Enhanced Residue Embeddings and Ligand-Integrated Model for Prediction of Accurate Protein-Ligand Binding Residue = 향상된 잔기 임베딩과 리간드 정보를 통합한 단백질-리간드 결합 잔기 예측 모델

다국어 초록 (Multilingual Abstract)

Accurately predicting ligand-binding residues in proteins plays a crucial role in understanding molecular interaction mechanisms and contributes significantly to identifying potential drug targets and narrowing down drug candidates during drug discovery. Recently, machine learning-based models have been leveraged for protein-ligand binding residue prediction, offering time and cost savings compared to traditional experimental approaches. Methods for predicting protein-ligand binding residues can be roughly categorized into sequence-based and structure-based methods. Although structure-based methods have shown better performance than sequence-based methods for predicting protein residues where ligands bind. Structure-based methods rely on protein structure data, making their application to large-scale protein datasets both time- and resource-intensive. Due to this limitation, sequence-based methods have recently gained attention. However, most sequence-based methods exclude ligand information, even though ligand-binding residues in protein are determined by interactions with specific ligands. Therefore, to achieve more accurate prediction of protein-ligand binding residues, we propose a novel protein-ligand binding residue prediction model that constructs enhanced residue embeddings by combining a protein language model with 1D-CNN and BiLSTM, and integrates this with atom-level ligand embedding. The proposed model outperformed existing methods across all evaluation metrics. Futhermore, the results of the visualization experiments demonstrated close alignment between the predicted and actual binding residues. This highlights the model's reliability in accurately identifying ligand-binding residues.

목차 (Table of Contents)

• List of Tables ⅲ
• List of Figures ⅳ
• List of Equations ⅴ
• Abstract ⅵ
• I. Introduction 1
• II. Previous studies 4
• III. Proposed method 6
• A. Residue embedding construction 6
• 1. Initial residue embedding construction 6
• 2. Enhanced residue embedding construction 7
• B. Ligand embedding construction 8
• 1. Ligand representation 8
• 2. Embedding construction 9
• C. Interaction modeling and prediction 10
• IV. Dataset and preprocessing 12
• A. Description of the train dataset 12
• B. Description of the test dataset 12
• C. Description of the data preprocessing 13
• 1. Protein sequence length restriction 13
• 2. Protein sequence similarity 14
• 3. Filtering by binding residue count 15
• 4. Ligand parsability 15
• V. Evaluation and results 16
• A. Evaluation metrics 16
• B. Experimental results 17
• 1. Benchmark models 17
• 2. Performance comparison 18
• C. Additional analysis 20
• 1. Ablation study 20
• 2. Visualization of predicted binding residues 22
• VI. Conclusion 25
• Reference 26
• Korean Abstract 31