Conditional similarity triplets enable covariate-informed representations of single-cell data

Area of Science:

• Computational Biology
• Immunology
• Machine Learning

Background:

• Single-cell technologies provide detailed immune cell profiling.
• Machine learning is used to create immunological summaries for diagnostics.
• Current methods predict only one outcome, ignoring other clinical data.

Purpose of the Study:

• To develop a novel machine learning approach for incorporating clinical covariates into immune signature analysis.
• To create per-sample encodings that reflect both immune profiles and clinical information.

Main Methods:

• Introduced CytoCoSet, a set-based encoding method.
• Formulated a loss function with a triplet term to penalize disparate embeddings for similar covariates.
• Optimized model parameters to integrate immune signatures and clinical data.

Main Results:

• CytoCoSet effectively incorporates measured covariates into per-sample encodings.
• The method learns featurizations that capture both immune and clinical information.
• Optimized encodings improve the prediction of clinical outcomes.

Conclusions:

• Integrating clinical covariates enhances the accuracy of per-sample encodings.
• This approach leads to more robust predictions of clinical phenotypes.
• The method has potential for improving diagnostic and treatment strategies.