Enzyme-Unified: Learning Holistic Representations of Enzyme Function with a Hybrid Interaction Model

ICLR 2026 Conference Submission3226 Authors

Keywords: AI for science, Enzyme function prediction, biotechnology, cross attention

Abstract: Predicting diverse functional properties of enzymes is a crucial challenge in biotechnology. Current machine learning approaches often fall short due to two key limitations: they predict properties in isolation using simplistic feature concatenation, thereby missing crucial inter-property relationships, and their performance is frequently overestimated on biased, homology-unaware datasets. To overcome these challenges, we introduce Enzyme-Unified, a unified framework built upon a single, powerful architecture. We train distinct instances of our model to predict a comprehensive suite of five key properties: turnover number, Michaelis constant, catalytic efficiency, optimal temperature, and optimal pH. At its core is our novel Hybrid Interaction Model, which dynamically integrates fine-grained local interactions via cross-attention with global feature representations through a trainable gate, enabling a more holistic representation of enzyme function. For robust evaluation, we developed three new large-scale, sequence-dissimilar datasets. Our experiments show that Enzyme-Unified achieves state-of-the-art results and significantly outperforms previous models on our rigorously curated datasets, demonstrating the powerful synergy of its hybrid architecture. The code and dataset will be available upon acceptance.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 3226