ESM2-driven transmembrane domain selection for optimising plant P450s in microbial hosts

Background. Specialised plant metabolites, such as terpenoids, are valuable compounds which have applications in fields such as food colouring and flavouring, pharmaceuticals and fragrance. Plant cytochrome P450 enzymes play a crucial role in the biosynthesis of these, by catalysing oxygenation reactions. For this, P450s are anchored with an N-terminal transmembrane domain (TMD) to the membrane of the endoplasmic reticulum (ER). These enzymes require membrane anchoring for proper folding and functionality, and to allow interaction with their redox partners. However, bacteria, such as Cereibacter sphaeroides, lack an ER, creating a major challenge for functional P450 expression. Current strategies rely on replacing the native TMDs with bacterial or synthetic alternatives, but the pool of available TMDs is limited and often results in inefficient inner membrane targeting and low enzyme activity. Approach. To address this limitation, we employ a data-driven strategy to identify novel TMDs that enhance inner membrane localisation and functionality of plant P450s in microbial hosts. Our approach leverages extensive sequence mining and deep learning-powered selection, using the protein language model ESM2. With the latter, TMD sequences are represented as numerical vectors, which allows to access sequence relationships through Euclidean distances. As a case study, we focus on CYP71AV8, a valencene oxidase from chicory that converts valencene to nootkatone, and test its membrane localisation in C. sphaeroides through confocal microscopy, and its functionality. Results. Protein sequences from plant P450s and bacterial membrane proteins were collected from UniProt. After preprocessing, the TMDs, together with their flanking regions, were extracted. ESM2 embeddings were calculated using the Barcelona supercomputer and the first two dimensions are visualised in Figure 1. A clear separation between bacterial and plant P450 TMDs is revealed, highlighting key structural and physicochemical differences. A diverse set of 88 TMDs was selected using multiple criteria and fused to the CYP71AV8-sfGFP fusion protein for membrane localisation studies. Initial experiments optimising microscopy studies showed that four out of seven TMD candidates directed sfGFP to the membrane, demonstrating that our selection approach is effective. Conclusion. By leveraging protein language models, we systematically explore TMD diversity and efficiently select sequences with high potential for membrane integration. Our early findings validate this approach, but still have to be confirmed for targeting CYP71AV8 to the membrane of C. sphaeroides. In the future, selected candidates will also be evaluated for their impact on nootkatone production to assess the link between membrane targeting and functionality. Additionally, this framework for TMD selection and testing is designed for broad application, allowing to use it across different P450s and microbial hosts.