Information bottleneck-driven Gaussian PID enables modality-level interpretability in multimodal survival prediction

Area of Science:

• Computational biology
• Bioinformatics
• Cancer research

Background:

• Multimodal survival analysis is crucial for precision oncology.
• Current interpretability methods struggle to differentiate shared, unique, or synergistic prognostic information from different data types.
• Partial Information Decomposition (PID) provides a theoretical basis but is challenging to apply to complex survival data.

Purpose of the Study:

• To introduce an information bottleneck-driven Gaussian PID (IB-GPID) framework for interpreting multimodal survival prediction.
• To enable modality-level interpretation of prognostic information.
• To quantify shared, unique, and synergistic molecular mechanisms in cancer prognosis.

Main Methods:

• Developed an information bottleneck-driven multimodal feature selection approach.
• Implemented continuous survival-target construction using Gaussianization.
• Applied bias-corrected Gaussian PID to decompose survival-predictive information into redundancy, uniqueness, and synergy.
• Benchmarked the framework on TCGA-BRCA and TCGA-LUSC, and validated across 17 TCGA cohorts.

Main Results:

• The IB-GPID framework achieved competitive survival prediction performance.
• Generated stable and biologically interpretable PID profiles.
• Demonstrated that redundancy and synergy, rather than unique information, predominantly encode cancer prognosis.
• Observed synergy in RNA-methylation and RNA-miRNA pairs, suggesting regulatory coordination.
• Found redundancy in RNA-RPPA and methylation-RPPA pairs, indicating shared pathway activity.

Conclusions:

• IB-GPID offers a principled and interpretable method for dissecting multimodal prognostic information.
• The findings highlight the importance of shared and synergistic molecular mechanisms in cancer.
• This framework advances our understanding of the complex interplay of molecular data in predicting patient outcomes.