New Framework Learns Cell Representations for Single-Cell Transcriptomics

Single-cell transcriptomic data is rapidly expanding, leading to the development of foundation models primarily trained by reconstructing masked gene expression values. While this approach helps models understand gene dependencies, it doesn't directly optimize for whole-cell representations, which are crucial for many subsequent analytical tasks. To address this, a new contrastive pretraining framework has been introduced to learn comprehensive cell representations. This framework adapts standard contrastive learning for single-cell data through three key innovations: partitioning genes based on co-expression structures to create complementary views of each cell, constructing "hard negatives" by permuting expression values while retaining gene identities to prevent shortcuts, and implementing a competence-aware controller to manage the application of the contrastive objective. Evaluations on tasks like cell-type annotation and gene regulatory network inference demonstrate the competitive transferability of this method. It achieved the highest mean AUROC and AUPRC point estimates in a six-network GRN evaluation, suggesting that complementary-view contrastive learning is a highly effective direction for single-cell pretraining beyond simple gene reconstruction.