Thai researchers developed MIF-MAPMS to classify candidate myelin autoantigenic peptides in multiple sclerosis by fusing peptide sequences, SMILES, descriptors, fingerprints and biomolecular-model embeddings. It improved MCC on benchmark and independent tests and released code/data, but remains an in-silico screen without T-cell or patient validation.
Key findings
- MCC ranged from 0.931-0.968 on main benchmark datasets and 0.812-0.928 on alternative/independent tests, improving on the existing method by 5.78-8.04% and 1.22-2.98%. Ablation supported multimodal fusion, and code/data were released on GitHub.
Why this matters globally
Automated peptide prioritisation could reduce costly experiments and accelerate MS immunology. Multimodal representation may generalise to vaccine and immune research, but clinical relevance requires functional assays and population-diverse validation.
Thai researcher contribution
Mahidol University, Kasetsart University and Chandrakasem Rajabhat University jointly developed the method, bioinformatics and open software, demonstrating reproducible Thai biomedical AI.
Limitations to consider
Performance depends on positive/negative definitions and benchmark independence. Homologous peptides across splits may inflate scores; an 'independent' set may still share database biases. MCC does not show calibration or experimental yield, and no wet-lab binding, T-cell or patient validation was reported.
Verify the original sources
Protein ScienceRead the original article↗DOI: 10.1002/pro.70695