DVNDTA: a dual virtual node based heterogeneous interaction model for accurate prediction of drug-target affinity Authors: Guoqiang Zhou, Ben Li, Weifeng Zhang, Shudi Bao, Huaming Chen, Haoran Li, Huan Wang, Jiacong Mi, Jun Shen Published: 2026 ABSTRACT AbstractEfficient prediction of drug-target affinity (DTA) is crucial for accelerating drug discovery. Recently, deep learning approaches leveraging 3D complexes have shown strong potential for drug-target affinity prediction. However, most of these approaches focus solely on local structural information at the binding site to learn non-covalent interactions, which may limit their ability to capture complex binding mechanisms. This paper introduces a dual virtual node-based heterogeneous interaction model for DTA (DVNDTA), which incorporates global contextual information of both the drug and the target pocket into the affinity prediction by utilizing two virtual nodes. This dynamically weighted global–local cooperative perception mechanism enhances DVNDTA’s ability to capture drug-target binding patterns. Moreover, the states of the virtual nodes are dynamically updated as the network layers deepen, enabling a bidirectional feedback mechanism between global and local information. This iterative refinement of feature representations serves as an abstract computational proxy for the mutual adjustments in the drug-target binding process. Experimental results across multiple benchmark datasets demonstrate that DVNDTA outperforms the state-of-the-art affinity prediction models, exhibiting good performance and generalization ability.Keywords: Binding affinity, Non-covalent message passing, Global information, Dual virtual nodes Introduction Drug development is a costly and time-consuming process, usually requiring more than 10 years and billions of dollars, with a low success rate [1]. A critical step in the early stages of drug development is to find drugs that interact strongly with specific protein targets, and drug-target affinity (DTA) is an essential indicator for measuring such interaction [2]. However, traditional experimental techniques, such as surface plasmon resonance and isothermal titration calorimetry [3, 4] necessitate testing thousands of complexes and consume a lot of resources. Computer simulation technology provides a more efficient and cost-effective approach to DTA calculation, significantly reducing the drug development timeline and benefiting the overall pharmaceutical industry [5]. In recent years, the accumulation of high-quality DTA data has facilitated the advancement of data-driven approaches. Both traditional machine learning [6, 7] and more recent deep learning methods have demonstrated significant advantages in DTA prediction, with deep learning being particularly effective indebted to its powerful automatic feature extraction capabilities [5]. Based on the input formats of targets and drugs, deep learning models can be broadly categorized into two types: complex-free models and complex-based models.In this study we focus on structure-based DTA models that condition on receptor or binding-site structural information. Ligand-based approaches that rely solely on ligand descriptors are not considered here and are discussed as a complementary line of work. Complex-free models typically employ convolutional neural networks (CNNs) or graph neural networks (GNNs) to extract features of both targets and drugs from 1D or 2D inputs [8], which are then used for affinity prediction. While this approach is straightforward, it fails to fully capture the non-covalent interactions between targets and drugs, either the structural information of the complex, both of which are crucial for determining DTA [9, 10]. Among these factors, non-covalent interactions play a fundamental role in biomolecular recognition [11], influencing not only binding strength but also binding specificity [12, 13]. Structural information is equally essential, as it directly reflects the spatial arrangement and interaction patterns between molecules. Consequently, disregarding this information might cause the model to memorize targets and drugs features rather than develop a deep understanding of their actual interactions [9], ultimately limiting its generalization ability. In contrast, complex-based models leverage spatially encoded GNN [9, 10] to directly extract features from 3D complexes, enabling them to better capture non-covalent interactions and complex structural information, thereby enhancing the model’s learning capacity and prediction accuracy. However, while complex-based models have addressed some limitations of complex-free models, they still fail to incorporate the global information of target pockets and drugs when learning non-covalent interactions. This omission might result in two key challenges that constrain the performance of such models: The model has a limited understanding of non-covalent interactions. Existing spatially encoded GNN-based methods, which learn non-covalent interactions between targets and drugs, are constrained by the message-passing mechanism. Each node can only aggregate features from its neighboring nodes [14], preventing the full integration of global target pocket and drug features into non-covalent interaction learning. As illustrated in Fig. 1a, each drug node receives information only from nearby pocket nodes at the binding site, failing to capture the overall physicochemical properties of the entire target pocket. Similarly, each pocket node lacks a comprehensive understanding of the drug’s global characteristics. This limitation may result in an incomplete representation of non-covalent interactions [15]. The model has a limited understanding of non-covalent interactions. Existing spatially encoded GNN-based methods, which learn non-covalent interactions between targets and drugs, are constrained by the message-passing mechanism. Each node can only aggregate features from its neighboring nodes [14], preventing the full integration of global target pocket and drug features into non-covalent interaction learning. As illustrated in Fig. 1a, each drug node receives information only from nearby pocket nodes at the binding site, failing to capture the overall physicochemical properties of the entire target pocket. Similarly, each pocket node lacks a comprehensive understanding of the drug’s global characteristics. This limitation may result in an incomplete representation of non-covalent interactions [15]. A single information source reduces robustness. Constructing non-covalent edges requires an artificially defined distance threshold (typically 5–8 Å) to infer implicit non-covalent interactions [16]. However, this approach inevitably introduces false positives and false negatives [17]. As illustrated in Fig. 1a, relying solely on local information can lead to errors in message transmission when deviations occur in the local structure. In such cases, a node updates itself with incorrect information without access to a supplementary reliable source, ultimately impairing affinity prediction. A single information source reduces robustness. Constructing non-covalent edges requires an artificially defined distance threshold (typically 5–8 Å) to infer implicit non-covalent interactions [16]. However, this approach inevitably introduces false positives and false negatives [17]. As illustrated in Fig. 1a, relying solely on local information can lead to errors in message transmission when deviations occur in the local structure. In such cases, a node updates itself with incorrect information without access to a supplementary reliable source, ultimately impairing affinity prediction. To fully capture the global information of target pockets and drugs and address the two key challenges mentioned above, we proposed a deep learning model, DVNDTA. In DVNDTA, two virtual nodes are introduced, with one for the pocket and the other for the drug, using a global pooling operation. These virtual nodes maintain the same feature dimensions as ordinary nodes and serve as fingerprints of target pockets and drugs [18], encoding their overall physicochemical properties and potential interaction patterns [19, 20]. As a result, they function as effective carriers of global information [21]. The motivation and solution of the proposed method. a drug node D can only receive part of the information from the target pocket. Node B is considered not to interact with node D because the distance between them is greater than the threshold, although they actually do interact. Conversely, nodes C and D meet the distance requirements but do not actually interact. b The virtual node carries the global information of the target pocket, and drug node D can simultaneously receive both global and local messages from the target pocket according to specific weights To enhance the model’s comprehension and robustness in capturing non-covalent interactions, we developed a dynamically weighted global–local cooperative perception mechanism. During the non-covalent message transmission stage, as illustrated in Fig. 1b, each drug (pocket) node updates its features by extracting atomic-level geometric and chemical information from the local neighborhood while also incorporating additional global information from the corresponding pocket (drug) virtual node. In this process, the model autonomously adjusts the fusion ratio of local and global information using learnable gating weights. Specifically. When the model determines that the local structure confidence is high, it prioritizes fine-grained local information for updates. Conversely, if local deviations are significant, it increases the weight of global information provided by the virtual node, forming a self-correcting mechanism. This dynamic balance strategy enables the model to integrate global and local information simultaneously, thereby enhancing its comprehension of non-covalent interactions and improving overall robustness. Additionally, since drug-target binding is a dynamic process and non-covalent interactions can induce conformational changes in targets [22], the virtual node states in DVNDTA are dynamically updated when the number of network layers increases, thus better simulating this complex binding process. This design enables a more accurate [22], the virtual node states in DVNDTA are dynamically updated as network layers deepen. This mechanism iteratively refines global and local feature representations, providing an abstract computational framework to capture the reciprocal adaptations in the binding process. The contributions of this work are summarized as follows: We thoroughly analyzed the role of global information in learning non-covalent interactions and propose a deep learning model, DVNDTA, capable of effectively capturing and incorporating such global information. We thoroughly analyzed the role of global information in learning non-covalent interactions and propose a deep learning model, DVNDTA, capable of effectively capturing and incorporating such global information. We developed a dynamically weighted global–local cooperative perception mechanism that innovatively integrates local structural information with global features through a gated self-regulatory process, significantly enhancing DVNDTA’s understanding and robustness in learning non-covalent interactions. We developed a dynamically weighted global–local cooperative perception mechanism that innovatively integrates local structural information with global features through a gated self-regulatory process, significantly enhancing DVNDTA’s understanding and robustness in learning non-covalent interactions. We designed a dynamic update mechanism that continuously incorporates the latest global and local information as the network layers deepen, thereby better simulating the conformational changes of targets during the binding process. We designed a dynamic update mechanism that continuously incorporates the latest global and local information as the network layers deepen, thereby better simulating the conformational changes of targets during the binding process. The experimental results of DVNDTA on three major benchmark datasets demonstrated its superior predictive performance and generalization ability, surpassing current state-of-the-art DTA prediction models. The experimental results of DVNDTA on three major benchmark datasets demonstrated its superior predictive performance and generalization ability, surpassing current state-of-the-art DTA prediction models. Related work In the research field of DTA prediction, existing deep learning methods can be broadly categorized into two types: complex-free models and complex-based models. We will now briefly review these methods, analyzing their respective advantages and disadvantages. Complex-free models The complex-free model typically uses the SMILES string or 2D molecular graph of the drug as input and combines it with the target’s amino acid sequence or its residue contact map for feature extraction. For instance, DeepDTA [23] employs a dual-channel CNN to process the drug SMILES and protein sequence separately, enabling end-to-end affinity prediction. Building on this, AttentionDTA [24] introduces an attention mechanism to enhance model interpretability. To preserve the drug’s 2D structure, GraphDTA [25] uses a GNN to process the molecular graph, although its shallow architecture struggles to capture complex ring structures, while deep GNNs also often suffer from over-smoothing issues [26, 27]. To address this, MGraphDTA [28] proposes a dense connection strategy and designs a deep GNN with 27 graph convolutional layers to simultaneously capture both local and global drug structures. On the other hand, DGraphDTA [29] replaces the amino acid sequence input with a protein residue contact map to mitigate the loss of target structure information. However, these models cannot learn the 3D structure of the complex and its non-covalent interactions, limiting their effectiveness in predicting affinity. Complex-based models Unlike models that do not incorporate complexes, models using 3D complexes as input are better equipped to capture the spatial structure and non-covalent interactions of these complexes. Earlier methods placed complexes within a 3D grid and used 3D-CNN to extract voxelized feature vectors [30, 31]. However, this approach lacks rotational invariance, and the large number of voxel grids often only contain limited useful information, resulting in low computational efficiency [18]. Fortunately, some spatial encoding-based GNN methods have addressed these issues [32]. For example, Jiang and Morrone’s teams [16, 33] modeled target and drug atoms in the complex as graph nodes, covalent bonds as edges, and determined implicit non-covalent edges through manually defined a cutoff radius. This approach successfully converted the 3D complex into a format suitable for GNN processing, preserving the spatial structure information and enhancing computational efficiency. To further improve model understanding, Li et al. [10] proposed SIGN, which added geometric information, such as interatomic distances and angles, to the node features. However, to simplify modeling, these methods fail to distinguish between covalent and non-covalent interactions, treating both types of bonds as the same without recognizing their specificity. Jiang et al. [34] introduced IGN, which models the complex in greater detail by decomposing it into a target graph, a drug graph, and a drug-target interaction graph. Through two stages of message passing, this method separately addresses covalent and non-covalent interactions, enhancing the understanding of both types. However, given that the induced fit theory suggests that drug-target binding is a dynamic process, methods like IGN still struggle to fully capture this complex change. To address this, Yang, Zhou et al. separately proposed GIGN [18] and GEMF [35], which designed multiple heterogeneous interaction layers, each representing a different state of the complex during the binding process. These layers simulate the complex’s changes at various binding stages, improving the model’s understanding and prediction capabilities regarding the dynamic binding process. Unfortunately, by far these methods fail to fully incorporate global information when learning non-covalent interactions. Although some approaches, such as EGNA [36] and DSBIN [37], attempt to extend the cutoff distance to capture a broader range of information, this merely expands the local structure rather than truly modeling global information. Additionally, such an extension of the cutoff distance leads to significant increases in computational and storage overhead. As shown in Table 1, for the complex with PDB ID 4wvu, when the cutoff radius increases from 4 to 10Å, the number of drug-target atom pairs that need to be stored and processed will grow substantially, potentially introducing more false positives and exacerbating model bias. In contrast, our work focuses on how to effectively obtain and utilize global information from target pockets and drugs, enhancing the model’s understanding of non-covalent interactions. This approach addresses the limitations of existing methods and further improves model performance and generalization. Table 1Number of drug-target atom pairs requiring storage and calculation as the distance threshold increases by 1 Å incrementsThreshold4 Å5 Å6 Å7 Å8 Å9 Å10 ÅAtom pairs913257191216176423683038 Number of drug-target atom pairs requiring storage and calculation as the distance threshold increases by 1 Å increments Methodology This section presents a comprehensive overview of the datasets used for training and testing DVNDTA, along with data preprocessing, graph construction, representation, and our model’s backbone architecture. Datasets PDBbind [38], as the authoritative benchmark dataset in the field of DTA prediction, is widely recognized for its stringent data filtering criteria. This dataset not only provides experimentally determined DTA metrics, such as dissociation constant Kd, inhibition constant Ki, or half-maximal inhibitory concentration IC50 [39], but also includes the corresponding 3D complex structural information. To meet the needs of machine learning models, these raw DTA values are often converted into logarithmic form, such as -logKi, -logKd, or -logIC50, to improve numerical stability and enhance model performance. In this study, we utilized the widely adopted PDBbind v.2016 dataset as the training foundation, which includes both the general and refined sets. The general set comprises 13,285 drug-target complexes, while the refined set is a subset containing 4,057 complexes. To ensure data quality, we preprocessed all 13,285 complexes using RDKit [40] and ultimately selected 12,904 high-quality samples. Among these, 1,000 samples were randomly chosen as the validation set, while the remaining 11,904 complexes constituted the training set. To comprehensively evaluate the model’s performance, we employed four independent test sets: CASF-2013 (197 complexes) [41], CASF-2016 (285 complexes) [42], Community Structure-Activity Resource (CSAR-HiQ) [43], and a carefully curated 2019 holdout set from PDBbind v.2019 (4366 complexes). The 2019 holdout set not only entirely excludes any complexes present in the training, validation, and previous test sets but is also significantly larger than the traditional CASF-2013 and CASF-2016 test sets. This larger scale allows for a more robust evaluation of our model’s generalization ability on novel drug-target complexes.CSAR is a dataset specifically designed for protein-ligand docking and scoring method evaluation. Its high-quality subset, CSAR-HiQ, contains 343 complexes divided into two sub-libraries. We first merged these two sub-libraries into a single set and then removed samples that overlapped with the PDBbind v.2016 training set. This yielded 49 completely independent complexes for model testing, serving as a non-PDBbind external test set to further measure the model’s generalization performance.The PDB IDs of these 49 complexes are listed in the Supporting Information and are available for reference.However, we found that this dataset is not suitable for strict cold-start evaluation. When we attempted to filter out homologous targets, only a few complexes remained, which is insufficient for statistical analysis. Therefore, to rigorously evaluate generalization, we constructed a larger Strict Cold-Start Holdout set from PDBbind v.2019 as described below. In addition to these standard benchmarks, we constructed a Strict Cold-Start Holdout set to rigorously evaluate the model’s generalization ability on unseen protein families. While the standard PDBbind v.2019 holdout set is temporally separated from the training data, it still contains targets with high sequence homology. To address this, we performed a pairwise sequence alignment between the v.2019 holdout set and the v.2016 training set using the Smith-Waterman algorithm. We removed any complex sharing >30% sequence identity with the training data. This process yielded a challenging subset of 699 complexes. The PDB IDs of these 699 complexes are provided in the Supplementary Materials to ensure reproducibility. Data preprocessing In this study, we systematically preprocessed the raw data to construct a DTA prediction dataset suitable for deep learning models. First, we used PyMOL [44] to load the target and drug structure files, remove all water molecules and hydrogen atoms, and extract amino acid residues within a 5 Å radius of the drug to define the binding pocket. These residues were then saved as a new PDB file. Next, we used Open Babel [45] to convert drug structures from non-PDB formats to PDB format (if necessary), ensuring consistency with the extracted binding pocket. Finally, we used RDKit to parse the PDB files of both the drugs and binding pockets, load them as molecular objects, and store them as tuples in pickle files for complex data. Graph construction Given that molecules are non-Euclidean data composed of atoms and chemical bonds, atoms can be naturally defined as nodes in a graph [46]. We use RDKit to extract the physicochemical properties of each atom in the target pocket and drug molecules to construct a feature vector that reflects the chemical characteristics of the atoms. We represent each atom as a 35-dimensional feature vector using one-hot encoding. This vector integrates six types of information: atomic symbol, atom degree, implied valence, hybridization, aromaticity, and the number of neighboring hydrogen atoms. Specifically, we encode common chemical elements (including C, N, O, S, F, P, Cl, Br, I) and an “others” category for unknown elements as part of the features. Meanwhile, atom degree and implied valence are encoded in a range from 0 to 6, while hybridization status covers five different types, from SP to SP3D2. Additionally, the feature vector includes a Boolean value to indicate whether the atom is part of an aromatic ring. Further, when hydrogen is explicitly considered, the number of hydrogen atoms directly connected to the atom, ranging from 0 to 4. To define edges in the graphs, we draw on previous research [18] and construct the complex as a graph with two types of edges: covalent and non-covalent interactions. The nodes in the graph are shared by both edge types. Covalent interaction edges represent covalent bonds between atoms in the target pocket and drug molecules. The adjacency matrix for covalent interactions is defined as follows: where and represent the set of nodes belonging to the target pocket and drug, respectively. For non-covalent interaction edges, we infer potential interactions using a fixed distance threshold. Based on previous studies, we set the threshold to 5 Å, as this distance encompasses most non-covalent interactions without increasing computational cost. Specifically, if the Euclidean distance between an atom in the target pocket and an atom in the drug molecule is less than 5 Å, we assume the presence of a non-covalent interaction and establish a corresponding edge. The resulting non-covalent interaction graph forms a bipartite structure. The adjacency matrix for non-covalent interactions is defined as follows: Overall pipeline of DVNDTA Our model’s architecture is illustrated in Fig. 2a. First, the constructed complex graph is processed through the representation layer to extract essential features for subsequent computations. This data is then fed into a series of Dual Virtual Node Heterogeneous Interaction Layers (DVNHIL) for feature extraction. Within these layers, the model learns covalent and non-covalent interaction features between the target and the drug. After passing through multiple DVNHIL layers, the integrated information from the entire complex graph is processed in the prediction layer, generating a feature vector that represents the graph and facilitates affinity prediction. The pipeline of our DVNDTA. a Overall pipeline of the Model. b Internal details of the virtual node fusion module in DVNHIL. c Internal details of the virtual node update module in DVNHIL Graph representation At this stage, additional processing is applied to the initially constructed complex graph. We first use a Multilayer Perceptron(MLP) to transform the original node features into higher-dimensional feature representations ,where n denotes the number of nodes in the graph. To further enhance the model’s ability to capture inter-atomic spatial relationships [10], we also calculate the distances between each atom and its neighboring nodes in three-dimensional space. Here, represents the set of distances between node i and its neighboring nodes j. To precisely regulate the fusion ratio between each node and the virtual nodes during the message-passing process, the non-covalent degree of each node should be determined. As shown in Eq. (3), this degree is computed from the non-covalent interaction adjacency matrix. A logarithmic transformation is then applied to these values to obtain a smoother data distribution.The logarithmic transformation is applied to compress the dynamic range of non-covalent degrees and stabilize gradient propagation during training, especially when some atoms participate in numerous interactions. where represents the non-covalent degree information of node i. The virtual node is a key component in this work. It is a special type of node in GNNs designed to aggregate global information within the graph and facilitate information exchange among all nodes [47, 48]. This is particularly beneficial when the graph structure has incomplete information, as it significantly enhances the overall effectiveness and completeness of message passing. In our approach, we introduce virtual nodes and for the target and drug, respectively. For their initialization, we first perform a global average pooling operation on the target pocket and drug within the complex to obtain their overall feature representations [49]. Then, an MLP is used to transform these overall feature representations into the initial embeddings of the virtual nodes. Dual virtual node heterogeneous interaction layer In DVNHIL, we designed a novel approach to effectively capture the intricate interactions between targets and drugs. Briefly, DVNHIL consists of three main modules: a virtual node fusion module, a covalent and non-covalent interaction message passing module and a virtual node update module. In the following sections, we will thoroughly provide a comprehensive and detailed explanation of the specific implementation of these three important modules. Virtual node fusion module In this module, we developed a dynamically weighted global–local cooperative perception mechanism. Specifically, each drug (or target) node dynamically fuses its local neighborhood information with the corresponding target pocket (or drug) virtual node in varying proportions. This mechanism enables each node to capture both the overall characteristics of the drug and target pockets while learning local structural information during non-covalent message transmission. By utilizing both global and local information sources to update node features, our method overcomes the limitations of previous approaches that relied solely on local information. Before entering the virtual node fusion module, the non-covalent degree information is merged with the original node features to form a new feature representation. This allows the model to identify atoms that may participate in more non-covalent interactions. The process is expressed as follows: In Eq. (4), represents stacking two feature vectors along the last dimension. Equation 5 denotes summing along the last dimension, where is a learnable parameter that controls the fusion ratio between each node and its covalent degree information. represents the node features integrated with non-covalent degree information. By incorporating non-covalent degree information, we can clearly identify which atoms participate in more non-covalent interaction exchanges. Subsequently, as shown in Fig. 2b, we input the node features integrated with non-covalent degree information into an MLP to obtain a weight α , which controls the fusion ratio between each ordinary node and the virtual node. Specifically, the fusion is performed as follows: where and represent the target pocket virtual node and drug virtual node, which node i needs to fuse with during the t-th round of dual virtual node heterogeneous interaction. represents the node features after fusion with the virtual node. Covalent and non-covalent interaction message passing module For different types of interactions, we adopt distinct message-passing strategies. For covalent interactions, message passing is performed directly using the previous layer’s node features . For non-covalent interactions, message passing is conducted using node features processed by the virtual node fusion module. Notably, for the definition of edge features, we employ a unified approach regardless of the interaction type. As shown in Fig. 2a, the edge feature between two nodes is obtained by applying a radial basis function (RBF) transformation [50] to their pairwise distance. The complete message-passing process is formulated as follows: where N(i) represents the neighbor set of node i, and represent covalent and non-covalent messages, respectively. Equation 11 shows that the results obtained from the two types of message passing are input into two separate MLPs for transformation and then summed to obtain the final node feature representation. It is also worth noting that we use a residual connection to add the node features from the previous layer to further enhance the model’s expressiveness. Virtual node update module After each round of node feature updates, we update the virtual nodes for the target pocket and drug to ensure that these virtual nodes always contain the most up-to-date global information. Specifically, as shown in Fig. 2c, we first perform a global average pooling operation on all target and drug nodes separately to obtain a feature vector, capturing the overall characteristics of the target pocket or drug in the current state. Then, we add this feature vector to the corresponding virtual node features from the previous round, allowing the virtual nodes to incorporate the overall state from the previous round. Next, we use an MLP to process the combined feature vector and generate a new virtual node feature representation. To flexibly adjust the update intensity, we introduce learnable scaling parameters (for the target pocket virtual node) and (for the drug virtual node), which scale the output virtual node features from the MLP, thus controlling the balance between new and old information. Finally, through residual connections, the virtual node features from the previous round are added, a practice that not only helps stabilize the training process but also retains historical information from the previous round. The specific update rules for the target and drug virtual nodes are as follows: where, and represent the pooled features of the target pocket and drug, respectively, after the t-th round of DVNHIL interaction updates. Prediction layer As shown in Fig. 2a, after processing through a series of DVNHIL layers, we obtain node features that effectively represent the interactions between the target and the drug. These features are subsequently used in the DTA prediction phase, which involves two key steps: first, applying a global sum pooling operation on the complex graph to generate a comprehensive feature representation of the entire graph; and second, feeding this feature into an MLP to predict the DTA. Result and discussion Evaluation metrics In this study, we use root mean square error (RMSE), mean absolute error (MAE), Pearson correlation coefficient (R), and standard deviation (SD) to comprehensively evaluate the performance of the model. Detailed definitions of evaluation metrics (RMSE, MAE, R, SD) are provided in the Supporting Information (Section S3). Comparison with baselines To evaluate the performance of DVNDTA, we compare it with the following advanced baseline models: DeepDTA [23] extracts features separately from the protein sequence and the drug SMILES string, concatenates them, and then performs DTA prediction. DeepDTA [23] extracts features separately from the protein sequence and the drug SMILES string, concatenates them, and then performs DTA prediction. MGraphDTA [28] uses a deeper GNN to extract features from the drug molecular graph. MGraphDTA [28] uses a deeper GNN to extract features from the drug molecular graph. Pafnucy [31] uses 3D-CNN to extract interaction features from the 3D structure of the complex. Pafnucy [31] uses 3D-CNN to extract interaction features from the 3D structure of the complex. GNN-DTI [51] uses GNN methods to process 3D complexes and designs a distance-aware attention mechanism to calculate interactions between targets and drugs. GNN-DTI [51] uses GNN methods to process 3D complexes and designs a distance-aware attention mechanism to calculate interactions between targets and drugs. IGN [34] learns the covalent and non-covalent interactions of complexes sequentially. IGN [34] learns the covalent and non-covalent interactions of complexes sequentially. GEMF [35], our previous work, simultaneously captures covalent and non-covalent interactions across heterogeneous interaction layers while integrating geometric features, such as distances and bond angles, to enhance the model’s expressive power. GEMF [35], our previous work, simultaneously captures covalent and non-covalent interactions across heterogeneous interaction layers while integrating geometric features, such as distances and bond angles, to enhance the model’s expressive power. MFE [52], A new multimodal feature extraction framework that integrates information from protein surface, 3D structure, and sequence for the first time and uses a cross-attention mechanism to compare features between different modalities. MFE [52], A new multimodal feature extraction framework that integrates information from protein surface, 3D structure, and sequence for the first time and uses a cross-attention mechanism to compare features between different modalities. All the above models are run on the same training, validation, and test datasets, and all experiments are repeated three times, each with a different random seed. Table 2 presents a performance comparison of our proposed DVNDTA model with multiple benchmark methods on the CASF-2013 and CASF-2016 test sets. The best results for each metric are bold. The results indicate that DVNDTA performs exceptionally well on both datasets. Specifically, on CASF-2013, DVNDTA outperforms the best benchmark model by 3.48%, 1.3%, and 1.4% in RMSE, R, and SD, respectively. On CASF-2016, DVNDTA achieves the best results across all four evaluation metrics, surpassing the best benchmark model by 4.2%, 2.8%, 1.3%, and 3.1%, respectively. Notably, in the field of DTA prediction, an R-value above 0.8 is considered an excellent performance, while models with R-values between 0.6 and 0.8 are also deemed acceptable. Observing the data in Table 2, we find that all models achieve R-values above 0.7, with many approaching or exceeding 0.8. However, this strong performance may be partly attributed to the high overlap between targets in the CASF-2013 and CASF-2016 datasets and those in the training set, which could lead to an overly optimistic evaluation of model performance [18]. The consistent improvements of DVNDTA over the best baseline models were further confirmed to be statistically significant (p < 0.05) by two-tailed paired t-tests. Detailed statistical analysis results, including t-values and p-values, are provided in Section S4 of the Supplementary Materials. Table 2Comparison results of DVNDTA with other baseline models on three benchmark datasetsTest setModelRMSE↓MAE↓R↑SD↓CASF-2013DeepDTA1.572 ± 0.0251.243 ± 0.0230.741 ± 0.0041.357 ± 0.015MGraphDTA1.619 ± 0.0151.263 ± 0.0280.724 ± 0.0071.612 ± 0.017Pafnucy1.518 ± 0.0131.186 ± 0.0320.768 ± 0.0141.493 ± 0.019GNN-DTI1.467 ± 0.0041.159 ± 0.0310.795 ± 0.0081.420 ± 0.022IGN1.510 ± 0.0161.164 ± 0.0280.771 ± 0.0051.487 ± 0.017GEMF1.380 ± 0.0261.056 ± 0.0250.816 ± 0.0081.349 ± 0.027MFE1.385 ± 0.0181.062 ± 0.0150.805 ± 0.0061.355 ± 0.016DVNDTA1.332 ± 0.0141.071 ± 0.0180.827 ± 0.0031.330 ± 0.010CASF-2016DeepDTA1.361 ± 0.0131.065 ± 0.0190.780 ± 0.0061.358 ± 0.017MGraphDTA1.440 ± 0.0161.108 ± 0.0200.749 ± 0.0041.438 ± 0.012Pafnucy1.455 ± 0.0331.137 ± 0.0230.761 ± 0.0041.442 ± 0.015GNN-DTI1.381 ± 0.0201.105 ± 0.0250.781 ± 0.0071.352 ± 0.012IGN1.308 ± 0.0121.004 ± 0.0130.810 ± 0.0041.296 ± 0.009GEMF1.195 ± 0.0160.917 ± 0.0110.840 ± 0.0041.179 ± 0.014MFE1.220 ± 0.0170.935 ± 0.0140.825 ± 0.0051.205 ± 0.015DVNDTA1.145 ± 0.0130.891 ± 0.0120.851 ± 0.0051.142 ± 0.012holdout-2019DeepDTA1.506 ± 0.0231.191 ± 0.0200.572 ± 0.0161.465 ± 0.019MGraphDTA1.620 ± 0.0131.267 ± 0.0150.503 ± 0.0141.547 ± 0.013Pafnucy1.433 ± 0.0191.130 ± 0.0230.620 ± 0.0041.395 ± 0.015GNN-DTI1.440 ± 0.0181.132 ± 0.090.618 ± 0.0101.398 ± 0.011IGN1.455 ± 0.0221.150 ± 0.0110.613 ± 0.0071.405 ± 0.010GEMF1.420 ± 0.0051.111 ± 0.0050.628 ± 0.0031.388 ± 0.004MFE1.400 ± 0.0081.085 ± 0.0100.640 ± 0.0051.390 ± 0.007DVNDTA1.369 ± 0.0071.064 ± 0.0110.658 ± 0.0061.368 ± 0.009 Comparison results of DVNDTA with other baseline models on three benchmark datasets In contrast, the holdout-2019 dataset has a target overlap rate of less than 70% with the training set, making it a more reliable measure of the model’s actual capabilities. As shown in Table 2, while DVNDTA’s performance slightly declines on the holdout-2019 test set, its advantage over other benchmark models becomes even more pronounced, with improvements of 3.6%, 4.2%, 4.7%, and 3.0% in all evaluation metrics compared to the best benchmark model. This demonstrates that DVNDTA possesses stronger generalization ability and robustness. Furthermore, Fig. 3 provides a more intuitive view of DVNDTA’s excellent regression performance across different datasets. Regression lines of predicted versus actual values for DVNDTA on the validation set, CASF-2013, CASF-2016, and Holdout-2019 Further analysis of the experimental results shows that, with the same number of training samples, the overall performance of the 2D molecular graph-based method MGraphDTA is inferior to that of the 3D complex-based approach. This demonstrates that incorporating 3D structural information can provide a higher-quality inductive bias for the model. Surprisingly, despite GEMF enhancing the model’s understanding of molecular structures by introducing additional geometric features, it still fails to surpass DVNDTA. On the one hand, this might be due to insufficient training caused by the limited number of training samples. On the other hand, it suggests that when designing deep learning models for drug discovery and drug-target binding affinity prediction, beyond simply increasing the dimensionality and complexity of input features, effectively modeling molecular interactions is a more crucial factor. Ablation study To evaluate whether the individual components of our proposed DVNDTA could positively contribute to DTA prediction, we designed five variants based on our model and conducted experimental comparisons on the three aforementioned test sets. For each variant, experiments were conducted three times, and the results were averaged. Specifically, the five different variants are as follows: w/o VN: In this variant, we do not use any virtual nodes and only rely on the regular message passing mechanism to learn covalent and non-covalent interactions. w/o VN: In this variant, we do not use any virtual nodes and only rely on the regular message passing mechanism to learn covalent and non-covalent interactions. w/o VN-D: In this variant, we omit the drug virtual node, and during non-covalent message passing, only allow the regular drug nodes to fuse information from the target pocket virtual node, while keeping the other components unchanged. w/o VN-D: In this variant, we omit the drug virtual node, and during non-covalent message passing, only allow the regular drug nodes to fuse information from the target pocket virtual node, while keeping the other components unchanged. w/o VN-P: Similar to the previous variant, but in this variant, we omit the target pocket virtual node and choose to retain the drug virtual node. w/o VN-P: Similar to the previous variant, but in this variant, we omit the target pocket virtual node and choose to retain the drug virtual node. SVN: In this variant, we employ only a single virtual node to simultaneously integrate the global information of drugs and targets SVN: In this variant, we employ only a single virtual node to simultaneously integrate the global information of drugs and targets w/o Deg: To investigate the importance of non-covalent degree information, we designed a variant without non-covalent degree information for comparison. w/o Deg: To investigate the importance of non-covalent degree information, we designed a variant without non-covalent degree information for comparison. Table 3 presents a performance comparison between the complete model and five variants across three datasets. Overall, our complete model demonstrates strong performance across all metrics, confirming that each module in DVNDTA is essential. A detailed analysis of the variant performances reveals that when only a single protein or ligand virtual node is used (i.e., w/o VN-D or w/o VN-P), the model’s performance not only fails to outperform the variant without virtual nodes (w/o VN), but in some cases, it even performs worse. This suggests that introducing a single virtual node may introduce noise or imbalance in information flow, failing to effectively enhance the model’s ability to learn protein-ligand interactions. In contrast, the SVN variant, which employs a single virtual node to simultaneously integrate global information from both the drug and the target pocket, demonstrates modest improvements over the w/o VN baseline across all three test sets. This indicates that incorporating global context—even through a shared virtual node—can benefit DTA prediction by providing a more comprehensive view of the interaction landscape. Table 3Ablation study results of DVNDTA on three test setsTest setModelRMSE↓MAE↓R↑SD↓CASF-2013w/o VN1.407 ± 0.0201.085 ± 0.0190.810 ± 0.0061.406 ± 0.016w/o VN-D1.430 ± 0.0151.129 ± 0.0240.799 ± 0.0081.429 ± 0.017w/o VN-P1.412 ± 0.0231.082 ± 0.0220.808 ± 0.0101.406 ± 0.021SVN1.385 ± 0.0181.070 ± 0.0200.815 ± 0.0051.382 ± 0.015w/o Deg1.392 ± 0.0161.077 ± 0.0190.810 ± 0.0071.396 ± 0.014DVNDTA1.332 ± 0.0141.071 ± 0.0180.827 ± 0.0031.330 ± 0.010CASF-2016w/o VN1.214 ± 0.0300.946 ± 0.0210.834 ± 0.0091.212 ± 0.014w/o VN-D1.227 ± 0.0140.946 ± 0.0180.828 ± 0.0061.226 ± 0.018w/o VN-P1.220 ± 0.0250.926 ± 0.0200.830 ± 0.0041.216 ± 0.013SVN1.185 ± 0.0160.912 ± 0.0170.842 ± 0.0071.181 ± 0.015w/o Deg1.202 ± 0.0211.894 ± 0.0270.835 ± 0.0051.196 ± 0.009DVNDTA1.145 ± 0.0130.891 ± 0.0120.851 ± 0.0051.142 ± 0.012holdout-2019w/o VN1.394 ± 0.0321.089 ± 0.0240.635 ± 0.0071.392 ± 0.016w/o VN-D1.403 ± 0.0181.094 ± 0.0170.636 ± 0.0091.399 ± 0.014w/o VN-P1.404 ± 0.0191.096 ± 0.0160.637 ± 0.0101.403 ± 0.015SVN1.375 ± 0.0121.072 ± 0.0100.648 ± 0.0051.372 ± 0.011w/o Deg1.382 ± 0.0141.084 ± 0.0090.641 ± 0.0081.381 ± 0.013DVNDTA1.369 ± 0.0071.064 ± 0.0110.658 ± 0.0061.368 ± 0.009 Ablation study results of DVNDTA on three test sets However, SVN still underperforms compared to the full DVNDTA model. The performance gap highlights a key limitation of the single-virtual-node design: it conflates the distinct global characteristics of the drug and the target pocket into a single representation. This fusion may lead to information interference or ambiguity, making it difficult for the model to differentiate and appropriately weight drug-specific versus pocket-specific global features during message passing. In contrast, DVNDTA’s dual virtual nodes preserve the semantic separation between drug and pocket global information, allowing each node to adaptively and independently regulate the integration of its counterpart’s global context through the dynamically weighted fusion mechanism (Eq. 6). This architectural design enables a more precise and flexible modeling of the bidirectional feedback between global and local information, ultimately leading to superior performance. Furthermore, when non-covalent degree information is incorporated (as in our complete model), performance improves further, indicating that node-specific interaction intensity awareness enhances the model’s ability to dynamically balance local and global cues. Performance evaluation on external and strict cold-start test sets To rigorously evaluate the generalization capability of DVNDTA on external test sets, we conducted a comprehensive evaluation on the CSAR-HiQ dataset, which serves a non-PDBbind external test set. This evaluation tests the model’s performance on complexes from different origins. The specific experimental results are shown in Table 4. Table 4Performance evaluation on the non-PDBbind external test set (CSAR-HiQ)Test setModelRMSE↓MAE↓R↑SD↓CSAR-HiQDeepDTA1.721 ± 0.0411.365 ± 0.0320.512 ± 0.0211.702 ± 0.038MGraphDTA1.683 ± 0.0371.324 ± 0.0290.538 ± 0.0191.658 ± 0.033Pafnucy1.546 ± 0.0281.210 ± 0.0240.601 ± 0.0151.523 ± 0.025GNN-DTI1.588 ± 0.0311.247 ± 0.0270.573 ± 0.0171.562 ± 0.029IGN1.512 ± 0.0251.189 ± 0.0200.624 ± 0.0131.490 ± 0.022GEMF1.506 ± 0.0081.183 ± 0.0100.588 ± 0.0061.462 ± 0.020MFE1.502 ± 0.0121.178 ± 0.0110.592 ± 0.0091.459 ± 0.018DVNDTA1.485 ± 0.0231.167 ± 0.0190.637 ± 0.0121.455 ± 0.012 Performance evaluation on the non-PDBbind external test set (CSAR-HiQ) Table 4 presents the performance of DVNDTA and state-of-the-art baselines on the non-PDBbind external test set (CSAR-HiQ). As expected, traditional complex-free models such as DeepDTA and MGraphDTA exhibit the weakest performance, with RMSE values exceeding 1.68 and R below 0.54. This significant performance drop highlights their heavy reliance on learned chemical substructures and protein sequences, making them highly sensitive to the distribution shift between the PDBbind training set and external test sets from different sources. making them highly sensitive to novel compounds not seen during training. In contrast, 3D structure-based models demonstrated superior generalization, with RMSE values ranging from 1.48 to 1.59 and R exceeding 0.60. This suggests that explicitly modeling spatial interactions can provide more transferable knowledge for affinity prediction under data-scarce conditions. Notably, DVNDTA outperformed all baseline models across all metrics. This superior performance is attributed to the dual virtual nodes designed into our model and the unique update mechanism, which enables a progressive optimization of the model’s understanding of interactions. These results demonstrate that DVNDTA not only exhibits high accuracy on standard benchmarks but also demonstrates excellent generalization on external test sets, making it a powerful tool for de novo drug discovery and target identification. Furthermore, to verify the model’s robustness under a rigorous cold-start scenario, we evaluated DVNDTA on the Strict Cold-Start Holdout set (N=699) constructed in Section 3.1. As shown in Table 5, we compared the performance of DVNDTA on the full v.2019 holdout set (N=4366) versus the strictly filtered subset. Table 5Performance comparison of DVNDTA on the PDBbind v.2019 Holdout set before and after strict sequence similarity filteringDatasetNRMSE↓MAE↓R↑SD↓Holdout-2019 (Full)43661.369 ± 0.0071.064 ± 0.0110.658 ± 0.0061.368 ± 0.009Holdout-2019 (Strict)6991.464 ± 0.0151.134 ± 0.0180.526 ± 0.0121.461 ± 0.016 Performance comparison of DVNDTA on the PDBbind v.2019 Holdout set before and after strict sequence similarity filtering Although removing homologous targets naturally leads to a performance drop (RMSE increasing from 1.369 to 1.464), the model maintains a Pearson correlation coefficient (R) of 0.526. This indicates that even in the absence of sequence homology, DVNDTA effectively captures the underlying drug-target binding mechanisms. This comparison demonstrates that the model’s predictive capability is driven by learned physical interactions rather than memorization of target templates. Visualization analysis To evaluate the effectiveness of our model in identifying potential patterns in drug-target complex data, we randomly selected 256 complexes from the CASF-2016 dataset for experimentation. Each complex was treated as a node in a graph and normalized based on its DTA label. Different colors were used to mark nodes with different normalized scores, allowing us to visually distinguish them. Next, the graph structure of each complex was processed through the dual virtual node heterogeneous interaction layer and transformed into a high-dimensional feature vector using global summation pooling techniques to represent the unique features of each complex. Detailed results of the high-dimensional feature visualization (including t-SNE analysis and corresponding figures) are provided in Section S5 of the supplementary materials to focus the main text on core performance conclusions. Parameter analysis In this section, we mainly explored the impact of the number of layers and the size of hidden layers in the DVNHIL model on DTA prediction. Figure 4(left) shows the performance of DVNHIL models with different numbers of layers on the validation set and three test sets after training. It can be observed that as the number of layers increases, the RMSE on the validation set continually decreases, and the prediction accuracy improves. However, when the number of layers reaches five, despite the continued reduction in RMSE on the validation set, the model’s performance on all three test sets declines, indicating overfitting. In contrast, the model with four layers of DVNHIL performed the best on the test sets. Figure 4(right) illustrates the effect of different hidden layer dimensions on model performance. The experimental results show that when the hidden layer dimension is 256, the model achieves optimal performance. However, when the dimension is further increased to 320, the model performance decreases, approaching the performance observed when the dimension is 192. This suggests that each node’s features do not need excessively high dimensions for representation, and too high a dimension might introduce unnecessary complexity, negatively impacting the model’s generalization ability. Additionally, from the RMSE’s trend in both figures, it is clear that increasing the number of DVNHIL layers is more likely to lead to overfitting than increasing the hidden layer dimensions. Therefore, when optimizing the model structure, careful selection of the number of layers and hidden layer dimensions is necessary to avoid overfitting and ensure the model’s best performance. The impact of different parameters on DVNHIL performance across three test sets. (Left) Different numbers of DVNHIL layers. (Right) Different hidden layer dimensions Conclusion In summary, this paper introduces an efficient deep learning model for DTA prediction, named DVNDTA. By incorporating virtual nodes for both target pockets and drug molecules, DVNDTA effectively captures local structural features at binding sites while preserving the global physicochemical properties of both the target pocket and the drug. Moreover, these virtual nodes are dynamically updated as network layers deepen, establishing a bidirectional feedback mechanism between global and local information. This design iteratively refines the integration of global and local information, serving as an abstract proxy for the mutual adjustments observed in the drug-target binding process.Experimental results across multiple test sets demonstrate that DVNDTA achieves superior predictive accuracy and generalization, highlighting its potential as a powerful tool for DTA prediction. Notably, the model maintains robust performance even on a strictly filtered cold-start dataset, validating its ability to generalize to novel targets with low sequence identity. In the future, as more high-quality 3D complex data will become available, we can further design more complex and precise gating strategies to optimize the fusion ratio between each node and the virtual nodes. This could not only further enhance model performance but also allow us to gain deeper insights into the interaction mechanisms between targets and drugs. Supplementary Information Supplementary Material 1. Author contributions B.L (Conceptualization, Methodology, Software, Data Curation, Writing—Original Draft), G.Z (Resource, Writing—Review & Editing, Supervision), H.C (Writing—Review & Editing), W.Z (Writing—Review & Editing),H.L (Methodology & Software), H.W (Data Curation & Writing Review),J.M (Writing Review & Methodology Validation), S.B (Writing—Review& Editing), J.S (Writing & Editing) Funding This work was supported by the Open Research Fund of National Mobile Communications Research Laboratory, Southeast University (No.2023D15); the National Program on Key Basic Research Project (2020YFA0713600); and the National Natural Science Foundation of China (62272214). Data availability The PDBbind dataset used in this study can be downloaded from https://www.pdbbind-plus.org.cn/. The source code can be freely downloaded from https://github.com/liben200012/dvndta. Declarations Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Competing interests The authors declare no conflict of interest. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Guoqiang Zhou and Ben Li contributed equally to this work. Supplementary Information The online version contains supplementary material available at 10.1186/s12859-026-06481-5. References Associated Data Supplementary Materials Supplementary Material 1. Data Availability Statement The PDBbind dataset used in this study can be downloaded from https://www.pdbbind-plus.org.cn/. The source code can be freely downloaded from https://github.com/liben200012/dvndta.