Single-cell foundation models are coming – are we ready?

By Sukalp Muzumdar

It was a privilege to take the stage in Bern at the 10x Genomics Single-cell & Spatial Discovery Symposium to discuss a question that increasingly defines the trajectory of our field: are we ready for single-cell foundation models?

I joined a fantastic panel moderated by Nigel Delaney alongside Thomas Wieland, Marianna Rapsomaniki, and Phil Cheng to speak about AI-grade data for biomarker discovery and precision medicine.

The market signal is loud

The market validates the timing. Recent moves such as the NOETIK ↔ GSK partnership for cancer foundation models, valued at over $50M, clearly demonstrate the demand. The industry is betting big on AI-driven drug discovery.

But the science demands caution

As we debated on stage, while great progress is being made toward the holy grail of a true virtual cell — where perturbations and drug responses are modeled in silico — these models are only as good as the data they are fed.

In fact, many cutting-edge perturbation models today still struggle to beat simpler linear models in the benchmarks. This is a sobering reality check for the field and underscores a fundamental truth: model architecture alone is not the bottleneck — data quality is.

Data quality is the real differentiator

At Scailyte, we see this daily. While by now we have vast amounts of public single-cell data, heterogeneity requires rigorous harmonization: both of data and, importantly, the associated metadata to ensure data is truly “AI-grade”.

The quality of representation achieved by foundation models is a direct function of the ability of the underlying data to represent clinical reality. Without proper curation, harmonization, and metadata standardization, even the most sophisticated architectures will learn noise rather than signal.

Looking ahead

The convergence of single-cell biology, spatial transcriptomics, and foundation models holds transformative potential for precision medicine and drug discovery. But realizing that potential requires the community to invest as heavily in data infrastructure as it does in model development.

---

Huge thanks to the 10x Genomics team for organizing this event — especially Nigel Delaney, Nico Dunkel, Gaye Tanriöver, and Yamina Bellaltia for seamless execution!

#SingleCell #SpatialBiology #PrecisionMedicine #GenAI #FoundationModels #DrugDiscovery #10XGenomics #DigitalBiology