Contrastive learning in both structure and function spaces improve drug-target interaction prediction

Abstract

Background

Identifying drug–target interactions (DTIs) is essential in drug discovery and repositioning. Recently, deep learning has become the mainstream methodology for DTI prediction. However, the scarcity of three-dimensional structural data has forced almost all methods to predict drug-target interactions with low-dimensional data, thereby constraining their overall performance.

Methods

In tackling this challenge, we introduce a novel approach, CLSF-DTI. CLSF-DTI incorporates high-dimensional structural and functional information into the drug and protein features through contrastive learning during the feature extraction stage. This ensures that the model no longer solely focuses on sequence information, leading to a more precise modeling outcome.

Results

Experiments on five benchmark datasets demonstrate that CLSF-DTI achieves the best overall performance among five state-of-the-art baselines. Through ablation studies, we further prove that the contrastive module enhances the predictive performance and generalization ability of CLSF-DTI. Moreover, CLSF-DTI successfully identified some ligands for the protein PKA-C in drug screening experiments.

Conclusions

This study proposed a contrastive learning model CLSF-DTI that integrates structural and functional similarity. It outperforms existing methods in drug-target interaction prediction and has stronger generalization ability. However, the handling of unbalanced data and long-distance dependencies still needs to be improved in the future. The data and source code are available at https://github.com/ZhangLab312/CLSF_DTI.

Introduction

Drug-target interactions (DTIs) involve the binding of drugs to target proteins (e.g., enzymes, receptors), enabling disease treatment and physiological modifications. Identifying DTIs is crucial for drug discovery and repositioning. While traditional experimental screening is costly and time-consuming, computational approaches offer efficient alternatives. However, a fundamental limitation persists across methods: Because most proteins lack experimentally determined three-dimensional structures, reliance on low-dimensional data representations (e.g., amino acid sequences, SMILES strings, molecular graphs) that lack critical structural and functional information, potentially compromising model performance and generalizability [1–4].

Traditional prediction paradigms include docking-based methods and ligand-based approaches [5, 6]. Docking-based methods require protein 3D structures, limiting applicability to targets without solved structures. Ligand-based methods depend on known ligand data, performing poorly for targets with few ligands. Chemogenomics emerged to address these gaps by integrating chemical and genomic spaces [7–11], with subsequent machine learning implementations [12–14]. Nevertheless, these methods require manual feature engineering, risking information loss and expert bias.

Deep learning now dominates DTI prediction, broadly categorized into network-based and structure-based methods. Network-based approaches [15, 16] leverage similarity principles ("similar drugs bind similar targets") using graph convolutional network(GCN) on biological networks [17–21]. For instance, Peng et al. [18] used GCN on heterogeneous graphs for feature extraction. These methods struggle with isolated nodes, are confined to network entities, and risk false positives due to empirical similarity rules [15, 16, 22]. Structure-based methods [23–25] follow the "structure determines function" principle, typically involving three stages: data representation, feature extraction, and prediction. Early works used protein descriptors and molecular fingerprints [26–29], but these handcrafted features often lacked completeness. Later approaches like DeepConv-DTI [23] used raw sequences with embeddings and Convolutional Neural Networks (CNNs), while others integrated descriptors with sequences using autoencoders and CNNs [30]. MolTrans [24] employed Transformer substructure encoding but overemphasized local interactions at the expense of global context. CoaDTI [31] used co-attention for interaction modeling but faced Transformer’s O() complexity limitations. Bian et al. [32] demonstrated CNN superiority over Transformer for current DTI data scales using cross-attention.

Existing methods prioritize model architecture innovation and interpretability but critically overlook limitations from low-dimensional data inputs. Figure 1 illustrate the molecules with distinct 2D topologies (A,B) exhibit similar bioactivity against VEGFR2, while structurally similar molecules (B,C) differ functionally due to 3D conformational differences [33]. This highlights the insufficiency of 1D/2D representations alone. Although high-dimensional annotation information is available during training, many proteins lack complete structural or functional annotations in real-world applications.

Fig. 1.

Molecular 2D structure similarity versus activity

To address this, we propose CLSF-DTI, an integrated framework that leverages contrastive learning to incorporate high-dimensional structural and functional knowledge. Our method inputs protein amino acid sequences and drug SMILES strings but enriches feature extraction through: 1) Protein annotations (Molecular Function, Structural Domain, Sequence) capturing functional, 3D substructural, and sequential information [34]; and 2) Drug fingerprints (MACCS, RDKF, 2D-PMF) encoding substructural, topological, and pharmacophore features [35]. Unlike heterogeneous graph-based methods, it uses a decoupled, modular contrastive learning loss function to treat multiple structural similarities as independent supervision signals. Crucially, contrastive learning aligns these multimodal representations during training, enabling prediction using only sequences and SMILES at inference.

We conducted a series of experiments on CLSF-DTI to showcase its superiority. Initially, CLSF-DTI was evaluated on five benchmark datasets and consistently outperformed existing advanced methods. Following that, we devised an unseen dataset to assess the generalization capabilities of different models. In the generalization comparison experiment, CLSF-DTI again exhibited outstanding performance. Subsequently, a drug screening experiment was conducted on the protein PKA-C in the ChEMBL database to validate the utility of CLSF-DTI. Remarkably, CLSF-DTI successfully identified 22 effective ligands. Lastly, we visualized the attention weight distribution of the Cross-Attention module to show how CLSF-DTI works. The experimental findings strongly suggest that CLSF-DTI offers specific insights into DTI prediction tasks.

Our key contributions are:

• We are proposing a novel approach to tackle limited information in low-dimensional data. To the best of our knowledge, this is the first instance where contrastive learning has been proposed to incorporate structural and functional annotation information, addressing the issue of limited information in low-dimensional data.

• Presenting a module that employs contrastive learning to combine various structural and functional annotations, enabling the model to indirectly learn information about proteins and drug molecules in high-dimensional space.

• Designing CLSF-DTI as a DTI prediction model with inherent interpretability and the potential for drug screening of specific targets.

Methods

Problem formulation

Drug-Target Interaction prediction aims to determine whether a given drug compound and target protein will interact. Formally, the problem is defined as learning a mapping function that predicts the interaction probability for a drug-protein pair . The drug compound is represented as a fixed-length encoded SMILES sequence , where each denotes an encoded SMILES symbol (0 = padding) and is the fixed maximum sequence length. Similarly, the target protein is represented as a fixed-length encoded amino acid sequence , where each denotes an encoded amino acid (0 = padding) and is the fixed maximum sequence length. The joint feature space corresponds to the domain of encoded input pairs processed by CLSF-DTI’s embedding layers: .

Overview of the CLSF-DTI

Our proposed CLSF-DTI’s overall architecture is shown in Figure 2. The model comprises five parts: "Encode-Embed Layers, CNN Blocks, Cross-Attention Module, Contrast Modules, and dense layers for DTI prediction. First, two distinct CNN Blocks initially extract features from amino acid sequences and drug SMILES. Then, a cross-attention module enhances protein and drug molecule representation. Subsequently, maximum pooling obtains features for proteins and drug molecules. Next, the Contrast Modules introduce structural similarity information separately into protein and drug molecule features. Finally, a fully connected layer merges these features, producing predicted probabilities for drug-target interactions. It is worth noting that CLSF-DTI introduces structural and functional similarity information through contrastive learning during the training phase, but relies solely on protein amino acid sequences and drug SMILES strings during the inference phase.

• Encode-embed layers translate protein and drug sequences into numerical representations for computation.

• CNN blocks employ one-dimensional convolution to extract features from protein and drug molecule sequences.

• Cross-attention module utilizes attention mechanisms to model the interaction between proteins and drug molecules.

• Contrast modules perform contrastive learning on the structural and similarity matrices of drugs and proteins separately and incorporate high-dimensional structural and functional information into the extracted features.

• DTI prediction merges protein and drug molecule features using fully connected layers to predict drug-target interactions.

Fig. 2.

Overview of the CLSF-DTI model architecture. The model encodes drug and protein sequences through Encode-Embed layers, extracts features via convolutional blocks, and captures drug–target interactions using a cross-attention module. Structural and functional information is integrated through contrastive learning. Finally, fully connected layers generate the drug–target interaction predictions

Encode-embed layers

In this work, we input the model’s amino acid sequences and SMILES strings. Following established approaches [23, 32, 36], we set fixed lengths for protein and drug strings. Characters are then numerically encoded, with amino acid sequences ranging from 1 to 24 and SMILES characters from 1 to 64. The number 0 is used to denote empty characters. Lastly, inspired by word embedding, we create trainable embedding matrices for proteins and for drug molecules, where and represent the embedding dimensions for amino acid characters and SMILES characters, respectively. By referencing the corresponding embedding matrix, proteins and drugs within a batch are defined as embedding tensors and , respectively, where N indicates the number of DTIs in a batch, and and are the maximum lengths of truncated amino acid sequences and SMILES strings, respectively.

CNNs blocks

Convolutional Neural Networks excel at capturing intricate local residue features. In this work, we utilize two CNN blocks to extract distinctive features from the embedding matrices of proteins and drugs, respectively. Each block consists of three one-dimensional convolutions, with a ReLU activation function applied after each convolutional layer. The CNNs extraction layers for proteins and drug molecules are as follows:

1

2

In this context, takes values from the set . and are the weight matrix and bias vector in the -th layer of the protein CNN Block. represents -th hidden protein representation, and . Similarly, and are the weight matrix and bias vector in the -th layer of the drug CNN Block. is the drug representation in the -th hidden layer, and . Here, denotes the non-linear activation function, and we use the ReLU activation function in this work.

Cross-attention module

The Cross-Attention mechanism and the self-Attention mechanism in the Transformer[37] exhibit subtle differences in structure, specifically with the Query and Key in the Cross Attention Module originating from different data sets. However, their functionalities diverge significantly. The self-Attention mechanism allows the model to capture relational information across various positions sequentially. In contrast, the Cross-Attention mechanism excels in integrating information from two different sources, enabling more precise modeling. Following the acquisition of protein and drug feature matrices, we employ the Cross Attention Module to learn the interactive information between proteins and drug molecules, thereby enhancing both representations. The equation for the Cross-Attention Module is as follows:

3

4

Where and represent the enhanced protein and drug feature matrices, respectively, and denote the protein and drug representations generated by the last layer of the CNN blocks. , , and are the mapping functions for query, key, and value, respectively, and refers to the multi-head attention function in the Transformer.

Next, the enhanced protein and drug molecule representations are fed into the max-pooling layer. This calculates, for each batch, the feature vectors for proteins and for drugs, where represents the number of drug-target interaction pairs in a batch, and denotes the feature dimension.

5

6

Contrastive modules

Contrastive learning, initially applied in computer vision, revolves around minimizing contrastive loss to decrease the distance in the feature space for positive pairs and maximizing the distance for negative pairs. This aims to facilitate the model acquiring a more compact representation of samples. Compared to directly inputting features, a key aspect of contrastive learning is identifying positive and negative sample pairs. The common practice in image processing is to construct these pairs through data augmentation, but this approach is only suitable for some biological data. In this study, we gathered three types of similarity matrices reflecting both the structural and functional properties of proteins and three types reflecting the structural and functional properties of drugs. Subsequently, we established thresholds for each type of similarity. When the similarity between two proteins surpasses the threshold, they are categorized as positive pairs; otherwise, they are considered negative. It is crucial to note that proteins and drugs are not regarded as positive or negative pairs with themselves.

Next, we will illustrate the contrastive learning module using proteins as an example. The contrastive learning module for proteins is depicted in Figure 3.

Fig. 3.

Diagram of Protein Contrastive Module. It first utilizes two fully connected layers to map protein features into a specific contrastive space, then calculates the contrastive loss based on negative and positive pairs

For a given batch of protein representations , three independent networks, each equipped with two fully connected layers, are employed to map the protein features to the contrast spaces of molecular function, structural domain annotation, and sequence annotation, respectively.

7

Here, represents the contrast space, and , represents the hidden features of proteins in the contrast space for the current batch. and denote the weights and biases of the first linear layer for the mapping of contrast space , while and represent the weights and biases of the second linear layer. denotes the activation function. in this context, LeakyReLU is employed.

Subsequently, positive and negative pairs are constructed based on the respective similarity matrices. The InfoNCE [38] method is utilized to compute the contrastive loss within a given batch.

8

Where represents the contrastive loss of proteins within the current batch in contrast space , indicates the positive pairs of protein in contrast space within the current batch. denotes the hidden features of protein in contrast space , and represents the function for similarity calculation, with cosine similarity utilized in this context. is the temperature coefficient.

Finally, aggregating the contrastive losses of proteins from the three contrast spaces to derive the protein’s comprehensive contrastive loss, denoted as .

9

Following the same procedure, we calculate the current batch’s total contrastive loss for drugs. Combining the contrastive losses for proteins and drugs results in the overall contrastive loss .

10

While experimenting with similarity thresholds, we treat different similarity thresholds as independent. We create six contrastive learning modules, each exclusively using one type of similarity matrix. Then, we conduct experiments on the DrugBank dataset with a series of discrete values. Finally, we selected the threshold with the highest accuracy as the optimal threshold for each corresponding similarity matrix. This optimal threshold is the mean of the five random partitioning experiments. The experimental results, shown in Figure 4, clearly indicate that integrating any similarity data for contrastive learning improves the model’s accuracy, with the most notable enhancements observed in protein sequence similarity (PrePSSM) and molecular structural similarity (MACCS).

Fig. 4.

Model performance under different similarity thresholds in the single contrastive space. A threshold of 0 indicates the absence of contrastive learning. Subfigure (a) shows the results of similarity threshold selection for drugs, while subfigure (b) presents the results for proteins

DTI prediction

During contrastive learning, we concatenate the protein features () with drug features () and feed them into a fully connected layer for prediction. We employ cross-entropy to calculate the loss () between the predicted results and the true labels.

11

Here, represents the actual DTI labels, and denotes the probability of predicting a positive DTI.

Finally, integrate the contrastive loss and the cross-entropy loss to obtain the ultimate loss that needs optimization.

12

Results

We conducted extensive experiments on several datasets to demonstrate the effectiveness of CLSF-DTI. These experiments primarily address the following research questions:

• Q1: Can the performance of our model surpass state-of-the-art DTI prediction methods?

• Q2: How does the model perform on unseen data?

• Q3: Does the contrastive learning strategy truly contribute to improving DTI prediction?

• Q4: Does our model perform well in practical drug screening processes?

In the following sections, we introduce the experimental setup covering the benchmark datasets utilized for evaluation and some experiment details. Subsequently, a comparison between CLSF-DTI and various advanced DTI prediction methods is presented. Following this, an analysis of how individual modules influence the model’s performance is conducted. Then, we work on a case study using CLSF-DTI, examining its performance in an actual drug screening process. Finally, we visualize the attention weights to analyze the rationality of the model predictions.

Experimental setup

Datatset

We employed five widely used datasets: DrugBank[39], KIBA[40], Enzymes, Ion Channels, and GPCRs. Specifically, DrugBank and KIBA were sourced from their respective public databases, DrugBank and KIBA. Enzymes and Ion Channels were obtained from KEGG BRITE[41], which includes data from DrugBank and Supertarget[42]. The GPCRs dataset was sourced from GLASS[43].

• The DrugBank dataset is a comprehensive collection encompassing a variety of protein-drug interactions and comprising 35,022 DTI records formed by 42,924 proteins and 6,655 drugs.

• The KIBA dataset, generated by partitioning continuous activity data with an activity threshold of 12.1, is imbalanced and dense. It includes 116,350 DTI records involving 225 kinases and 2,068 inhibitors.

• The Enzymes dataset is a smaller collection, containing 5,840 DTI records formed by 660 enzyme proteins and 444 drugs.

• The GPCRs dataset, a more compact dataset, comprises 6,197 DTI records formed by 296 G protein-coupled receptor (GPCR) proteins and 567 drugs.

• The Ion Channels dataset, also relatively compact, includes 2,950 DTI records formed by 204 ion channel proteins and 210 drugs.

They were partitioning activity data obtained from all samples in the KIBA dataset. Positive samples were extracted from the corresponding public datasets for the other datasets, and negative samples were randomly generated following the HyperAttention[36] method. Protein and drug representations were expressed in each dataset using amino acid sequences and SMILES strings, respectively. Statistics for these datasets are presented in Table 1, where "Pos" and "Neg" are abbreviations for "Positive" and "Negative," respectively.

Table 1.

Statistics of Benchmark Datasets

Dataset	Protein	Drug	Pos DTI	Neg DTI	Total DTI
DrugBank	4294	6655	17511	17511	35022
KIBA	225	2068	22154	94196	116350
Enzymes	660	444	2920	2920	5840
GPCRs	296	567	3098	3099	6197
Ion Channels	204	210	1475	1475	2950

Then, we collected structural and functional annotation information for proteins and drugs and computed their similarities. In the construction of the drug similarity matrices, we initially employed RDKit[44] to calculate Molecular Access Systems keys fingerprint (MACCS)[35], RDKFingerprint (RDKit Fingerprint), and 2D pharmacophore fingerprints for drugs. These metrics reflect substructure information, topological structure information, and pharmacophoric functional information of drug molecules[45]. Subsequently, we used the Tanimoto function to compute fingerprint similarities.

In constructing the protein similarity matrices, we initially gathered molecular function annotations (MF) and structural domain annotations for proteins from uniprot[34] and then computed a pseudo-position specific scoring matrix (PsePSSM)[46] for each protein as a sequence annotation, in which the amino acid sequence of each protein is mapped to a 420-dimensional vector. Next, following the methodology of Wang[47], we assessed molecular function similarity between proteins. Then, we determined structural domain similarity using the formula 13, where denotes the structural domain annotation set of protein i, and represents the structural domain annotation set of protein j. Finally, we calculated sequence similarity between proteins using formula 14, with and denoting the PrePSSM vectors of protein i and protein j, respectively.

13

14

Baselines

We compared our proposed model, CLSF-DTI, with state-of-the-art DTI prediction models, which include DeepConv-DTI[23], Moltrans[24], TransformerCPI[12], HyperAttentionDTI[36], and MCANet[32]. DeepConv-DTI employs a multi-scale single convolutional layer to extract protein features. HyperAttentionDTI and MCANet, based on DeepConv-DTI, use more complex CNN architectures for feature extraction and incorporate attention mechanisms to enhance interpretability by capturing interaction information between proteins and drugs. In contrast, Moltrans and TransformerCPI are DTI prediction models built on the Transformer architecture. Moltrans uses the encoder part of the Transformer for feature extraction, while TransformerCPI utilizes the decoder part of the Transformer to simulate protein-drug interaction, thus improving model interpretability. To ensure fairness, all baseline models were given their best hyperparameters and used the exact same dataset as our model CLSF-DTI.

Parameter settings

CLSF-DTI is implemented in PyTorch 1.13 and trained on an NVIDIA GeForce RTX 2080Ti with 12GB RAM. To maintain consistency, we used the same set of hyperparameters across all datasets, where the batch size was set to 64, the total number of epochs was set to 200, the optimizer was AdamW, the learning rate was 0.0001, the weight decay was 0.0001, and the dropout size was 0.1. To ensure model consistency, we used the same hyperparameters across all datasets. In all subsequent experiments, all datasets were split into training and test sets in an 8:2 ratio. The training set was then further divided into five parts for 5-fold crossover experiments. Finally, the average performance of the model on the test set was used as the experimental result.

Performance comparison (Q1)

We comprehensively compared the larger datasets, DrugBank and KIBA, using five metrics: Accuracy, AUC, AUPR, and Precision. The experimental results are presented in Table 2. Our approach consistently outperforms other methods on four datasets, showcasing the effectiveness of integrating structural and functional information through contrastive learning in enhancing DTI prediction modeling. Compared to the MCANet model, our method demonstrates notable improvements on the Enzymes dataset with increases of 4.3, 4.9, and 4.4 in accuracy, AUC, and AUPR, respectively. Similarly, on the DrugBank dataset, we observe enhancements of 2.5, 2.4, and 2.5 in accuracy, AUC, and AUPR. For the GPCRs dataset, our approach achieves improvements of 3.1, 3.5, and 2.9 in accuracy, AUC, and AUPR. Because the ratio of positive to negative samples in the KIBA dataset is approximately 2:8, the performance differences among all methods are not significant on the KIBA dataset. This indicates that current methods face certain difficulties in effectively modeling highly imbalanced datasets like KIBA.

Table 2.

Comparison Results of CLSF-DTI and Baselines on DrugBank and KIBA Datasets

Dataset	Method	Accuracy	Precision	AUC	AUPR
DrugBank	DeepConv-DTI[23]	0.792(0.004)	0.803(0.005)	0.858(0.006)	0.851(0.009)
	MolTrans[24]	0.789(0.003)	0.776(0.012)	0.862(0.003)	0.870(0.004)
	TransformerCPI[12]	0.782(0.012)	0.773(0.017)	0.853(0.011)	0.859(0.015)
	HyperAttentionDTI[36]	0.811(0.008)	0.799(0.021)	0.889(0.008)	0.884(0.007)
	MCANet[32]	0.826(0.005)	0.824(0.015)	0.897(0.004)	0.905(0.005)
	CLSF-DTI	0.851(0.003)	0.843(0.006)	0.921(0.002)	0.930(0.002)
KIBA	DeepConv-DTI[23]	0.891(0.007)	0.805(0.035)	0.907(0.003)	0.804(0.016)
	MolTrans[24]	0.889(0.014)	0.704(0.028)	0.923(0.002)	0.795(0.002)
	TransformerCPI[12]	0.885(0.018)	0.721(0.024)	0.907(0.007)	0.765(0.013)
	HyperAttentionDTI[36]	0.878(0.012)	0.673(0.021)	0.916(0.009)	0.772(0.008)
	MCANet[32]	0.899(0.001)	0.735(0.007)	0.931(0.002)	0.811(0.005)
	CLSF-DTI	0.903(0.001)	0.742(0.016)	0.934(0.003)	0.815(0.009)

The best results are marked bold

As for the smaller datasets, Enzymes, Ion Channels, and GPCRs, where each model’s performance was unsatisfactory, we employed three comprehensive metrics, Accuracy, AUC, and AUPR, to compare the models. The experimental results are shown in Table 3. TransformerCPI and HyperAttentionDTI models perform well on the Drugbank and KIBA datasets but exhibit poor performance on the smaller datasets, such as Enzymes, Ion Channels, and GPCRs. In contrast, DeepConv-DTI, MCANet, and CLSF-DTI perform well across all five datasets. This suggests that, for smaller datasets, CNN-based models are more competitive than Transformer-based models.

Table 3.

Comparison Results of CLSF-DTI and Baselines on Enzymes, Ion Channels and GPCRs Datasets

Dataset	Method	Accuracy	AUC	AUPR
Enzymes	DeepConv-DTI[23]	0.743(0.006)	0.807(0.011)	0.750(0.018)
	MolTrans[24]	0.728(0.014)	0.789(0.017)	0.735(0.020)
	TransformerCPI[12]	0.551(0.005)	0.582(0.002)	0.591(0.002)
	HyperAttentionDTI[36]	0.551(0.008)	0.518(0.010)	0.535(0.005)
	MCANet[32]	0.761(0.011)	0.804(0.012)	0.757(0.023)
	CLSF-DTI	0.804(0.007)	0.853(0.006)	0.801(0.012)
Ion channels	DeepConv-DTI[23]	0.688(0.020)	0.730(0.016)	0.694(0.008)
	MolTrans[24]	0.715(0.016)	0.717(0.020)	0.642(0.026)
	TransformerCPI[12]	0.562(0.017)	0.557(0.012)	0.524(0.015)
	HyperAttentionDTI[36]	0.538(0.019)	0.554(0.013)	0.560(0.008)
	MCANet[32]	0.718(0.022)	0.762(0.017)	0.702(0.014)
	CLSF-DTI	0.729(0.024)	0.764(0.002)	0.716(0.011)
GPCRs	DeepConv-DTI[23]	0.608(0.009)	0.641(0.009)	0.595(0.011)
	MolTrans[24]	0.606(0.013)	0.639(0.012)	0.602(0.009)
	TransformerCPI[12]	0.531(0.008)	0.567(0.021)	0.541(0.019)
	HyperAttentionDTI[36]	0.586(0.014)	0.608(0.017)	0.589(0.005)
	MCANet[32]	0.624(0.008)	0.666(0.007)	0.615(0.011)
	CLSF-DTI	0.655(0.012)	0.701(0.011)	0.644(0.013)

The best results are marked bold

Performance on unseen data (Q2)

To assess the model’s generalization capability, we reorganized the DrugBank dataset into three subsets: unseen drug, unseen protein, and unseen both datasets. Initially, we allocated 10 of the DrugBank dataset as the testing set for the three datasets above. Next, we excluded the drug-target interactions that involved the drugs in the testing set from the remaining data to form the training set for the unseen drug dataset. Similarly, we constructed the training sets for the unseen protein and datasets. Specifically, the training set for the unseen drug dataset lacks drug data from the testing set, the training set for the unseen protein dataset lacks protein data from the testing set, and the training set for the unseen dataset lacks drug and protein data from the testing set. Subsequently, we evenly divided the training above sets into five parts for 5-fold cross-validation. The experimental results were derived by averaging the model’s performance on the testing set and were detailed in Table 4.

• Compared to the full DrugBank dataset, all models show a notable performance decrease on these unseen datasets, particularly evident in the unseen datasets. Despite this, CLSF-DTI maintains optimal performance, exhibiting its better generalization capability.

• All models perform better on the unseen drug dataset than the unseen protein dataset, indicating that removing protein data significantly impacts the models. This is because, compared to drug molecules, proteins occupy a more extensive chemical space, making modeling proteins more challenging. Therefore, the model’s generalization capability for proteins tends to be slightly inferior to that for drugs.

Table 4.

Comparison Results of CLSF-DTI and Baselines on Unseen Datasets

Dataset	Method	Accuracy	AUC	AUPR
unseen drug	HyperAttentionDTI[36]	0.686(0.006)	0.777(0.001)	0.796(0.002)
	TransformerCPI[12]	0.719(0.008)	0.779(0.007)	0.813(0.005)
	MolTrans[24]	0.648(0.023)	0.748(0.007)	0.769(0.007)
	DeepConv-DTI[23]	0.693(0.005)	0.753(0.006)	0.761(0.008)
	MCANet[32]	0.722(0.006)	0.795(0.004)	0.813(0.007)
	CLSF-DTI	0.751(0.008)	0.824(0.008)	0.847(0.009)
unseen protein	HyperAttentionDTI[36]	0.567(0.006)	0.624(0.005)	0.649(0.007)
	TransformerCPI[12]	0.559(0.009)	0.603(0.003)	0.611(0.006)
	MolTrans[24]	0.508(0.003)	0.611(0.006)	0.651(0.004)
	DeepConv-DTI[23]	0.582(0.004)	0.611(0.004)	0.643(0.003)
	MCANet[32]	0.605(0.009)	0.656(0.014)	0.689(0.008)
	CLSF-DTI	0.634(0.014)	0.717(0.013)	0.733(0.012)
unseen both	HyperAttentionDTI[36]	0.507(0.007)	0.567(0.005)	0.551(0.011)
	TransformerCPI[12]	0.499(0.006)	0.561(0.004)	0.546(0.026)
	MolTrans[24]	0.507(0.009)	0.549(0.015)	0.535(0.008)
	DeepConv-DTI[23]	0.503(0.005)	0.526(0.013)	0.521(0.005)
	MCANet[32]	0.511(0.009)	0.569(0.016)	0.559(0.019)
	CLSF-DTI	0.527(0.009)	0.631(0.018)	0.611(0.017)

The best results are marked bold

Model ablation

To evaluate the efficacy of each module in our model, we deconstructed the CLSF-DTI model into the following variants:

• NoAttention(NoA): Removing the attention mechanism from the CLSF-DTI model.

• NoDrugContrast(NoDC): Excluding the drug contrastive learning modules from CLSF-DTI.

• NoProteinContrast(NoPC): Omitting the protein contrastive learning module from CLSF-DTI.

• NoContrasts(NoCs): Eliminating all contrastive learning modules from CLSF-DTI.

We evaluated the variant above models on the DrugBank dataset, and the experimental results are presented in Figure 5. From these results, we can draw several conclusions:

• Contrast learning modules and attention mechanisms improve the model’s performance in drug-target interaction prediction. However, the impact of the attention mechanism on model performance is relatively minor. After removing the attention mechanism, the model’s accuracy decreases by 0.6. In contrast, removing all contrastive learning modules results in a more substantial 2.7 accuracy drop. Among them, the NoDC variant showed a 1.8 decrease in AUC on DrugBank, indicating that drug contrast learning module is crucial for performance improvement. MACCS fingerprints encode key substructures of drug molecules, which are often directly related to biological activity and may contribute the most. The NoPC variant resulted in a larger AUC decrease of 2.4, suggesting that higher-order protein annotation provides a stronger supervisory signal. This indicates that incorporating functional and structural information from proteins and drug molecules through contrastive learning can significantly enhance drug-target interaction prediction.

• The Model NoDrugContrast outperforms NoProteinContrast, suggesting that removing the protein contrastive learning module leads to a more significant performance decline. This may be due to the challenging nature of extracting protein features, and including the protein contrastive learning module alleviates this issue, resulting in greater improvement in model performance.

Fig. 5.

Performance of CLSF-DTI and its variants on the DrugBank dataset

Drug screening for target PKA-C (Q4)

To evaluate the practical utility of CLSF-DTI in real-world scenarios, we conducted drug screening on the cAMP-dependent protein kinase catalytic subunit alpha (PKA-C, UniProt ID: P17612) from the KIBA dataset. PKA-C, encoded by the PRKACA gene, is known for phosphorylating numerous substrates in both the cytoplasm and nucleus, with prior research linking it to cardiovascular diseases and breast cancer formation[48].

Firstly, we gathered all small molecule data from the ChEMBL database[49], resulting in 1,913,075 remaining molecules after excluding those that couldn’t be represented using the drug embedding matrix Ed. Subsequently, numerical encoding was applied to these drug molecules and proteins. Finally, the data were separately evaluated by each of the five models trained through five-fold cross-validation on the KIBA dataset. We considered compounds with predicted probabilities above 0.5 and absent from the KIBA dataset as potential active drugs. To assess the reliability of the screening, we verified the predicted interactions using evidence from the ChEMBL database. The statistical results are shown in Table 5 (detailed results in Supplementary Material 1). In this study, "Verified Actives" specifically refers to compounds in the ChEMBL database that exhibit clearly defined binding or inhibitory activity against human PKA-C protein or its highly homologous proteins, as determined experimentally. All model instances correctly predicted some verified active compounds, but there are large differences between the predictions of different instances. This may be due to the small number of drugs in the KIBA dataset, which leads to variations in the drug features learned by each model.

Table 5.

Summary of drug screening results for PKA-C Using CLSF-DTI Trained on KIBA

Model instance	Predicted positives	Verified actives
Model from Fold 1	4299	3
Model from Fold 2	7594	9
Model from Fold 3	5746	14
Model from Fold 4	34808	57
Model from Fold 5	3483	11

Model interpretation (Q5)

To further demonstrate the efficacy of CLSF-DTI screening results, we visualized the model’s predictions on the protein PKA-C. Initially, we searched for the 3D structural analysis results of PKA-C in the PDB database[50] and identified two relevant PDB entries: 2GU8 and PDB: 5N23, which include two distinct molecules selected during the screening process. Subsequently, we computed the attention weights of the protein on drug molecules within the Cross-Attention Module. Finally, we utilized Pymol[51] for visualization, considering amino acids within a 5Å radius of the drug molecules as binding sites. The visualization results are shown in Figure 6. The fragments marked in red represent the binding sites where the model assigns high-weight scores, indicating correctly predicted binding regions. The blue areas denote true binding sites that did not receive high-weight scores. The yellow segments, on the other hand, represent non-binding site regions but received relatively high weight scores. It can be observed that our model accurately predicts some binding sites (red areas). Additionally, some yellow areas are close to the binding site in three-dimensional space, and the model may have captured this structural correlation, suggesting the possible existence of binding regions. Moreover, despite having the same protein in two PDB complexes, our model produces notably different attention weight patterns for different drugs. These findings indicate that CLSF-DTI indeed demonstrates a level of insight in predicting drug-target interactions.

Fig. 6.

Visualization of weight scores in Cross-Attention module. (a) PDB: 2GU8. (b) PDB: 5N23

Discussion

CLSF-DTI addresses a key limitation of existing deep learning methods in drug-target interaction (DTI) prediction: their reliance on low-dimensional representations, which fail to capture important high-order structural and functional information. By integrating contrastive learning within structural and functional spaces, CLSF-DTI significantly enhances prediction accuracy and generalization capability. Our comprehensive experiments demonstrate this improvement. This work presents three key contributions:

First, CLSF-DTI leverages contrastive learning to integrate high-dimensional structure and function annotations while requiring only low-dimensional input during inference. This approach mitigates the data scarcity issue associated with 3D protein structures without sacrificing model usability. Second, the contrastive module explicitly enforces similarity constraints derived from distinct biological perspectives. This enables CLSF-DTI to achieve SOTA performance on sparse datasets and in unseen scenarios, exhibiting AUC and AUPR improvements of 2.4–4.9 over Transformer- and CNN-based baseline models. Finally, in a virtual screening targeting PKA-C, CLSF-DTI successfully identifies 22 validated ligands from the ChEMBL database. Visualization of attention weights further confirms its ability to localize binding regions, supporting the model’s interpretability for real-world drug discovery applications.

Although this paper does not directly compare it with methods that rely on 3D structures, the core advantage of CLSF-DTI lies in its ability to achieve competitive performance using only the sequence, without relying on the three-dimensional structure of the protein. In scenarios where structural information is complete, 3D methods may be more advantageous; however, when the structure is unknown or scarce, CLSF-DTI provides a reliable and efficient alternative.

Conclusion

This work presents a contrastive learning model called CLSF-DTI, fusing structural-functional similarity to predict drug-target interactions. We argue that the current dependence on low-dimensional information constrains the model’s performance. To alleviate this issue, we introduce a new method combining contrastive learning to integrate diverse structural and functional annotations for proteins and drugs. This allows the model to concentrate on the semantic information of low-dimensional sequence data and grasp higher-dimensional structural and valuable features, ultimately improving the model’s overall performance. Compared to benchmark models on five datasets, CLSF-DTI outperforms them overall and exhibits more robust generalization capabilities to unknown data. During the drug screening, each instance of CLSF-DTI identified some effective binding compounds for PKA-C. Visualizations further support CLSF-DTI’s ability to capture specific drug-binding sites on proteins. While our CLSF-DTI demonstrates competitiveness in DTI prediction, it has some limitations. In future research, we plan to improve its performance on imbalanced datasets like KIBA and design better feature extractors to tackle challenges in capturing long-distance dependencies that CNNs may struggle with.

Author contributions

Y.Z., L.C. Q.Z. and W.Z. conceived this work and designed the experiments. L.C.,H.L.,T.L. collected the data and carried out the experiments. Y.Z., S.X., Z.W. developed the machine learning algorithm, analyzed the results. Y.Z. L.C. H.L. and W.Z. wrote the manuscript. Y.Z. and Q.Z. obtained funding. All authors read and approved the final manuscript.

Funding

This work was supported by the National Natural Science Foundation of China under Grant Nos. 62272067, 62131004, and the Sichuan Science and Technology Program under Grant No. 2026NSFSC0419.

Data availability

The data and source code are available at https://github.com/ZhangLab312/CLSF_DTI. The source data for the datasets used to train and evaluate the models are available at https://pan.baidu.com/s/13-fLZc-jODsvn2xVpsiTnw?pwd=09we, http://redpoll.pharmacy.ualberta.ca/drugbank/ and https://pubs.acs.org/doi/10.1021/ci400709d.

Declarations

Ethical approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Clinical trial number

Not applicable.

Conmpeting interest

The authors declare no Conflict of interest.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12859-026-06377-4.