TMO-Net: an explainable pretrained multi-omics model for multi-task learning in oncology.

Area of Science:

• Oncology
• Bioinformatics
• Artificial Intelligence

Background:

• Cancer involves complex, multi-scale systemic alterations.
• Integrating diverse biological data is crucial for understanding cancer.
• Existing models often struggle with incomplete multi-omics datasets.

Purpose of the Study:

• To develop a novel deep learning framework for integrating multi-omics data in cancer research.
• To enable joint representation learning and inference from incomplete omics datasets.
• To enhance the interpretability of multi-omics data in predicting clinical outcomes.

Main Methods:

• Development of the Tumor Multi-Omics pre-trained Network (TMO-Net).
• Integration of pan-cancer multi-omics datasets for model pre-training.
• Application of interpretable learning techniques to analyze omics feature contributions.

Main Results:

• TMO-Net effectively integrates multi-omics data, improving sample representation.
• The model facilitates cross-omics interactions and joint representation learning.
• Interpretable learning identified key omics features influencing clinical outcomes.

Conclusions:

• TMO-Net provides a versatile framework for cross-modal multi-omics learning in oncology.
• The model enhances the utility of incomplete multi-omics datasets for downstream tasks.
• This work paves the way for tumor omics-specific foundation models.