Identifying Cell Type-Specific Chemokine Correlates with Hierarchical Signal Extraction from Single-Cell
Sherry Chao1, Michael P Brenner, Nir Hacohen
1Department of Biomedical Informatics, Harvard University, Boston, MA, United States, schao@g.harvard.edu.
Pacific Symposium on Biocomputing. Pacific Symposium on Biocomputing
|December 10, 2021
Summary
This study introduces an interpretable neural network for analyzing single-cell gene expression data. The method accurately identifies cell and gene subsets, outperforming traditional classifiers in biological sample analysis.
Area of Science:
- Computational Biology
- Genomics
- Immunology
Background:
- Biological data, particularly from single-cell sequencing, is complex, high-dimensional, and heterogeneous.
- Identifying specific cell and gene subsets linked to biological labels is a significant challenge in transcriptomics.
Purpose of the Study:
- To develop an interpretable and robust method for classifying tissue samples using single-cell gene expression profiles.
- To identify cell type-specific chemokine correlates for predicting treatment response and disease status.
Main Methods:
- Integration of a signal-extractive neural network architecture with axiomatic feature attribution.
- Classification of tissue samples based on single-cell gene expression data, requiring minimal gene and cell subsets.
Main Results:
- The approach achieved >70% accuracy in distinguishing signal from noise using only 5% of genes and 23% of cells in silico.
- Demonstrated effectiveness in predicting immune checkpoint inhibitor response and classifying DNA mismatch repair status in colorectal cancer.
- Significantly outperformed traditional machine learning classifiers.
Conclusions:
- The developed method provides an interpretable and robust framework for analyzing complex single-cell genomics data.
- It offers actionable biological insights into chemokine-mediated tumor immunogenicity and identifies potential biomarkers for clinical applications.
