Manufacturing-aware generative models enable petascale synthesis of designed DNA

Generative modeling offers a robust framework for designing functional DNA, RNA and protein sequences. However, physical synthesis of these sequences at scale is prohibitively expensive. We introduce a method to efficiently synthesize DNA designs from generative models. The method integrates machine learning and wet lab procedures, implementing generative sampling algorithms physically through controlled stochastic chemical reactions using DNA oligosynthesis. We synthesize ~10¹⁶ designs from a generative model of human antibodies, with realism and diversity comparable to state-of-the-art protein language models, and verify the designed DNA by sequencing. Translation and expression of the designed antibody scFvs in human cell lines, combined with high-throughput screening against multiplexed human leukocyte antigen (HLA)-presented intracellular proteins, yields potential therapeutic chimeric antigen receptors. We further synthesize ~10¹⁶ DNA designs from models of Taq polymerase and the HLA-presented peptidome, confirming the method's generalizability.