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

AI summary

Overview

• Research area: Computational drug discovery, specifically machine-learning models for drug-target interaction (DTI) and drug-target affinity (DTA) prediction, with extensions to peptide-protein, protein-RNA, and drug-RNA interactions.
• Technical level: Advanced — the paper assumes familiarity with contrastive learning, siamese dual-encoder architectures, graph neural networks, protein language models, and molecular docking.
• Scope: The paper introduces Tensor-DTI, a contrastive multimodal framework that combines molecular graphs, protein language models, and binding-site predictions, and evaluates it across classification benchmarks, affinity benchmarks, cryptic-pocket case studies, and billion-scale virtual screening on CDK2, AChE, and MAO-A.

What This Paper Is About

Most existing DTI models rely on a single type of molecular representation — either predefined descriptors, sequence embeddings, or whole-protein representations — and often ignore localized binding-site information, which limits their ability to generalize to unseen drugs and targets. Tensor-DTI addresses this by fusing several complementary representations (molecular graphs, protein language model embeddings, and predicted binding-pocket embeddings) into one shared latent space and training a siamese dual-encoder with a contrastive objective so that interacting pairs are pulled together and non-interacting pairs are pushed apart. The goal is a model that is both more accurate and more interpretable for large-scale virtual screening, including when the target of interest was withheld from training.

Key Contributions

1. A multimodal contrastive framework for DTI/DTA. Tensor-DTI integrates embeddings from molecular graphs, protein language models (including ESM-2 and SaProt in the experiments), and binding-site predictions from PickPocket into a shared latent space, trained with a siamese dual-encoder under a contrastive objective.
2. Binding-site-aware prediction via pocket embeddings. Pocket embeddings derived from PickPocket, a hybrid of protein language models and GearNet-based structural message passing, are combined with full-protein representations so the model can distinguish functionally relevant sites (including cryptic and allosteric pockets) from nonspecific contacts.
3. Benchmarking under both standard and leakage-controlled conditions. The model is evaluated on BIOSNAP, BindingDB, DAVIS, unseen-drug and unseen-target splits, DUD-E, TDC-DG, PLINDER, LP-PDBBind, and PDBBind-Opt.
4. Prospective, large-scale screening evaluation. A virtual screen of the Enamine REAL 5B library against CDK2, plus enrichment comparisons against Glide docking and Boltz-2 on CDK2, AChE, and MAO-A, with a reliability filter based on an "unfamiliarity" metric from a molecular autoencoder.

Main Findings

• Top AUPR on all standard DTI benchmarks: Tensor-DTI reached 0.903 ± 0.003 on BIOSNAP, 0.699 ± 0.002 on BindingDB, and 0.547 ± 0.006 on DAVIS. On BindingDB the margin over alternatives was +7.1 over ConPLex, +10.1 over MolTrans, and +9.7 over EnzPred-CPI. Test-set imbalance was roughly 1:1 for BIOSNAP, 1:6 for BindingDB, and 1:19 for DAVIS.
• Generalization to unseen entities: On unseen drugs Tensor-DTI scored 0.888 ± 0.002 AUPR; on unseen targets it scored 0.839 ± 0.003, which the authors describe as comparable to ConPLex (0.842 ± 0.006) since the difference lies within the margin of error.
• Contrastive training produces structured embeddings: On the DUD-E kinase family test, average AUPR was 0.686 ± 0.006 across five runs. A t-SNE visualization for B-raf Kinase showed actives clustering closer to their protein targets after contrastive training than before.
• Competitive DTA performance: On the TDC-DG benchmark Tensor-DTI achieved a PCC of 0.580 ± 0.004, versus 0.538 ± 0.008 for ConPLex, 0.433 ± 0.010 for MMD, 0.432 ± 0.010 for CORAL, 0.427 ± 0.012 for ERM, and 0.425 ± 0.010 for MTL. Morgan fingerprints for small molecules and ESM-2 embeddings for proteins gave the strongest performance on this dataset.
• Pocket similarity can act as a shortcut: On PLINDER (positive pairs only, negatives constructed by the authors), random-negative evaluation gave 0.785 ± 0.002 AUPR, while requiring structural dissimilarity between pockets dropped it to 0.754 ± 0.005. Ablating pocket embeddings in the dissimilar-negative setting reduced it further to 0.739, indicating the model relies on pocket cues.
• Low-leakage affinity results: On LP-PDBBind, Tensor-DTI reached PCC 0.565 ± 0.004 and RMSE 1.620 ± 0.024 for Kd, but lower PCC (0.528 ± 0.013) and higher RMSE (2.122 ± 0.032) for ΔG. On PDBBind-Opt, peptides reached PCC 0.679 ± 0.014 and RMSE 1.175 ± 0.020, while molecule-protein interactions reached PCC 0.750 ± 0.005 and RMSE 1.335 ± 0.011 on a random split — dropping to PCC 0.493 ± 0.005 and RMSE 1.545 ± 0.006 under the no-leakage split of Li et al. (2024).
• Cryptic-pocket behavior is mixed: Tensor-DTI correctly rejected ATP binding to the closed ATP site in CDK2 structure 3FWQ and correctly predicted CAM4066 as a binder in the cryptic pocket of 5CU3. For RET it correctly predicted LOXO-292 and BLU-667 binding to the cryptic site in 7JU5 with higher confidence than for the active site, but incorrectly predicted binding in the active site 2IVS, and failed to identify AMP as an active-site binder of 2IVS. The authors describe a consistent bias toward cryptic-site recognition.
• Billion-scale CDK2 screening transfers without CDK2 in training: Against the Enamine REAL 5B library, the model recovered the expected activity landscape whether or not CDK2 was in the training data, with predicted actives showing Glide gscores that overlapped experimental ligands and shifted left relative to random compounds. Predicted positives also showed a right-shift toward higher ligand efficiency. Populations were drawn from 100,000 top-scoring and 100,000 bottom-scoring molecules, alongside 85,000 random compounds and 817 experimental ligands, then filtered by Glide gscore availability and unfamiliarity < 1.0.
• The pocket-aware variant was unstable at screening scale: A parallel pocket-aware screening campaign did not converge, showing broader, noisier Glide gscore distributions and systematically higher unfamiliarity values; the authors attribute this to an insufficient pocket-level dataset for generalization at inference scale.
• Enrichment versus docking and co-folding: On CDK2, Boltz-2 gave the strongest early enrichment for recovering 1%, 5%, 20%, and 50% of actives, with Tensor-DTI-c consistently second best; Tensor-DTI required markedly fewer compounds than either Glide or Boltz-2 to recover the full set of actives. Removing CDK2 from training weakened early enrichment but Tensor-DTI-nc still beat random ranking and stayed close to Glide at moderate recall. On AChE, all cholinesterase-family interactions were removed from training; the three methods behaved similarly at the lowest recall thresholds, Glide required the smallest fraction of the library to recover larger fractions of actives, with Tensor-DTI following closely and Boltz-2 lagging. On MAO-A, with all oxidase-family interactions removed, Boltz-2 slightly outperformed at the very lowest recall thresholds, but Tensor-DTI was the most efficient option from roughly 5% to 50% recall. The authors note that no true active appeared in the SMPBind I training set used for the enrichment models.
• Performance extends beyond small molecules: Tensor-DTI reached 0.953 ± 0.001 AUPR on the Propedia peptide-protein dataset and 0.916 ± 0.008 AUPR on the CoPRA protein-RNA dataset. On PRA310 it reached PCC 0.631 ± 0.111 for Kd and 0.621 ± 0.052 for ΔG, with RMSEs of 1.443 ± 0.232 and 1.910 ± 0.212. On drug-RNA pairs from PDBBind it reached PCC 0.792 ± 0.015 and RMSE 1.684 ± 0.038, outperforming a one-hot baseline (PCC 0.633 ± 0.018, RMSE 1.738 ± 0.036).

Methodology in Plain English

The researchers built a model that looks at a drug-target pair from several angles at once. For the drug, it uses molecular graph embeddings and pretrained molecular embeddings; for the protein, it uses embeddings from protein language models; and when available it adds a "pocket" embedding that describes the specific binding site rather than the whole protein, generated by PickPocket from ESM-2 features and GearNet structural message passing. These representations are combined into a shared latent space.

The architecture is a siamese dual-encoder — two copies of the same encoder, one for each side of the pair — and training uses a contrastive objective, which means the model is explicitly taught to make interacting drug-target pairs land close together in the embedding space and non-interacting pairs land far apart. Ablation studies (Appendix D) found that pretrained molecular embeddings for drugs plus structural embeddings for proteins worked best for DTI, while the best embeddings for DTA depended on the dataset. The authors also trained models that deliberately excluded the target (or the target's whole protein family) from training, so that screening results reflect genuine generalization rather than memorization. For the CDK2 screen they added a reliability filter: a molecular autoencoder computes an "unfamiliarity" score for each compound, following van Tilborg et al. (2025), and only compounds with unfamiliarity < 1.0 were retained for the main analyses.

Why This Matters

The paper argues that DTI models which only see whole-protein sequences miss the localized binding-site information that determines selectivity, and that models trained on structural data suffer from scarce and leaky datasets. Tensor-DTI's combination of multimodal embeddings, a contrastive objective, and pocket-aware representations is presented as a way to improve both accuracy and reliability while giving screening results that are interpretable and quality-controlled rather than opaque.

Real-world applications:

• Early-stage hit discovery: rapidly ranking billion-scale make-on-demand libraries such as Enamine REAL (over 70 billion readily synthesizable compounds) and ZINC22 (more than 97 billion molecules) where docking alone scales to only around a few million compounds.
• Knockout-target screening: using models trained without the target to score novel proteins when experimental or structural data for that target is unavailable.
• Alternative binding-mode drug design: prioritizing allosteric and cryptic pockets, as examined for CDK2 and RET kinases, where ATP-competitive inhibitors face selectivity and resistance problems.
• Beyond small molecules: triaging peptide-protein, protein-RNA, and drug-RNA interactions for biologics and RNA therapeutics.

Industry relevance: the authors are affiliated with Nostrum Biodiscovery, the Barcelona Supercomputing Center, the University of Barcelona, ICREA, and Almirall S.A., and the study explicitly benchmarks Tensor-DTI against Glide docking and Boltz-2 co-folding in terms of screening budget — the practical cost of finding actives — which is the metric that matters for industrial virtual screening campaigns.

Future Directions

• Improving the pocket-aware variant: the large-scale CDK2 screen with pocket embeddings failed to converge, and the authors attribute this to an insufficient pocket-level dataset. Scaling that data is an explicit open problem.
• Balancing cryptic-site and canonical-site performance: Tensor-DTI shows a consistent bias toward cryptic-site recognition and mispredicted AMP binding in the RET active site, which the authors say requires further refinement.
• Addressing shortcut learning: the PLINDER results suggest the model can exploit pocket similarity as a heuristic, so better negative-sampling regimes or objectives are needed to force reliance on genuine interaction features.
• Extending and strengthening the non-small-molecule work: the authors note that evaluating peptide, protein-RNA, and drug-RNA interactions mostly used one-hot baselines, and that RMSE values were comparable in some cases despite better correlations — leaving room for stronger baselines and broader validation.

Target Audience

Machine-learning researchers working on molecular representation learning and contrastive objectives; computational chemists and structural biologists involved in virtual screening; and drug-discovery teams evaluating whether a learned interaction model can replace or complement physics-based docking and co-folding tools at billion-compound scale. The paper is most useful to readers who already understand DTI/DTA terminology and benchmark conventions, since it reports many metrics without explaining them in detail. Note that several results — including ablation studies, dataset descriptions and sizes, and full benchmark tables — are placed in appendices (D, E, F, G, H, and sections B.1 and C) that are referenced but not included in the available content, and the provided text ends mid-sentence in the conclusion.

Authors’ 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.