Scholarlink

The discovery of new drugs for tuberculosis remains a significant challenge, particularly due to the rising number of multidrug-resistant cases. Traditional molecular docking approaches, while effective, are computationally expensive, especially when employing Fully-Flexible Receptor (FFR) models. This study proposes a novel deep learning-based approach to filter candidate ligands before docking, thereby reducing computational costs without compromising accuracy. We introduce a 3D Deep Neural Network (DNN) trained on molecular docking results obtained from an FFR model of the InhA enzyme. The model utilizes structured 3D charge density data from ligand-receptor interactions to predict Free Energy of Binding (FEB). Training was conducted using docking results from 15 ligands co-crystallized with InhA structures available in the Protein Data Bank (PDB), for model optimization, and 61 ZINC ligands for final evaluation. The proposed 3D DNN achieved a Spearman correlation coefficient of 0.6133 with actual docking results and a mean absolute error (MAE) of 1.8392 kcal/mol. Compared to traditional docking workflows, our method significantly reduces computational costs, predicting FEB values in around 3.73 s per ligand. This work demonstrates the feasibility of integrating deep learning with molecular docking to improve rational drug design.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

From Docking to Deep Learning: A 3D Neural Network for Accelerating Ligand Screening

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Duncan Dubugras A. Ruiz,
  • Renata De Paris,
  • Christian Vahl Quevedo,
  • Osmar Norberto de Souza

摘要

The discovery of new drugs for tuberculosis remains a significant challenge, particularly due to the rising number of multidrug-resistant cases. Traditional molecular docking approaches, while effective, are computationally expensive, especially when employing Fully-Flexible Receptor (FFR) models. This study proposes a novel deep learning-based approach to filter candidate ligands before docking, thereby reducing computational costs without compromising accuracy. We introduce a 3D Deep Neural Network (DNN) trained on molecular docking results obtained from an FFR model of the InhA enzyme. The model utilizes structured 3D charge density data from ligand-receptor interactions to predict Free Energy of Binding (FEB). Training was conducted using docking results from 15 ligands co-crystallized with InhA structures available in the Protein Data Bank (PDB), for model optimization, and 61 ZINC ligands for final evaluation. The proposed 3D DNN achieved a Spearman correlation coefficient of 0.6133 with actual docking results and a mean absolute error (MAE) of 1.8392 kcal/mol. Compared to traditional docking workflows, our method significantly reduces computational costs, predicting FEB values in around 3.73 s per ligand. This work demonstrates the feasibility of integrating deep learning with molecular docking to improve rational drug design.