A pan-cancer somatic mutation embedding using autoencoders.

Area of Science:

• Computational Biology
• Genomics
• Machine Learning in Oncology

Background:

• Next-generation sequencing generates vast amounts of cancer genome data, necessitating advanced computational methods for interpretation.
• Machine learning approaches are increasingly employed to unravel the complexity of cancer disease using large tumor datasets.
• High-dimensional tumor samples with germline and somatic mutation data require sophisticated computational modeling.

Purpose of the Study:

• To develop and apply a neural network-based pipeline for analyzing complex somatic mutation data in cancer.
• To discover lower-dimensional representations of tumor mutation profiles for enhanced data interpretation.
• To accurately classify tumor subtypes using machine learning models applied to mutation data.

Main Methods:

• An autoencoder model was utilized to learn compressed representations (embeddings) of somatic mutation data.
• Kernel learning and hierarchical cluster analysis were employed to evaluate the quality of the learned embeddings.
• Support vector machine models were trained on the embeddings for tumor subtype classification.

Main Results:

• The autoencoder successfully generated lower-dimensional representations of somatic mutation data from 40 tumor types and subtypes.
• The learned embeddings preserved significant biological signals from the original high-dimensional data.
• Support vector machine models achieved accurate classification of tumor subtypes based on the reduced-dimension embeddings.

Conclusions:

• The developed pipeline effectively maps high-dimensional tumor mutation data into a lower-dimensional latent space.
• The resulting embeddings facilitate easier exploration of intra- and inter-tumor heterogeneity.
• This approach enables accurate classification of tumor samples within the pan-cancer somatic mutation landscape.