GenCircuit-RL: Reinforcement Learning from Hierarchical Verification for Genetic Circuit Design

Designing genetic circuits, which are biological systems capable of programmed behaviors within living cells, remains a laborious, expert-driven process despite decades of progress in synthetic biology. We introduce GenCircuit-RL, a reinforcement learning framework that trains language models to reason about genetic circuit design through code generation, where models produce Python code using PySBOL to construct circuits in the standardized Synthetic Biology Open Language (SBOL) format. Our approach addresses the challenge of sparse feedback in biological design through hierarchical verification rewards that decompose correctness into five levels, from code execution through structural validity to functional behavior, providing dense learning signal while multiplicative dependencies prevent reward hacking. We contribute SynBio-Reason, a benchmark of approximately 4,753 circuits spanning six canonical circuit types and nine tasks from code repair to de novo design, with held-out biological parts enabling rigorous out-of-distribution evaluation. A four-stage curriculum progressively shifts optimization pressure from basic code generation toward functional correctness, enabling models to acquire compositional reasoning capabilities incrementally. Our framework demonstrates that hierarchical verification combined with curriculum learning enables compact language models to generate functionally correct genetic circuits, including generalization to novel biological parts and rediscovery of canonical designs from synthetic biology literature.

Lay Summary Synthetic biologists design "genetic circuits," which are DNA networks that program living cells to sense pollutants, produce medicines, or detect disease. The work is slow, expensive, and requires rare expertise, concentrating a powerful technology in a few well-resourced labs. This makes the field a natural target for AI assistants, though standard training methods struggle here: a generated circuit can fail in many distinct ways, and reducing those failures to a single right-or-wrong score gives the model almost nothing to learn from. We trained AI models to write code that builds genetic circuits, scoring each attempt across five linked criteria that range from whether the code runs to whether the circuit performs its intended biological function. We paired this graded feedback with a "curriculum" that walks the model from basic coding up through structural assembly, functional reasoning, and finally designing new circuits from a description alone. Our trained models rediscover landmark designs from the synthetic biology literature and work successfully with biological parts they had never seen during training. By lowering the expertise barrier, this work could broaden who builds with biology, with applications spanning biosensors to targeted therapies. More broadly, our recipe of graded rewards plus staged learning may help train AI for other scientific problems whose correctness can be checked step by step.