SCimilarity

Summary

SCimilarity is an autoencoder-based metric learning model designed to enable rapid search and annotation of single-cell datasets. It learns a unified embedding space where cells of the same type are projected close together, regardless of technical batch effects, using a contrastive triplet loss function. In systematic evaluations (as of 2026-06-09), SCimilarity was the top-performing model on cell-type classification tasks and batch integration, though it showed no improvements from increasing the pretraining dataset size beyond a fraction of the corpus.

Architecture and Training

SCimilarity's design is optimized for cell-type representation: - Network: An autoencoder architecture with hidden layer sizes of (1024, 1024, 1024) for the encoder (reversed for the decoder) and a 128-dimensional latent space. - Loss Function: Employs a triplet loss objective that pulls together cells of similar types (positive pairs) and pushes away cells of different types (negative pairs). - Normalization: Preprocessed cell expression profiles are normalized to 10,000 counts per cell and log-transformed.

Benchmarking Results

During comparative studies of Single-Cell Foundation Models, SCimilarity outperformed all other models, including simple baselines, on zero-shot cell-type classification. Like other models, its learning saturated at 1% of the pretraining data (~200,000 cells), indicating that the representations learned from a subset are sufficient for cell-type comparison.

Citations

• Heimberg, G. et al. A cell atlas foundation model for scalable search of similar human cells. Nature 638, 1085–1094 (2025). DOI: 10.1038/s41586-025-08637-2
• Key Evaluation: DenAdel, A. et al. (2026). Evaluating the role of pretraining dataset size and diversity on single-cell foundation model performance. Nature Methods, 23, 1447–1457. DOI: 10.1038/s41592-026-03120-y. Source paper: s41592-026-03120-y.pdf