Enhancers are gene regulatory elements—short DNA sequences that control when and where genes are active. Deep learning models can help decode this “regulatory code”, but existing approaches are often tailored to one dataset or one task, making them hard to reuse or extend.

To address this, Prof. Stein Aerts and his team developed CREsted, a new framework that turns enhancer modeling from a collection of one-off analyses into a more systematic and reusable workflow.

“We wanted to move beyond one-off models,” says Niklas Kempynck, PhD student in the Aerts lab. “CREsted allows researchers to systematically study enhancer logic across biological systems, starting from cell-by-cell maps of accessible regulatory DNA and going all the way to sequence design.”

CREsted brings together several steps that are usually handled separately: preprocessing, model training, interpretation, and synthetic enhancer design. It is also built to fit into existing single-cell analysis workflows, making it easier for researchers to adopt and use. “With CREsted, we give researchers a complete workflow,” says Dr. Seppe De Winter, who shares first authorship with Kempynck. “You can train deep learning models on chromatin accessibility data, interpret which regulatory features they capture, and then use those models to design new DNA sequences with predicted cell-type-specific activity.”

To show its versatility, the team applied CREsted to multiple systems, including mouse brain tissue, human immune cells, cancer cell states, and zebrafish development. Across these settings, the framework identified regulatory patterns, predicted enhancer activity, and enabled the design of synthetic enhancers, which were validated in vivo in zebrafish.

For Prof. Stein Aerts, Scientific Director of VIB.AI, the strength of CREsted lies in making a powerful development more coherent and reusable.

“CREsted makes it much easier to train, interpret, and compare enhancer models across datasets,” he says. “That is important if we want these approaches to become broadly useful, not just for understanding regulatory DNA, but also for designing and testing new sequences in a systematic way.”

Taken together, the work shows how AI can help move the field from describing regulatory DNA to actively exploring and designing it. With applications ranging from fundamental biology to biotechnology and medicine, CREsted lays the groundwork for more systematic and programmable control of gene regulation.

Funding
This research was funded in part by Aligning Science Across Parkinson’s through the Michael J. Fox Foundation for Parkinson’s Research (MJFF); the European Research Council, the Chan-Zuckerberg Initiative, the Research Foundation Flanders - FWO, and the Foundation Against Cancer. For all funders, see the paper’s acknowledgements.