Leveraging AI and Directional Phosphorylation Network for Drug Response Prediction In BRAF-Mutant Colorectal Cancer
Abstract
Purpose
Kinase-substrate interactions are inherently directional: kinases phosphorylate substrates, not vice versa, yet most network-based methods ignore this asymmetry, potentially discarding predictive signal.1 We investigated whether respecting biological directionality improves the prediction of targeted therapy response in BRAF-mutant metastatic colorectal cancer (mCRC). Using PhosphoAtlas, a curated kinase-substrate database that expanded 4-fold between 2016 (PA1) and 2023 (PA2), we evaluated: (1) whether directed networks outperform undirected networks, and (2) whether network expansion further improves topology-based feature selection.2
Methods
We analyzed kinase activity profiles from 44 patient-derived mCRC tumor samples: 24 untreated and 20 treated with Encorafenib plus Panitumumab (BRAF/EGFR inhibition). Site-level phosphorylation interactions were collapsed to gene-level edges (PA1: 1,996 edges; PA2: 8,306 edges). We constructed directed (kinase → substrate) and undirected (symmetric) networks, applying heat kernel diffusion to propagate kinase activity signals through network topology, allowing upstream kinase states to inform downstream substrate features. We compared three feature selection strategies: all 192 measured kinases, top 142 hubs plus 1-hop neighbors, and top 105 hubs plus 1-hop neighbors. Classification used Random Forest with leave-one-out cross-validation (n=44), reporting accuracy and AUC.
Results
Directed networks outperformed undirected versions, achieving 95.5% accuracy (AUC=0.990) compared to 93.2% (AUC=0.982) using all 192 kinases. This +2.3% improvement confirms that biological directionality encodes predictive information lost in symmetric representations. Network expansion further enhanced topology-based feature selection: PA2 achieved 93.2% accuracy (AUC=0.985) versus PA1 at 88.6% (AUC=0.939), a +4.6% gain.
Conclusion
Respecting kinase-substrate directionality yields 95.5% accuracy in predicting Encorafenib-Panitumumab response, outperforming symmetric phosphorylation network representations. Because kinase signaling also mediates radiation response, this directionality-aware framework may extend to predicting radiosensitivity in mCRC and other kinase-driven malignancies.