Machine learning to predict enzymes’ optimal catalytic temperature

Abstract

Enzymes are proteins which operate as biological catalysts in chemical processes, for instance in biofuel production. The efficiency and sustainability of these processes may be greatly improved by knowing the optimal catalytic temperature (Topt) of the enzymes. However, determining these temperatures experimentally is timeconsuming and instead a machine learning approach for predicting Topt is suggested. In a previous approach, sequential features were used to predict Topt. In this thesis, new structural features which account for various structural properties in the enzymes were used alongside the sequential features. Test scores from the models show that structural features combined with sequential features improve previous R2 scores from 0.4 to 0.48. Furthermore, in the case where there is a pair of similar enzymes, but one has a colder and one a hotter temperature, the models correctly predicts the temperature order of the enzymes 83% of the time. By gathering more data and fine-tuning the structural features, it is anticipated that scores will improve even further.