2528 - Network-Informed Functional Kinase Activity Profiling in a BRAF-Mutant Colorectal Cancer Patient-Derived Xenograft Model
Presenter(s)
B. Karimi1, H. Lin2, C. Carroll3,4, and J. P. Coppé1; 1Department of Radiation Oncology, University of California, San Francisco, San Francisco, CA, 2Department of Radiation Oncology, University of California San Francisco, San Francisco, CA, 3Data Institute, University of San Francisco, San Francisco, CA, 4Department of Mathematics and Statistics, University of San Francisco, San Francisco, CA
Purpose/Objective(s): Adaptive reprogramming of kinase signaling cascades contributes to therapeutic resistance in cancer. While functional kinase activity measurements can capture these changes, methods integrating kinase activity with signaling network topology have been limited. We hypothesize that integrating kinase activity with signaling topology improves the stability and reproducibility of treatment-responsive kinase identification, enabling more reliable prioritization of resistance targets.
Materials/Methods: A BRAF-mutant colorectal cancer Patient-Derived Xenograft (PDX) model treated with BRAF+EGFR inhibition or vehicle control yielded 44 tumor samples profiled via High-Throughput Kinase Activity Mapping, measuring ATP consumption of 192 kinases. The kinase signaling graph was derived from PhosphoAtlas, a rigorously curated map of the human phospho-reactome. Classifiers (Ridge/ElasticNet Logistic Regression/Multi-Layer Perceptron) were fit and evaluated 1)using direct activity measurements (DAMs) only or 2)using DAMs augmented with embeddings created using kinase activity, PhosphoAtlas network topology, and proteinBERT amino acid sequence embeddings. Performance was evaluated using leave-one-mouse-out cross-validation to prevent leakage. Kinase importance was assessed via cross-fold ranking stability.
Results: Classifiers using network-projected features distinguished treated from untreated tumors (LR-Ridge: ACC 0.78 ± SE 0.2; AUC 0.8 ± 0.2, LR-EN: ACC 0.8 ± 0.2; AUC 0.8 ± 0.2, MLP: Accuracy (ACC) 0.82 ± 0.16; AUC 0.81 ± 0.22), with DAM-only models achieving comparable metrics but with higher variance (LR-Ridge: ACC 0.74 ± 0.33; AUC 0.88 ± 0.2, LR-EN: ACC 0.75 ± 0.30; AUC 0.86 ± 0.2, MLP: ACC 0.56 ± 0.16; AUC 0.56 ± 0.32), confirming that network features increase prediction consistency and stability. DAM-only classifier performance ranged from near-random to near-perfect across folds. Network-projected features yielded more stable kinase rankings, consistently identifying the same kinases as treatment-responsive regardless of hold-out mouse. Stable differentiators included CDK1, PRKACA, SRC, AKT1, & MAPK pathway members.
Conclusion: Network-informed kinase activity profiling produces more stable and reproducible identification of treatment-responsive kinases than DAMs alone in a BRAF-mutant colorectal cancer PDX model. This approach may improve prioritization of kinase targets for combination therapy.