Topical hidden genome: discovering latent cancer mutational topics using a Bayesian multilevel context-learning...

Area of Science:

• Genomics
• Computational Biology
• Cancer Research

Background:

• Inferring cancer-type specificities from ultra-rare, genome-wide somatic mutations is challenging due to high dimensionality and data sparsity.
• Existing statistical methods struggle with the scale and complexity of mutation data.

Purpose of the Study:

• To develop a novel framework for dimension reduction of mutation contexts using topic models.
• To enable interpretable, decorrelated meta-feature topics for cancer type characterization.
• To facilitate rigorous, scalable full Bayesian inference for ultra-rare mutation analysis.

Main Methods:

• Leveraging topic models from computational linguistics for dimension reduction.
• Developing an efficient Markov Chain Monte Carlo (MCMC) algorithm for Bayesian inference.
• Applying the model to the Pan Cancer Analysis of Whole Genomes (PCAWG) dataset.

Main Results:

• The model generates interpretable, decorrelated mutation context topics.
• Identified somatic mutational topics linked to UV exposure (skin cancer), aging (colorectal cancer), and epigenome organization (liver cancer).
• Achieved highly competitive predictive performance against random forest and deep learning methods.

Conclusions:

• The proposed framework effectively addresses the challenges of analyzing ultra-rare somatic mutations.
• The approach provides biological insights and demonstrates strong predictive capabilities in cancer type specificity.
• Enables scalable, rigorous Bayesian inference for large-scale genomic mutation data.