Original title: Foundation-model-guided radiogenomic discovery linking cancer genomes to cancer scans
Frederik Hauke, Jeremias Krause, Patrick Wienholt, Christiane Kuhl +5 · July 2026
Doctors struggle to learn what rarely mutated genes do in cancer because standard tools only notice mutations that appear frequently. To solve this, researchers connected cancer genetics to standard clinical scans. They used an artificial intelligence model called Evo 2, which reads DNA. For every mutation, Evo 2 scores how damaging it is without needing special training. The team then checked whether severe gene damage lined up with distinct physical features seen in tumor scans, while adjusting for a patient's overall mutation count. In kidney cancer scans, this approach detected known cancer drivers and found 46 significant new genes missing from curated cancer panels, including genes linked to rare inherited cellular diseases. This shows scans can uncover hidden gene functions.
You are trying to find which rare ingredient in a recipe makes a cake rise. Normally you only notice ingredients used in many recipes, so rare ones are overlooked. Here, you use a machine that reads the ingredient's molecular structure and predicts its effect, then matches that to photos of the cakes. In this paper, the ingredients are gene mutations, the machine is the genomic language model, and the photos are medical scans.
Most cancer genes are found by looking at frequently mutated genes, missing rare ones. This work pairs a genomic language model with routine clinical scans to find gene-imaging associations without any training. It discovered 46 new genes linked to kidney cancer, including genes for rare genetic diseases, showing that existing medical images can reveal hidden genetic links.
The abstract only reports results for clear cell renal cell carcinoma (cRCC) with 162 patients, not for the other two cancer types. The associations are correlational and not experimentally validated. The method depends on the quality of tumor segmentations and the Evo~2 severity scores, which have not been independently verified for this task.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Foundation-model-guided radiogenomic discovery linking cancer genomes to cancer scans — Frederik Hauke, Jeremias Krause, Patrick Wienholt et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.