Scholarlink

Predicting drug-target affinity (DTA) is crucial for drug discovery, yet existing deep learning approaches are often constrained by overly simplified feature fusion and reliance on single information sources. We introduce VoluAlign-DTA—a two-stage multi-knowledge alignment pretraining framework. It systematically integrates seven information sources across different scales: For drugs, the model processes not only SMILES sequences, molecular graphs, and fused fingerprints, but also innovatively incorporates text descriptions rich in expert knowledge. For proteins, it simultaneously encodes amino acid sequences, n-gram features, and hierarchical taxonomic annotations (HTA) defining biological functions. The core pre-training stage employs a Gramian volume-based global alignment mechanism. By minimizing the geometric volume formed by all modal embeddings, it achieves deep semantic alignment across heterogeneous information sources within a unified representation space, thereby capturing higher-order dependencies. The advanced Mamba encoder efficiently parses semantic sequences, supplemented by a gradient-driven adaptive modality selection mechanism to enhance model robustness. Extensive experiments across seven benchmark datasets demonstrate VoluAlign-DTA achieves superior overall performance compared to other baseline methods. Furthermore, we successfully identified three potential lung cancer inhibitors using VoluAlign-DTA, validated by corresponding clinical trials.

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

VoluAlign-DTA: Enhancing Prediction of Drug-Target Binding Affinity by Integrating Geometric Alignment with Dynamic Multimodal Management

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Mengfan Yuan,
  • Yuan Zhang,
  • Bing Jia,
  • Baoqi Huang

摘要

Predicting drug-target affinity (DTA) is crucial for drug discovery, yet existing deep learning approaches are often constrained by overly simplified feature fusion and reliance on single information sources. We introduce VoluAlign-DTA—a two-stage multi-knowledge alignment pretraining framework. It systematically integrates seven information sources across different scales: For drugs, the model processes not only SMILES sequences, molecular graphs, and fused fingerprints, but also innovatively incorporates text descriptions rich in expert knowledge. For proteins, it simultaneously encodes amino acid sequences, n-gram features, and hierarchical taxonomic annotations (HTA) defining biological functions. The core pre-training stage employs a Gramian volume-based global alignment mechanism. By minimizing the geometric volume formed by all modal embeddings, it achieves deep semantic alignment across heterogeneous information sources within a unified representation space, thereby capturing higher-order dependencies. The advanced Mamba encoder efficiently parses semantic sequences, supplemented by a gradient-driven adaptive modality selection mechanism to enhance model robustness. Extensive experiments across seven benchmark datasets demonstrate VoluAlign-DTA achieves superior overall performance compared to other baseline methods. Furthermore, we successfully identified three potential lung cancer inhibitors using VoluAlign-DTA, validated by corresponding clinical trials.