Mut2Vec: distributed representation of cancerous mutations.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Embedding techniques convert high-dimensional sparse data into low-dimensional distributed representations, widely used in deep learning.
• Existing methods lack effective embedding for sparse mutation profiles, limiting applications like novel driver mutation discovery.
• Binary representations of mutations fail to capture biological context, hindering advanced analysis.

Purpose of the Study:

• To introduce Mut2Vec, a novel computational pipeline for generating distributed representations of cancerous mutations.
• To address the challenge of limited biological data for training mutation embeddings.
• To enhance the biological context captured by mutation representations for improved downstream applications.

Main Methods:

• Mut2Vec utilizes the Skip-Gram model trained on cancer profiles, leveraging co-occurring mutations.
• The pipeline integrates biomedical literature and protein-protein interaction networks to augment sparse biological data.
• This approach compensates for data insufficiency inherent in training mutation embeddings.

Main Results:

• Mut2Vec successfully generated distributed representations for mutations, enabling clear visualization distinguishing driver and passenger mutations.
• The study identified novel driver mutation candidates using a clustering method, validated through literature surveys.
• Pre-trained mutation vectors and candidate driver mutations are publicly accessible for further research.

Conclusions:

• Mut2Vec provides an effective method for generating distributed mutation representations, validated by experimental tasks.
• The generated mutation representations can be applied to various deep learning models for cancer classification and drug sensitivity prediction.
• This work advances the use of embedding techniques in cancer genomics and precision medicine.