Intelligence on the graph: Graph neural networks for mechanistic drug target discovery

Abstract

Drug discovery is increasingly challenged by rising costs, long development cycles and high attrition rates, with accurate target identification remaining a critical bottleneck. Although artificial intelligence (AI) has demonstrated transformative potential, the systematic application of graph neural networks (GNNs) to drug target discovery remains underexplored. To address this gap, this paper provides a comprehensive and structured analysis of recent advances in GNN-based methods for drug-target interaction (DTI) and drug-target affinity (DTA) prediction. We dissect the methodological foundations of representative architectures including graph convolutional networks (GCNs), graph attention networks (GATs) and graph autoencoders (GAEs), and compare their mechanisms, advantages and applicable scenarios in modeling complex molecular and biological systems. Also, we synthesize frontier paradigms such as multimodal data fusion, high-order graph reasoning and dynamic GNNs, which enable the capture of atom-residue interactions, multi-target coordination mechanisms and cross-scale biological features. By systematically mapping methodological innovations to biological applications, this paper offers both theoretical guidance and translational insights. The key contributions of this paper include: (1) establishing a comparative framework that clarifies when and how different GNNs architectures can be applied in drug target discovery; (2) integrating cutting-edge paradigms rarely addressed in prior reviews, such as multimodal fusion and high-order graph modeling; and (3) highlighting representative case studies that bridge algorithmic innovation with practical drug discovery outcomes. Collectively, this work provides an authoritative and forward-looking reference, promoting the development of AI-driven, efficient and interpretable drug discovery pipelines.

Graphical abstract

Highlights

• •We comprehensively summarize recent advances in Graph Neural Networks for drug–target interaction and affinity prediction.
• •Analysis of GCNs, GATs and GAEs reveals their strengths in modeling molecular and biological network complexities.
• •Emphasis on multimodal data fusion, high-order graph reasoning and dynamic learning enhances DTI/DTA predictive performance.
• •Key challenges such as data sparsity, computational load and interpretability are discussed with proposed solutions.
• •The paper advocates a paradigm shift toward mechanism-informed AI to advance precision medicine and rational drug design.

1. Introduction

Drug development is a highly complex and resource-intensive process that spans multiple stages, which is from target identification and compound screening to preclinical evaluation, clinical trials, regulatory approval and post-marketing surveillance. Despite decades of progress, this process continues to face formidable challenges, including exorbitant costs, lengthy development timelines and alarmingly low success rates. Current estimates suggest that developing a single new drug requires an average investment of approximately USD 2.6 billion and 12–15 years, with clinical failure rates approaching 90% [[1], [2], [3], [4]]. The challenges stem from the biological complexity of diseases, the vastness of chemical space (10⁶⁰–10⁸⁰ compounds) and stringent regulatory requirements [5,6]. These realities underscore the urgent need for transformative approaches to reduce costs, accelerate timelines and improve success rates across the drug discovery pipeline.

Drug target discovery, as the starting point of drug development, critically influences downstream success. Traditionally, computational tools such as molecular docking, quantitative structure-activity relationship (QSAR) modeling and pharmacophore analysis, have laid the foundation for computer-aided drug design [[7], [8], [9]]. With the development of artificial intelligence (AI), many approaches like machine learning and deep learning have enabled automatic extraction of complex patterns from large-scale biomedical data [10] (Fig. S1). Classical machine learning models, e.g., random forests and support vector machines, further advanced predictive capabilities for drug-target interactions (DTI) and drug-target affinities (DTA) [11,12]. Particularly, graph neural networks (GNNs) stand out by operating directly on graph-structured data, allowing atoms, residues, proteins and even signaling pathways to be represented as nodes and edges. This unique ability makes GNNs more powerful for modeling molecular structures, protein flexibility and multi-target interactions that traditional methods struggle to capture. Indeed, GNNs operate directly on graph-structured data where atoms are represented as nodes and chemical bonds as edges and can integrate high-order topological information with multimodal biological features (Fig. S2). All these capabilities have significantly enhanced the predictive power and interpretability of models in drug development, particularly in drug target discovery (Fig. 1).

Fig. 1.

Multi-dimensional applications of graph neural networks (GNNs) in drug development. GNNs have been increasingly adopted across various stages of the drug discovery and development pipeline, leveraging their ability to model complex molecular and biological structures. Key application domains include: (1) Molecular property prediction; (2) Virtual screening; (3) Drug response prediction; (4) Drug-drug interaction prediction and (5) Drug target prediction. These capabilities highlight the transformative potential of GNNs in enabling data-driven, mechanism-aware drug development. ADMET: absorption, distribution, metabolism, excretion, toxicity.

In addition, conventional approaches including high-throughput screening, structure-based drug design and phenotype-based assays, have yielded important breakthroughs [[13], [14], [15]], but are often limited by their single-target focus, cost and labor intensity [16]. By contrast, GNNs enable network-based modeling of complex biological systems, simultaneously predicting multiple DTI and elucidating atom-residue interactions or polypharmacological effects [[17], [18], [19]]. Recent models such as Siamese spectral convolutional networks and self-supervised GNNs demonstrate the potential of these methods to improve predictive accuracy, interpretability and scalability in drug discovery. While several reviews have introduced the potential of GNNs for drug discovery, existing work suffers from two major limitations: (1) a predominant focus on narrow application scenarios, e.g., DTI or DTA alone, without a systematic comparison of GNNs architectures; and (2) insufficient discussion of the mechanisms, design principles and applicable conditions of representative models such as graph convolutional networks (GCNs), graph attention networks (GATs) and graph autoencoders (GAEs) [20]. These gaps hinder the practical adoption of GNNs in biomedical research and drug development.

Currently, to address these gaps and to provide a dedicated, in-depth and forward-looking examination of GNNs specifically for drug target discovery, we systematically analyze and contrast major GNNs architectures, providing a principled framework for selecting appropriate models under different biological contexts. Additionally, we incorporate recent advances in multimodal data fusion, high-order graph modeling, and dynamic learning, which are essential for capturing the complexity of real-world drug discovery. Besides, the representative studies where GNNs have yielded biologically and pharmacologically meaningful predictions are also highlighted, demonstrating how algorithmic innovations can directly impact therapeutic discovery. A detailed description of the methodologies for drug target discovery and their practical applications is also provided in the Supplementary materials. By combining theoretical rigor with practical insights, this paper not only consolidates current knowledge but also provides actionable guidance for computational biologists, chemists and drug developers, as well as promotes the development of next-generation, AI-driven pipelines for efficient and interpretable drug target discovery.

2. GNN overview

GNNs have emerged as a powerful class of deep learning architectures designed to operate directly on graph-structured data by iteratively propagating and aggregating information across nodes and edges [21,22]. Despite their relatively recent inception, GNNs have rapidly outperformed traditional machine learning and generic deep learning models on a variety of graph-based tasks [23]. Their ability to learn rich, topology-aware embeddings makes them particularly well suited for drug target discovery, where the underlying molecular and biological interaction networks are naturally represented as graphs.

To place these advances in context, several seminal reviews have laid the theoretical and methodological foundations for GNN development. For instance, Bronstein and colleagues [24] situated GNNs within the broader paradigm of geometric deep learning, outlining key challenges and potential solutions for learning in non-Euclidean domains. Building on this, Zhang et al. [25] provided an in-depth analysis of GCNs, dissecting the mathematical formulation of convolution operators on irregular structures. Complementarily, Wu et al. [26] proposed a taxonomy that classifies GNNs into recurrent, convolutional, autoencoder, and spatio-temporal variants, each addressing distinct facets of graph learning. More recently, Zhang et al. [27] expanded this classification to include reinforcement learning and adversarial approaches, offering a comprehensive survey of models organized by architectural motif and training paradigm.

Additionally, the application of GNNs to drug discovery has become particularly compelling. By modeling small molecules as atom-bond graphs and protein targets as residue-interaction graphs, GNNs can effectively capture the multiscale dependencies underlying DTIs. Consequently, in the following sections, we focus on three foundational GNN architectures, i.e., GCNs, GATs and GAEs, and discuss how their distinct mechanisms of neighborhood aggregation and representation learning enable accurate prediction of DTI and DTA.

2.1. GCNs

CNNs have long excelled at extracting hierarchical representations from grid-structured data such as images. However, many real-world systems including social networks, molecular graphs and biological interaction maps are inherently non-Euclidean, thereby defying the direct application of conventional CNNs [24]. To address this limitation, GCNs generalize the convolution operation to arbitrary graph topologies, enabling end-to-end learning of node representations that integrate both local connectivity and node-specific attributes. Conceptually, each GCN layer updates a node's embedding by aggregating and transforming features from its neighboring nodes, thus propagating information across the graph in a process analogous to message passing (Fig. 2A). This formulation can be traced back to the seminal GNNs framework introduced by Gori et al. [28], who first formalized graph-based recurrent message propagation.

Fig. 2.

Graph convolutional networks (GCNs) principle and architecture diagram. (A) The processing flow of graph topology and features in graph convolutional neural networks. (B) The network architecture of GCN. ReLU: rectified linear unit.

Building on this foundation, subsequent advancements replaced costly iterative propagation with spectral and spatial formulations. Spectral GCNs exploit the eigendecomposition of the graph Laplacian to define convolution operations in the Fourier domain, whereas spatial GCNs perform localized message aggregation directly within each node's neighborhood. Earlier unsupervised graph embedding approaches such as DeepWalk and node2vec employed random walks and skip-gram models to learn node embeddings [29]. However, these methods lacked inductive generalization and suffered from scalability limitations. In contrast, modern GCNs combine neighbor aggregation with trainable weight matrices and non-linear activations, yielding highly expressive, scalable models that generalize across graphs. As illustrated in Fig. 2B, a typical GCN stacks multiple convolutional layers, interleaved with nonlinearities and normalization, to capture increasingly global graph structure.

By jointly encoding both topological structure and node-level features, GCNs have become a foundational technique in graph-based drug-target prediction. They enable fine-grained modeling of atom-level interactions in small molecules and residue-level contacts in proteins, thereby enhancing the prediction of DTIs, binding affinities, and virtual screening performance. The efficiency and representational flexibility of GCNs have enabled breakthroughs in predicting DTIs, affinity estimation and virtual screening, setting the stage for more advanced GATs and GAEs architectures.

Actually, GCNs have revolutionized the application of convolutional operations to graph-structured data by enabling localized, parameter-efficient message passing. The foundational spectral formulation by Bruna et al. [30] defines graph convolution via the eigendecomposition of the graph Laplacian, thereby projecting node features into a Fourier basis, but suffers from cubic complexity. However, Defferrard et al. [31] overcame this limitation with ChebNet, which approximates spectral filters using truncated Chebyshev polynomials to restrict convolutions to a fixed neighborhood radius, eliminating expensive eigenvalue computations. Further, Kipf et al. [23] simplified this approach to a first-order approximation that blends each node's own features with those of its immediate neighbors, while applying symmetric normalization to balance contributions across nodes of differing degrees.

Meanwhile, spatial-based approaches, such as GraphSAGE, circumvent spectral theory entirely by performing inductive feature aggregation through neighborhood sampling. These methods aggregate information from sampled neighboring nodes and combine their embeddings using simple yet effective functions, such as mean, pooling, or even learnable sequence models, thereby enabling seamless generalization to previously unseen nodes [32]. More recent architectures, exemplified by Graph features and pharmacophores augmented cross-attention networks based drug-target binding affinity prediction (GPCNDTA), further advance this paradigm by integrating domain-specific constraints (e.g., pharmacophore or physicochemical information) with attention mechanisms to emphasize critical binding residues within drug-target complexes. This integration not only enhances predictive accuracy but also improves the interpretability of the learned representations [33].

In addition, during the process of drug discovery, GCNs naturally capture the non-Euclidean structure of molecular graphs (atoms connected by bonds) and protein interaction networks (residues linked by spatial or biochemical contacts) [34]. They excel at diverse tasks including molecular property prediction, DTI scoring, drug-drug interaction modeling, and individualized drug response forecasting. For example, Torng et al. [35] and Liu et al. [36] demonstrated a Graph-CNN framework for protein-ligand affinity prediction and DeepCDR combined multi-omics profiles with chemical structure graphs to forecast cancer drug sensitivity.

By iteratively aggregating neighbor information while preserving each node's original attributes, GCNs capture both local topology and node-specific features without diluting critical signals. This combination of expressive power, scalability, and inductive generalization has firmly established GCNs as a core tool for deep learning on non-Euclidean data, with profound implications for social network analysis, recommendation systems, and, critically, the accurate modeling of complex biological systems in drug development.

2.2. GATs

GATs represent a major advancement in GNNs by incorporating a dynamic, context-sensitive attention mechanism that overcomes the rigidity of fixed, uniform neighbor aggregation in standard GCNs. Inspired by attention models in natural language processing and transformer architectures, GATs enable each node to learn how much importance to assign to each of its neighbors, based on their feature representations [37,38]. For instance, Velicković et al. [39] pioneered this approach by introducing self-attention into graph learning: they first project all node features into a shared latent space, then compute pairwise attention scores for connected node pairs using a small neural network, and finally normalize these scores across each node's neighborhood so that they sum to one.

This adaptive weighting allows nodes to selectively emphasize informative neighbors while attenuating noisy or irrelevant connections, greatly enhancing representational power for heterogeneous and complex graph structures. In practice, a GAT layer applies the same linear transformation to every node's feature before computing attention coefficients. These coefficients are then used to weight neighbor features, which are aggregated and passed through a nonlinearity to form refined embeddings (Fig. 3). Further, multi-head attention variants stabilize learning by averaging or concatenating multiple independent attention mechanisms.

Fig. 3.

Workflow of a graph attention network (GAT). A GAT projects node features into a shared embedding space, computes pairwise attention coefficients via a learnable mechanism, and normalizes them across each neighborhood using softmax. Node representations are then updated through weighted aggregation of neighbors' features followed by non-linear activation. Multi-head attention can be employed in parallel, with outputs concatenated or averaged, to enhance stability and capture diverse relational patterns.

Indeed, in drug discovery, GATs excel at capturing subtle molecular interactions by assigning higher importance to atoms or residues that are most predictive of binding or activity. For example, GAT-based models have demonstrated improved accuracy in predicting drug-target binding affinities by focusing attention on key pharmacophore interactions and allosteric residues. Their ability to learn context-dependent importance makes GATs particularly well suited for modeling cross-scale, multi-modal biological graphs where different node types and interaction strengths must be weighed differently.

Building on the foundational GAT architecture, Brody et al. [40] identified a tendency for static, rather than dynamic, attention patterns in the original formulation and introduced GATv2, which reorders the internal operations and applies a non-linear activation (e.g., LeakyReLU) to the attention computation, thereby restoring full expressiveness to the learned attention weights. Extending the attention paradigm to hypergraphs, Bai et al. [41] proposed the hypergraph attention network, which captures higher-order relationships such as multi-node interactions far beyond pairwise edges and demonstrates superior semi-supervised classification performance relative to standard GATs. More recently, Weng et al. [42]'s graph attention&interaction network (GAIN) integrates multi-head attention with an additional “aggregator-level” attention mechanism and explicit feature interaction modules, improving inductive learning on large-scale graphs and excelling in user response prediction tasks.

Actually, in drug discovery, GATs have proved particularly adept at highlighting the most relevant atoms or residues within complex molecular and protein interaction graphs. For example, Xiong et al. [43]'s AttentiveFP employs a layered attention scheme to learn powerful molecular fingerprints, achieving state-of-the-art performance on diverse drug datasets. Moreover, Lv et al. [44] introduced Meta-GAT, combining GATs with meta-learning to predict molecular properties in low-data regimes typical of early-stage drug development. In the realm of polypharmacology, Su al. [45]'s drug-drug interactions prediction of knowledge graphs (DDKG) framework applies attention-augmented knowledge graph embeddings to forecast drug-drug interactions. Further, Huang et al. [46] refined this approach in hierarchical and dynamic graph attention network (HDGAT), a hierarchical, dynamic attention network that outperforms prior methods in drug-disease prediction tasks.

Despite these advances, challenges remain. Deeper GAT stacks still suffer from performance degradation, although residual connections (ResGAT) mitigate some issues, fully preserving attention dynamics in deep models demands further innovation. Most current GAT variants also underutilize edge attributes and directionality though directional GATs partially address this and the dense attention computations continue to hinder scalability on very large graphs. Nonetheless, the ongoing evolution of GAT architectures from basic self-attention to hypergraph and hierarchical designs continues to expand their applicability across graph-structured domains, promising ever more precise and interpretable models for complex systems such as molecular networks in drug discovery.

2.3. GAEs

GAEs adapt the powerful encoder-decoder framework of classical autoencoders—renowned for unsupervised representation learning tasks such as denoising, dimensionality reduction, and anomaly detection, to graph-structured data. By mapping nodes into a compact latent space and reconstructing the graph topology, GAEs learn continuous embeddings that preserve both node features and network connectivity [[47], [48], [49], [50], [51]]. Additionally, variants such as variational graph autoencoders (VGAEs) and adversarial autoencoders introduce probabilistic and adversarial regularization, respectively, yielding more robust, expressive models capable of capturing uncertainty and enforcing distributional priors. These advances enable GAEs to tackle key graph-analysis challenges—node embedding, link prediction, graph generation, and relational reasoning—while faithfully retaining the semantic and topological complexity of real-world networks. The general architecture of a graph autoencoder is illustrated in Fig. 4.

Fig. 4.

The graph autoencoders (GAEs) comprises five key components: (1) Input, where node features and graph structure (adjacency) are provided; (2) Encoder, typically a stack of graph convolutional layers that transforms inputs into a compact latent embedding; (3) Latent space, representing each node by a low-dimensional vector or distribution; (4) Decoder, which reconstructs graph connectivity (and optionally node features) from the latent embeddings—often via inner-product or neural predictors; and (5) Output, the reconstructed adjacency (and feature) matrices. This encoder-decoder framework enables unsupervised learning of node representations that capture both topological and attribute information.

In fact, GAEs bring the power of unsupervised representation learning to graph-structured data by embedding nodes into a low-dimensional latent space from which the original graph topology can be reconstructed. Early efforts applied sparse autoencoders directly to the adjacency matrix, compressing node features through an L₂ reconstruction loss but failing to account for explicit graph structure [52]. The introduction of variational autoencoder (VAEs), which learn a probabilistic latent distribution via variational inference laid the groundwork for probabilistic GAEs. For instance, Kipf and Welling [53]'s VGAE replaces the standard encoder with a graph convolutional network to extract node embeddings and uses an inner-product decoder to reconstruct edges, enabling effective link prediction while preserving topological and feature information.

Additionally, subsequent variants have enhanced expressivity and robustness. For instance, the Graph2Gauss represents each node as a Gaussian distribution, leveraging Kullback-Leibler (KL) divergence to encode pairwise constraints and capture uncertainty [54], whereas deep variational network embedding (DVNE) adopts a Wasserstein distance metric to maintain similarity transitivity in high-dimensional spaces and overcome KL-divergence limitations [27]. Adversarially regularized GAEs further impose prior consistency by integrating a generative adversarial network in latent space, leading to more robust and generalizable embeddings [55]. To scale these models to large graphs, GraphSAGE employs neighborhood sampling to enable inductive inference on unseen nodes [32], while Cluster-GCN partitions graphs into subgraphs for minibatch training, achieving linear scalability [56].

In drug discovery, GAEs excel at modeling the high-dimensional and heterogeneous nature of biomedical networks. Hybrid models combining GAEs with GANs have demonstrated robust performance on sparse protein-protein interaction and molecular graphs [57]. For drug-disease association, multi-similarity GAE frameworks aggregate heterogeneous drug and disease similarities, yielding significant gains through adaptive feature fusion [58]. Knowledge-graph-embedded Wasserstein autoencoders preserve semantic relationships in drug-drug interaction prediction while enforcing distributional consistency [59]. Moreover, attention-augmented VGAEs like those in the GVDTI model leverage attribute-level attention to extract discriminative features from protein sequences and molecular graphs, further boosting DTI accuracy [60].

Despite these advances, GAEs development continues to face challenges posed by graph scale, density, and heterogeneity. Large-scale graphs impose significant memory and computational burdens, and sparse, diverse node features complicate reliable embedding learning [61]. Addressing these issues will require more efficient architectures, dynamic sampling strategies, and algorithmic innovations to fully realize GAEs’ potential in computational drug discovery.

3. GNNs in drug-target discovery

With the continuous advancement of biomedical research, drug target prediction, as a central aspect of drug development, has garnered increasing attention [62]. Accurate drug target prediction not only accelerates the process of drug discovery but also provides strong support for personalized medicine. Traditional approaches for predicting drug targets primarily rely on experimental data and prior biological knowledge. However, these methods often suffer from issues such as data scarcity and the high dimensionality of biological information.

In recent years, GNNs, as a powerful deep learning technology, have demonstrated excellent performance across various bioinformatics tasks, particularly in the field of drug target prediction [63]. Drug target prediction tasks can generally be classified into two main types: DTI and DTA predictions. DTI prediction aims to determine whether an interaction exists between a drug and a target, whereas DTA prediction focuses on estimating the binding strength between them.

In both tasks, GNNs can fully exploit the complex topological relationships between molecules by modeling molecular structures and biological networks as graphs, thereby enabling more accurate and robust prediction. Numerous GNN-based models have been developed for these tasks. According to the strategies used for learning protein and drug features, these models can be broadly categorized into four types. Type A is the symmetric GNN architecture, where both proteins and drugs are encoded using GNNs. Type B is the asymmetric encoding architectures, which use different modalities for encoding proteins and drugs. Type C is the multi-channel or heterogeneous feature fusion models that integrate multiple types of data. Type D is the non-traditional GNN approaches that incorporate novel architectural or algorithmic innovations. A summary of representative DTI and DTA prediction models based on GNNs is provided in Table 1 [[63], [64], [65], [66], [67], [68], [69], [70], [71], [72], [73], [74], [75], [76], [77], [78], [79], [80], [81], [82], [83], [84], [85], [86], [87], [88], [89], [90], [91], [92], [93], [94], [95], [96], [97], [98], [99], [100]] and Table 2 [33,67,77,78,81,82,85,86,[101], [102], [103], [104], [105], [106], [107], [108], [109], [110], [111], [112], [113], [114], [115], [116], [117], [118], [119], [120], [121], [122], [123], [124], [125], [126], [127]].

Table 1.

Summary of drug-target interaction prediction models based on graph neural networks (GNNs).

Model name	Model type	Protein feature learning	Drug feature learning	Prediction method	Datasets used in building models
GCN-DTI [64]	A	GCN	GCN	DNN	Yamanashi et al. [65]'s dataset, HIPPIE [66], and DrugBank (5.0) [67]
DTIHNC [68]	A	DAE, and GAT	DAE, GAT	CNN	DTINet [69]
MHGNN [70]	A	GAT	GAT	GCN	DTINet [69], DrugBank (5.0) [67], and UniProtKB [71]
AMGDTI [72]	A	Node2Vec and GCN	Node2Vec and GCN	–	DTINet [69], and Zheng et al. [73]'s dataset
DTI-HETA [74]	A	GCN, and GAT	GCN, GAT	Inner product decoder	He et al. [75]'s dataset, and Yamanashi et al. [65]'s dataset
SSLDTI [76]	A	Self-supervised, and GCN	Self-supervised and GCN	–	NeoDTI [77], and deepDR [78]
DT-DHG [79]	A	Dynamic heterogeneous GCN	Dynamic heterogeneous GCN	–	Yamanashi et al. [65]'s dataset, and Shao et al. [74]'s dataset
EmbedDTI [80]	B	CNN	GCN with attention module	FC	Kinase dataset, Davis et al. [81]'s dataset, and KIBA dataset [82]
Transfer learning and BNN DPI [83]	B	Transformer + CNN	GraphNet [84]	FC	BindingDB dataset [85], Human [86] and C. elegans datasets [86]
GNN and CNN CPI [87]	B	CNN	GNN	Attention + Softmax	CPI datasets [86], DrugBank 4.1 [88], Matador [89]
TransformerCPI [90]	B	Gated Convolutional network and word2vec [91]	GCN	FC	BindingDB dataset [85], Human dataset [86], Caenorhabditis elegans dataset [87]
HampDTI [92]	C	GNN (multi-channel embeddings)	GNN (multi-channel embeddings)	Channel attention	DTINet [69]
DTI-MGNN [93]	C	Multi-channel GNN (GCN, CAT, and GAE)	Multi-channel GNN (GCN, CAT, and GAE)	MLP	DrugBank database (Version 3.0) [94], HRDO database [95]
GanDTI [96]	C	Attention module	Residual GNN	MLP	Human datasets [86] and Binding Database [85]
GSRF-DTI [97]	D	Deepwalk, and GraphSAGE	Deepwalk, and GraphSAGE	Random forest classifer	DTINet [69]
GeNNius [63]	D	Amino acid ratio + Residual2vec	Molecular descriptors (RDKit/SMILES)	SAGEConv + neural network	DrugBank [98], BioSNAP [99], Binding Database [85], Davis et al. [81]'s dataset, and Yamanashi et al. [65]'s dataset
DrugormerDTI [100]	D	Residual2vec + Transformer	Graph Transformer	Transformer decoder + FC lLayers	Human [86], C.elegans [86], Davis et al. [81]'s dataset, and GPCR [90].

DNN: deep neural network; GCN: graph convolutional network; DAE: denoising autoencoder; GAT: graph attention network; CNN: convolutional neural network; FC: fully connected layer; MLP: multilayer perceptron.

Table 2.

Summary of drug-target affinity prediction models based on graph neural networks.

Model name	Model type	Protein feature learning	Drug feature learning	Prediction method	Datasets used in building models
MGraphDTA [101]	A	MCNN	MGNN	MLP	Metz [102], KIBA [82], Davis et al. [81]'s dataset, Human and C. elegans [86]
WGNN-DTA [103]	A	WGNN	WGNN	FC	Davis et al. [81]'s dataset and KIBA [82], Human [86], C.elegans [86] and DUD-E dataset [104]
DGraphDTA [105]	A	GNN (GCN/GAT)	GNN (GCN/GAT)	FC	Davis et al. [81]'s dataset and KIBA datasets [82]
GSAML-DTA [106]	A	GNN (GCN/GAT)	GNN (GCN/GAT)	FC	Davis et al. [81]'s dataset and KIBA datasets [82]
S2DTA [107]	A	CNN, and GCN	CNN, and GCN	FC	PDBbind database (2016 version) [108]
DeepMGT-DTI [109]	A	CNN	MCGCN and modified transformernetwork	Fully connected neural network	Tang et al. [110] and DrugBank database [67], KEGG database [111] and PubChem database [112]
GraphDTA [113]	A	CNN	GNN (GCN/GAT/GIN/GAT-GCN)	FC	NeoDTI [77] and deepDR [78]
SAG-DTA [114]	B	CNN	GNN with SAG	FC	Davis et al. [81]'s dataset and KIBA datasets [82]
DTA-HYGCN [115]	B	CNN with TF-IDF	GNN (GCN/GAT/GIN)	FC	Davis et al. [81]'s dataset, KIBA [82], Binding Datasets [85] and Human datasets [86]
SSR-DTA [116]	B	BiGNN	Multi-head GAT	MLP	Davis et al. [81]'s dataset and KIBA datasets [82]
NHGNN-DTA [117]	C	BiLSTM	BiLSTM	GIN	Davis et al. [81]'s dataset, KIBA [82], Metzet al. [102]'s dataset and BindingDB datasets [85]
MSGNN-DTA [118]	C	GNN and Gated skip-connection mechanism and GNN	GNN and Gated skip-connection mechanism and GNN	GNN	Davis et al. [81]'s dataset and KIBA datasets [82]
GPCNDTA [33]	C	CensNet [119], EW-GCN	CensNet, and EW-GCN	MLP	Davis et al. [81]'s dataset and KIBA datasets [82]
CGraphDTA [120]	D	Multiscale CNN and variant GNNs	Multiscale CNN and variant GNNs	FC	Davis et al. [81]'s dataset, KIBA [82], Filtered Davis [121], Metz et al. [102]'s dataset and ToxCast dataset [122]
GRA-DTA [123]	D	Attention based BiGRU	GraphSAGE [124]	FC	PDBbind (version 2019) [125], and CASF2016 datasets [126]
TDGraphDTA [127]	D	Multi-scale CNN	Molecular Graph + Diffusion	GNN + Transformer	Davis et al. [81]'s dataset, and KIBA datasets [82]

MCNN: multiscale convolutional neural network; MGNN: multiscale graph neural network; MLP: multilayer perceptron; WGNN: weighted graph neural network; FC: fully connected layer; GNN: graph neural network; GCN: graph convolutional network; GAT: graph attention network; CNN: convolutional neural network; MCGCN: molecular complementary graph convolutional neural network; Bi-LSTM: bidirectional long short-term memory; GIN: graph isomor-phism network; EW-GCN: edge-weighted GCN; GRU: gated recurrent unit; BiGRU: bidirectional gated recurrent unit.

Additionally, the quantitative analysis of DTI and DTA prediction models based on GNNs are summarized in Table 3 [33,63,64,68,70,72,74,76,79,80,83,87,[91], [92], [93],96,97,100,101,103,[105], [106], [107],109,[113], [114], [115], [116], [117], [118],120,123,127]. Obviously, Classical methods such as GCN-DTI and heterogeneous networks and cross-modal similarities for drug-target interaction (DTiHNC) already achieved high performance (AUC ≈ 0.96−0.98), while more recent architectures, including MHGNN (metapath-aggregated heterogeneous graph neural network), metapath-aggregated heterogeneous graph neural network (DT-DHG) and adaptive meta-graph-based method for drug-target interactions (AMGDTI), reached near-optimal accuracy (AUC > 0.98, F1 up to 0.96). In DTA tasks, models such as MGraphDTA, DGraphDTA and GraphEmbDTA achieved the lowest error rates (MSE ≈ 0.126−0.128), underscoring the advantage of graph-embedding strategies. By contrast, some approaches (e.g., deep learning model with transformer network incorporating multilayer graph information for drug-target interaction (DeepMGT-DTI)) showed imbalanced precision-recall or relatively modest gains. Collectively, these results highlight that advanced GNN variants integrating heterogeneous graphs, attention mechanisms, or embedding techniques markedly enhance both predictive accuracy and regression robustness, offering a powerful paradigm for drug discovery.

Table 3.

Quantitative analysis of drug-target interaction (DTI) and drug-target affinity (DTA) prediction models based on graph neural networks (GNNs).

Methods	AUPR	AUC	F1 value	MSE
GCN-DTI [64]	0.98	0.98	–	–
DTIHNC [68]	0.969	0.963	0.910	–
MHGNN [70]	0.9799	0.9893	0.9563	–
AMGDTI [72]	0.977	0.977	–	–
DTI-HETA [74]	0.94722	0.93224	–	–
SSLDTI [76]	0.8637	0.9346	–	–
DT-DHG [79]	0.986	0.990	–	–
EmbedDTI [80]	–	–	–	0.133
Transfer learning and BNN DPI [83]	–	0.981	–	–
GNN and CNN CPI [87]	–	0.95	–	–
TransformerCPI [90]	–	0.973	–	–
HampDTI [92]	0.9263	0.9273	0.8689	–
DTI-MGNN [93]	0.9683	0.9665	–	–
GanDTI [96]	–	0.983	–	–
GSRF-DTI [97]	0.9839	0.9818	–	–
GeNNius [63]	0.9349	0.9340	–	–
DrugormerDTI [100]	–	–	–	–
MGraphDTA [101]	–	0.983	–	0.128
WGNN-DTA [103]	–	0.994	0.973	0.149
DGraphDTA [105]	–	–	–	0.126
GSAML-DTA [106]	–	–	–	0.201
S2DTA [107]	–	–	–	–
DeepMGT-DTI [109]	0.7711	0.9024	0.7931	–
GraphDTA [113]			–	–
SAG-DTA [114]	0.986	0.985	–	0.131
DTA-HYGCN [115]	–	–	–
SSR-DTA [116]	–	–	–	0.121
NHGNN-DTA [117]	–	–	–	0.124
MSGNN-DTA [118]	–	–	–	0.117
GPCNDTA [33]	–	–	–	0.165
CGraphDTA [120]	–	–	–	-
GRA-DTA [123]	–	–	–	0.142
TDGraphDTA [127]	–	–	–	0.121

AUPR: area under the precision versus recall curve; AUC: area under the receiver operating characteristic curve; MSE: mean squared error.

3.1. GNNs in DTI prediction

DTI refers to the specific binding between drug molecules such as small-molecule compounds or biologics and in vivo targets, typically proteins like enzymes, ion channels, G protein-coupled receptors, or nuclear receptors. These interactions exert pharmacological effects by modulating the biological functions of targets, such as activating, inhibiting, or regulating signaling pathways, ultimately contributing to disease treatment [65]. Traditional approaches for predicting DTI, including molecular docking and QSAR modeling, often suffer from high computational cost and time-consuming experimental validation.

In recent years, GNNs have emerged as powerful tools for DTI prediction due to their inherent capacity to model non-Euclidean data structures, such as molecular graphs and protein-protein interaction networks. The general pipeline of using GNNs for DTI prediction typically involves four key stages. First, GNNs are employed to extract feature representations of drug molecules and protein targets (Fig. 5A). Next, the extracted features are used to construct both a topological graph and a feature matrix (Fig. 5B). Subsequently, GCNs are used to learn multi-view graph representations of drug-protein pairs (DPPs) (Fig. 5C). Finally, these learned representations are passed through a multilayer perceptron (MLP) or fully connected (FC) layer to predict DPPs interaction scores (Fig. 5D).

Fig. 5.

Schematic diagram of drug-target interaction analysis and feature extraction process based on graph neural networks (GNNs). (A) The acquisition process of drug and protein features. (B) The graph construction and feature representation of drug-protein pairs (DPPs). (C) The multi-view graph representation learning process. (D) The final prediction process.

Despite the promising performance of GNN-based models, DTI prediction still faces two major challenges. The first lies in the complexity of heterogeneous biological relationships. Biomedical networks typically consist of multiple types of nodes and edges, such as drug-target, drug-drug, and target-target interactions. Effectively capturing and integrating both topological and semantic information in such heterogeneous graphs remains difficult. The second challenge is the sparsity of interaction data and the cold start problem. The limited availability of known DTI data hampers the ability to predict interactions involving novel drugs or targets with no prior interaction history.

To address these issues, researchers have developed a series of innovative computational frameworks that enhance model performance and generalization. These frameworks can be broadly categorized into four paradigms: multimodal fusion (integrating diverse data types), dynamic learning (capturing temporal or contextual changes), high-order modeling (considering more complex topological dependencies) and cross-technology integration (combining GNNs with other machine learning or deep learning techniques). These advancements highlight the increasing sophistication and flexibility of GNN-based models in tackling the inherent challenges of DTI prediction.

3.1.1. Multimodal fusion for DTI prediction

Multimodal fusion integrates heterogeneous data sources such as chemical structures, gene expression profiles, protein sequences and biomedical knowledge graphs to enhance predictive performance through comprehensive modeling [128]. The fundamental premise is that leveraging the complementarity, redundancy, and robustness of multiple modalities can overcome the limitations inherent in single-modality representations. Conventional DTI prediction approaches, including molecular docking and network-based inference, often rely on isolated data modalities, which are insufficient to capture the complex and multifaceted nature of drug–target relationships.

Recent advances in GNNs have enabled effective fusion of structural, bioinformatic, and semantic modalities, yielding more accurate and generalizable models [129]. For instance, the DTI-multiscale graph neural network (MGNN) framework proposed by Li et al. adopts a dual-channel architecture that combines GATs and GCNs to simultaneously capture topological drug-protein interactions and semantic features (e.g., molecular fingerprints, protein sequences) [93]. Additionally, independent GATs encode distinct graph types, while shared-weight GCNs extract common latent representations. This architecture achieved an AUC of 0.9665, demonstrating the efficacy of attention mechanisms in prioritizing informative node-edge interactions while mitigating noise. However, the reliance on predefined topologies and static feature maps may constrain its adaptability in dynamic or evolving biological networks.

Moreover, Shao et al. [74] further extended this approach by proposing the drug-drug interactions prediction of end-to-end model based on HETerogeneous graph with attention mechanism (DTI-HETA) model, which integrates drug-drug, target-target, and DTI subnetworks into a unified heterogeneous graph using a GCN-GAT hybrid structure. The attention mechanism enables selective aggregation of neighborhood information, enhancing predictive accuracy in sparse interaction settings. While its end-to-end architecture improves implementation efficiency, its scalability to highly heterogeneous datasets with diverse node types and edge attributes remains a concern. Nonetheless, both DTI-MGNN and DTI-HETA underscore the importance of jointly modeling network topology (e.g., adjacency matrices) and node semantics (e.g., molecular descriptors) for robust DTI prediction.

Besides, to address the challenge of modeling higher-order semantic relations, Li et al. [70] introduced MHGNN, a meta-path-based heterogeneous GNN that constructs drug-target pair (DTP) graphs and aggregates information along biologically meaningful paths (e.g., drug-pathway-protein). This design captures indirect associations often overlooked by traditional GNNs. By explicitly modeling DTP nodes and context-aware embeddings, MHGNN enhances interaction discovery and outperforms 17 baseline models. However, the model's performance is contingent on carefully curated meta-paths, which may require domain-specific knowledge.

Furthermore, to overcome the limitations of manually defined meta-paths, Su et al. [72] proposed AMGDTI, which leverages adaptive metagraph learning to automatically discover optimal semantic paths within heterogeneous networks. This approach eliminates the dependency on prior biological knowledge and allows dynamic integration of diverse data sources such as drug similarity networks and protein-protein interactions. By learning metagraph weights in an end-to-end fashion, AMGDTI effectively captures fine-grained topological semantics. Both MHGNN and AMGDTI highlight the critical role of meta-path or metagraph design in uncovering non-obvious but biologically relevant associations, such as polypharmacology or cross-pathway interactions.

Collectively, these representative models exemplify the transformative potential of multimodal fusion in DTI prediction. Innovations in attention mechanisms, metagraph learning, and heterogeneous graph representation have advanced the state-of-the-art. Nonetheless, challenges remain, particularly regarding the interpretability and biological validation of automatically generated semantic structures, which will be critical for translational applications.

3.1.2. Dynamic learning

In the context of machine learning and deep learning, dynamic learning refers to an algorithm's capacity to adaptively modify its parameters or structural representations in response to temporal evolution or the arrival of new data. This adaptive behavior is particularly crucial for modeling complex, time-dependent biological processes. For instance, dynamic learners have been employed to track state transitions of nodes within evolving graph-structured data, enabling more effective modeling of dynamic molecular interactions [130].

To address data sparsity and temporal variation in DTI prediction, Xu et al. [79] developed a DTI prediction model based on dynamic heterogeneous graph (DT-DHG), which constructs dynamic heterogeneous graphs (DHGs) and incorporates progressive learning to adaptively adjust the receptive field of each node. This strategy enhances information propagation in sparse networks and captures complex temporal patterns in drug–target associations. Notably, DT-DHG demonstrates superior robustness in imbalanced datasets, where positive samples are scarce.

DT-DHG emphasizes time-aware graph reconstruction to address data sparsity. It shows the importance of dynamic and adaptive mechanisms in capturing the non-static characteristics of biological systems. However, a limitation lies in their reliance on predefined topologies, which constrains generalization to previously unseen drug–target pairs or novel biological contexts.

Collectively, dynamic learning frameworks highlight the necessity of temporal modeling and structural adaptivity in DTI prediction, especially under conditions of data scarcity, imbalance, or temporal drift. Future advances may benefit from integrating continuous-time graph models, self-evolving architectures, or reinforcement learning paradigms to further enhance adaptability and generalization.

3.1.3. Higher-order modeling

Traditional GNNs are inherently constrained by their local message-passing mechanisms, limiting their ability to capture complex, high-order interactions that extend beyond immediate node neighborhoods. Higher-order modeling addresses this limitation by directly incorporating global or relational contexts, enabling the representation of more intricate biological phenomena, such as DTP diffusion and transmission processes [131].

To overcome the locality bottleneck, Zhao et al. [64] introduced GCN-DTI, a model that conceptualizes each DTP as an independent node in the graph, with edges encoding relationships between DTPs. By performing graph convolution over the DTP-centric network, GCN-DTI extracts context-aware features that are subsequently refined using a deep neural network for interaction prediction. This architecture enables the explicit modeling of indirect drug-target relationships and outperforms conventional methods that treat drugs and targets in isolation.

Building upon this framework, Zhu et al. [97] developed GraphSAGE and random forest for drug-target interactions (GSRF-DTI), which integrates GraphSAGE with a random forest classifier to further enhance both representation learning and model interpretability. By leveraging the DTP network, GSRF-DTI captures both direct and indirect dependencies, while its inductive learning paradigm facilitates generalization to previously unseen drug-target combinations. The incorporation of a random forest layer offers transparent feature importance analysis, addressing a key limitation of deep learning models in biomedical applications.

Higher-order modeling thus represents a paradigm shift from node-centric to interaction-centric inference, allowing for a more systemic understanding of drug–target interplay. While GCN-DTI emphasizes end-to-end feature extraction, GSRF-DTI introduces interpretability and inductive scalability. Nevertheless, both approaches face significant computational challenges when applied to large-scale DTP graphs, particularly in terms of memory efficiency and training time. Future directions may include the development of hierarchical pooling mechanisms or sparse graph optimization techniques to enable scalable high-order inference.

3.1.4. Cross-technical integration

Cross-technical integration refers to hybrid frameworks that couple GNNs with complementary deep learning architectures such as Transformers or attention-based modules to exploit cross-domain strengths via unified representation learning. These integrative approaches aim to bridge modality gaps by leveraging the structural sensitivity of GNNs and the global context modeling capacity of sequence-aware architectures.

For instance, DeepMGT-DTI, proposed by Zhang et al. [109], fuses GCNs with Transformer modules, wherein the GCN captures molecular substructures while the Transformer encodes long-range dependencies in protein sequences. This dual-pathway architecture enables joint optimization of structural and sequential features, substantially improving performance in drug repositioning tasks.

The mutual transformer-drug target affinity (MT-DTA) method proposed by Zhu et al. [132] significantly improved the accuracy of DTA prediction by cleverly integrating variational autoencoders, convolutional neural networks, attention mechanisms, and innovative intermolecular interaction modules. It not only achieves high-performance prediction at the technical level, but also enhances the credibility and interpretability of the model through theoretical proof and attention mechanism. At the same time, it shows the potential in generating new drug molecules and provides a powerful computing tool for accelerating the drug discovery process.

Cross-technical models demonstrate significant potential in unifying multimodal biological knowledge, particularly by integrating GNNs' local topology modeling with Transformers’ global attention mechanisms. Such combinations bridge traditionally siloed representations, e.g., sequence-based embeddings and structure-informed graphs, offering a more holistic understanding of drug-target landscapes.

Nevertheless, these frameworks often introduce increased model complexity, posing challenges in terms of convergence stability, computational overhead, and hyperparameter optimization. Efficient strategies, such as pre-training, transfer learning, or parameter-efficient fine-tuning, may alleviate these issues and improve generalizability without incurring prohibitive computational costs.

3.2. GNNs in drug-target affinity (DTA) prediction

DTA prediction is a pivotal task in computational drug discovery that aims to quantitatively assess the binding strength between drug molecules and target proteins, commonly expressed by dissociation constants or half-maximal inhibitory concentrations [133]. As depicted in Fig. 6, the typical DTA prediction workflow involves first converting a drug's simplified molecular input line entry system (SMILES) representation into a molecular graph, while constructing a corresponding protein graph through prediction of the protein contact map based on its amino acid sequence.

Fig. 6.

Schematic diagram of drug-protein target affinity prediction process based on graph neural network (GNN).

GNNs are then employed to extract hierarchical features from these molecular and protein graphs. The resulting graph-based embeddings are integrated with sequence-derived representations via attention mechanisms, enabling robust and context-aware affinity prediction.

Despite significant progress, several challenges persist. Functional groups in drug molecules (e.g., benzene rings, hydroxyl groups) and structural domains in proteins (e.g., α-helices, β-sheets) hold critical biological meaning at multiple scales, requiring adaptive hierarchical representation learning. Furthermore, the lack of experimentally resolved three-dimensional protein structures for most targets necessitates accurate prediction of contact maps or pseudo-structural features directly from primary sequences. Moreover, precise identification of key binding sites specific interactions between drug atoms and protein residues is essential for guiding rational drug optimization.

To overcome these challenges, advanced GNN architectures have been developed. For instance, the groundbreaking GraphDTA model systematically demonstrated that GNNs outperform traditional machine learning methods and prior deep learning models such as DeepDTA, establishing GNNs as the new standard for DTA prediction [113,134]. This milestone has propelled the widespread adoption of graph representation learning in drug discovery.

Current research in this area can be categorized into four major paradigms: (1) graph structure optimization and feature enhancement models, (2) multi-modal data fusion models, (3) attention mechanisms and interpretability enhancement models, and (4) multi-scale feature extraction and deep architectures.

3.2.1. Graph structure optimization and feature enhancement models

Graph structure optimization focuses on enhancing model performance by refining the topology or relational architecture of molecular and protein graphs [135]. Such methods improve the construction of molecular and protein maps through strategies including weighted edges, contact graph refinement, and diffusion-based graph optimization. Advanced architectures, such as depth-wise convolutional networks and GATs, further enrich feature representations by capturing non-covalent interactions and long-range dependencies critical for DTA prediction.

For example, Jiang et al. [103] proposed a weighted graph neural network (WGNN) framework that predicts DTA using only protein sequences and drug molecular formula strings. By directly constructing molecular and protein graphs from sequences, WGNN-DTA circumvents computationally expensive procedures like multiple sequence alignment (MSA). The model assigns weighted edges to atom-atom interactions within drugs and residue-residue contacts within proteins, enabling the GNN to prioritize structurally significant relationships. Benchmark evaluations demonstrate that WGNN-DTA achieves competitive accuracy while maintaining high computational efficiency, making it well-suited for large-scale virtual screening campaigns.

Building upon these structural insights, Jiang et al. [105] further developed DGraphDTA, which leverages protein contact maps predicted by tools such as AlphaFold to construct residue-level protein graphs from sequence data. This approach bypasses the dependence on experimentally resolved protein structures, thereby expanding applicability to novel targets. GNNs then learn hierarchical representations from the integrated drug molecular and protein contact graphs to enhance affinity prediction accuracy. By effectively bridging sequence and structural information, DGraphDTA addresses the common challenge of missing 3D data, aligning with current trends in exploring uncharacterized drug targets. Nonetheless, predicted contact maps may propagate inaccuracies inherent to prediction tools like AlphaFold, potentially affecting downstream performance. Incorporating hybrid approaches that combine evolutionary and co-evolutionary signals can mitigate such errors.

Extending the scope of graph optimization, Zhu et al. [127] introduced TDGraphDTA, which integrates a transformer-based multi-scale interaction module with a diffusion model-driven graph refinement strategy. The transformer component captures long-range dependencies between molecular substructures, while the diffusion module iteratively refines graph topology to enhance feature interpretability. Evaluation on benchmark datasets including Davis, Metz and KIBA demonstrates that TDGraphDTA achieves state-of-the-art regression performance (concordance index (CI), and coefficient of determination (R²)). Importantly, the diffusion-based refinement offers a novel mechanism to track critical substructures, facilitating hypothesis generation for lead compound optimization. However, despite circumventing MSA, the added complexity of transformer and diffusion modules increases computational demands. Developing lightweight architectures or employing model distillation techniques could offer a favorable trade-off between predictive accuracy and efficiency.

In order to solve the problem that the existing DTA prediction model ignores the complex interaction between the internal substructures of the molecule, resulting in insufficient information representation, Zhu et al. [136] proposed a new network model called rotary retention graph drug-target affinity (RRGDTA). The core contribution of this model is the introduction of the rotation encoding mechanism (ROE) to more accurately capture the local structural features of drugs and targets, and the design of a multi-scale interaction module (MSI) to enhance the correlation between different scale features. At the same time, the key information is retained to the maximum extent through its association prediction module (APM) and internal mask retention module (IMR) to prevent information loss. Finally, the model surpasses the existing baseline model on multiple standard datasets, significantly improves the prediction accuracy, and provides a more powerful computing tool for drug discovery.

Additionally, to solve the problem of ignoring the importance of molecular graph structure and simplifying the complex interaction mechanism in the current DTA prediction, Zhu et al. [137] proposed a new prediction model called drug-target binding affinity prediction model utilizing multi-scale (MDCT-DTA). The main contributions are as follows: (1) A multi-scale graph diffusion convolution (MGDC) module is designed to effectively capture the complex local and global interactions between nodes in the drug molecular graph and optimize the drug feature extraction; (2) The CNN-Transformer network (CTN) module is introduced, which combines the local feature extraction of CNN and the global dependence modeling ability of Transformer to enhance the characterization of protein sequences; (3) The local inter-layer information interaction structure is constructed, and the characteristics of drugs and targets are deeply fused, which improves the representativeness and robustness of the model. Through experimental verification on multiple public benchmark datasets such as Davis and KIBA, the model significantly improves the prediction accuracy, provides a new perspective for DTA prediction tasks, and promotes the development of this field.

Collectively, these graph-based frameworks mark a paradigm shift in DTA prediction, moving from static structural inputs toward dynamic, sequence-driven graph representations. Through elegant simplicity (WGNN-DTA), structural integration (DGraphDTA), and interpretability enhancement (TDGraphDTA), they furnish versatile tools to accelerate drug discovery pipelines [103,105,127]. Nonetheless, further validation is necessary across diverse target classes, particularly membrane proteins and allosteric sites, where structural dynamics critically influence binding interactions.

3.2.2. Multimodal data fusion models

Multimodal data fusion models integrate heterogeneous data modalities, such as chemical structures, protein sequences, molecular graphs, and gene expression profiles, to extract complementary features and improve the predictive performance of DTA models. Given the complex interplay between drug molecules and target proteins, multimodal fusion provides a powerful framework to capture diverse biological signals.

For example, Zeng et al. [107] proposed S2DTA, a heterogeneous deep learning framework that synergistically combines sequence features (drug molecular formulas, target/pocket sequences) with structural graph representations. By employing GNNs to encode molecular and protein structural graphs and semantic networks to embed sequence information, S2DTA effectively captures complementary features across modalities. Compared to DeepDTA and GraphDTA, this fusion approach reduces mean absolute error (MAE) by 25.2% and root mean square error (RMSE) by 20.1%, while exhibiting superior performance across regression metrics including Pearson correlation coefficient (PCC), CI and R². Notably, the model's interpretable architecture links key substructures to affinity outcomes, establishing a meaningful sequence-structure relationship that aids prediction on novel targets.

Similarly, Zhang et al. [33] introduced GPCNDTA, which integrates graph features, pharmacophore information, and cross-attention mechanisms to overcome challenges of side-information loss and effective multimodal fusion. Their framework incorporates residual GNNs, residual CensNet for drugs and edge-weighted graph convolutional networks (EW-GCN) for proteins, to enhance feature extraction. The cross-attention mechanism operates at two levels: intramolecular attention fuses sequence and graph features within the same biomolecule, while intermolecular attention aligns drug-protein interactions in a shared latent space. Across five benchmark datasets, GPCNDTA outperformed state-of-the-art models, with ablation studies confirming the critical contributions of pharmacophore integration and attention modules. Both S2DTA and GPCNDTA effectively unify sequence, structure, and domain knowledge (e.g., pharmacophores), simulating the multifaceted nature of DTI. Their cross-attention modules provide valuable biological insights that facilitate hypothesis generation in lead optimization.

However, models such as convolutional neural networks and graph neural networks of drug-target binding affinity (CGraphDTA) and S2DTA depend on predicted protein structures (e.g., from AlphaFold), which may propagate prediction errors. Integrating experimental structural data for mixed training can mitigate these inaccuracies. Wang et al. [120] proposed CGraphDTA, a dual-input framework that processes both target sequences and structures. Multi-scale convolutional neural networks capture hierarchical patterns from sequences, while GNNs encode drug and target structural graphs. Leveraging structural priors from AlphaFold enriches sequence-based representations, enabling robust affinity predictions even for proteins lacking resolved structures. Compared to GPCNDTA's emphasis on domain-specific knowledge (pockets, and pharmacophores), CGraphDTA highlights the complementary nature of structure-sequence duality.

Unlike graph optimization-focused models such as WGNN-DTA and TDGraphDTA [103,127], multimodal fusion approaches emphasize cross-modal synergy rather than solely refining intra-graph features. This paradigm shift transcends the limitations of single-modality methods by integrating sequence, structure, and expert knowledge, thereby achieving enhanced accuracy and biological interpretability. Future directions should prioritize lightweight architectures, error-resilient data integration strategies, and comprehensive validation across diverse target classes, including G protein-coupled receptors (GPCRs) and ion channels. Furthermore, incorporating dynamic binding information like molecular dynamics trajectories could narrow the gap between computational predictions and experimental observations, advancing the translational relevance of DTA models.

3.2.3. Attention mechanisms and interpretability-enhanced models

Attention mechanisms including self-attention, cross-attention and mutual information-based approaches have emerged as powerful tools to uncover critical drug-protein interaction patterns. By leveraging gradient attribution or attention weight visualization, these approaches can highlight key binding residues and functional groups, thereby aligning computational predictions with established biophysical principles and enhancing model interpretability.

Building on this concept, Liao et al. [106] introduced an interpretable drug-target binding affinity prediction model based on graph neural networks with self-attention mechanism and mutual information (GSAML-DTA), which integrates GNNs with a self-attention mechanism to jointly model drug and target structures while employing mutual information (MI) to filter redundant features. Specifically, the GNN captures atom- and residue-level interactions within drugs and proteins, while self-attention prioritizes crucial structural motifs. MI optimization further refines feature selection by ensuring that only relevant information contributes to affinity prediction. When evaluated on benchmark datasets, GSAML-DTA outperforms existing methods in accuracy and offers atomic- and residue-level interpretability by identifying interaction “hot spots” through attention weight analysis. This dual emphasis on predictive performance and interpretability renders GSAML-DTA a valuable tool for guiding experimental validation.

Meanwhile, Zhang et al. [114] proposed self-attention graph pooling drug target affinity (SAG-DTA), which applies a self-attention mechanism directly on molecular graphs to optimize drug representation. By weighting atomic node features prior to aggregation, SAG-DTA enables the model to focus on chemically significant regions such as functional groups. Moreover, the authors systematically compared global (whole-molecule) and hierarchical (substructure-based) pooling strategies, demonstrating that hierarchical pooling more effectively captures multi-scale features. As a result, SAG-DTA achieves robust performance in both regression tasks (e.g., KIBA, Davis datasets) and classification tasks, consistently surpassing sequence-based and earlier graph-based models. Importantly, attention-driven visualization highlights key atoms consistent with known pharmacophores, affirming the model's interpretability. Furthermore, hierarchical pooling enhances adaptability to variable molecular sizes, addressing a common limitation of graph-based approaches. Taken together, both GSAML-DTA and SAG-DTA link computational attention weights with medicinal chemistry insights; notably, MI-based filtering in GSAML-DTA reduces noise beyond what pure attention mechanisms can achieve.

Extending the application of attention mechanisms, Tang et al. [123] developed a drug-target affinity prediction based on graph sample and aggregate, bidirectional gated recurrent units (BiGRU) and attention neural network (GRA-DTA), which combines BiGRU with a soft attention mechanism to encode protein sequences and capture long-range dependencies and contextual motifs. For drug molecules, GraphSAGE aggregates molecular graph features with an emphasis on discriminative substructures. A cross-modal attention neural network dynamically aligns drug and target representations, adjusting their relative contributions. This hybrid architecture decouples sequence and graph feature contributions, enhancing model transparency. Although explicit metric improvements were not detailed, attention maps provide interpretable insights by highlighting key residues and molecular substructures driving predictions. GRA-DTA's integration of sequence-based BiGRU and graph-based encoders balances contextual and structural modeling effectively.

In contrast to multimodal fusion models such as S2DTA, which primarily emphasize integration across heterogeneous data sources [107], these attention-driven approaches focus on refining intramolecular features to improve predictive precision and interpretability. Collectively, their advancements have yielded breakthroughs in DTA prediction by not only improving accuracy but also delivering actionable biological insights. Future research directions include exploring multi-headed attention to capture diverse binding modes, incorporating dynamic structural data from molecular dynamics simulations, and experimentally validating interpretability claims in collaboration with wet-lab partners. Additionally, integrating attention- and interpretability-focused frameworks with pharmacophore-aware models such as GPCNDTA could establish a comprehensive approach that balances prediction accuracy, computational efficiency, and biological plausibility [33].

3.2.4. Multi-scale feature extraction and deep architectures

Hierarchical GNNs incorporating residual connections or graph pooling operations enable the extraction of features at multiple biological scales ranging from atomic and motif-level to whole molecular representations. Moreover, ultra-deep architectures, such as those exceeding 10 GNN layers combined with skip connections, mitigate over-smoothing issues while capturing complex interactions across these scales.

Building on this principle, Liu et al. [116] introduced substructure-aware multi-layer graph neural networks for drug-targetbinding affinity prediction (SSR-DTA), a substructure-aware multi-layer GNN that hierarchically extracts features through stacked graph convolutional layers, progressing from atomic motifs (e.g., functional groups) to macroscopic ring structures (e.g., ring systems). To overcome limitations inherent in sequence-only models, SSR-DTA integrates a bidirectional GNN (BiGNN) to concurrently process protein primary sequences and predicted tertiary structures (e.g., AlphaFold-derived graphs). This dual-stream design improves robustness to structural prediction noise by balancing sequence and structure information. Benchmark evaluations demonstrate that SSR-DTA outperforms sequence-based baselines, especially in scenarios where target structures are noisy or incomplete. The BiGNN architecture specifically mitigates the impact of structural inaccuracies.

Additionally, Yang et al. [101] proposed MGraphDTA, a deep multi-scale GNN grounded in chemical intuition for DTA prediction. MGraphDTA employs an ultra-deep GNN with 27 graph convolutional layers and dense skip connections, enabling the model to capture local atomic interactions (such as bond angles) and global molecular topologies (e.g., skeleton alignment). This depth allows iterative refinement of features across scales, effectively identifying subtle structural determinants of affinity. Additionally, the authors introduced gradient-weighted affinity activation mapping (Grad-AAM), a chemically interpretable visualization tool that highlights atomic contributions to the model's predictions. Grad-AAM bridges deep learning outputs with medicinal chemistry insights, representing a significant advance in translational applications. However, the high computational cost of training a 27-layer GNN underscores the need for efficiency improvements, such as neural architecture search (NAS) or dynamic graph sparsification. Grad-AAM annotates key pharmacophores through atomic-scale heat maps and supports the identification of residues in protein binding pockets. It is proposed to combine the mutual information filtering to optimize the interpretation results, emphasizing that the technology can be transformed into experimental instructions (such as fixed-point mutation), and promote the ‘prediction-verification-feedback’ cycle to iterate the model.

Taken together, unlike attention-driven models (e.g., GSAML-DTA), SSR-DTA and MGraphDTA emphasize hierarchical structural representation over dynamic feature weighting, offering complementary strengths in capturing conserved molecular motifs. The depth of MGraphDTA contrasts with shallower architectures such as WGNN-DTA, illustrating a trade-off between interpretability (via Grad-AAM) and computational scalability.

4. Challenges and prospects

GNNs, as an emerging deep learning paradigm, have gained widespread application across various domains, particularly in drug discovery and target prediction. Owing to their inherent advantage in handling non-Euclidean data such as molecular structures and protein interaction networks, GNNs have become indispensable tools in DTI prediction. Nevertheless, despite their promise, GNN-based methods face several significant limitations and challenges that hinder their practical deployment.

4.1. Limitations and challenges

Although GNNs demonstrate considerable potential in drug-target prediction, key obstacles remain in real-world applications. One fundamental issue is the scarcity and quality of DTI data available from public repositories. Existing datasets are often plagued by annotation biases and experimental noise, which compromise the training of robust GNN models. In low-data scenarios, models are particularly susceptible to overfitting spurious correlations, undermining generalizability.

Moreover, GNNs are computationally intensive due to their message-passing mechanisms across graph edges, resulting in high time and memory complexity. This challenge is especially pronounced for large-scale molecular graphs, such as proteins containing over 1000 residues. Standard architectures, including GCNs and GATs, struggle to scale effectively even with graphics processing unit (GPU) acceleration, limiting their feasibility in high-throughput virtual screening applications.

The known high-quality DTI data (positive samples) are far less than the unknown interaction (potential negative sample pool), resulting in a serious imbalance in the proportion of positive and negative samples [69]. This class imbalance problem tends to make the model tend to predict ‘no interaction’ to obtain higher overall accuracy, but at the same time, it will significantly reduce the recognition sensitivity to real interaction pairs, resulting in an increase in false negative rate. Traditional solutions usually randomly select samples from unknown interactions as negative examples, but such samples may contain positive examples that have not been experimentally confirmed, thus introducing label noise and training data contamination [138].

In order to alleviate this problem, existing research has proposed a variety of improvement strategies. For example, some methods regard unlabeled samples as potential negative samples, and gradually screen out negative samples with high confidence through iterative training process, which can reduce the deviation caused by random negative sampling in the absence of verified negative samples [139]. The other method fuses multi-source information of drugs and targets (such as side effects, and disease associations), and aggregates neighborhood topological features with the help of GNNs to identify highly reliable negative samples that are structurally ‘highly unlikely to interact’ [140].

Despite the above progress, current DTI prediction still faces a fundamental challenge. For example, the limited scale and quality of drug-target datasets restrict the model generalization and the computational demands of large-scale graph modeling hinder the efficiency of high-throughput screening. Moreover, model interpretability remains insufficient, particularly for analyzing complex higher-order mechanisms such as allosteric effects. To overcome these barriers, future work should emphasize integrating multimodal data sources (e.g., molecular dynamics and single-cell omics) to bolster robustness, developing lightweight architectures (e.g., sparse GNNs and distributed training) to enhance scalability, and combining interpretability tools with experimental validation to establish closed-loop “prediction–optimization-validation” pipelines that accelerate rational drug design.

Indeed, during the drug discovery research, when the protein target does not appear at all in the training set, the model cannot learn its structural characteristics and the interaction mode with the ligand. Traditional QSAR methods rely heavily on the training data of known targets, so it is difficult to predict new targets [141]. For example, the method based on direct QSAR modeling has obvious limitations, that is, it cannot be effectively extended to protein targets that do not appear in the training phase [141]. Similarly, if the candidate drug molecule or its key pharmacophore does not appear in the training set, it is difficult for the model to accurately infer its chemical properties and interaction mechanism with the target. In response to this problem, Liu et al. [142] divided the cold-start scenario system into four categories: known drugs and known target pairs (S1), new drugs and known target pairs (S2), known drugs and new target pairs (S3), and new drugs and new target pairs (S4). Among them, the S2 and S4 cases especially highlight the generalization challenge of the model in the face of new molecular entities [142].

In order to deal with the cold start problem, some solutions have been proposed. For instance, Nguyen et al. [143] conducted pre-training on large-scale protein-protein interaction and cell-cell interaction (CCI) tasks to enable the model to learn a universal ‘intermolecular interaction’ representation, thereby improving the generalization ability of drug and target characterization. Chu et al. [144] proposed a hierarchical graph representation learning framework, which simultaneously captures the local structural attributes and global topological similarity of drugs and targets through a message passing mechanism, and introduces a similarity-based embedding mapping method to infer the potential representation of unseen entities.

The essence of the cold start problem is the lack of real interactive data of new entities (drugs or targets). Future research can further integrate three-dimensional structure information (such as protein structure predicted by AlphaFold) and combine paradigms such as multi-task learning to improve the prediction performance and reliability of the model in more complex cold start scenarios.

4.2. Future directions

To overcome these challenges, future research should focus on several promising directions. Integrating multimodal data combining GNNs with heterogeneous biological data sources, such as various omics datasets, can enhance prediction robustness by leveraging complementary information. Self-supervised and contrastive learning techniques hold promise for alleviating data scarcity by exploiting large unlabeled molecular datasets.

Generative models capable of synthesizing novel drug-target pairs may further augment training corpora, improving model generalization. From a computational perspective, the development of sparse or hierarchical GNN variants can substantially reduce memory and time requirements, improving scalability. Hardware-aware optimizations, such as quantization-aware training for deployment on specialized accelerators like Google Cloud TPU, offer additional avenues for accelerating inference.

Feature extraction from protein data can be streamlined by employing pretrained residue-level embeddings instead of raw sequences, enhancing efficiency. Additionally, integrating interpretable AI tools such as GNNExplainer can elucidate key subgraph motifs driving model predictions, improving transparency and biological insight [91].

Lastly, hybrid modeling approaches that combine GNNs with MD simulations may link data-driven predictions with biophysical principles, such as free energy calculations. This integration could advance mechanistic understanding and improve the rationality of DTI predictions.

5. Conclusion

GNNs have emerged as a core technology in drug target prediction, owing to their unique ability to process non-Euclidean data, such as molecular graphs and biological networks. Due to the difficulties in accurately identifying and characterizing drug targets, the current drug discovery pipelines remain prohibitively costly, time-consuming and inefficient. Presently, we systematically analyze GNN architectures and their applications, providing a timely and comprehensive perspective that addresses this critical bottleneck and offering practical guidance for accelerating artificial intelligence (AI)-driven drug discovery.

The importance of this study lies in filling key gaps left by prior works. Instead of treating GNNs as a monolithic tool, we establish a principled classification paradigm that dissects the design principles, mechanisms and applicable conditions of representative models such as GCNs, GATs and GAEs. Beyond summarization, we highlight how multimodal fusion, high-order graph reasoning and dynamic learning strategies can overcome real-world challenges such as data sparsity, limited interpretability and the complexity of polypharmacology.

Overall, this work consolidates the fast-evolving field of GNN-based drug target discovery, establishes a unified methodological foundation, and advocates a paradigm shift from purely data-driven to mechanism-informed AI. By highlighting both theoretical innovations and practical applications, it not only clarifies the current landscape but also charts a forward-looking roadmap. We believe this work will serve as a comprehensive reference for computational biologists, medicinal chemists and drug developers, and will ultimately contribute to advancing precision medicine and rational drug design.

CRediT authorship contribution statement

Jing Chen: Writing – original draft, Visualization. Nini Fan: Visualization. Yuqing Lu: Visualization. Jianhua Yang: Visualization. Wenchao Song: Visualization. Haiyang Sheng: Investigation, Formal analysis. Yinfeng Yang: Supervision, Conceptualization. Shengxi Chen: Supervision, Conceptualization. Jinghui Wang: Visualization, Supervision, Project administration, Funding acquisition, Conceptualization.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

The following is the Supplementary data to this article: