Generative Models in Synthetic Biology

Abstract

Synthetic biology aims to design and construct biological systems with novel or enhanced functions -- a goal that traditionally required exhaustive experimental iteration over vast combinatorial sequence spaces. Generative AI has opened a qualitatively different approach: learning the statistical structure of biological sequences and structures from natural evolution, then sampling from that learned distribution to propose functional designs without exhaustive search. This study evaluates seven generative model architectures for synthetic biology applications across four design tasks: protein sequence generation (for thermostability and binding affinity), gene regulatory element design (promoters and enhancers), metabolic pathway optimisation, and CRISPR guide RNA design. Models evaluated include ESMFold-based sequence sampling, ProtGPT2, RFdiffusion, EvoDiff, DiffSBDD (structure-based drug design), DNABERT-2 fine-tuned for regulatory element design, and a purpose-built Transformer-VAE (TVAE) for metabolic pathway optimisation. Experimental validation was performed for a subset of AI-designed proteins (n=84) and regulatory elements (n=240) using cell-free expression systems and luciferase reporter assays respectively. RFdiffusion achieves the highest experimental success rate for protein designs (62.4% of designs showing detectable target function). DNABERT-2 fine-tuned designs achieve 48.6% experimental validation rate for synthetic promoters. EvoDiff provides the most diverse design library (intra-design sequence similarity index 0.42 vs. 0.74 for ProtGPT2). A generative model selection framework for synthetic biology practitioners is proposed