Generative Models in Synthetic Biology

Abstract

Synthetic biology designs biological systems -- proteins, genetic circuits, metabolic pathways, and whole organisms - with functions not found in nature. The design-build-test-learn (DBTL) cycle is the central workflow: design a candidate sequence computationally, build it by DNA synthesis, test its function experimentally, and learn from the results to improve the next design. Generative AI models are transforming the design phase by learning the distribution of functional biological sequences and generating novel candidates with desired properties. Protein language models generate functional enzymes, diffusion models design protein structures with specified folds, generative adversarial networks produce synthetic promoters with tunable expression, and reinforcement learning optimises metabolic pathway configurations. We present the Generative Synthetic Biology Assessment Framework (GSBAF), evaluating five generative model classes -- autoregressive protein generators, structure-conditioned diffusion models, regulatory sequence GANs, metabolic pathway optimisers, and whole-circuit design agents -- across four synthetic biology design tasks (enzyme engineering, de novo protein design, promoter library construction, and genetic circuit design). Our Generative Synthetic Biology Score (GSBS) measures sequence novelty, functional success rate, design diversity, computational efficiency, and experimental validation rate. Structure-conditioned diffusion models achieve the highest GSBS (0.928) through physics-informed generation that constrains designs to physically realisable protein folds, while autoregressive generators achieve the highest experimental validation rate (0.940) for enzyme engineering through evolutionary-scale learned fitness landscapes.