Terpene synthases (TPSs) produce terpenes, useful molecules for fragrances, flavors, and medicine. Designing new TPS by hand is hard: minor active-site changes can completely change the product, and there is little usable sequence–function relationship to guide edits. Generative AI is starting to change this. A recent example, TpsGPT (Ramanathan et al., 2025), fine-tuned a protein language model (ProtGPT2 (Ferruz et al., 2022)) on ~79,000 TPS sequences from UniProt (“UniProt,” 2024) and generated tens of thousands of de novo candidates; after strict filtering, seven passed all checks and at least two were experimentally confirmed to be active enzymes. 

This project builds directly on that result but takes a different route: it works from protein structure rather than sequence. AI models such as AlphaFold (Jumper et al., 2021) and ESMFold (Lin et al., 2023) can predict a 3D structure for any TPS sequence, and this structural information encodes the active-site geometry that actually governs the carbocation cyclization cascade determining product outcome. Based on an already assembled dataset of predicted TPS structures (unpublished) you will create structure-aware generative models, for example through structure-conditioned approaches that generate sequences fitted to a target fold or active-site shape.

Concretely, you will (I) train or fine-tune a generative model on structures, building on approaches such as RFdiffusion (Watson et al., 2023) and Chroma (Ingraham et al., 2023); (II) generate variants and screen them with computational metrics, e.g. fold confidence, sequence diversity, domain detection, and structural alignment (Foldseek (van Kempen et al., 2024)); and (III) shortlist the most promising designs for experimental validation.