Elyas Heidari

I am a PhD researcher in AI for Biology, in Oliver Stegle’s and Moritz Gerstung’s labs. I’m interested in scalable, academic-budget AI, and in realist evaluation: whether a method survives contact with real biological data, which most AI doesn’t, out of the box. That’s why I care less about the fanciness of an architecture than whether it’s usable, accurate, robust, fast, and scales. That’s what the benchmarks and metrics I build measure, and why I still run the single-cell and spatial analysis by hand. The devil is in the details.

My main project, Segger, turns cell segmentation into link prediction on a graph and assigns 30 million transcripts in about 10 minutes, roughly 1,000× faster than the tools before it; it’s the tokenizer the spatial foundation models above it are built on. I’m now building one of those, Laminar: a self-supervised model that turns a tumour into a cross-scale embedding field, an AlphaEarth for tissues rather than the planet, trained on 50 billion transcripts and 500 million cells at the German Cancer Research Center.

Before Heidelberg, I did a double bachelor’s in computer engineering and mathematics at Sharif University of Technology in Tehran, where I worked with Ali Sharifi-Zarchi and built MUVis. I then did a master’s in computational biology at ETH Zürich, where my thesis won the ETH Medal. Along the way: a summer at EMBL with Wolfgang Huber, where I built scPotter; single-cell pipelines with Mark Robinson in Zurich; and SageNet with John Marioni and Shila Ghazanfar at the Cancer Research UK Cambridge Institute.