No sequencing required from histology to transcriptomics! One of the winning teams at Autoimmune Disease ML Crunch 2 in partnership with @Schmidt_Center built a pipeline that predicts 2,000 immune gene expressions directly from tissue images. Let’s break it down 🧵

The Problem Can we infer gene expression from tissue slides alone? That’s the kind of signal compression AI is now unlocking, turning stained images into molecular insights for autoimmune disease. This team built one of the most elegant systems in the challenge.

Their Stack • Patches extracted from histology slides • Vision Transformer (ViT) for 1024-d embeddings • Multi-headed attention to predict 200 top-variable genes • Variational encoder-decoder reconstructing all 2,000 gene expressions Trained on Crunch’s normalized scRNA-seq data.

The Architecture Image → Embedding → Signal → Profile This system maps inputs to outputs by modeling the underlying biology. Each layer extracts a different kind of information: visual, statistical, and biological. The final result is accurate, interpretable, and grounded in the structure of the data.

The Impact This kind of system shrinks cost, accelerates diagnosis, and bridges imaging with omics. especially valuable in diseases where immune activity hides beneath the surface. A step forward for computational pathology. And for patients.

The Team • Dr. Sukrit Gupta @YoSukrit • Amit Kumar • Maninder Kaur • Dr. Michele Ceccarelli @mceccarelli • Dr. Raghvendra Mall @MallRaghvendra Read their full write-up here: