Artificial intelligence perceives marginal gains from MHC class I haplotype data in antigen presentation predictions.

Artificial Intelligence offers valuable tools for scientific discovery, but when used improperly, it can cause blunders. In this paper, we report findings related to the role of Major Histocompatibility Complexes (MHCs) in epitope prediction. Through a serendipitous programmatic error, we observed that methods like TransPHLA yield similar results on both training and testing datasets when only peptides are used, without knowing the specifics of MHC alleles. To further investigate, we developed a new dataset free of the artifacts present in the original. Results from experiments on this new dataset would suggest a field led astray by AI hype, as understanding the MHC allele may not be as critical for epitope prediction as previously thought. Yet, further experiments with synthetic datasets reveal the limitations of current AI applications in biology, paving the way for a stricter multidisciplinary approach.