Graph-based drug–target interaction modeling: from representation learning to output-driven drug discovery

Abstract

Graph-based deep learning has emerged as a powerful framework for modeling drug–target interactions (DTIs), enabling the integration of molecular, structural, and systems-level information within a unified representation. In this review, we survey graph-based DTI models across biomedical network-, sequence/hybrid-, and structure-based paradigms, which form a continuum from large-scale association inference to structure-resolved interaction modeling with increasing mechanistic specificity. Beyond architectural advances, we introduce an output-driven perspective in which models are evaluated according to how well their predictions align with the informational and decision-making requirements of different stages of the drug discovery pipeline. Within this framework, attention mechanisms and semi-supervised learning are discussed as key developments that enhance feature prioritization and data efficiency in data-limited settings. We further examine how model outputs support applications ranging from target identification and drug repurposing to structure-guided lead optimization. Finally, we analyze key benchmarking challenges, including data leakage, sequence redundancy, and structural bias, and discuss emerging directions such as multimodal integration and the use of predicted protein structures. Together, this review provides a unified perspective on the design, evaluation, and translational application of graph-based DTI models.

Introduction

Artificial intelligence has become a central component of modern drug discovery. This progress is driven by advances in machine learning, the rapid expansion of chemical and biological datasets, and the need to model complex molecular systems that are difficult to capture using traditional approaches. Among artificial intelligence (AI) methods, graph-based models have emerged as a powerful framework for representing molecular structures and biological interactions. By naturally encoding relational information, these models enable the learning of chemically meaningful features through message passing. More recently, the field has shifted toward large-scale pretraining and transformer-based architectures, including heterogeneous graph transformers [1] and molecular foundation models such as Uni-Mol [2]. These approaches aim to learn generalizable representations from diverse chemical and biological data. In drug–target interaction (DTI) modeling, such advances provide new opportunities to integrate heterogeneous modalities, including molecular graphs, protein sequences, 3D structures, and biomedical networks. This integration may improve cross-target generalization and enable more robust prediction in data-limited settings.

While early graph neural networks were primarily developed for ligand-based property prediction, recent advances have extended their use to DTI modeling by incorporating protein information and enabling target-aware prediction [3]. In this review, DTI models involving graph representations are organized according to where graph structure is introduced, including biomedical networks, ligand molecular graphs, 3D protein or pocket structures, and explicit protein–ligand interaction graphs. Sequence-based approaches are therefore described more precisely as models that combine ligand representations with protein sequence encoders, rather than as fully graph-based models of both drugs and targets.

Biomedical network models, including knowledge graph-based approaches, operate at a systems level by integrating drugs, proteins, diseases, pathways, and other biomedical entities to infer associations and support drug-repurposing hypotheses [4, 5]. However, they generally provide limited mechanistic insight into molecular binding. Sequence-based and hybrid models combine ligand representations with protein amino acid sequence encoders, enabling scalable affinity prediction when structural data are unavailable [6, 7]. Their generalization to underrepresented or low-homology target families, however, remains limited. In contrast, structure-based models incorporate 3D ligand, protein, pocket, or protein–ligand complex representations. These models allow more explicit modeling of binding context, interaction specificity, and pose-dependent effects relevant to structure-guided screening and early optimization [8–10]. Within this broad structure-based group, some models place greater emphasis on explicit atom–residue contacts, ligand–protein interaction edges, pocket-level context, or contact-level weighting. These interaction-focused approaches should therefore be viewed as part of the structure-based modeling continuum, rather than as a completely separate paradigm.

These paradigms form a continuum from large-scale association inference to structure-resolved interaction modeling, with each model class suited to different data regimes and stages of drug discovery [11–13]. However, these categories should not be viewed as strictly separate. Recent hybrid and multimodal models combine ligand graphs, protein sequences, structural features, and heterogeneous biomedical networks, thereby bridging network-, sequence-, and structure-based approaches [14, 15]. Accordingly, this review surveys graph-based DTI models across major representation frameworks, while highlighting emerging hybrid, multimodal, and interaction-focused developments. To improve narrative coherence, the review first organizes DTI models by their primary representation framework and then discusses attention mechanisms, contrastive or self-supervised learning, and semi-supervised learning as methodological advances that span multiple model categories.

Representation frameworks in graph-based drug–target interaction modeling

To understand how graph-based DTI models differ in practice, it is useful to formalize how drugs, targets, binding pockets, and biomedical networks are represented. In graph-based DTI modeling, a system can be represented as a graph , where denotes the set of nodes and denotes the set of edges [16]. Depending on the modeling paradigm, nodes may correspond to atoms, residues, drugs, proteins, diseases, or other biomedical entities, whereas edges may represent covalent bonds, spatial contacts, residue–residue relationships, drug–target associations, or heterogeneous biomedical relationships [3, 17, 18]. This flexible representation allows graph-based models to encode chemical, structural, and systems-level information within a unified framework [11, 19, 20] (Fig. 1).

Figure 1.

Representation frameworks in graph-based DTI models; graph-based models encode DTIs across multiple levels of biological abstraction; these representations reflect a continuum from large-scale association inference to structure-resolved interaction modeling, with increasing mechanistic specificity and data requirements.

Each node is initialized with a feature vector , and each edge may be associated with an edge feature vector . In molecular graphs, node features may include atom type, aromaticity, charge, hybridization, or residue identity, whereas edge features may include bond type, distance, contact category, or geometric orientation [21]. A common message-passing formulation updates node embeddings by aggregating information from neighboring nodes:

where denotes the neighbors of node , is the aggregated message at layer , is a learnable message function, and is a learnable update function. After message-passing layers, node embeddings can be pooled into a graph-level representation:

where READOUT may be implemented as sum pooling, mean pooling, max pooling, or attention-based pooling. In DTI models, ligand and protein representations may be encoded separately and then fused, or jointly encoded within a unified ligand–protein graph. This formulation clarifies how graph-based models transform initial atom, residue, or biomedical entity features into latent embeddings for downstream DTI prediction.

Although ligand-based graph models are not DTI models per se, they provide the foundational representation learning framework upon which many DTI approaches are built. Architectures such as Graph Isomorphism Network [22], Directed Message Passing Neural Network [23], Graph Attention Networks (GAT) [24], and AttentiveFP [25] are commonly used to construct ligand-only models, where 2D molecular graphs are used to learn structure–activity relationships without explicit target context. While these approaches are highly scalable and effective for quantitative structure activity relationship (QSAR) and virtual screening, their lack of protein information limits their ability to capture target specificity and binding mechanisms [26–28].

In contrast, DTI modeling explicitly incorporates protein information to predict how a specific ligand interacts with a specific target, requiring more complex and multimodal inputs. Beyond molecular-level representations, network-based DTI models adopt a systems-level perspective by representing drugs, targets, and related biomedical entities as nodes in heterogeneous graphs. Models such as NeoDTI [29], DTI based on Heterogeneous Network Embedding (DTI-HeNE) [30], NASNet-DTI [9], H²GnnDTI [31], EEG-DTI [32], and iGRLDTI [33] integrate heterogeneous biomedical relationships to support large-scale interaction inference, association discovery, and drug-repurposing hypotheses, but generally provide limited direct molecular-level interpretability. Sequence-based and hybrid DTI models, including DeepDTA [34] and GraphDTA [35], combine drug representations with protein amino acid sequence encoders to enable structure-agnostic affinity prediction across diverse targets. Because the protein component is typically encoded as a 1D sequence rather than an explicit residue graph or 3D structure, these methods provide limited direct information about binding geometry unless structural or interaction-level features are incorporated. More recent hybrid and multimodal models further integrate complementary information from molecular graphs, protein sequences, structural features, and biomedical networks. For example, CSCo-DTA integrates molecule-scale and network-scale drug–target features through cross-scale graph contrastive learning, reflecting a shift toward multiscale representation learning for binding-affinity prediction [14]. AttentionMGT-DTA [15] further combines molecular graph and binding-pocket graph representations using multimodal attention. More broadly, unified frameworks such as DTIAM [36] extend DTI modeling beyond binary interaction prediction by jointly considering interaction likelihood, binding affinity, and activation/inhibition mechanisms. Nevertheless, despite their scalability and flexibility, these models provide limited direct information about binding geometry unless structural or interaction-level features are explicitly incorporated.

Structure-based DTI models increase representational resolution by integrating 3D ligand, protein, pocket, or protein–ligand complex information [18, 37]. These models differ not by whether they use structural context, but by how explicitly they encode ligand–protein interactions. Some approaches represent the binding environment through spatial neighborhoods, geometric descriptors, or staged message passing, as in PotentialNet [38], SIGN [39], EquiBind [40], and PLANET [41]. Others, including GIGN [42], InteractionGraphNet [43], and PIGNet [44], more directly encode protein–ligand contacts or interface graphs derived from structural proximity. Despite these differences, structure-based models share several common limitations, including strong dependence on the availability and quality of protein–ligand complex structures, sensitivity to docking accuracy and pose uncertainty, limited capacity to capture protein conformational flexibility, and increased computational cost [45–48].

Recent structure-based models have moved beyond conventional protein–ligand complex graphs toward specialized scoring and screening frameworks. RTMScore [49] introduced residue–atom graphs with graph transformer layers for protein–ligand scoring, whereas GenScore [50] further developed this direction into a generalized framework balancing scoring, docking, ranking, and screening performance. PBCNet [51] emphasized relative binding-affinity ranking among congeneric ligands using a physics-informed graph attention mechanism, and PIGNet2 [52] updated physics-informed interaction modeling for binding-affinity scoring and virtual screening. In parallel, docking-oriented methods such as TANKBind [53], DiffDock [54], FABind [55], DynamicBind [56], and SurfDock [57] highlight the growing role of equivariant learning and diffusion-based generative modeling in binding-site localization, ligand pose prediction, and protein–ligand complex modeling. More recent biomolecular foundation models, such as AlphaFold3 [58] and Boltz-2 [59], further extend this trend toward joint complex prediction and affinity-aware structural modeling, although they should be viewed as sources of structural priors rather than conventional DTI predictors.

In this review, the term interaction-focused refers to structure-based models that place particular emphasis on explicit atom–residue interaction edges, contact-level weighting, global pocket context, or ranking-aware objectives. For example, Drug-Target Interaction Graph Neural Network (DTIGN) incorporates 3D ligand structures, localized binding pockets, and spatially defined ligand–protein interaction edges to represent pocket-specific and pose-dependent interaction patterns [60]. LigoSpace further explores broader geometric context, global pocket representations, and ranking-aware learning objectives for modeling pocket occupancy and functional binding relevance [61]. Together, these examples illustrate how interaction-focused models refine structure-based DTI modeling by emphasizing contact-level interactions and broader pocket context. Key differences in input representation across models are summarized in Table 1.

Table 1.

Representative DTI models and methodological advances across modeling frameworks.

Model category	Representative models	Main representation	Recent methodological advance
Network-based	NeoDTI, DTI-HeNE, EEG-DTI, iGRLDTI, NASNet-DTI, H2GnnDTI	Heterogeneous biomedical networks	Heterogeneous graph learning, graph contrastive learning, knowledge-based regularization
Sequence/hybrid	DeepDTA, GraphDTA, AttentionDTA, DTITR, CAT-DTI	Ligand representation + protein sequence	Attention, transformer encoders, sequence pretraining
Multimodal/cross-scale	CSCo-DTA, AttentionMGT-DTA, DTIAM	Molecular graphs, sequences, pockets, networks	Cross-scale contrastive learning, multimodal attention, unified prediction
Structure-based	GIGN, PIGNet, SIGN, PLANET, RTMScore, GenScore, PBCNet	3D pockets, protein–ligand complexes, residue–atom graphs, interaction graphs	Geometric learning, physics-informed descriptors, interaction graph encoding
Interaction-focused structure-based	DTIGN, LigoSpace	Local pocket context and protein–ligand contact features	Contact-level weighting, pose-sensitive prediction, ranking-aware learning

Attention mechanisms in graph-based drug–target interaction modeling

Attention mechanisms should not be viewed as a separate model class; rather, they function as cross-cutting weighting components across network-based, sequence/hybrid, and structure-based DTI models. Their role depends on the underlying input representation. In network-based models, attention prioritizes heterogeneous biomedical relationships. In sequence-based models, it highlights informative sequence or molecular tokens. In structure-based models, it can assign greater weight to binding-site regions or ligand–protein contacts. As summarized in Fig. 2, their function depends strongly on the underlying representation used by each model class.

Figure 2.

Roles of attention mechanisms across graph-based DTI models; attention mechanisms enable selective weighting of relevant information across different representation levels; their function varies by model class, operating over nodes and edges in network-based models, informative sequence tokens in sequence-based models, spatial binding regions in structure-based models, and explicit atom–residue interactions in interaction-aware frameworks.

In network-based DTI models, attention primarily improves representation learning over heterogeneous biomedical entities, such as drugs, targets, diseases, and pathways, rather than directly modeling molecular recognition [29]. Related examples include DDAGDL, which applies geometric deep learning and attention-based representation learning over heterogeneous information networks for drug repositioning [62]. In contrast, attention plays a more direct role in sequence-based, structure-based, and interaction-focused models. It can highlight informative sequence regions, binding-site features, spatial neighborhoods, or ligand–protein contacts that contribute to predicted interactions.

In sequence-based DTI models, attention mechanisms are primarily used to learn alignment and compatibility patterns between ligand features and protein amino acid sequences, rather than explicit spatial interactions. Models such as AttentionDTA [63] and Graph-sequence attention and transformer for predicting drug–target affinity (GSATDTA) [64] apply attention-based or transformer-based architectures to capture informative dependencies within drug and protein sequence representations. Similarly, TransformerCPI [65], DeepAffinity [66], Drug–Target Interaction TRansformer (DTITR) [67], and CAT-DTI [68] use self-attention or cross-attention mechanisms to model residue-level dependencies and ligand–protein compatibility from protein sequences, SMILES strings, or learned sequence-derived molecular representations. Accordingly, attention in these models usually operates over sequence- or token-derived representations rather than explicit protein structures or physical interaction graphs. This paradigm offers clear advantages in scalability and broad applicability, making it well suited for proteome-wide screening when 3D structures are unavailable [69]. However, because attention does not directly resolve spatial ligand–protein interfaces, the resulting representations remain abstract with respect to binding geometry [70]. As a result, attention weights in these models should not be directly interpreted as physical interaction sites, despite their potential value for highlighting informative residues, molecular tokens, or compatibility patterns [71].

In structure-based DTI modeling, many established approaches, including GIGN, PotentialNet, PLANET, and EquiBind, encode ligand–protein interactions using predefined binding pockets, geometric heuristics, or implicit message passing [18, 38, 40–42]. Although effective, such representations may not always explicitly learn the relative importance of individual binding-site regions or intermolecular contacts. To address this limitation, attention mechanisms have been incorporated into structure-aware graph models. For example, AttentionSiteDTI uses protein binding-site information together with self-attention to identify binding-site contributions to DTI prediction [72]. GAT-score represents protein–ligand complexes as graph structures and applies GATs for binding-affinity prediction [73]. More recently, Protein–Ligand binding Affinity with the Graph Cross-Attention mechanism (PLAGCA) introduced graph cross-attention to learn local 3D features of protein–ligand pockets, while integrating global sequence and molecular features [74]. These mechanisms support bidirectional information exchange and improve representation capacity compared to fixed or heuristic interaction encoding. However, their ability to consistently identify binding pockets or discriminate native-like poses remains architecture- and dataset-dependent [75], and attention weights do not consistently correspond to physically meaningful interaction sites [63, 76, 77].

Within structure-based DTI modeling, interaction-focused attention architectures place greater emphasis on atom–residue contacts and contact-level weighting within 3D space [60]. By dynamically weighting ligand–protein contact contributions, such models aim to improve pose discrimination and binding-pocket representation compared with fixed or heuristic interaction encodings. Their apparent advantages remain dependent on the datasets, docking protocols, target classes, and evaluation designs used in the original studies.

Building on these examples, attention mechanisms can be formalized as extensions of standard message passing that assign different importance weights to neighboring nodes, sequence tokens, binding-site residues, or ligand–protein interaction edges [24].

where denotes the normalized attention weight, is a learnable scoring function, and are node embeddings at layer , and represents edge features when available. The updated node embedding can then be expressed as:

where is a learnable transformation matrix and is a non-linear activation function [21].

To further formalize the interaction-aware setting, the ligand–protein complex can be represented as a joint interaction graph:

where and denote ligand and protein nodes, and denote intra-ligand and intra-protein edges, and denotes ligand–protein interaction edges [42–44]. These interaction edges are commonly defined according to spatial proximity:

where is the Euclidean distance between ligand atom and protein atom or residue , and is a predefined distance cutoff. Each ligand–protein interaction edge can be associated with an edge feature vector:

where and denote ligand and protein node types, denotes an interaction category such as hydrogen bonding, hydrophobic contact, electrostatic interaction, or van der Waals contact, and denotes geometric descriptors such as relative orientation or position [73, 74].

In interaction-aware frameworks, including DTIGN-like or LigoSpace-like models [60, 61], the contribution of each ligand–protein contact can be represented using a learnable interaction-edge weighting function:

where denotes the learned interaction-edge weight and is an edge-scoring network. The ligand–protein interaction representation can then be aggregated as:

where represents the learned interaction embedding, and maps atom–residue or atom–atom contacts into latent interaction features.

Taken together, attention mechanisms provide flexible weighting functions that can enhance representation learning across graph-based DTI models. However, their biological interpretation remains context-dependent, and attention weights should not be treated as direct mechanistic explanations without additional structural or experimental validation.

Learning paradigms in graph-based drug–target interaction modeling

In parallel with architectural advances, modern DTI models increasingly rely on learning strategies such as contrastive learning, self-supervised pretraining, and semi-supervised learning to improve data efficiency, representation quality, and generalization. These strategies are not confined to a single representation category; rather, they can be applied across network-based, sequence/hybrid, and structure-based models, depending on the available data sources, such as labeled interactions, unlabeled molecular data, protein sequences, docking poses, or biomedical networks [13, 36, 78]. This is particularly important because experimentally validated DTIs and binding affinities remain limited, unevenly distributed across target families, and biased toward well-studied regions of chemical and biological space. By contrast, unlabeled chemical structures, protein sequences, predicted structures, docking poses, and biomedical networks are increasingly abundant. Supervised, unsupervised, and semi-supervised learning should therefore be viewed as complementary strategies for leveraging different data regimes in DTI prediction. Figure 3 provides a schematic comparison of these learning strategies.

Figure 3.

Learning paradigms in graph-based DTI models; schematic comparison of supervised, unsupervised, and semi-supervised learning strategies for leveraging labeled and unlabeled data in DTI prediction.

Supervised learning remains the dominant paradigm when curated experimental annotations are available. In sequence-based models such as GraphDTA [35] and DeepDTA [34], supervised learning enables direct optimization of DTI-related endpoints, such as interaction probability, binding affinity, or activity labels. These models can achieve high predictive accuracy when trained on sufficiently large and consistent datasets, making them well suited for tasks such as affinity prediction and compound prioritization within well-characterized target families [79]. However, their performance is fundamentally constrained by the scale and bias of labeled data, which are often limited to a small subset of proteins and chemical space [80, 81]. This limitation is particularly pronounced in structure-based DTI modeling, where high-quality protein–ligand complexes and experimentally measured affinities are both scarce, restricting the applicability of purely supervised approaches. In supervised DTI prediction, the final output is typically obtained by passing ligand, protein, and, when available, ligand–protein interaction representations through a prediction function [34, 35]:

where denotes the ligand representation, denotes the protein or pocket representation, denotes the ligand–protein interaction representation, and is a learnable prediction network. In models that do not explicitly construct an interaction graph, may be omitted or replaced by a fused ligand–protein representation. For binary DTI classification, represents the predicted probability of interaction, whereas for affinity prediction, represents a continuous binding-affinity value.

For binary interaction prediction, models are commonly optimized using binary cross-entropy loss:

where is the true interaction label, is the predicted interaction probability, and is the number of drug–target pairs. For binding-affinity regression, a common objective is mean squared error [34, 35]:

where is the experimentally measured affinity and is the predicted affinity. In ranking-aware settings, the objective may prioritize the correct ordering of candidate ligands rather than only minimizing absolute prediction error. A margin-based ranking loss can be written as:

where ligand is expected to rank above ligand , ligand is a lower-ranked candidate, and is a margin hyperparameter. Such objectives are particularly relevant for virtual screening and lead prioritization, where the relative ordering of compounds may be more important than exact affinity estimation [51, 52].

Unsupervised learning plays an indirect but foundational role by enabling representation learning from large unlabeled datasets. In graph-based DTI contexts, this includes learning molecular, protein, or network embeddings through autoencoders, contrastive learning, or self-supervised pretraining. For example, autoencoder-based frameworks such as AutoDTI [82] leverage unsupervised representation learning to extract latent features from heterogeneous biological data, which can subsequently support downstream interaction prediction. More broadly, such approaches are widely used for chemical space exploration, scaffold discovery, and feature extraction from omics or structural data [83–85]. However, because unsupervised methods do not directly optimize DTI objectives, their outputs are not immediately predictive of binding affinity or efficacy without downstream integration [86]. Consequently, unsupervised learning is seldom used as a standalone strategy for DTI prediction, but instead serves as a preparatory stage that improves the expressiveness and generalizability of downstream models [83, 87].

Semi-supervised learning has emerged as an important data-efficient paradigm for graph-based DTI modeling, as it directly addresses the imbalance between limited labeled data and abundant unlabeled data. By combining self-supervised or weakly supervised representation learning with task-specific fine-tuning, semi-supervised approaches are designed to capture generalizable patterns from large datasets, while retaining sensitivity to experimentally validated interactions [88, 89]. In network-based DTI models such as NeoDTI [29] and SHGCL-DTI [90], semi-supervised learning leverages sparse known interactions within heterogeneous biomedical graphs to learn robust node embeddings and infer new associations. In sequence-based DTI models such as DeepAffinity [66], it typically involves pretraining protein and ligand encoders on large sequence corpora, followed by fine-tuning on limited affinity datasets to improve cross-target generalization. In structure-based DTI models, semi-supervised strategies are particularly valuable, as they allow models to learn interaction patterns from large collections of protein–ligand complexes or docking poses, capturing features such as atom–residue contributions and pose-dependent effects that are difficult to supervise directly [60]. However, semi-supervised approaches introduce additional training complexity and may be sensitive to the quality and distribution of unlabeled data, potentially propagating biases or learning spurious correlations if not carefully designed. Formally, semi-supervised DTI learning can be described using a labeled dataset and an unlabeled dataset. The labeled dataset contains experimentally annotated drug–target pairs, whereas the unlabeled dataset may include unannotated drug–target pairs, molecular graphs, protein graphs, or protein–ligand complexes. The total training objective can then be expressed as [88, 90]:

where denotes the supervised classification or regression loss, denotes the consistency regularization loss, and controls the contribution of the semi-supervised objective. A common consistency objective encourages stable predictions under perturbations of unlabeled inputs:

where denotes an unlabeled input and denotes a perturbed or augmented version of the same input. In graph-based DTI modeling, such perturbations may include node masking, edge dropout, subgraph sampling, conformer perturbation, or alternative docking poses. This objective encourages the model to learn representations that remain stable under biologically or structurally plausible variation.

More broadly, the growing adoption of semi-supervised and self-supervised strategies reflects a shift toward data-efficient learning in graph-based DTI modeling. The objective is not only to maximize predictive accuracy, but also to improve robustness across targets, chemical space, and experimental conditions [91]. Although these approaches introduce additional modeling complexity and remain sensitive to data quality, they offer a practical balance between scalability and mechanistic resolution. As such, they are increasingly relevant to translational drug discovery settings, where labeled data are scarce but reliable generalization and interaction-level insight are essential [80].

Linking model outputs to applications across the drug discovery pipeline

The practical utility of graph-based DTI models depends largely on the nature of their outputs. These outputs reflect the level of interaction detail encoded during model design and influence the model’s suitability for different stages of drug discovery [17]. As illustrated in Fig. 4, model outputs range from system-level association scores to structure-resolved and interaction-focused predictions, each supporting distinct downstream tasks.

Figure 4.

Applications of graph-based DTI models across the drug discovery pipeline; representative use cases of different DTI model classes at distinct stages of drug discovery and translation.

At the system level, network-based DTI models generate drug–target association scores or interaction likelihoods derived from global network topology. These outputs capture relational patterns across heterogeneous biomedical data, including drug–drug, protein–protein, and disease associations, enabling large-scale inference across the proteome or phenome. Recent heterogeneous-network models further extend this systems-level paradigm by incorporating richer biological connectivity patterns [92, 93]. For example, RGLDR [94] uses regulation-aware graph learning and meta-path-based representations to model drug–protein–disease regulatory mechanisms, whereas FuHLDR [95] integrates higher- and lower-order biological information for drug repositioning. These approaches illustrate how graph-based representations can support indication discovery and translational hypothesis generation beyond pairwise DTI prediction. However, because these outputs are not grounded in molecular or structural representations, they provide limited insight into binding mechanisms and are not directly applicable to structure-guided optimization or mechanistic interpretation [29, 96].

Sequence-based DTI models increase output specificity by producing quantitative interaction or affinity predictions for defined ligand–protein pairs using protein sequence information. These models enable scalable target-aware prediction across diverse proteomes without requiring experimentally resolved structures. Given their typical training datasets and structure-agnostic representations, they are particularly useful for drug repurposing, target profiling, off-target liability assessment, polypharmacology analysis, and first-pass compound prioritization [13, 97]. They may also support virtual screening when applied to defined targets and calibrated against relevant chemical series, but their predictions generally lack explicit information about binding sites and interaction geometry. Therefore, sequence-based outputs are most appropriate for broad prioritization rather than structure-guided optimization or detailed binding-mode interpretation [98, 99].

Structure-based DTI models increase the resolution of outputs by incorporating 3D information at the protein–ligand interface. Depending on the underlying representation, these models can generate affinity scores, docking scores, pose-dependent predictions, or interaction-aware latent features that reflect spatial relationships between atoms and residues. In practice, such outputs are most useful for structure-guided virtual screening, docking rescoring, distinguishing likely active from inactive compounds, binding-mode analysis, and prioritizing hits within defined binding pockets [18, 37]. Although structure-based predictions can inform early hit-to-lead hypotheses, most models are not sufficiently accurate on their own for definitive lead optimization, which requires integration with structure–activity relationship (SAR), medicinal chemistry judgment, orthogonal assays, and experimental structural validation [81].

Within structure-based DTI modeling, interaction-focused and pose-sensitive models produce outputs that more directly describe molecular recognition at the binding interface [42, 44, 60]. By integrating ligand and protein representations into unified interaction graphs or attention-based frameworks, these models can generate contact-aware and pose-dependent predictions that capture heterogeneous interaction contributions. Such outputs may support SAR interpretation, binding-site hypothesis generation, mutation or resistance analysis, and prioritization of candidates for experimental validation. However, these outputs should be interpreted as mechanistic hypotheses rather than direct evidence of the binding mechanism or translational efficacy. Their utility in late-stage discovery depends on validation by biochemical, biophysical, mutagenesis, cellular, or structural assays.

Taken together, these observations emphasize that model outputs should be interpreted as decision-support evidence rather than direct substitutes for experimental validation. The appropriate use of a DTI model depends on the resolution of its output, the assumptions embedded in its training data, and the level of uncertainty acceptable for a given stage of drug discovery. Lower-resolution association or affinity scores are most useful for broad prioritization, whereas structure-resolved and interaction-focused outputs can support more specific mechanistic hypotheses when reliable structural inputs are available. An output-driven perspective therefore requires matching each model not only to a discovery stage, but also to the type of decision being made, the confidence required, and the experimental evidence needed for validation.

Benchmarking challenges and evaluation pitfalls in graph-based drug–target interaction modeling

Despite their increasing impact across multiple stages of drug discovery, benchmarking graph-based DTI models remains challenging. Reported performance is shaped by more than model architecture. Dataset construction, splitting strategy, negative-sampling procedure, evaluation metric, and biological bias can also strongly influence benchmark results. As a result, performance estimates can be overly optimistic and difficult to compare across studies [100, 101].

Reliable benchmarking therefore requires transparent reporting of both the data-splitting strategy and the evaluation metrics used. Random splits or random k-fold cross-validation mainly assess interpolation within known ligand and target spaces. In contrast, scaffold-based, cold-drug, cold-target, sequence-identity-controlled, pocket-similarity-controlled, and external test-set evaluations provide more stringent tests of generalization [79, 102]. Different prediction tasks require different evaluation metrics. Binary DTI classification is commonly assessed using the area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC) [103, 104]. Affinity regression is evaluated using the root mean squared error (RMSE), mean absolute error (MAE), or correlation-based metrics [105]. Virtual screening is better assessed using enrichment- or ranking-based metrics [106]. Common evaluation strategies, metrics, and reporting details are summarized in Table 2.

Table 2.

Common evaluation strategies and metrics for graph-based DTI models.

Evaluation aspect	Common options	Main purpose
Data split	Random split, -fold cross-validation, scaffold split, cold-drug split, cold-target split, sequence-identity-controlled split, pocket-similarity-controlled split, external test set	Assess whether the model generalizes beyond known ligands, targets, or datasets
Binary classification metrics	AUROC, AUPRC, F1-score, MCC	Evaluate interaction/non-interaction prediction
Affinity regression metrics	RMSE, MAE, Pearson, Spearman, concordance index	Evaluate continuous binding-affinity prediction
Ranking/screening metrics	Enrichment factor, top-(k) recall, BEDROC	Evaluate compound prioritization for virtual screening
Key reporting details	Split strategy, class balance, negative sampling, unseen ligands/targets	Support fair comparison across studies

Building on these evaluation considerations, Fig. 5 summarizes four major sources of benchmarking bias in graph-based DTI modeling, including information leakage, sequence/structural redundancy, structural bias, and dataset imbalance.

Figure 5.

Benchmarking pitfalls in graph-based DTI models; (a) information leakage from shared graph components; (b) sequence and structural redundancy leading to homology- or fold-driven predictions; (c) structural bias from idealized training poses; (d) dataset imbalance resulting in biased representation and limited generalization.

A major source of bias arises from information leakage in network-based models (Fig. 5a). Because drug–drug, protein–protein, and drug–target relationships are highly interconnected, shared nodes or edges can introduce correlations between training and test sets. Under random or edge-level splits, models may exploit these latent connections rather than learn generalizable interaction patterns, leading to inflated performance estimates [9]. Leakage can be mitigated by using stricter inductive settings, such as cold-drug or cold-target splits, withholding unseen nodes, separating highly connected graph components, or evaluating on external networks when available. Moreover, reliance on relational features limits mechanistic interpretability, making it difficult to determine whether predictions reflect genuine biochemical interactions [107].

In sequence-based and hybrid DTI models, benchmarking is strongly affected by target-side sequence redundancy and ligand-side scaffold similarity (Fig. 5b). Proteins with high sequence similarity may appear across both training and test sets, enabling models to rely on homology-driven signals rather than learning transferable interaction principles [108, 109]. Similarly, structurally related ligands can inflate apparent performance when they are shared across splits. Scaffold-aware evaluation, such as Bemis–Murcko scaffold partitioning, provides a more stringent assessment of ligand-side generalization by separating compounds according to their core chemical frameworks [110]. Sequence-identity-controlled or protein-family-based splits are similarly important for assessing target-side generalization. Without sequence- or scaffold-aware evaluation, model performance may be overestimated, particularly for novel targets or unexplored chemical space [79, 91].

Structure-based DTI models face additional challenges associated with structural bias, sequence or structure redundancy, and idealized evaluation conditions (Fig. 5c). Many approaches are trained and evaluated on co-crystal structures or near-native docking poses, which may not reflect the uncertainty encountered in practical screening workflows [111]. Homologous proteins or structurally similar binding pockets may also appear across training and test sets. In such cases, models may benefit from conserved folds, pocket geometry, or interaction patterns rather than learning transferable structure–activity principles. This issue is particularly relevant when benchmark datasets contain multiple complexes from the same protein family, highly similar binding sites, or closely related protein–ligand complexes. As a result, model performance may be overestimated under random complex-level splits and may decrease under cross-target, cross-family, cross-pocket, or cross-docking settings [13, 112]. Robustness under realistic structural uncertainty can be assessed using sequence-identity-controlled, protein-family-aware, and pocket-similarity-controlled evaluation. Additional strategies include cross-docking, pose perturbation, testing multiple protein conformations, and transparent reporting of docking and pocket-definition protocols [113].

Across all paradigms, dataset imbalance and negative-sample uncertainty introduce further limitations (Fig. 5d). DTI datasets are often skewed toward well-characterized target classes, such as kinases and G protein-coupled receptors (GPCRs), restricting generalization to underexplored regions of chemical and biological space [79, 114]. In addition, negative interactions in public datasets are frequently inferred from the absence of known interactions rather than experimentally confirmed non-binding events [102]. Some presumed negative pairs may represent unobserved or weak interactions, which can bias both training and evaluation [115]. Transparent reporting of negative-sampling procedures, class balance, and validation settings is therefore essential for fair model comparison.

Addressing these challenges requires stronger dataset curation, standardized benchmarking resources, and evaluation protocols aligned with intended applications. In practice, benchmarking should combine similarity-controlled data splits, leakage-aware network evaluation, transparent negative-sampling procedures, and structure-aware tests such as cross-docking or pose perturbation [13]. These practices are essential for distinguishing genuine model generalization from performance gains driven by chemical similarity, target homology, graph connectivity, or idealized structural inputs.

Conclusion and future perspectives

Graph-based DTI modeling has evolved into a diverse and versatile framework that spans multiple levels of biological abstraction, from systems-level association inference to structure-resolved and interaction-focused modeling. Network-based approaches enable large-scale hypothesis generation by integrating heterogeneous biomedical data [99, 116], whereas sequence-based models support scalable, target-aware affinity prediction in the absence of structural information [34, 66, 117]. Beyond model development, practical platforms such as DRHIN illustrate how heterogeneous-network-based prediction methods can be implemented as accessible web servers, supporting translational applications such as drug repositioning and hypothesis prioritization [118]. Structure-based models further increase mechanistic resolution by incorporating 3D context at the protein–ligand interface, enabling pose-sensitive and pocket-specific modeling [37, 119, 120]. Interaction-focused approaches add contact-level, pocket-aware, and pose-sensitive information, providing a more detailed description of molecular recognition. Together, these paradigms address complementary biological questions across distinct data regimes and stages of drug discovery, from early target identification to late-stage lead refinement and translational decision-making. Accordingly, the choice of model should be guided by data availability, structural confidence, interpretability requirements, and the specific objectives of the discovery program [121, 122].

Despite these advances, a central challenge in graph-based DTI modeling lies not in predictive performance alone, but in aligning model outputs with the informational and decision-making requirements of real-world drug discovery. Models that achieve high accuracy under benchmark conditions may fail to provide actionable insights if their outputs do not match the level of mechanistic detail required for downstream tasks. This highlights the need for an output-driven perspective, in which model design and evaluation are guided by the intended role of the model within the discovery pipeline, rather than by aggregate performance metrics alone [12, 123–125].

Looking forward, several strategies are needed to improve the translational reliability of graph-based DTI models. First, evaluation practices should become more realistic and application-specific. Rather than relying mainly on random splits, future benchmarks should assess ligand-side, target-side, network-level, and structure-based generalization using protocols appropriate to the intended deployment setting. This includes similarity-controlled evaluation, leakage-aware graph partitioning, and validation on external or prospective datasets where feasible. Such practices can reduce overestimation of model performance and provide a clearer view of whether predictions are robust beyond familiar ligands, targets, or structural contexts [126, 127].

Second, multimodal integration should preserve the uncertainty associated with each input modality rather than treating all information sources as equally reliable. Future models may benefit from combining ligand graphs, protein sequences, experimental or predicted structures, docking poses, omics profiles, and biomedical networks within unified representations [128, 129]. However, predicted protein structures and docking poses should be treated as uncertain inputs rather than fixed ground truth. A practical direction is to prioritize predictions that remain stable across alternative protein conformations, pocket definitions, docking poses, conformer ensembles, and model seeds.

Third, computational prediction should be embedded within iterative experimental workflows. Model-ranked candidates can be tested using biochemical, biophysical, cellular, mutagenesis, or structural assays. The resulting data can then be used to refine training sets, improve negative-label quality, recalibrate confidence estimates, and guide subsequent rounds of prediction. Such feedback loops are particularly important for interaction-focused and structure-based models, where predicted binding modes and atom–residue contacts require experimental support before being interpreted as mechanistic explanations.

Taken together, the future of graph-based DTI modeling will depend not only on advances in model architecture, but also on realistic benchmarking, uncertainty-aware structural modeling, multimodal data integration, and tighter feedback between computational prediction and experimental validation. By aligning model design with biological questions and translational objectives, graph-based approaches have the potential to move beyond predictive tools toward becoming integral components of data-driven drug discovery workflows.

Key points

• Graph-based models provide a unified framework for drug–target interaction (DTI) prediction across biomedical network-, sequence/hybrid-, and structure-based paradigms.

• Model outputs differ in mechanistic resolution and should be matched to specific stages of the drug discovery pipeline.

• Attention mechanisms and semi-supervised learning enhance representation quality and data efficiency in DTI modeling.

• Benchmarking remains challenging due to data leakage, sequence redundancy, structural bias, and unreliable negative samples.

• Future advances will rely on multimodal integration, improved evaluation strategies, and closer alignment between computational predictions and experimental validation.

Author contributions

Hien Minh To (conceptualization, data curation, writing and revising the manuscript for publication), Duy Anh Nguyen (conceptualization, data curation, writing and revising the manuscript for publication), Duy Trieu (conceptualization, data curation, writing and revising the manuscript for publication), Thanh Nguyen (conceptualization, data curation, writing and revising the manuscript for publication), and Giang Nguyen (project administration, funding acquisition, writing and revising the manuscript for publication). All authors reviewed and approved the final manuscript.

Conflicts of interest

All authors declare that they have no conflicts of interest.

Funding

This work was supported by Nanyang Biologics [No. A-25-10404 (03159–25)].

Data availability

No new data were generated or analyzed in support of this research.