mSigHdp: hierarchical Dirichlet process mixture modeling for mutational signature discovery

Area of Science:

• Genomics and Bioinformatics
• Cancer Research
• Computational Biology

Background:

• Mutational signatures reveal patterns of DNA damage from endogenous or exogenous mutational processes.
• Current methods for signature discovery primarily rely on non-negative matrix factorization (NMF).
• Hierarchical Dirichlet Process (HDP) mixture models offer an alternative, less-explored approach for signature discovery.

Purpose of the Study:

• To introduce mSigHdp, an enhanced method utilizing HDP mixture models for discovering mutational signatures.
• To compare the performance of mSigHdp against state-of-the-art NMF-based methods.
• To evaluate the accuracy and efficiency of mSigHdp in identifying various mutation types and signatures.

Main Methods:

• Development and implementation of mSigHdp, an improved HDP mixture model approach.
• Benchmarking mSigHdp against NMF-based signature discovery tools using four synthetic datasets.
• Analysis of large-scale mutation data, including 3.5 × 10^7 single-base substitutions and 6.1 × 10^6 small insertion/deletion mutations across 18 cancer types.

Main Results:

• mSigHdp demonstrated superior performance in discovering mutational signatures compared to NMF methods across synthetic datasets.
• The method achieved the highest positive predictive value for signature discovery in three out of four datasets.
• mSigHdp exhibited the best true positive rate for signature discovery across all four tested datasets, with comparable CPU usage to NMF approaches.

Conclusions:

• mSigHdp represents a significant advancement in the computational tools available for mutational signature analysis.
• The HDP-based approach offers improved accuracy and reliability for identifying mutational processes in cancer genomics.
• mSigHdp is a practical and effective tool for researchers studying the etiology and mechanisms of cancer development.