araCNA: somatic copy number profiling using long-range sequence models

Area of Science:

• Genomics
• Computational Biology
• Cancer Research

Background:

• Somatic copy number alterations (CNAs) are key indicators in cancer development.
• Existing computational methods for CNA detection from whole-genome sequencing (WGS) data face scalability challenges with deep learning.

Purpose of the Study:

• To introduce araCNA, a novel deep learning approach for accurate CNA prediction from WGS data.
• To overcome computational limitations in deep learning for genomic-scale sequence analysis.

Main Methods:

• Developed araCNA, a deep learning model utilizing transformer alternatives like Mamba for long-range genomic interactions.
• Trained araCNA exclusively on simulated WGS cancer genome data.
• Employed a zero-shot learning approach for application to real cancer WGS samples.

Main Results:

• Achieved high accuracy in predicting CNAs on simulated data.
• Demonstrated performance comparable to existing methods on 50 Cancer Genome Atlas WGS samples.
• Required only tumor samples, not matched normal samples, for analysis.
• Showcased rapid inference times (minutes) and reduced overfitting markers.

Conclusions:

• araCNA effectively leverages simulated data and modern machine learning for biological applications.
• The approach offers a computationally efficient and accurate method for CNA detection in cancer WGS.
• Domain knowledge integration in simulation is key to harnessing deep learning for genomic analysis.