Glass Box Machine Learning to Aid the Design of Small Peptide Agonists from Very Sparse Data: The Example of Erythropoietin Analogues

Abstract

This paper presents a perspective on interpretable approaches to peptide design when, as is often the case, initial datais very sparse. It provides a worked example with discussion for the very earliest stage that one may can consider a“hunch phase”, when data is sparse and often circumstantial, and suspected activity and inactivity for many peptidesis more correctly described as unknown. The very earliest steps involve bioinformatics and related aspects ofcomputational chemistry approaches that are essentially standard, but their purpose is to generate all possibleinformation for a final explanatory Glass Box AI approach involving use of a Theory of Expected Informationdeveloped for sparse data which is much less commonly described. They still have some novel or unusual featureswhich make them well suited to that task, and glass box AI applied to peptides can itself be considered a recent formof bioinformatics. The present discussion can also be considered as an early step in the design of traditionaltherapeutics, i.e. small organic “in-a-pill” drugs. This is because biologically active peptides can provide clues fordesign of small organics, help establish laboratory assays, and provide important information as to the action ofagonists and antagonists, and as to safety. Also providing early data are cases where peptides have reached themarketing stage, but still have disadvantages, even being withdrawn, so increasing the demand for more traditionaltherapeutics that may not have the same problems. In other cases, there is data from viable peptide therapeuticswhere use is confined to one or few countries. Erythropoietic peptides for treatment of anemia provide an exampleof all these cases.