MS-ConTab: multi-scale contrastive learning of mutation signatures for Pan-Cancer representation and stratification

Area of Science:

• Computational Biology
• Genomics
• Machine Learning

Background:

• Understanding pan-cancer mutational landscapes is crucial for deciphering tumorigenesis.
• Cohort-level cancer clustering traditionally relies on statistical methods, limiting advanced analyses.

Purpose of the Study:

• To develop a novel unsupervised contrastive learning framework for cohort-level cancer type clustering.
• To group 43 cancer types based on coding mutation data from the COSMIC database.

Main Methods:

• Constructed dual mutation signatures: gene-level and chromosome-level profiles.
• Employed TabNet encoders and multi-scale contrastive learning (NT-Xent loss) for unified embeddings.
• Applied the framework to coding mutation data from the COSMIC database.

Main Results:

• Achieved biologically meaningful clusters of cancer types based on learned latent representations.
• Clusters align with known mutational processes and tissue origins.
• Demonstrated the framework's ability to reveal mutation-driven cancer subtypes.

Conclusions:

• This study presents the first application of contrastive learning for cohort-level cancer clustering.
• The proposed framework offers a scalable and interpretable approach for cancer subtyping.
• The findings provide new insights into the molecular mechanisms underlying cancer heterogeneity.