LCL: Contrastive Learning for Lineage Barcoded scRNA-seq Data

We present LCL (Lineage-aware Contrastive Learning), a contrastive deep learning algorithm designed to learn low-dimensional embeddings from lineage-barcoded single-cell RNA-seq data. LCL facilitates the identification of cell fate-determining gene signatures by leveraging lineage information as a self-supervised training signal.

Key contributions:

• Contrastive learning framework that incorporates lineage barcode information into scRNA-seq embedding
• Comprehensive evaluation using Calinski-Harabasz Index, KNN classification, UMAP visualization, and a multi-label neural network classifier
• Local autocorrelation analysis to identify gene pathways involved in cell fate determination

Recommended citation: Yang, S. J., Wang, Y., & Lin, K. Z. (2024). "LCL: Contrastive learning for lineage barcoded scRNA-seq data." bioRxiv. (to appear in PMLR 2026)
Download Paper