Interpretable modeling of time-resolved single-cell gene-protein expression with CrossmodalNet

Area of Science:

• Single-cell biology
• Computational biology
• Immunology

Background:

• Cell-surface proteins are crucial for cell function and therapeutic targeting.
• CITE-seq measures gene and surface protein expression simultaneously but is costly and complex.
• Existing computational methods for predicting protein expression from gene data are resource-intensive and lack interpretability.

Purpose of the Study:

• To develop an interpretable machine learning model for predicting surface protein expression from single-cell RNA sequencing (scRNA-seq) data.
• To address the computational demands and lack of interpretability in current prediction methods.
• To enable deeper insights into molecular mechanisms involving cell-surface proteins.

Main Methods:

• Developed CrossmodalNet, an interpretable machine learning model.
• Utilized a customized adaptive loss function for accurate prediction of surface protein abundances.
• Incorporated temporal information encoding for time-point-specific predictions and causal relationship discovery.

Main Results:

• CrossmodalNet accurately predicts surface protein abundances from scRNA-seq data.
• The model effectively encodes temporal information for time-resolved analysis.
• Identified noise-free causal gene-protein relationships, enhancing interpretability.
• Validated performance against benchmarking methods on three public CITE-seq datasets.

Conclusions:

• CrossmodalNet provides an accurate and interpretable method for profiling surface protein expression using scRNA-seq data.
• The model enhances the analytical power of CITE-seq experiments.
• Facilitates the investigation of molecular mechanisms involving cell-surface proteins.