Graph Neural Networks for Predicting Drug-Target Interactions

1. Introduction

Drug-target interaction (DTI) prediction is a fundamental problem in computational pharmacology and drug discovery. Accurate identification of the binding between a small molecule (drug) and a macromolecular target (typically a protein or RNA) accelerates the prioritization of candidates for experimental validation and reduces the cost of late-stage failures [30]. In veterinary medicine, DTI models are applied to identify antiviral compounds targeting animal pathogens, to repurpose approved drugs for zoonotic diseases, and to predict off-target toxicity in food-producing species [1]. Graph neural networks (GNNs) have emerged as a dominant machine learning paradigm for DTI prediction, as they naturally represent both drugs as molecular graphs and targets as protein graphs or residue-contact graphs [2, 3, 4].

A drug molecule is a set of atoms (nodes) connected by covalent bonds (edges), and a protein is a sequence of amino acids whose three-dimensional packing can be captured as a residue interaction graph [5, 6, 7]. GNNs learn node and edge representations through iterative message passing, aggregating information from local neighborhoods to produce global graph-level embeddings that are then used for interaction classification or binding affinity regression [8, 9, 10]. The field has progressed from simple graph convolutional networks to sophisticated architectures that incorporate attention mechanisms, multi-modal fusion, contrastive learning, and physics-informed constraints [11, 12, 13, 14, 15].

2. Graph Construction and Feature Engineering for DTI

2.1 Molecular Graphs for Drugs

The first step in any GNN-based DTI method is the construction of graph representations. A drug compound is mapped to a graph G_d = (V_d, E_d) where each node v in V_d corresponds to a heavy atom (e.g., carbon, nitrogen, oxygen) and each edge (u,v) in E_d represents a chemical bond [16, 17]. Node features typically include atom type, hybridization, formal charge, degree, and chirality. Edge features may encode bond type (single, double, triple, aromatic) and bond length. Many frameworks use RDKit or Open Babel to precompute these features before feeding them into the GNN [18, 19, 20].

Several studies emphasize the importance of three-dimensional (3D) conformations of drugs, rather than relying on two-dimensional topology alone. The Deep3D-DTA method incorporates 3D structural representations of both drugs and targets by using spatial coordinates of atoms from docking poses or crystallographic structures to build distance-based graphs [5]. Similarly, bond-aware GNNs integrate docking scores, pharmacophore features, and molecular dynamics (MD) simulation snapshots to enrich the input graph [13, 27]. These physics-informed inputs improve the generalizability of the model to unseen chemical space.

2.2 Protein Graphs for Targets

Protein targets can be represented in multiple ways. The most straightforward is to treat the amino acid sequence as a linear chain and apply a 1D convolutional network or a recurrent network to obtain residue embeddings [3, 21]. However, GNNs explicitly model residue-residue contacts. A protein graph G_p = (V_p, E_p) has nodes for each residue, with node features derived from the residue type, physicochemical properties (e.g., hydrophobicity, charge), and positional encodings. Edges are defined either by sequence adjacency (k-nearest neighbors in the sequence) or by spatial proximity (e.g., C_alpha atoms within 8 Ã…) when a 3D structure is available [11, 6, 7].

Predicting DTI without an experimentally solved structure is common in veterinary virology, where many animal viral targets lack high-resolution crystal data. In such cases, predicted structures from AlphaFold or homology models can be used to define residue contact graphs [6, 34]. The MMTF-DTI framework exploits multi-modal feature extraction by combining sequence-derived descriptors, predicted secondary structure, and evolutionary profiles (e.g., position-specific scoring matrices) to construct robust protein graphs [10].

2.3 Heterogeneous Graphs for Interaction Spaces

An alternative strategy is to build a single heterogeneous graph that includes drug nodes, target nodes, and known interaction edges. The graph convolutional network then learns embeddings for both entity types in a unified latent space [2, 4, 22]. The heterogeneous biological graph convolution approach of Zhu et al. constructs a graph where nodes correspond to drugs, proteins, diseases, and side effects, while edges represent known associations. Convolutional layers propagate information across the graph to predict new edges (i.e., novel DTIs) [2]. Multi-domain heterogeneous networks further incorporate drug-drug similarity and protein-protein interaction networks to provide contextual information [4]. Hyperbolic graph neural networks, as used in MML-DTI, embed the heterogeneous graph in a hyperbolic space to better capture hierarchical relationships inherent in biological data [22].

3. GNN Architectures for DTI Prediction

3.1 Message Passing and Graph Convolution

In a typical GNN for DTI, the message passing operation updates the representation of a node v by aggregating features from its neighbors N(v). The graph convolutional network (GCN) uses a normalized adjacency matrix to propagate signals, while the graph attention network (GAT) applies a learned attention weight to each neighbor [3, 20]. The TransGAT-DTI model combines a transformer-based sequence encoder for the protein with a graph attention mechanism for the molecular graph, using a cross-attention layer to align drug and target representations [3]. TransGAT-DTA extends this to a multi-task framework that simultaneously predicts binding affinity and generates conditional molecular structures [32].

Several models apply residual connections and parallel branches to improve convergence and expressiveness. The Parallel Residual Graph Network (PRGNet) uses multiple residual graph convolutional blocks to capture multi-scale topological features, and then concatenates the outputs before the final regression layer [21]. The LapGAT framework introduces a semi-supervised learning strategy by using graph Laplacian regularization to enforce smoothness of predictions over the interaction graph [20].

3.2 Attention and Transformer Mechanisms

Attention mechanisms have become integral to DTI prediction. The GATv2-TransDTI model uses GATv2 layers (a dynamic attention variant) on the drug graph and a transformer encoder on the protein sequence, then fuses the two branches with a fine-grained attention mechanism [17]. TransGAT-DTI also employs a transformer encoder for the protein but substitutes a GAT for the drug graph, allowing the model to focus on those atomic neighborhoods that are most relevant for binding [3].

The MolXProt architecture uses a cross-attention transformer that operates over both molecular graph nodes and protein residue nodes, computing attention between all pairs of atoms and residues [18]. This all-pairs approach resembles a learnable docking scoring function but is computationally expensive; group lasso regularization is applied to sparsify the attention weights and improve interpretability [19]. The GATv2-TransDTI and similar models use multi-head attention to capture different interaction patterns (e.g., hydrogen bonding, hydrophobic contacts) [17].

3.3 Multimodal and Multi-Task Frameworks

Recent work emphasizes the integration of multiple data modalities. The CM-MTL-DTI framework uses cross-modal alignment between drug molecular graphs and protein sequence embeddings through a multi-task learning objective: the model simultaneously predicts the interaction type (activation vs. inhibition) and the binding affinity [12]. The FGAIM model goes a step further by using inductive graph neural networks that learn distinct patterns for activation and inhibition mechanisms based on fine-grained interaction strategies derived from known functional responses [15].

The SynerDTI framework introduces a global feature coordinated attention mechanism that fuses features from the drug graph, the target graph, and a co-attention tensor that encodes the pairwise interaction features [23]. Another approach, MMTF-DTI, employs dynamic fusion of features from three modalities: drug molecular graph, protein sequence, and protein structural graph [10]. This multimodal fusion is also central to the work of Xu et al., who combine graph features with structural modeling from AutoDock Vina to improve affinity prediction [24].

3.4 Contrastive and Self-Supervised Learning

Label scarcity is a persistent challenge in DTI prediction, as experimental binding data are available for only a small fraction of possible drug-target pairs. Contrastive learning methods mitigate this by leveraging unlabeled data through a pretext task that maximizes mutual information between different views of the same graph. The method of Kang et al. applies heterogeneous graph contrastive learning to enhance cross-scale feature mutual information between drug molecular graphs and protein residue graphs [11]. The CMMSCL-DPI framework uses cross-modal multi-structural contrastive learning, constructing separate contrastive objectives for drug substructures and protein domains [28].

The HopWD-DTA model fuses multi-hop neighborhood information from both drugs and targets and uses deep feature extraction layers that are pre-trained in a self-supervised manner to initialize the weights [25]. Similarly, KANPM-DTA replaces the final feed-forward layers with Kolmogorov-Arnold networks, which are more parameter-efficient and learnable activation functions, and the model is pre-trained on large-scale chemical databases to capture general molecular knowledge [7].

3.5 Physics-Informed and Structure-Aware GNNs

Incorporating physical and biophysical priors into GNNs improves both accuracy and interpretability. The structure-aware compound-protein affinity prediction model of Shi et al. uses group lasso regularization to enforce that the attention weights correspond to known interaction contacts from docking or crystallography [19]. The bond-aware GNN of Nezhad et al. integrates docking scores and pharmacophore features directly into the edge features of the drug graph, and then uses graph convolution to propagate these physics-encoded signals [13].

A physics-informed GNN for docking-based binding affinity approximation was developed by Gider and Budak, who trained a GNN to reproduce the scoring function of AutoDock Vina for the DYRK2 kinase, enabling rapid virtual screening for drug repurposing [27]. This approach is particularly relevant for veterinary applications where high-throughput docking of thousands of approved animal drugs against a viral target (e.g., the capsid of African swine fever virus) would be computationally prohibitive without a surrogate GNN model.

4. Applications in Veterinary Drug Discovery

4.1 Antiviral Target Identification

GNN-based DTI models have been applied to identify inhibitors of viral proteins that are targets for veterinary antiviral development. For example, the JAK2 kinase is a host cell factor exploited by certain animal retroviruses, and the bond-aware GNN integrated docking and pharmacophore modeling to predict JAK2 inhibitors that could be repurposed for veterinary oncology and antiviral therapy [13]. The deep hybrid framework DeepHybridCPI was validated on compound-protein interaction datasets including host-viral interactions, demonstrating the ability to predict binding between small molecules and viral proteins such as influenza neuraminidase [31].

Repurposing of existing veterinary drugs is a cost-effective strategy for emerging zoonotic threats. The machine learning-driven repurposing for GPR17, a G-protein-coupled receptor involved in neuroinflammation, used a GNN to predict activity of approved drugs and then validated through multistage computational methods including molecular docking and MD simulations [14]. Such pipelines can be adapted to predict the activity of animal health drugs against conserved viral targets (e.g., viral proteases, helicases).

4.2 Predicting Antimicrobial Resistance in Zoonotic Pathogens

The emergence of antimicrobial resistance (AMR) in bacteria that infect both animals and humans poses a One Health challenge. DTI models can predict whether a given antibiotic will bind to a mutated target protein, such as a modified penicillin-binding protein or an efflux pump. Domain adversarial graph networks, such as those described by Lv et al., are designed for cross-domain DTI prediction, learning representations that are invariant across different protein families and bacterial species [26]. This capability is essential for predicting the activity of veterinary antibiotics against resistant strains identified in livestock.

4.3 Host Range and Spillover Risk

DTI prediction also informs host range assessment. By predicting the binding affinity of a viral attachment protein (e.g., bat coronavirus spike) to the ACE2 receptor of different animal species, GNN models can estimate zoonotic spillover risk [28]. The CMMSCL-DPI framework, with its cross-modal contrastive learning, was specifically designed to handle drug-protein interactions where the protein target is a viral surface glycoprotein and the drug is a humanized monoclonal antibody [28]. While the original work focused on human coronaviruses, the methodology is directly transferable to veterinary coronaviruses such as porcine epidemic diarrhea virus (PEDV) and canine coronavirus.

5. Workflow Diagram and Summary Table

The following Mermaid diagram illustrates a generalized pipeline for GNN-based DTI prediction.

flowchart TD
    A[Compound SMILES + 3D Conformer], > B[Generate Molecular Graph]
    B, > C[Atom Node Features + Bond Edge Features]
    D[Protein Sequence / Structure], > E[Generate Protein Residue Graph]
    E, > F[Residue Node Features + Contact Edge Features]
    C, > G[Graph Neural Network Encoder <br> (GCN / GAT / Transformer)]
    F, > H[Graph Neural Network Encoder <br> (GCN / GAT / Transformer)]
    G, > I[Drug Embedding h_d]
    H, > J[Target Embedding h_t]
    I & J, > K[Interaction Layer <br> (Concatenate / Bilinear / Cross-Attention)]
    K, > L[Prediction Head <br> (Binary Classifier or Regression)]
    L, > M{Prediction}
    M, > N[DTI Probability / Binding Affinity (pKd or IC50)]

Table 1 summarizes representative GNN architectures for DTI, highlighting their key features.

6. FAQ

What is a graph neural network in the context of drug-target interaction prediction?

A graph neural network is a deep learning model that operates on graph-structured data, learning representations of drug molecules (atoms as nodes, bonds as edges) and protein targets (residues as nodes, contacts as edges) through iterative message passing, and then fusing these representations to predict whether the drug binds to the target and with what affinity [8, 9, 2, 30].

How are drug molecules represented as graphs for GNN input?

Drug molecules are represented as graphs where nodes are atoms (with features such as atom type, hybridization, charge) and edges are chemical bonds (with features such as bond type). Three-dimensional conformations can add distance-based edge features [5, 13, 16]. This graph is then fed into a GNN encoder.

How are protein targets represented as graphs?

Proteins can be represented as graphs where nodes are amino acid residues (with features from sequence, evolutionary profiles, or physicochemical properties) and edges represent spatial proximity (C_alpha distances) or sequence adjacency. Predicted structures from AlphaFold can be used when experimental structures are unavailable [11, 19, 6, 34].

What is the advantage of using GNNs over traditional machine learning for DTI?

GNNs capture the topological and geometric structure of molecules and proteins directly, rather than relying on hand-crafted fingerprints that may lose local interaction patterns. They also allow end-to-end learning from raw graphs and can incorporate higher-order neighborhood information through iterative message passing [2, 21, 30].

Which GNN architectures are most commonly used for DTI?

Graph convolutional networks (GCN), graph attention networks (GAT), and graph isomorphism networks are common. More recent work uses transformers combined with GNNs, hyperbolic GNNs, and Kolmogorov-Arnold networks for improved representation learning [3, 17, 7, 22, 32].

How do multimodal approaches improve DTI prediction?

Multimodal approaches combine drug molecular graphs, protein sequence features, protein structural graphs, and sometimes docking scores or pharmacophore features. By fusing these complementary modalities, the model learns more robust representations that capture both the chemical and biophysical determinants of binding [10, 12, 24, 28].

Are GNN-based DTI models interpretable?

Some GNN architectures incorporate attention mechanisms or group lasso regularization to highlight atoms or residues that contribute most to the interaction prediction. Explainability methods can detect important drug atoms by analyzing attention weights or gradient-based saliency maps [16]. However, full structural interpretability remains an active research area.

Can GNNs predict inhibition versus activation mechanisms?

Yes, specialized frameworks such as FGAIM can predict whether a drug-target interaction leads to enzyme inhibition or activation by learning fine-grained interaction strategies from known functional annotations and graph features [15].

What are the limitations of current GNN-DTI models?

Limitations include dependence on the quality of predicted protein structures, difficulty in modeling water-mediated interactions and solvent effects, lack of large-scale labeled data for veterinary species, and limited generalizability to novel chemical scaffolds that are far from the training distribution [1, 30]. Physics-informed GNNs partially address the generalizability issue [13, 27].

How are GNN-DTI models validated in the context of veterinary drug discovery?

Validation typically involves cross-validation on curated datasets (e.g., BindingDB, PDBbind), and for veterinary applications, experimental binding assays using target proteins from animal pathogens (e.g., viral proteases, bacterial penicillin-binding proteins) are used. Some studies employ molecular docking as a surrogate for experimental validation when wet-lab resources are limited [14, 27].

7. Conclusion

Graph neural networks have become a cornerstone methodology for predicting drug-target interactions, offering a flexible and powerful framework to encode the molecular and structural features of both drugs and protein targets. The diversity of architectures, from basic GCNs to multimodal attention-based and physics-informed models, allows researchers to tailor solutions to specific problems in veterinary medicine, including antiviral drug repurposing, AMR prediction, and host range assessment. Continued improvements in pre-training strategies, contrastive learning, and integration with large language models are expected to further enhance the accuracy and applicability of GNN-DTI predictions in animal health [29, 35]. The transition of these computational tools from academic benchmarks to routine use in veterinary drug discovery pipelines will require careful validation against experimental data from target species.

Disclaimer: This article is for educational and informational purposes only. It is not intended to substitute for professional veterinary advice, diagnosis, treatment, or regulatory guidance. Always consult a licensed veterinarian or qualified specialist regarding animal health, disease diagnosis, and therapeutic decisions.

References

[1] Mondal H, Bishal A, Debnath B, et al. AI-driven drug-target interaction prediction: current progress, challenges, and future roadmap for precision medicine. J Comput Aided Mol Des. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/41910880/

[2] Zhu H, Wang J, Hua Z, et al. Heterogeneous biological graph convolutional network for drug-target interaction prediction. PLoS One. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42154765/

[3] Zhou C, Wang S, Zhong Y, et al. TransGAT-DTI: Transformer and Graph Attention Network for Drug-Target Interaction Prediction. J Comput Biol. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42100938/

[4] Zhou C, Liu Y, Yu L, et al. Multi-domain based heterogeneous network for drug-target interaction prediction. Artif Intell Med. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/41966585/

[5] Zhou H, Shi X, Wang L. Deep3D-DTA: A Tri-Modal Deep Learning Framework for Binding Affinity Prediction Leveraging 3D Structural Representations of Drugs and Targets. Interdiscip Sci. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42295634/

[6] Jiang M, Hang H, Cui Z, et al. A novel method for drug-target affinity prediction by integrating predicted evolutionary information and multi-scale protein graphs. BMC Biol. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/41913214/

[7] Rakib MDYK, Alamin MH, Li J, et al. KANPM-DTA: improving drug-target affinity prediction with Kolmogorov-Arnold networks and pretrained models. Brief Bioinform. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/41818722/

[8] Li Y, Jia C, Li M. CFM-DTI: Protein-conditioned feature modulation for drug-target interaction prediction. Comput Biol Chem. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42341701/

[9] Murala DK, Panda SK. Multimodal machine learning and deep graph neural networks for the prediction of molecular inhibitory activity and disease associations. J Comput Aided Mol Des. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42334640/

[10] Zeng P, Li Z, Tang X, et al. MMTF-DTI: Drug-target interaction prediction via multimodal feature extraction and dynamic fusion. J Biomed Inform. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42323037/

[11] Kang X, Zhang L, Wang Z, et al. Enhancing Cross-scale Feature Mutual Information via Heterogeneous Graph Contrastive Learning for Drug-Target Binding Affinity Prediction. IEEE J Biomed Health Inform. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42268754/

[12] Zhao Y, Gao S, Liu Y, et al. CM-MTL-DTI: Drug-Target Interaction Prediction via Cross-Modal Alignment and Multi-Task Learning. J Chem Inf Model. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42224430/

[13] Nezhad SK, Solout MV, Meknatkhah S, et al. A bond-aware graph neural network integrated with docking, pharmacophore modeling and molecular dynamics for JAK2 inhibitors affinity prediction. Biochim Biophys Acta Proteins Proteom. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42218952/

[14] Agha Babaie N, Farnia M, Ghasemi JB. Machine learning-driven drug repurposing for GPR17: activity prediction via graph neural networks and multistage computational validation. J Biomol Struct Dyn. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42207497/

[15] Tang Y, Fan Y, Sun G, et al. FGAIM: Identifying Drug-Target Activation and Inhibition Mechanisms via Inductive Graph Neural Networks Based on Fine-Grained Interaction Strategies. IEEE Trans Comput Biol Bioinform. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42166265/

[16] Mahindran M, Liu Q, Kadambalithaya VM, et al. Explainability Methods from Machine Learning Detect Important Drugs' Atoms in Drug-Target Interactions. J Chem Inf Model. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/41985080/

[17] Li X, Tong G. GATv2-TransDTI: A graph and sequence hybrid model for fine-grained drug-target interaction prediction. Anal Biochem. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/41932420/

[18] Cucco B. MolXProt: A Cross-Attention Transformer-Based Graph Neural Network for Protein-Ligand Binding Affinity Prediction. J Chem Theory Comput. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42081530/

[19] Shi Z, Wang Y, Weerawarna PM, et al. Structure-Aware Compound-Protein Affinity Prediction via Graph Neural Networks with Group Lasso Regularization. Comput Struct Biotechnol J. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42057778/

[20] Song L, Yuan W, Pei X. LapGAT: A Semi-Supervised Learning Framework for Drug-Target Interaction Prediction. Mol Inform. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42011802/

[21] Liu J, Mehmood A, Liu H, et al. PRGNet: a Parallel Residual Graph Network for enhanced drug-target binding affinity prediction. BMC Genomics. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42050415/

[22] Guan H, Bai T, Yang C, et al. MML-DTI: Multimanifold Learning with Hyperbolic Graph Neural Networks for Enhanced Drug-Target Interaction Prediction. J Chem Inf Model. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/41802299/

[23] Liao Y, Lin M, Peng J, et al. SynerDTI: a synergistic deep learning framework for drug-target interaction prediction via global feature coordinated attention mechanism. Mol Divers. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/41739380/

[24] Xu Y, Xiao X, Lin W. A drug-target affinity prediction model integrating multimodal feature fusion and structural modeling. Comput Biol Chem. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/42001823/

[25] Liu X, Zhou M, Huang X, et al. HopWD-DTA: a novel framework for drug-target affinity prediction fusing multi-hop neighborhoods and deep features. J Mol Model. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/41973256/

[26] Lv W, Zhang Q, Liu L. Domain adversarial gated bilinear attention networks for cross domain drug target interaction prediction. Biochem Biophys Res Commun. 2026. URL: https://pubmed.ncbi.nlm.nih.gov/41687302