Interpretable Machine Learning of Amino Acid Patterns in Proteins: A Statistical Ensemble Approach

Explainable and interpretable unsupervised machine learninghelpsone to understand the underlying structure of data. We introduce anensemble analysis of machine learning models to consolidate theirinterpretation. Its application shows that restricted Boltzmann machinescompress consistently into a few bits the information stored in asequence of five amino acids at the start or end of & alpha;-helicesor & beta;-sheets. The weights learned by the machines reveal unexpectedproperties of the amino acids and the secondary structure of proteins:(i) His and Thr have a negligible contribution to the amphiphilicpattern of & alpha;-helices; (ii) there is a class of & alpha;-helicesparticularly rich in Ala at their end; (iii) Pro occupies most oftenslots otherwise occupied by polar or charged amino acids, and itspresence at the start of helices is relevant; (iv) Glu and especiallyAsp on one side and Val, Leu, Iso, and Phe on the other display thestrongest tendency to mark amphiphilic patterns, i.e., extreme valuesof an effective hydrophobicity, though they are notthe most powerful (non)hydrophobic amino acids.