Katy Beckermann, Medical Director of GU Clinical Research at Tennessee Oncology, shared on X:
“Congratulations to Tauhidul Islam and team on scVision, a single-cell AI model: instead of reading a cell as a string of gene tokens, it renders each cell as an image.
scGenomics clearly informative in tumor biology and why I’m watching it as a GU oncologist: on held-out clear-cell RCC immune cells, it read the microenvironment better than every other foundation model, with no RCC-specific training. [PREPRINT, not yet peer reviewed]
- Most accurate annotator of 7 ccRCC immune cell types across 4 foundation models
- Genes placed so co-expressed genes sit as neighbors and gene programs read as texture
- Vision transformer, 72M human cells
Resolving immune cell states in the RCC microenvironment is part of the challenge behind checkpoint-inhibitor response. Improved ability to separate the biology and layer it with the network is exciting.
Could a vision-based model sharpen how we map the RCC immune landscape?”
Other articles about GU Oncology on OncoDaily.
