Scholarlink

Virtual screening (VS) in drug discovery faces a key challenge: balancing computational efficiency with predictive accuracy. We propose EquiMapLE, a hierarchical framework combining speed and precision. It introduces Equivariant Fingerprints (EFPs) derived from SE(3)-equivariant transformers via invariant attention pooling, encoding 3D molecular geometry into compact vectors. The framework uses a 2D fingerprint pre-screening stage followed by EFP-based ranking, optimized by a Learning-to-Rank (LTR) model that fuses multi-metric scores. Experiments on CASF-2016 and DUD-E benchmarks show EquiMapLE outperforms 2D/3D similarity and docking methods in ranking accuracy (EF1%, NDCG@10) and computational throughput, offering a scalable solution to the speed-accuracy trade-off in VS.

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

EquiMapLE: High-Fidelity and High-Throughput Molecular Analogue Screening with Equivariant Fingerprints and Learned Fusion

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Xinyu Lu,
  • Lingge Gao,
  • Yuhao Zhang,
  • Yanhui Gu,
  • Weiguang Qu,
  • Peirong Ma,
  • Yafei Li,
  • He Huang

摘要

Virtual screening (VS) in drug discovery faces a key challenge: balancing computational efficiency with predictive accuracy. We propose EquiMapLE, a hierarchical framework combining speed and precision. It introduces Equivariant Fingerprints (EFPs) derived from SE(3)-equivariant transformers via invariant attention pooling, encoding 3D molecular geometry into compact vectors. The framework uses a 2D fingerprint pre-screening stage followed by EFP-based ranking, optimized by a Learning-to-Rank (LTR) model that fuses multi-metric scores. Experiments on CASF-2016 and DUD-E benchmarks show EquiMapLE outperforms 2D/3D similarity and docking methods in ranking accuracy (EF1%, NDCG@10) and computational throughput, offering a scalable solution to the speed-accuracy trade-off in VS.