Scholarlink

In the pursuit of structure-based drug discovery, the goal is to find small molecules capable of binding to a particular target protein and altering its function. Recently, deep learning (DL) has emerged as a promising approach for crafting drug-like molecules. It excels in creating compounds possessing precise biochemical characteristics while being influenced by structural features. Yet, their typical shortfall lies in the neglect of a critical element: the intrinsic physics that governs the structure and binding of molecules within real-world contexts. In this study, we explore and build on deep generative models informed by physics principles for drug discovery. These models not only consider the binding site but also incorporate physics-derived features that describe the interaction mechanism between a receptor and a ligand. We tested the proposed models by generating corresponding drug molecule candidates for a variety of protein-ligand complexes from the PDBBind dataset. On average, more than 75% of the structures generated by our hybrid model were stronger binders than the original experimental reference ligands to the protein. In addition, they had higher values of \(\varDelta G_{bind}\) (binding affinity) than molecules generated by the baseline methods by an average margin of 1.39 kcal/mol. Moreover, drug-like attributes of the generated molecules are evaluated in accordance with the Lipinski rules. To extend the analysis, their synthesizability is evaluated using ASKCOS, elevating the evaluation to a more comprehensive level. This revealed that the hybrid models notably excel in generating synthesizable molecules, with scores suggesting a higher likelihood of successful synthesis. Adherence to the Lipinski Rule of Five was also high, with compliance of 98.9%, suggesting favorable drug-like properties and a reduced risk of development failure due to poor bioavailability. This approach outperforms previous works, indicating significant improvements in drug discovery by enhancing both binding affinity and synthesizability.

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

Enhancing Drug Discovery via Physics-Guided Deep Generative Models

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Dikshant Sagar,
  • Ari Jasko,
  • Negin Forouzesh

摘要

In the pursuit of structure-based drug discovery, the goal is to find small molecules capable of binding to a particular target protein and altering its function. Recently, deep learning (DL) has emerged as a promising approach for crafting drug-like molecules. It excels in creating compounds possessing precise biochemical characteristics while being influenced by structural features. Yet, their typical shortfall lies in the neglect of a critical element: the intrinsic physics that governs the structure and binding of molecules within real-world contexts. In this study, we explore and build on deep generative models informed by physics principles for drug discovery. These models not only consider the binding site but also incorporate physics-derived features that describe the interaction mechanism between a receptor and a ligand. We tested the proposed models by generating corresponding drug molecule candidates for a variety of protein-ligand complexes from the PDBBind dataset. On average, more than 75% of the structures generated by our hybrid model were stronger binders than the original experimental reference ligands to the protein. In addition, they had higher values of \(\varDelta G_{bind}\) (binding affinity) than molecules generated by the baseline methods by an average margin of 1.39 kcal/mol. Moreover, drug-like attributes of the generated molecules are evaluated in accordance with the Lipinski rules. To extend the analysis, their synthesizability is evaluated using ASKCOS, elevating the evaluation to a more comprehensive level. This revealed that the hybrid models notably excel in generating synthesizable molecules, with scores suggesting a higher likelihood of successful synthesis. Adherence to the Lipinski Rule of Five was also high, with compliance of 98.9%, suggesting favorable drug-like properties and a reduced risk of development failure due to poor bioavailability. This approach outperforms previous works, indicating significant improvements in drug discovery by enhancing both binding affinity and synthesizability.