Using machine learning to predict quantitative phenotypes from protein and nucleic acid sequences

The link between sequence and phenotype is essential to understanding the molecular mechanisms of evolution, and the design of proteins and genes with specific properties. However, it is difficult to describe the relationship between sequence and protein or organismal phenotypes, due to the complex relationship between sequence, protein folding and activity, and organismal physiology. Here, we use machine learning models trained on individual families of proteins or nucleic acids to predict the originating species9 optimal growth temperatures or other quantitative phenotypes. Trained multilayer perceptrons (MLPs) outperformed linear regressions in predicting the originating species growth temperature from protein sequences, achieving a root mean squared error of 3.6 °C. Similar machine learning models were able to predict the binding affinity of mutant WW domain sequences, brightness of fluorescent proteins, and enzymatic activity of ribozymes. Notably, the trained models are protein or nucleic acid family specific and therefore useful in the design of biopolymers with particular properties. This method provides a new tool for the in silico prediction of quantitative biophysical and organismal phenotypes directly from sequence.