GADTI: A Geometry-Aware Dual-Tower Framework for Drug-Target Interaction Prediction

Zeng, Pan, Liu, Tian, Meng, Yajie, Tang, Xianfang, Cui, Feifei, Zhang, Zilong, Xu, Junlin and Wu, Taoyang ORCID: https://orcid.org/0000-0002-2663-2001 (2027) GADTI: A Geometry-Aware Dual-Tower Framework for Drug-Target Interaction Prediction. Expert Systems with Applications, 334 (A). ISSN 0957-4174

Preview

PDF (GADTI_R3) - Accepted Version Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (13MB) | Preview

Abstract

Accurate prediction of drug–target interactions (DTI) is essential for virtual screening and early-stage drug discovery, as it helps identify potential therapeutic compounds and narrow the experimental search space. Existing geometric structure-based DTI models have attempted to incorporate three-dimensional coordinates or spatial relationships to better characterize molecular interactions. However, they still face challenges in fully characterizing the spatial constraints involved in real binding processes. First, the alignment between full-length protein semantics and local binding-pocket information remains insufficient, which may either introduce noise from non-binding regions or discard useful global evolutionary context. Second, chemical semantic features and three-dimensional geometric features lie in heterogeneous feature spaces, making simple feature concatenation inadequate for precise cross-modal fusion. Third, current geometric structure models may still lack rotational and translational consistency, limiting their ability to generalize to realistic binding patterns and novel scaffold samples. To address the above issues, this study proposes GADTI, a geometry-aware dual-tower framework for DTI prediction. First, GADTI introduces a pocket-centric global-local alignment strategy, which maps full-length protein semantics to predicted binding pocket residues, thereby preserving global evolutionary information while focusing on local structural regions. Second, GADTI adopts a semantic-geometric dual-tower architecture to represent the 1D/3D features of drugs and targets, and uses hierarchical bidirectional cross-attention to achieve modality alignment and interaction, rather than relying on simple concatenation. Meanwhile, a PaiNN-based E(3)-equivariant geometric encoder is introduced to learn spatial features from three-dimensional coordinates that are consistent under rotational and translational transformations. The experimental results show that GADTI achieves competitive predictive performance on the structurally filtered Davis and BindingDB benchmark samples and maintains relatively stable generalization under the scaffold-split setting. These findings indicate that, for DTI samples with complete structural and semantic inputs, pocket alignment, semantic–geometric fusion, and equivariant geometric encoding contribute to enhancing the model’s ability to learn structure-aware representations of drug–target interactions.

Item Type:	Article
Additional Information:	Data and software availability: The raw Davis and BindingDB datasets used in this study are publicly accessible. The curated datasets, including the optimized 3D drug conformations and the corresponding AF2 unbound protein structures, along with the source code for the GADTI framework, are freely available on GitHub at https://github.com/sharpuser1122/GADTI. Detailed documentation, including data preprocessing scripts, model training procedures, and inference guidelines, is provided within the repository to ensure full reproducibility.
Uncontrolled Keywords:	drug-target interaction,geometric deep learning,e(3)-equivariant geometric encoder,hierarchical cross-attention,3d spatial information
Faculty \ School:	Faculty of Science > School of Computing Sciences
UEA Research Groups:	Faculty of Science > Research Centres > Centre for Ecology, Evolution and Conservation Faculty of Science > Research Groups > Computational Biology Faculty of Science > Research Groups > Data Science and AI Faculty of Science > Research Groups > Health Computing
Related URLs:	https://www.sciencedirect.com/science/ar...
Depositing User:	LivePure Connector
Date Deposited:	25 Sep 2026 08:26
Last Modified:	25 Sep 2026 10:31
URI:	https://ueaeprints.uea.ac.uk/id/eprint/104638
DOI:

Downloads

Actions (login required)

View Item