High-order neural networks and kernel methods for peptide-MHC binding prediction.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning

Background:

• Accurate peptide-protein binding prediction is crucial for vaccine development.
• Existing computational methods struggle with nonlinear, high-order dependencies between amino acid positions, leading to suboptimal peptide rankings.
• This limitation hinders the effective search and design of clinical peptide vaccines.

Purpose of the Study:

• To develop advanced computational methods for predicting major histocompatibility complex-peptide binding.
• To address the limitations of current models in capturing nonlinear, high-order feature interactions.
• To improve the quality and accuracy of peptide binding predictions for vaccine design.

Main Methods:

• Implementation of nonlinear high-order machine learning approaches.
• Utilizing high-order neural networks (HONNs), including deep extensions.
• Employing high-order kernel support vector machines for prediction.
• Investigating the impact of pre-training with high-order semi-restricted Boltzmann machines on HONN performance.

Main Results:

• The proposed high-order methods demonstrate improved binding prediction quality compared to existing methods.
• A significant performance gain of 25-40% was observed on benchmark and reference datasets.
• Pre-training HONNs with high-order semi-restricted Boltzmann machines was shown to significantly enhance performance.
• Shallow HONNs outperformed popular pre-trained deep neural networks on most tasks, highlighting the effectiveness of modeling high-order interactions.

Conclusions:

• Nonlinear high-order machine learning methods, particularly HONNs, are effective for predicting major histocompatibility complex-peptide binding.
• Modeling high-order feature interactions is critical for improving prediction accuracy.
• These advanced methods offer a significant improvement for clinical peptide vaccine search and design.