Conditional Similarity Triplets Enable Covariate-Informed Representations of Single-Cell Data

Area of Science:

• Immunology
• Computational Biology
• Machine Learning

Background:

• Single-cell technologies provide deep immune cell profiling.
• Machine learning translates immune data into diagnostic features.
• Current methods optimize solely on outcome variables, ignoring patient covariates.

Purpose of the Study:

• To develop a machine learning approach that incorporates clinical covariates for improved immune profiling.
• To enhance per-sample representations by considering patient-specific information.
• To improve the prediction of clinical phenotypes using integrated data.

Main Methods:

• Introduction of CytoCoSet, a set-based encoding method.
• Formulation of a loss function with an additional triplet term.
• Penalizing disparate embedding results for samples with similar covariates.

Main Results:

• Incorporating clinical covariates directly informs learned per-sample representations.
• CytoCoSet demonstrates improved prediction of clinical phenotypes.
• Enhanced featurizations lead to more robust diagnostic models.

Conclusions:

• Integrating clinical covariates into machine learning models significantly improves immune profiling.
• CytoCoSet offers a novel approach for creating informative per-sample representations.
• This method advances the translation of immune profiling data into clinical diagnostics.