Existing methods lack biological grounding and interpretability.

Purpose of the Study:

• To develop a biologically constrained neural network, TF-GateNet, for accurate primary-metastatic state prediction.
• To integrate transcription factor (TF)-gene regulatory priors and dynamic gating for improved model performance.
• To evaluate TF-GateNet's efficacy and interpretability in prostate and breast cancer cohorts.

Main Methods:

• Developed TF-GateNet, a neural network incorporating TRRUST and DoRothEA TF-gene regulatory networks and Reactome pathways.
• Utilized mutation and copy-number data from multi-center prostate and breast cancer cohorts.
• Compared TF-GateNet against various baseline models including biologically informed networks and conventional machine learning algorithms.

Main Results:

• TF-GateNet achieved superior internal and external validation performance in prostate cancer (AUROC 0.954, AUPRC 0.925 internally; AUROC 0.952, AUPRC 0.898 externally).
• TF-GateNet demonstrated strong internal ranking performance in breast cancer (AUROC 0.893, AUPRC 0.835).
• Ablation studies confirmed the contribution of TF-aware integration and dynamic gating to model performance, with interpretability analysis revealing cross-level biological insights.

Conclusions:

• TF-GateNet provides a biologically grounded and interpretable framework for primary-metastatic state prediction.
• The model shows significant promise, particularly in prostate cancer, and favorable performance in breast cancer.
• TF-GateNet advances computational oncology by linking genomic perturbations to clinical phenotypes.