[MHC-I epitope presentation prediction based on transfer learning].

Area of Science:

• Computational biology
• Immunoinformatics
• Cancer immunotherapy

Background:

• Accurate prediction of T cell epitopes is crucial for developing neoantigen-based cancer immunotherapies.
• Mass spectrometry (MS) data aids in training epitope prediction models, but limited data for many human leukocyte antigens (HLAs) hinders model reliability.

Purpose of the Study:

• To develop a robust epitope presentation prediction model using transfer learning to overcome data limitations.
• To enhance the accuracy of predicting T cell epitopes for improved cancer immunotherapy targets.

Main Methods:

• Employed transfer learning to pre-train a model on common features from mixed allele-specific epitopes.
• Fine-tuned the pre-trained model using allele-specific epitopes to create the Pluto prediction model.
• Evaluated Pluto's performance against existing methods on validation and external datasets.

Main Results:

• Pluto demonstrated superior performance with an average 0.1% positive predictive value (PPV) of 0.4255 on external datasets, outperforming non-pretrained models and popular tools like MixMHCpred and NetMHCpan.
• The model achieved a 0.078 margin improvement over a non-pretrained model on the validation dataset.
• Pluto identified more neoantigens than other tools when predicting immunogenicity.

Conclusions:

• Pluto, a transfer learning-based model, significantly enhances epitope presentation prediction accuracy for neoantigen discovery.
• The model's improved performance holds promise for advancing the development of targeted cancer immunotherapies.
• Pluto is publicly available, facilitating further research and application in the field.