Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Tensor-DTI: Enhancing Biomolecular Interaction Prediction with Contrastive Embedding Learning

Manel Gil-Sorribes, Júlia Vilalta-Mor, Isaac Filella-Mercè, Robert Soliva, Álvaro Ciudad, Víctor Guallar, Alexis Molina

TL;DR

Tensor-DTI addresses the challenge of robust drug-target interaction (DTI) and drug-target affinity (DTA) prediction by integrating multimodal embeddings from molecular graphs, protein language models, and pocket predictions within a contrastive siamese framework. The model optimizes a combined objective that fuses contrastive embedding separation with supervised classification or regression losses, and it further incorporates pocket embeddings via PickPocket and GearNet to enable site-specific interaction modeling, guarded by confidence and unfamiliarity reliability signals. Across BIOSNAP, BindingDB, and DAVIS, Tensor-DTI achieves state-of-the-art or competitive results and demonstrates strong generalization to unseen drugs and targets, as well as scalable performance on billion-scale libraries. The approach also extends to peptide-protein and protein-RNA interactions, offering a versatile, reliability-aware tool for broad biomolecular discovery and virtual screening workflows.

Abstract

Accurate drug-target interaction (DTI) prediction is essential for computational drug discovery, yet existing models often rely on single-modality predefined molecular descriptors or sequence-based embeddings with limited representativeness. We propose Tensor-DTI, a contrastive learning framework that integrates multimodal embeddings from molecular graphs, protein language models, and binding-site predictions to improve interaction modeling. Tensor-DTI employs a siamese dual-encoder architecture, enabling it to capture both chemical and structural interaction features while distinguishing interacting from non-interacting pairs. Evaluations on multiple DTI benchmarks demonstrate that Tensor-DTI outperforms existing sequence-based and graph-based models. We also conduct large-scale inference experiments on CDK2 across billion-scale chemical libraries, where Tensor-DTI produces chemically plausible hit distributions even when CDK2 is withheld from training. In enrichment studies against Glide docking and Boltz-2 co-folder, Tensor-DTI remains competitive on CDK2 and improves the screening budget required to recover moderate fractions of high-affinity ligands on out-of-family targets under strict family-holdout splits. Additionally, we explore its applicability to protein-RNA and peptide-protein interactions. Our findings highlight the benefits of integrating multimodal information with contrastive objectives to enhance interaction-prediction accuracy and to provide more interpretable and reliability-aware models for virtual screening.

Tensor-DTI: Enhancing Biomolecular Interaction Prediction with Contrastive Embedding Learning

TL;DR

Tensor-DTI addresses the challenge of robust drug-target interaction (DTI) and drug-target affinity (DTA) prediction by integrating multimodal embeddings from molecular graphs, protein language models, and pocket predictions within a contrastive siamese framework. The model optimizes a combined objective that fuses contrastive embedding separation with supervised classification or regression losses, and it further incorporates pocket embeddings via PickPocket and GearNet to enable site-specific interaction modeling, guarded by confidence and unfamiliarity reliability signals. Across BIOSNAP, BindingDB, and DAVIS, Tensor-DTI achieves state-of-the-art or competitive results and demonstrates strong generalization to unseen drugs and targets, as well as scalable performance on billion-scale libraries. The approach also extends to peptide-protein and protein-RNA interactions, offering a versatile, reliability-aware tool for broad biomolecular discovery and virtual screening workflows.

Abstract

Accurate drug-target interaction (DTI) prediction is essential for computational drug discovery, yet existing models often rely on single-modality predefined molecular descriptors or sequence-based embeddings with limited representativeness. We propose Tensor-DTI, a contrastive learning framework that integrates multimodal embeddings from molecular graphs, protein language models, and binding-site predictions to improve interaction modeling. Tensor-DTI employs a siamese dual-encoder architecture, enabling it to capture both chemical and structural interaction features while distinguishing interacting from non-interacting pairs. Evaluations on multiple DTI benchmarks demonstrate that Tensor-DTI outperforms existing sequence-based and graph-based models. We also conduct large-scale inference experiments on CDK2 across billion-scale chemical libraries, where Tensor-DTI produces chemically plausible hit distributions even when CDK2 is withheld from training. In enrichment studies against Glide docking and Boltz-2 co-folder, Tensor-DTI remains competitive on CDK2 and improves the screening budget required to recover moderate fractions of high-affinity ligands on out-of-family targets under strict family-holdout splits. Additionally, we explore its applicability to protein-RNA and peptide-protein interactions. Our findings highlight the benefits of integrating multimodal information with contrastive objectives to enhance interaction-prediction accuracy and to provide more interpretable and reliability-aware models for virtual screening.
Paper Structure (42 sections, 14 equations, 7 figures, 25 tables)

This paper contains 42 sections, 14 equations, 7 figures, 25 tables.

Table of Contents

  1. Introduction
  2. Results and Discussion
  3. Benchmarking Tensor-DTI
  4. Effectiveness of the Contrastive Learning Approach and Evaluation on DUD-E Dataset
  5. Tensor-DTI Performs Well on Affinity Prediction Datasets
  6. DTI and DTA Assessments on Low-Leakage Datasets
  7. Tensor-DTI Allows Pocket Specificity with the Addition of Pocket Embeddings
  8. Assessment of Binding Across Cryptic Site Inhibitor Binding Predictions in CDK2 and RET Kinases
  9. Generalization and Reliability in a Large-Scale CDK2 Virtual Screening Case Study
  10. A Comparative Enrichment Analysis of Tensor-DTI, Glide, and Boltz-2
  11. Broadening the Scope of Biomolecular Interaction Predictions
  12. Conclusion
  13. Methods
  14. Model architecture
  15. Training procedure
  16. ...and 27 more sections

Figures (7)

  • Figure 1: t-SNE visualization of protein and drug embeddings before (left plot) and after (right plot) applying Tensor-DTI with contrastive learning. The visualization corresponds to the B-raf Kinase protein, one of the targets from the test split.
  • Figure 2: Structural arrangements from CDK2 and RET kinases in holo and apo states for their corresponding cryptic pockets.
  • Figure 3: CDK2 screening results with Tensor-DTI. (A-B) Glide gscore distributions for models trained with and without CDK2, compared against experimental and random ligands. (C-D) Unfamiliarity-based reliability distributions (filtered for unfamiliarity $<$ 1.0), showing that Tensor-DTI remains confident and chemically consistent even when CDK2 is excluded. (E-F) Ligand efficiency distributions (-gscore per heavy atom), illustrating that the model preserves balanced, size-normalized scoring behavior.
  • Figure 4: Tensor-DTI architecture with pocket embeddings. The model extends the base architecture by incorporating binding pocket representations, enabling site-specific interaction modeling. The protein shown is PDBID: 5ISX. The SaProt image is adapted from saprot, and the Pickpocket image is adapted from gearnet.
  • Figure 5: Tensor-DTI architecture. A siamese dual-encoder processes multimodal embeddings from drugs and proteins, using contrastive learning to refine the interaction space. The protein shown is 5ISX. The SaProt image is adapted from saprot.
  • ...and 2 more figures