Improving Predictions of MHC-Bound Peptides with Geometric Deep Learning
Author Information
Author(s): Marzella Dario F., Crocioni Giulia, Radusinović Tadija, Lepikhov Daniil, Severin Heleen, Bodor Dani L., Rademaker Daniel T., Lin ChiaYu, Georgievska Sonja, Renaud Nicolas, Kessler Amy L., Lopez-Tarifa Pablo, Buschow Sonja I., Bekkers Erik, Xue Li
Primary Institution: Radboud University Medical Center
Hypothesis
Structure-based predictors may be more generalizable to unseen alleles than sequence-based approaches due to ultra-conserved MHC structures.
Conclusion
Structure-based methods using geometric deep learning improve generalizability in predicting MHC-bound peptides, outperforming sequence-based models.
Supporting Evidence
- Structure-based methods showed better generalization than sequence-based methods on unseen alleles.
- 3D-SSL outperformed sequence-based methods trained on significantly more data.
- Geometric deep learning methods demonstrated resilience to biases in binding data.
Takeaway
This study shows that using 3D models and geometric deep learning helps predict how well peptides bind to MHC molecules, which is important for vaccines and therapies.
Methodology
The study developed end-to-end geometric deep learning methods to analyze structural features at atom or residue level for predicting peptide-MHC binding.
Potential Biases
Sequence-based methods may inherit biases from training data, affecting their predictions.
Limitations
The study primarily focuses on MHC class I and may not directly apply to MHC class II complexes.
Participant Demographics
The study involved data from 110 HLA alleles and peptides of varying lengths.
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website