Reframing enzyme function prediction as conditional generation
- 1AITHYRA, Research Institute for Biomedical Artificial Intelligence, Austrian Academy of Sciences, Vienna, Austria
Abstract
Enzymes frequently exhibit promiscuous activity beyond their native roles, providing starting-points for new functions. Finding these promiscuous enzymes, especially for non-native chemical transformations, is challenging but highly valuable, as they promise novel, sustainable solutions for chemistry and biotechnology. However, current machine learning methods are poorly suited to discovering unseen chemistry as they often frame function prediction as closed set classification or a retrieval task. Here, we present Fluxion, a generative deep learning framework that learns enzymatic catalysis by modeling dynamic electron flow trajectories across the enzyme's catalytic residues. By combining both synthetic chemistry and biochemical datasets with protein language model representations, Fluxion generates multi-step electron-flow trajectories analogous to the arrow-pushing representations used to describe enzyme reaction mechanisms. Generation is conditioned on enzyme context, including the enzyme sequence, catalytic residues, substrates, and cofactors. We show that this conditioning allows Fluxion to learn enzyme-dependent regioselectivity across cytochrome P450 enzymes with different sequences shifting the predicted reaction sites for the same substrate. We then demonstrate that Fluxion's embeddings are useful for downstream tasks, such as specificity prediction on two experimental datasets, with and without finetuning. Finally, we show that Fluxion has the potential to transfer synthetic chemical logic to biology; it can generate the observed non-native product from real-world non-native directed evolution screens. Our results establish a proof of concept that generative modeling through mechanistic representations of enzymes can shift enzyme function prediction beyond static database retrieval and closed set classification to function generation. This conceptual framework provides a stepping stone towards an in silico generative method to discover non-native biocatalysts.
Related
- A Two-Stage ESM-Based Machine Learning Pipeline for Robust Hierarchically Evaluated Enzyme Function Prediction
- DeepEnzyKat: Turnover number prediction through enzyme-substrate interaction modeling and adaptive expert ensembles
- RNAbpFlow: base pair-augmented SE(3) flow matching for conditional RNA 3D structure generation
- Joint enzyme–reaction retrieval and catalytic optima prediction via multimodal fusion
- Biosensor-enabled single-molecule enzyme activity profiling for function-based molecular counting
BibTeX
@misc{rieger2026reframing,
title = {{Reframing enzyme function prediction as conditional generation}},
author = {W. J. Rieger and S. Haeussermann and L. Herrmann and Z. Li and B. P. Frohn and M. Maluenda and G. Lobinska and S. Loomba and J. C. Reisenbauer and M. Boden and A. Tong and A. Mora},
howpublished = {bioRxiv},
year = {2026},
month = sep,
doi = {10.64898/2026.09.09.750355},
url = {https://www.biorxiv.org/content/10.64898/2026.09.09.750355v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Reframing enzyme function prediction as conditional generation
AU - W. J. Rieger
AU - S. Haeussermann
AU - L. Herrmann
AU - Z. Li
AU - B. P. Frohn
AU - M. Maluenda
AU - G. Lobinska
AU - S. Loomba
AU - J. C. Reisenbauer
AU - M. Boden
AU - A. Tong
AU - A. Mora
PY - 2026
DA - 2026/09/09
PB - bioRxiv
DO - 10.64898/2026.09.09.750355
UR - https://www.biorxiv.org/content/10.64898/2026.09.09.750355v1
AB - Enzymes frequently exhibit promiscuous activity beyond their native roles, providing starting-points for new functions. Finding these promiscuous enzymes, especially for non-native chemical transformations, is challenging but highly valuable, as they promise novel, sustainable solutions for chemistry and biotechnology. However, current machine learning methods are poorly suited to discovering unseen chemistry as they often frame function prediction as closed set classification or a retrieval task. Here, we present Fluxion, a generative deep learning framework that learns enzymatic catalysis by modeling dynamic electron flow trajectories across the enzyme's catalytic residues. By combining both synthetic chemistry and biochemical datasets with protein language model representations, Fluxion generates multi-step electron-flow trajectories analogous to the arrow-pushing representations used to describe enzyme reaction mechanisms. Generation is conditioned on enzyme context, including the enzyme sequence, catalytic residues, substrates, and cofactors. We show that this conditioning allows Fluxion to learn enzyme-dependent regioselectivity across cytochrome P450 enzymes with different sequences shifting the predicted reaction sites for the same substrate. We then demonstrate that Fluxion's embeddings are useful for downstream tasks, such as specificity prediction on two experimental datasets, with and without finetuning. Finally, we show that Fluxion has the potential to transfer synthetic chemical logic to biology; it can generate the observed non-native product from real-world non-native directed evolution screens. Our results establish a proof of concept that generative modeling through mechanistic representations of enzymes can shift enzyme function prediction beyond static database retrieval and closed set classification to function generation. This conceptual framework provides a stepping stone towards an in silico generative method to discover non-native biocatalysts.
ER -