A Novel Weighted Ensemble Framework of Transformer and Deep Q-Network for ATP-Binding Site Prediction Using Protein...

Area of Science:

• Computational Biology
• Biochemistry
• Bioinformatics

Background:

• Adenosine triphosphate (ATP) is crucial for cellular energy and signaling.
• Identifying ATP-binding sites is vital for understanding protein function and drug discovery.
• Current methods require improvement for accurate ATP-binding site prediction.

Purpose of the Study:

• To develop a novel hybrid deep learning framework for accurate ATP-binding site prediction using protein sequence information.
• To integrate Transformer-based and deep Q-network (DQN)-inspired models for synergistic prediction.
• To uncover biological mechanisms of protein-ATP interactions.

Main Methods:

• Utilized Evolutionary Scale Modeling 2 (ESM-2) for high-level protein embeddings.
• Implemented a Transformer model with a local-global dual-attention mechanism.
• Employed a DQN-inspired classifier for residue-level sequential prediction.
• Applied a weighted ensemble strategy with cross-validation for optimal performance.

Main Results:

• The hybrid deep learning framework demonstrated satisfactory performance on benchmark datasets.
• The model effectively captured both local and long-range dependencies in protein sequences.
• Analysis revealed insights into protein-ATP interaction mechanisms, including structural motifs and binding patterns.

Conclusions:

• The developed framework offers a significant advancement in predicting ATP-binding sites.
• The study provides a deeper understanding of protein-ligand recognition mechanisms.
• This research supports large-scale functional annotations critical for systems biology and drug target discovery.