Generative design of synthetic gene circuits for functional and evolutionary properties

In the past decades, a wide suite of design tools for biological systems has been developed, but using these to create biotechnologies that achieve reliable and predictable behaviour remains challenging. Modelling approaches have enabled researchers to traverse the vast search space of genetic circuits more efficiently, while machine learning has proven useful for designing parts and predicting their function or evolutionary properties. Generative algorithms have the potential to leverage these features to design entire genetic circuits from the sequence level, but have only recently begun to be applied to synthetic biological applications. Here, we show that even simple generative models like the conditional variational autoencoder (CVAE) can produce novel genetic circuits that match complex dynamic functions such as signal adaptation. Using in silico RNA simulation, we construct a dataset of RNA sequences and convert them to circuits via RNA interaction predictors, allowing us to estimate functional features alongside evolutionary stability and interpret model-learned features. Our model generates diverse distributions of circuits that match their target adaptation specification well, even when limited to small training data sets. Structures in the embedding space correspond to motifs previously identified as crucial for adaptation and reflect the design rules for adaptable circuits. Framing adaptation as a single design objective outperforms other input representations, reflecting the importance of choosing the correct data encoding for generating genetic circuits. Finally, we show that functional and evolutionary properties can be prompted simultaneously, providing a proof-of-concept for the combined design of phenotype and evotype.