Improved prediction of microbial optimal growth temperatures with neural networks and protein language models

Temperature is one of the strongest selective forces that determines the composition of microorganisms in the environment. Structural properties of proteins shape the thermal adaptation of an organism at the macro level, and aggregate protein features can be used for many types of predictions. Because most microorganisms remain uncultured, the inference of physiological traits, such as optimal growth temperature, has become essential for microbial ecology and biotechnology.

A large dataset of optimal growth temperatures for microorganisms was compiled from the literature and used to train several machine learning predictors. Our goal was also to test the usefulness of protein language models and to evaluate predictive performance on incomplete genomes.

We confirmed a strong correlation between protein sequence properties and optimal growth temperatures. The analysis showed that calculating better protein sequence features, specifically through protein language models, leads to more accurate predictions.

We used state-of-the-art tools and compared our models with several others developed for optimal growth temperature prediction over the past 2 decades. Our models showed excellent ability to generalize across a range of temperatures. We concluded that larger datasets and increased representation of psychrophiles and thermophiles will be needed to continue improving the predictors.