Reading and designing gene regulation

Genes are switched on and off by short stretches of regulatory DNA that surround them. We build interpretable deep-learning models — including our openly available deepCRE toolkit — that learn this regulatory "grammar" directly from the genome and predict, with over 80% accuracy, whether a gene will be active in a given species, tissue or condition. Because the models are explainable, they tell us which sequence features drive a prediction and how a single natural or engineered variant would change it. We have used this to trace regulatory variation across maize, tomato, sorghum, Arabidopsis and oilseed rape, to pinpoint stress-related variants such as those in the flooding-response gene RAP2.12, and to model transcription-factor binding genome-wide. We are now turning prediction into design: combining generative AI with synthetic-biology validation (Plant STARR-seq) to propose promoter edits and gene-editing strategies that tune gene activity toward a desired trait, in an iterative design–build–test–learn loop.