Welcome to CancerRCDPredictor. This platform integrates 96 validated predictive models across 33 TCGA Pan-Cancer cohorts, spanning the OS, DSS, PFI, and DFI survival endpoints, through a deterministically mapped, audit-compliant SuperLearner architecture coupled with advanced AI-driven biological synthesis.
Operating not only as a predictive diagnostic engine but also as an interactive Digital Molecular Tumor Board, the application translates high-dimensional multi-omic geometries into precision oncology insights. It shifts the analytical paradigm from population-level risk stratification toward individualized biological risk mapping, by interrogating precomputed survival topologies to generate high-fidelity, manuscript-grade interpretations.
At the core of the framework is an extensive machine-learning pipeline specifically designed to overcome the limitations of conventional linear proportional-hazards models. Through a non-linear Quadripartite ML Ensemble composed of Random Survival Forests (RSF), XGBoost, insulated Survival-Boruta feature selection, and Multi-Task Logistic Regression (MTLR), combined through an Elastic Net Multi-View Meta-Learner (MVL), the system evaluates more than 12,613 multi-omic prognostic signatures encompassing 25 forms of Regulated Cell Death (RCD) across seven distinct omic layers. Rather than relying on simple additive prognostic markers, this architecture isolates the predictive contribution of complex non-linear biological geometries. By extracting exact N-dimensional TreeSHAP interactions, the platform exposes the specific lethal and protective prognostic trajectories contributing to each patient's predicted outcome. These topological representations are subsequently interpreted by the integrated LLM engine through dynamic tumor-state reasoning. The system translates mathematical vulnerabilities, tumor stemness metrics (TSM), tumor mutational burden (TMB), microsatellite instability (MSI), pharmacogenomic associations, and multi-omic biological signals into a comprehensive translational molecular report designed to support precision oncology interpretation.
Importantly, all biological interpretations generated by the LLM are constrained by a multi-layer governance framework incorporating evidence-tiering, audit validation, uncertainty communication, ecological-fallacy prevention, hypothesis framing, and domain-boundary controls to ensure that patient-level conclusions remain strictly grounded in the available computational evidence.
Machine Learning survival geometries (like SHAP and LIME) can be dense. This section serves as a static guide to interpreting penalty vs. protection axes.
1. The SHAP BeeSwarm
A right-sided (positive) SHAP value indicates a lethality driver (increased non-proportional hazard). A left-sided (negative) value indicates a protective shield (decreased non-proportional hazard).
2. SHAP Topologies (Synergy, Antagonism, & Bifurcation)
By plotting two interacting genes against their SHAP values, we decode their biological synergy, antagonism, or bifurcation across the clinical non-proportional hazard domain.
3. Individual Patient Trajectories (Precision Oncology)
SHAP Waterfall/Force Plots decompile the predictive logic for a single patient against a population baseline. The visualized bars directly represent specific multi-omic signatures (both continuous expression vectors and discrete genomic states like mutations/CNVs) acting as vectors of non-proportional hazard or protection. The predictive weight of each signature is defined by its breadth (width), orientation along the axis, and its top-to-bottom sequence in the trajectory.
Orange bars represent omic signatures imposing aggressive non-proportional hazard penalties (lethality), while purple bars represent signatures forcing deep negative non-proportional hazard pushes (protective shields).
4. The Digital Molecular Tumor Board
The Digital Molecular Tumor Board acts as a multi-agent AI system designed for deep, context-aware precision oncology. It integrates a patient's geometric trajectory (SHAP) with pharmacogenomic vulnerability profiling, allowing users to dynamically interrogate the model. The AI synthesizes multi-omic signatures, correlates them with tumor stemness (TSM), and proposes targeted therapeutic interventions while providing an interactive dialogue for hypothesis testing.
Phase I: Harmonization
Raw multi-omic inputs strictly harmonized and dimensionally mapped via LiSHMOM logic.
Phase II: CANARY Protocol
Cohorts violating Proportional Hazards (PH) geometry structurally quarantined via CoxNet.
Phase III: Ensemble Synthesis
Topologies decoded using a Quadripartite Ensemble synthesized by a Multi-View ElasticNet SuperLearner.
Select a specific machine learning algorithmic framework to instantly project its specific mathematical Importance and Validation payload across the Golden 150 biological signatures.
Explore the Time-dependent ROC (AUROC) horizons for the 96 finalized models.
Figure 9: Dual TimeROC
HOW TO READ: AUROC PERFORMANCE
The X-axis represents False Positive Rate (1-Specificity), while the Y-axis tracks True Positive Rate (Sensitivity). The colored curves represent the discriminatory capability (AUC) of the MVL framework across three time horizons (1-year, 3-year, 5-year). The dotted diagonal represents random effect (AUC 0.50).
(A) Lush Multi-Omic Prognostic Stability (LGG DSS): Testing the algorithm's decentralized quad-core resilience, this panel demonstrates the framework generating and maintaining a high plateau of clinical discrimination across a deeply fragmented multi-omic terrain. By democratically balancing 25.0% trust across all four learning architectures, the SuperLearner successfully flattens prognostic entropy over time (maintaining an AUC of 0.931 at 1-year, 0.900 at 3-year, and 0.802 at 5-year progression nodes). This prevents the chronological predictive degradation typically seen in Lower Grade Gliomas.
(B) Supreme Algorithmic Convergence (READ OS): Powered by a near-total 95.7% sparsity-aware XGBoost hierarchy, this panel maps the temporal diagnostic trajectory when continuous omic parameters align perfectly into deterministic geometric axes. The framework achieves flawless instantaneous prognostic authority (AUC = 1.000 at 1-year) and successfully anchors a virtually impenetrable predictive barrier (AUC = 0.996) out to the 3-year tracking horizon, before encountering expected entropy drop-offs at extended boundaries (AUC = 0.842 at 5-year).
Explore the macro-level non-proportional hazard drivers isolated across the 96 finalized prognostic models.
Pedagogical Exemplars
Select a specific cohort exemplar from the manuscript to explore its macro-level non-proportional hazard drivers.
Explore specific multi-omic interactions mapping clinical non-proportional hazard interception at the cohort level.
Figure 8: Interaction Topologies
HOW TO READ: PANELS A, B, C
(A) Synergy (LUAD DSS): The partner mutation acts as a potent catalyst, violently accelerating the patient cloud upward into extreme lethality as the primary driver increases across the x-axis.
(B) Antagonism (LUAD DSS): The partner mutation acts as a functional buffer, physically rescuing the patient cloud by forcing the mortality trajectory back down into the protective zone.
(C) Bifurcation (SKCM OS): The primary driver dictates the absolute clinical hemisphere (x-axis), while the distinct modulating partner stratifies the internal density of the clouds.
This module isolates the 150 elite biological signatures (Table S12) that were universally retained by all four Machine Learning algorithms (RSF, XGBoost, Boruta, MTLR).
Explore specific multi-omic interactions mapping clinical non-proportional hazard interception at the individual patient level.
Clinical Probabilities: Analyze dynamic patient-specific survival and event probabilities across sequential 1, 3, and 5-year clinical landmarks using the Phase III algorithmic ensemble.
Non-Proportional Hazard Trajectory Selector
Select a patient trajectory to singularize the signatures driving the prediction toward lethality or pulling it back to safety.
CancerRCDPredictor Diagnostic Report (Precision Oncology)
Model Selection
CancerRCDPredictor Diagnostic Report (Precision Oncology)
Navigate the completely sequestered internal validation cohort. View the Dual-Track continuous algorithmic trajectories mapping 1, 3, and 5-year clinical probabilities.
A searchable datatable containing all prognostic signatures, categorized by their biological omic layer origin.
This engine converts the 12,613 high-dimensional prognostic signatures (Table S11) into human-readable clinical diagnostic reports.
Translate the mathematical trajectory (SHAP) of a specific patient into clinical vulnerability analysis.
Interrogate the Tumor Board AI on specific topological interactions or patient trajectory vulnerabilities.
Explore the video and audio presentations covering the core concepts and findings of the study.
Podcast
Why AI Ignores DNA in Cancer Survival: A Multi-Omic Paradox
Video Presentation
Insight Engine: Bypassing the Cox-Proportional Collapse in Personalized Survival Topologies
Navigate and download the supporting raw datasets, trained model bundles, and analytical matrices directly from the ZIMA server architecture.
Directory Contents
The comprehensive reference table of the 25 recognized forms of regulated cell death.
Open RCDome in New TabClick the button above to open the full reference table in a new browser tab.
If you use this prediction tool or the underlying ML methodologies in your research, please cite our manuscript:
@article {Rodrigues de Souza2026.05.29.728842,
author = {Rodrigues de Souza, Emanuell and Almeida Cordeiro Nogueira, Higor and dos Santos Lopes, Victor and Medina-Acosta, Enrique},
title = {A Pan-Cancer Multi-Omic SuperLearner for Regulated Cell Death Survival Topologies},
elocation-id = {2026.05.29.728842},
year = {2026},
doi = {10.64898/2026.05.29.728842},
publisher = {Cold Spring Harbor Laboratory},
URL = {https://www.biorxiv.org/content/early/2026/06/02/2026.05.29.728842},
eprint = {https://www.biorxiv.org/content/early/2026/06/02/2026.05.29.728842.full.pdf},
journal = {bioRxiv}
}
Meet the team behind the Phase I-III computational pipelines.
