A Novel Computational Machine Learning Pipeline to Quantify Similarities in Three-Dimensional Protein Structures

Abstract

Animal models are widely used during drug development. The selection of suitable animal model relies on various factors such as target biology, animal resource availability and legacy species. It is imperative that the selected animal species exhibit the highest resemblance to human, in terms of target biology as well as the similarity in the target protein. The current practice to address cross-species protein similarity relies on pair-wise sequence comparison using protein sequences, instead of the biologically relevant 3-dimensional (3D) structure of proteins. We developed a novel quantitative machine learning pipeline using 3D structure-based feature data from the Protein Data Bank, nominal data from UNIPROT and bioactivity data from ChEMBL, all of which were matched for human and animal data. Using the XGBoost regression model, similarity scores between targets were calculated and based on these scores, the best animal species for a target was identified. For real-world application, targets from an alternative source, i.e., AlphaFold, were tested using the model, and the animal species that had the most similar protein to the human counterparts were predicted. These targets were then grouped based on their associated phenotype such that the pipeline could predict an optimal animal species.