scBERT is a transformer model optimized for single-cell RNA sequencing (RNA-seq) data. It can help identify which cell populations express a potential drug target and might be affected by changing its activity. section 13.2.1
Single-cell RNA-seq data is often sparse: most genes have zero or near-zero counts in most cells. Measurements also contain technical noise, and data from different experiments can have batch effects. scBERT uses specialized normalization and masking during pretraining to address these challenges. section 13.2.1
Applications include:
- Identifying cell types in mixed samples to examine target expression at cellular resolution. section 13.2.1
- Predicting knockout effects for specific cell types and analyzing transitions between cell states relevant to disease progression. section 13.2.1
Predicted knockout effects are computational hypotheses, not validated phenotypes; they can help prioritize targets for experimental follow-up. section 13.2.1
