Tool: Improved prediction of microbial optimal growth temperatures

Microorganisms live and grow across a wide range of temperatures. Measuring optimal growth temperatures in the lab is slow and, in many cases, impossible when microbes cannot be grown in pure cultures. Our goal was to predict the temperature at which they grow best. This work builds on previous studies, which showed that using protein sequence information is a viable way to predict optimal growth temperatures. We gathered a large, carefully chosen set of complete genomes and metagenomes reconstructed from environmental samples. Next, we tested several modern machine learning methods, including neural networks. Our machine learning tools make predictions that closely match real lab results and work for many kinds of microbes. In addition, the predictions are fast and scale to large datasets, thus reducing the need for time-consuming experiments.

Our neural network predictor is based on protein language models. On a large test dataset with 1,430 (meta)genomes not used for training, its predictions have RMSE < 3 and R² > 0.9 on average.

PLM predictions

Importantly, model predictions are accurate even for metagenome-assembled genomes (MAGs), which are often reconstructed at varying degrees of completeness from environmental samples.

MAG predictions