A comprehensive foundation model for generalizable cytogenetics in precision oncology with CHROMA

Area of Science:

• Genomics
• Artificial Intelligence
• Computational Biology

Background:

• Automated cytogenomic analysis faces limitations in scope, annotation, and robustness.
• Existing methods struggle with the complexity of real-world cytogenomic data.

Purpose of the Study:

• Introduce CHROMA, the first single-chromosome foundation model for comprehensive cytogenomic analysis.
• Enable unified, cell-level detection of a wide spectrum of chromosomal abnormalities, including rare types.

Main Methods:

• Developed CHROMA, a foundation model for cytogenomics, pre-trained on over 4 million chromosomal images using self-supervised learning.
• Utilized a unified framework for detecting numerical and structural chromosomal abnormalities.

Main Results:

• CHROMA achieves robust and comprehensive detection of chromosomal abnormalities across diverse classes.
• Reduced expert annotation workload by 40% through efficient label utilization.
• Maintained state-of-the-art accuracy on imbalanced data and challenging imaging conditions.

Conclusions:

• CHROMA offers a scalable, accessible, and precise solution for cytogenomic analysis.
• The model can be deployed as a risk-aware screening and triage tool, particularly where expert availability is limited.
• An integrated risk-control strategy ensures safe application by flagging uncertain cases for expert review.