Scholarlink

Context <p>In this study, we introduce a meta-learning graph neural network (GNN) combined with molecular dynamics (MD)–guided attention mechanisms to predict small-molecule inhibitors of PdK1 for termite control, addressing the limited availability of species-specific data. The method redefines structure-based virtual screening by integrating cross-species kinase-inhibition knowledge with physically plausible binding constraints derived from MD simulations.</p> Methods <p>The GNN represents kinase-ligand complexes as heterogeneous graphs, employing dual-stream message passing to capture structural and chemical interactions. Meanwhile, the MD-guided attention mechanism highlights energetically favorable contacts using a mix of learned and physics-based components. Additionally, the meta-learning framework promotes strong generalization across kinase families by training on diverse tasks and fine-tuning with task-specific support sets. The scoring function replaces empirical docking scores for predicting binding affinity, improving thermodynamic consistency, though explicit pose-level geometric validation remains a direction for future work. Experimental results on the KinomeScan and PDBbind datasets demonstrate improved prediction accuracy for termite PdK1 inhibitors compared to traditional methods. This work bridges the gap between data-driven and physics-based approaches, providing a scalable solution for pesticide discovery when target-specific data are limited. Incorporating meta-learning and MD insights improves cross-species transferability and yields interpretable attention patterns based on biophysical principles.</p>

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

Meta-learning GNN with MD-informed attention for cross-species prediction of phosphoinositide-dependent kinase-1 (PdK1) inhibitors in termite control

  • 期刊论文

  • 发表日期：

    2026年7月13日

查看全文
  • Haroon

摘要

Context

In this study, we introduce a meta-learning graph neural network (GNN) combined with molecular dynamics (MD)–guided attention mechanisms to predict small-molecule inhibitors of PdK1 for termite control, addressing the limited availability of species-specific data. The method redefines structure-based virtual screening by integrating cross-species kinase-inhibition knowledge with physically plausible binding constraints derived from MD simulations.

Methods

The GNN represents kinase-ligand complexes as heterogeneous graphs, employing dual-stream message passing to capture structural and chemical interactions. Meanwhile, the MD-guided attention mechanism highlights energetically favorable contacts using a mix of learned and physics-based components. Additionally, the meta-learning framework promotes strong generalization across kinase families by training on diverse tasks and fine-tuning with task-specific support sets. The scoring function replaces empirical docking scores for predicting binding affinity, improving thermodynamic consistency, though explicit pose-level geometric validation remains a direction for future work. Experimental results on the KinomeScan and PDBbind datasets demonstrate improved prediction accuracy for termite PdK1 inhibitors compared to traditional methods. This work bridges the gap between data-driven and physics-based approaches, providing a scalable solution for pesticide discovery when target-specific data are limited. Incorporating meta-learning and MD insights improves cross-species transferability and yields interpretable attention patterns based on biophysical principles.