Scoring protein-ligand binding structures through learning atomic graphs with inter-molecular adjacency.

Area of Science:

• Biomolecular Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Artificial intelligence (AI) is increasingly applied across scientific fields, including biomolecular science.
• Accurately scoring protein-ligand binding strength is critical for computational drug discovery.
• Existing methods require improvement for efficiency and accuracy in binding affinity prediction.

Purpose of the Study:

• To develop an efficient deep learning framework for scoring protein-ligand binding structures.
• To enhance the accuracy of binding strength predictions in computational drug discovery.
• To provide interpretable insights into the AI model's predictions.

Main Methods:

• Representing protein-ligand binding structures as high-resolution atomic graphs.
• Focusing on inter-molecular interactions by defining graph edges based on multiple distance ranges.
• Employing deep learning techniques for rational graph learning and binding strength prediction.
• Conducting model-level and post-hoc analysis for interpretability.

Main Results:

• The proposed framework demonstrates competitive performance in binding structure scoring.
• The AI model effectively captures key atomic information and inter-molecular interactions.
• The framework shows promise for protein-ligand binding affinity prediction tasks.
• Interpretability analysis provides confidence in the predicted binding strengths.

Conclusions:

• The developed deep learning framework offers an efficient and accurate approach to scoring protein-ligand binding.
• This AI-driven method has the potential to significantly advance computational drug discovery.
• Further development and application of this framework are expected to benefit related scientific fields.