AdventML: Advanced Enzyme Temperature Prediction with Transformer-Based Embeddings and Resampling Strategies

Journal of Chemical Information and Modeling

Abstract

Abstract Accurate prediction of enzyme optimal catalytic temperature (Topt) is important for enzyme applications, particularly for identifying biocatalysts stable in extreme conditions, but experimental measurement remains costly and time-consuming. Here, we present AdventML, a machine learning framework for (Topt) prediction that expands the available models, evaluates protein representations and model architectures, and compares resampling strategies for improving performance at extremes. The base AdventML model achieved the best global held-out performance among the evaluated methods (R2 ≈ 0.64), MAE ≈7–8 °C, while tail-focused variants improved selected extreme-temperature regimes, reducing tail MAE by up to ∼1.8 °C. AdventML also showed that competitive (Topt) prediction is possible in the absence of organism growth temperature. Our Topt prediction models are made freely available as AdventML on GitHub.

Authors 3

1. Jérémie François

   KU Leuven

   Affiliation as printed

   KU Leuven , , ,

2. Bart De Moor

   KU Leuven

   Affiliation as printed

   KU Leuven , , ,

3. Vera van Noort Aachen

   Leiden University · KU Leuven

   Affiliation as printed

   KU Leuven , , ,

   Leiden University , , ,