CancerRCDPredictor: A Precision Oncology Prediction Platform Driven by Multi-Omics Super Learner

CancerRCDPredictor is a cancer regulated cell death prediction platform based on multi-omics data and super learner architecture. It integrates seven types of molecular-level data and explainable AI technologies to provide a transparent and auditable clinical decision support tool for precision oncology. Developed by the BioCancerInformatics team, this platform is open-sourced on GitHub and was released on May 28, 2026.

In the transition of cancer treatment to precision medicine, the assumptions of traditional Cox proportional hazards models often fail, and single-omics data cannot capture the full picture of cancer heterogeneity. Tumor genomic data faces challenges such as sparsity, high dimensionality, and nonlinear survival structures. Integrating heterogeneous data sources while maintaining model interpretability is a core issue.

CancerRCDPredictor adopts an R Shiny interactive design and is based on the Pan-Cancer Multi-Omic SuperLearner architecture. It integrates seven types of omics data from 33 tumor types: protein abundance, somatic mutations, copy number variations, miRNA expression, transcript isoform-specific expression, mRNA expression, and DNA methylation. The data is standardized and encoded through an 11-part tokenized naming system.

The core is the Multi-View Elastic Net SuperLearner (MVL), which dynamically integrates four base learners: Random Survival Forests, XGBoost, Survival-Boruta, and Multi-Task Logistic Regression. Through a four-fold validation framework, 150 "Golden Anchors" features are selected. It is found that continuous phenotypic layers (transcript isoforms, mRNA) dominate the prediction topology, and the retention rate of static genomic variations is 0%.

For missing values, 12 imputation methods and 372 lineage-specific coordinated multi-omics matrices are deployed. The Dual-Track Inference Engine is adopted: Path A generates continuous risk Z-scores for structurally complete records, and Path B routes fragmented records to XGBoost fallback, ensuring a 100% prediction penetration rate.

It integrates SHAP Beeswarm plots, LIME surrogate models, and TreeSHAP visualization tools to map 26,800 cross-feature dependencies. It follows a three-stage audit compliance architecture and calibrates prediction probabilities through IPCW and Time-Dependent Brier Scores to ensure clinical reliability.

It performed robustly in a clinical blind validation cohort of 1050 patients, especially in high-entropy environments such as low-grade glioma (LGG), overcoming the limitations of traditional single-biomarker prediction.

CancerRCDPredictor represents a new generation paradigm of AI tools for precision oncology, revealing key findings such as continuous phenotypic layers dominating survival prediction. It has both educational value (teaching sandbox) and benchmark significance for clinical translation, providing tools and research directions for computational biology researchers and clinical oncologists.