Machine learning powered enzyme development

Discovering Terpene Synthases Beyond Homology

Safa Mert Akmeşe

Terpene synthases (TPSs) generate the carbon scaffolds of terpenoids, the largest and most structurally diverse class of natural products, including front-line medicines such as Taxol. Yet for most terpenes the responsible TPS remains unknown, because conventional annotation tools cannot reliably recognise TPSs beyond close homology, a central bottleneck for terpenoid biosynthesis.

We address this by modelling TPS biochemistry with machine learning to predict enzymatic function de novo, enabling TPS discovery by screening large sequence databases. Using MARTS-DB, a manually curated database of experimentally characterised TPSs, we built EnzymeExplorer, which predicts TPS activity and substrate specificity. It combines protein language model representations of sequence with structural features extracted by a novel domain detection algorithm applied to AlphaFold3 structures. EnzymeExplorer detects TPS activity with near-perfect precision and predicts substrates with 81% mean average precision, outperforming existing methods. Screening UniProt proteins that lack functional annotations in resources such as InterPro in silico, EnzymeExplorer revealed the first terpene synthases reported in the domain Archaea confirmed in vitro, expanding the known taxonomic range of terpene cyclisation. Predicting the product of a TPS remains harder, as catalysis proceeds through cascades of carbocation rearrangements. To address this, we model the reaction mechanisms explicitly, tracing the cascade of carbocation intermediates to predict the final products. Preliminary results are promising but limited by the scarcity of stereochemically resolved training data.

De novo generation of terpene synthases with desired catalytic activity

Matouš Soldát

Designing terpene synthases with specific terpene products would enable sustainable production of both existing and new pharmaceuticals, fragrances, fuels, and materials. However, targeted enzyme generation is not achievable by biochemical intuition and combinatorial mutagenesis, and while recent advances in machine learning (ML) unlocked the generation of structurally plausible proteins, controlling enzymatic function remains a major challenge.

As the first step towards product-controlled generation of terpene synthases, we focused on generating novel terpene synthases that accept the selected substrate, farnesyl pyrophosphate, and produce sesquiterpenes. First, we trained a variational autoencoder1 on multiple sequence alignment (MSA) of plant sesquiterpene synthases from MARTS-DB2, our manually curated database of terpene synthase reaction mechanisms. Second, we used the MSA to steer the generation of a pretrained discrete diffusion model3.

We expressed 16 enzymes designed by each ML model, conducted in vitro assays, and identified terpene products via mass spectrometry. The experimental validation confirmed the targeted sesquiterpene-synthase activity in four generated enzymes and related terpene chemistry in eight additional designs. Moreover, the discrete diffusion model demonstrated control of substrate specificity with 19% success rate, while maintaining substantial novelty among successful designs, with only 23-32% sequence identity to the closest known terpene synthase.

Having succeeded in substrate-conditioned design of terpene synthases, we are now focusing on more precise control of the designed enzymatic activity by narrowing the definition of the desired products.

Pluskal Group IOCB