Merck & Co. and the BCG X AI Science Institute have developed TEDDY, a family of transformer-based foundation models for understanding single-cell biology and disease, the researchers said in their paper on 17 June 2025.

The researchers scaled pre-training data to a diverse collection of 116 million cells, larger than those used by previous single-cell foundation models. They also used large-scale biological annotations as supervision during pre-training.

The TEDDY family comprises six models with 70 million, 160 million and 400 million parameters. The models were tested on tasks including identifying the disease state of held-out donors and distinguishing diseased from healthy cells.

The authors said the models showed substantial improvement over existing work, and that performance improved predictably with both data volume and parameter count. Open-source code and configurations for the best-performing TEDDY models are available on Hugging Face under an Apache 2.0 licence.


Share this post
The link has been copied!