Horizyn: advanced deep learning for enzyme discovery and annotation

The Department of Energy (DOE) seeks innovative methods to address challenges in bioenergy, carbon cycling, and sustainable bioproducts. Accurate enzyme function prediction is essential but currently hampered by fragmented, underutilized datasets and reliance on outdated bioinformatics methods. Existing approaches struggle to scale with the size and complexity of modern biological datasets, limiting advances in bio-based industries. To address these challenges, this project will develop Horizyn, a multi-modal machine learning platform that integrates protein sequence and biochemical reaction data to predict enzyme functions and discover new biocatalysts. Using advanced neural networks and contrastive learning, Horizyn will enable scalable, accurate predictions, unlocking new pathways for industrial biocatalysis and bioenergy production. The initial phase will focus on developing and validating the dual-encoder model that maps protein sequences and chemical reactions into a shared embedding space. A proprietary dataset of 15 million protein-reaction pairs, along with public biochemical data, will be used to train and benchmark the model. Additional features like genomic context and metabolic network data will be integrated to improve the accuracy and generalization of the model. Experiments will validate Horizyn’s performance against industry-standard metrics, with the goal of significantly reducing the cost and time required for enzyme discovery. Horizyn will transform enzyme discovery for biofuels, pharmaceuticals, and chemical manufacturing by reducing the experimental burden, lowering costs, and improving the scalability of biocatalyst development. The platform supports DOE's mission by advancing sustainable bioenergy technologies and enabling greener industrial processes.