Cancer molecular subtyping using limited multi-omics data with missingness.

Area of Science:

• Oncology
• Bioinformatics
• Computational Biology

Background:

• Accurate cancer subtype diagnosis is crucial for effective treatment selection.
• Current multi-omics data fusion methods require extensive complete datasets, which are difficult to obtain clinically.
• Limited clinical samples and incomplete multi-omics data pose a significant challenge for developing robust diagnostic models.

Purpose of the Study:

• To develop a flexible integrative model, CancerSD, for diagnosing cancer subtypes using limited samples with incomplete multi-omics data.
• To address the data scarcity and incompleteness issues in clinical multi-omics data for cancer diagnosis.
• To improve the accuracy, authenticity, and interpretability of cancer subtype diagnosis.

Main Methods:

• Proposed CancerSD, a flexible integrative model incorporating contrastive learning and masking-and-reconstruction tasks for reliable omics imputation.
• Fused available and imputed omics data for accurate cancer subtype diagnosis.
• Extended meta-learning with a category-level contrastive loss to effectively transfer knowledge from external datasets for model pretraining, addressing limited clinical samples.

Main Results:

• CancerSD demonstrated accurate cancer subtype diagnosis on benchmark datasets.
• The model maintained high authenticity and interpretability in its diagnostic predictions.
• Identified key molecular characteristics associated with cancer subtypes and defined an Integrated CancerSD Score for patient prognosis prediction.

Conclusions:

• CancerSD offers a robust solution for cancer subtype diagnosis with limited and incomplete multi-omics data.
• The model's ability to impute missing data and leverage meta-learning enhances diagnostic accuracy and generalizability.
• The Integrated CancerSD Score provides a valuable independent predictive factor for patient prognosis, aiding clinical decision-making.