Reinforcement learning guided directed evolution of synthetic cis-regulatory elements

Here, we present GO-CRE (Guided Optimization of CRE), a reinforcement learning guided framework built on a hybrid Transformer-Mamba2 architecture for efficient iterative CRE design. Across K562, HepG2, and SK-N-SH cells, GO-CRE drives CRE generation along directional trajectories in a reconstructed sequence-grammar landscape, revealing a staged process of in silico directed evolution.

We resolve this process into three phases—search, commitment, and optimization. Productive trajectories in K562 and HepG2 progress through all three phases and converge on optimized regulatory programs, whereas SK-N-SH trajectories become trapped in local sequence-grammar basins. In HepG2, tuning k-mer composition redirects evolution away from low-complexity traps toward functional motif programs that retain HNF, FOXA, and related regulatory features.