Prediction of protein mononucleotide binding sites using AlphaFold2 and machine learning

Area of Science:

• Biochemistry
• Computational Biology
• Structural Biology

Background:

• Protein-ligand interactions are fundamental to cellular processes.
• Accurate prediction of nucleotide-binding sites is crucial for understanding protein function and drug discovery.
• Existing methods often lack comprehensive accuracy for diverse nucleotide types.

Purpose of the Study:

• To develop and evaluate a robust system for predicting protein binding sites for five key mononucleotides: AMP, ADP, ATP, GDP, and GTP.
• To compare the performance of machine learning and template-based prediction methods.
• To create an ensemble model combining multiple prediction strategies for improved accuracy.

Main Methods:

• Developed two machine learning (ML) predictors (convolutional neural network, gradient boosting machine) utilizing protein structure features.
• Created two template-based predictors leveraging sequence and structure alignment.
• Employed data augmentation with similar ligand structures for enhanced ML model training.
• Utilized AlphaFold2 for generating accurate protein structure models.
• Constructed an ensemble predictor integrating the outputs of the four individual predictors.

Main Results:

• The template-based predictor using structure alignment demonstrated superior performance among individual methods.
• The ensemble predictor achieved the highest overall performance, yielding an Area Under the Curve (AUC) of 0.958 across all mononucleotides.
• Data augmentation and AlphaFold2-derived features improved ML predictor performance.

Conclusions:

• The developed ensemble system provides a highly accurate method for predicting mononucleotide-binding sites in proteins.
• Combining diverse prediction strategies (ML and template-based) is effective for enhancing prediction accuracy.
• This system has significant potential for applications in structural biology, drug discovery, and functional genomics.