Discordant results among major histocompatibility complex binding affinity prediction tools

Background: Human leukocyte antigen (HLA) alleles are critical components of the immune system’s ability to recognize and eliminate tumors infections. A large number machine learning-based major histocompatibility complex (MHC) binding affinity (BA) prediction tools have been developed widely used for both investigational therapeutic applications, so it is important explore differences in tool outputs. Methods: We examined predictions four popular (netMHCpan, HLAthena, MHCflurry, MHCnuggets) across a range possible peptide sources (human, viral, randomly generated) MHC class I alleles. Results: uncovered inconsistencies BA, allele promiscuity relationship between physical properties peptides by source BA predictions, as well quality training data. We found amount data does not explain yet all tools, predicted quantities similar human viral proteomes. Lastly, we find associated with allele-specific predictions. Conclusions: Our work raises fundamental questions about fidelity peptide-MHC their real-world implications. The use these theoretical worrying, substantial differentiate potential foreign versus self-antigens. Evaluating more viruses – bacteria, fungi, other pathogens linking analyses metrics such evolutionary distance may give greater insight into HLA evolution disease.