Finetuning GearNet-Edge Embedding, utilising Various MLPs to Predict Peptide Binders to MHC Class 1 Complex

Abstract

The binding of peptides to the MHC class 1 complex allows the CTL cells in the immune system to eliminate the cell if the LTC receptor recognises it as foreign. Therefore, predicting the peptides that can bind to the MCH-I complex is a significant step in personalised vaccine development. However, finding the right approach to maximise the predicting capabilities takes time and effort. Here, we show an alternative approach to predict the binding of peptides to the MHC complex, finetuning GearNet-Edge embedding for the amino acids within the peptides binding to the MHC complex utilising various MLPs. Here, we show that the finetuning approach of the GearNet-Edge embedding and concatenated one-hot embedding do not outperform the baseline method of one-hot encoding. We found that reducing the embedding size with a linear transformation retains the most information; by concatenating sequence-based information, we further improved the performance of the binding prediction but not enough to outperform the baseline method. The findings suggest possibilities for improvement, such as exploring alternative hyperparameters and MLP structures, leveraging the entire HLA dataset, and retraining GearNet-Edge to this specific task. The study provides valuable insights into GearNet-Edge's predictive capabilities and could direct future research toward optimising its performance in predicting pMHC binders.