Contextual Representation Learning With ResNet Refinement and Deep Classification for Enhanced Drug Target Interaction Prediction

ABSTRACT

Accurate prediction of drug target interactions (DTIs) plays a pivotal role in drug discovery and repositioning. However, it remains challenging due to the structural complexity of proteins and small‐molecule compounds, along with the limited generalisation capability of existing computational approaches. Drug target interaction prediction is important in computer‐aided drug design and drug repurposing, especially for complex diseases where multiple targets are involved. This study aims to develop a robust deep learning framework that enhances DTI prediction accuracy by effectively capturing contextual and biochemical features from both protein and drug representations. Four standard DTI datasets and a combined drug repurposing dataset are used to learn interaction patterns across multiple targets. We propose a novel deep learning framework that leverages pre‐trained BERT‐based language models to extract contextual embeddings from protein and drug sequences. These modality‐specific representations are refined using a proposed dedicated ResNet‐based subnetwork to preserve intrinsic biochemical characteristics. The refined embeddings are subsequently integrated and passed through a proposed deep feedforward neural network for final DTI prediction. The proposed model was evaluated on four benchmark datasets, namely DrugBank, Caenorhabditis elegans , BindingDB and GPCR. Experimental results demonstrate consistent performance improvements over baseline methods, including an F1‐score gain of ~6.6% on the GPCR dataset and a 3.7% increase in classification accuracy on BindingDB. Stable F1‐score improvements were also observed on DrugBank (0.4%) and Caenorhabditis elegans (0.8%). Statistical validation using paired ω‐tests at a 5% significance level confirms the improvements are significant. Evaluation on an independent drug repurposing dataset achieved a 6.7% performance gain over existing approaches. The results demonstrate that the proposed framework effectively captures contextual and structural information, leading to improved prediction accuracy and generalisation, highlighting its robustness and practical applicability for drug discovery and drug repurposing.

A proposed deep learning framework for drug target interaction (DTI) prediction integrates transformer‐based representations of drug compounds and protein sequences using ChemBERT and ProtBERT. Extracted embeddings are refined via proposed subnetworks and fused, then processed by a proposed deep feed‐forward neural network, enabling accurate DTI prediction.

1. Introduction

Drug discovery remains one of the most complex and significant challenges in pharmaceutical research and cheminformatics. A crucial step in the drug discovery process is the identification of drug target interactions (DTIs). However, the discovery of novel DTIs through experimental wet‐lab experiments is both time‐intensive and financially demanding (Hinkson et al. 2020). Traditional drug development pipelines typically require 10–17 years from initial discovery to market approval and involve substantial financial investment, often estimated at USD 2–3 billion per approved drug (Hinkson et al. 2020). Moreover, the failure rate is extremely high, with more than 90% of drug candidates failing during preclinical and clinical trials. Computational DTI prediction models have emerged as valuable tools for helping to prioritise drug target pairs, thereby accelerating target validation and reducing experimental cost. The high failure rate in traditional drug development underscores the need for innovative, cost‐efficient strategies that accelerate, improve and ensure the reliability of the drug discovery process (Rognan 2017).

Computational models accelerate DTI prediction by significantly reducing experimental cost and time (Rognan 2017). Accurate DTI prediction plays a crucial role in drug discovery, particularly in drug repurposing (Corbett et al. 2013). Drug repurposing refers to the identification of new therapeutic applications for existing drugs (Cai et al. 2023; Corbett et al. 2013). The drug repurposing process involves identifying new therapeutic applications of an existing drug, often transforming previously unintended effects into primary clinical uses. Pharmaceutical companies are increasingly prioritising rare and incurable diseases, as drug repurposing involves significantly lower initial development costs (Li et al. 2016; Tarín‐Pelló et al. 2023). Although the human proteome comprises over 20,000 proteins, only about 5000 are considered druggable, enabling rapid prioritisation of candidate drug target pairs and significantly reducing the time, cost and risk associated with new drug discovery. Integrating drug repurposing with computational approaches provides a faster and more cost‐effective strategy for identifying potential drug target interactions, thereby accelerating the drug repurposing process (Tarín‐Pelló et al. 2023; Wu et al. 2013).

2. Literature Review

2.1. Drug Target Interactions

Machine Learning (ML) and Deep Learning (DL) techniques are increasingly utilised for encoding protein and drug data and for classifying DTIs, offering acceptable solutions to the intricate problem of predicting DTIs in drug discovery (Simon and Bankapur 2023).

• Similarity‐based techniques predict DTIs by assuming that chemically or biologically similar drugs and targets exhibit similar interaction patterns (He et al. 2017).

• Feature‐based techniques use handcrafted drug and protein descriptors with classifiers to predict DTIs (Simon and Bankapur 2023; Monteiro et al. 2021; Roy 2004).

• Kernel‐based techniques employ similarity kernels to capture non‐linear relationships between drugs and targets (Pathak and Cai 2017).

Pathak et al. (Pathak and Cai 2017) introduced the KronRLS machine learning algorithm that applies regularised least squares by leveraging similarity scores between drugs and targets and utilises Kronecker product kernel regularisation to identify interactions. The KronRLS model utilises predefined similarity matrices, making it sensitive to similarity quality and limiting its ability to capture complex, non‐linear relationships, unable to learn representations directly from raw data and lacks generalisability over large datasets. Tong et al. (He et al. 2017) introduced SimBoost, a gradient‐boosting framework that integrates similarity‐based features to compute a final interaction score for DTIs. SimBoost leverages handcrafted and similarity‐based features, limiting generalisation and reducing performance on sparse, noisy, or unseen data.

Deep learning‐based DTI prediction models can be classified according to their underlying architectures, including Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Graph Neural Networks (GNNs) and attention‐based networks.

• CNN‐based models apply one‐dimensional convolutions to drug SMILES and protein targets to learn chemical substructures and amino acid motifs (Ozturk et al. 2018; Lee et al. 2019; Wallach et al. 2015).

• RNN‐based models process sequences sequentially to capture long‐term dependencies in drug and protein representations (Peng et al. 2024).

• GNN‐based models represent drugs as molecular graphs and targets as biological or interaction graphs, enabling the learning of topological embeddings that capture structural information (Tsubaki et al. 2019; Nguyen et al. 2021; Son and Kim 2021).

• Attention‐based models incorporate self‐attention or cross‐attention mechanisms to selectively focus on critical drug atoms and protein residues (Yazdani‐Jahromi et al. 2022; Nguyen et al. 2021; Huang et al. 2021).

Ozturk et al. (2018) introduced DeepDTA, which encodes protein sequences and drug SMILES as embeddings by applying a CNN to compute binding affinity along with DTI prediction. DeepDTA does not incorporate attention mechanisms, limiting its ability to focus on the most informative regions of drug and protein sequences and reducing interpretability, as well as its capacity to capture long‐range and context‐specific interactions.

Lee et al. (2019) presented DeepConvDTI, which uses a CNN that extracts evolutionary information from sequences of protein targets and drug SMILES without relying on predefined molecular descriptors. DeepConv‐DTI relies on convolutional architectures without attention mechanisms, limiting its ability to capture long‐range dependencies and context‐aware interactions between drug and protein sequences and reducing its effectiveness in identifying the most relevant regions for interaction.

Ragoza et al. proposed AtomNet (Wallach et al. 2015), which applied a 3D CNN using molecular structures as input, learned spatial and chemical features for DTI prediction, and identified therapeutic possibilities for untreatable diseases through drug repurposing. AtomNet relies on 3D structural information and convolutional operations without attention mechanisms, limiting its ability to capture long‐range and context‐aware interactions and reducing effectiveness when high‐quality structural data is unavailable.

Tsubaki et al. (2019) developed CPIGNN, which employs GNNs to generate molecular graphs from drug SMILES and sequential and structural graphs from protein sequences. CPI‐GNN relies on graph‐based representations without attention mechanisms, limiting its ability to capture global and context‐aware interactions between drugs and proteins, and making it sensitive to the quality and completeness of input graph structures.

Nguyen et al. (2021) developed GraphBAR and GraphDTA, which utilised advanced graph convolutional networks to generate molecular structures instead of traditional 1D SMILES to predict DTI. GraphBAR employed Bayesian learning by representing drugs and proteins as graphs, capturing both structural and sequential features. GraphBAR requires extensive preprocessing and is not scalable for increasing molecular complexity. GraphDTA utilised graph convolutional networks and graph attention networks to generate meaningful molecular embeddings by leveraging the structural properties of drugs. GraphDTA uses graph neural networks without explicit attention mechanisms, limiting its ability to capture long‐range and context‐aware interactions, and restricting its effectiveness in highlighting the most relevant substructures for drug target interactions.

Yazdani‐Jahromi et al. (2022) proposed AttentionSiteDTI, integrating binding‐site information with graph attentional features and self‐attention modules to improve model accuracy. AttentionSiteDTI employs localised attention mechanisms, which limit its ability to capture global and long‐range interaction patterns between drugs and targets, leading to incomplete modelling of overall drug target interactions. Huang et al. (2021) developed a transformer model, MolTrans, which segments molecules into substructures and applies self‐attention to capture interactions for DTI prediction using drug and protein sequences. MolTrans employs transformer‐based attention on sequence‐derived representations, but may struggle to fully capture complex structural and long‐range interaction patterns, as it relies solely on sequence‐based inputs without incorporating explicit molecular or 3D structural information.

Huang et al. (2021) developed a transformer model, MolTrans, which segments molecules into substructures and applies self‐attention to capture interactions for DTI prediction using drug and protein sequences. MolTrans employs transformer‐based attention on sequence‐derived representations, but may struggle to fully capture complex structural and long‐range interaction patterns, as it relies solely on sequence‐based inputs without incorporating explicit molecular or 3D structural information.

Zhou et al. (Cheng et al. 2022) proposed IIFDTI, which enhances DTI prediction by combining molecular graph features and sequence‐based embeddings to provide more biologically meaningful information, but it increases computational cost due to its complex feature extraction. IIFDTI relies on feature fusion and predefined representations, which limit its ability to capture complex and long‐range interaction patterns, and its performance can be sensitive to the quality of input features and similarity information. Wen et al. (Peng et al. 2024) proposed BINDTI, leveraging deep learning techniques, namely CNNs and Transformers, to generate highly accurate regression‐based affinity predictions along with DTI prediction. BINDTI relies heavily on high‐quality labelled data but has limited ability to capture complex interactions in noisy or incomplete datasets. BINDTI relies on predefined features and binding‐related representations, which limit its ability to capture complex and global interaction patterns; this highlights the importance of feature refinement, as insufficiently optimised features can reduce the model's ability to learn more informative and discriminative representations. Li et al. (Zeng et al. 2024) proposed CATDTI, a cross‐attention based framework that enables selective focus on relevant feature pairs and improves DTI prediction. CAT‐DTI employs attention‐based mechanisms but may still rely on predefined or coarse‐grained feature representations, limiting its ability to fully capture fine‐grained and complex interaction patterns. Zheng et al. (Bai et al. 2023) introduced DrugBAN, a network employing separate drug and protein encoders with bilinear attention to capture atom residue interactions. DrugBAN employs bilinear attention for interaction modelling, but depends on predefined feature representations, which can limit its ability to capture fine‐grained and complex interaction patterns. Additionally, it struggles with modelling long‐range dependencies, can be sensitive to input data quality, and requires substantial training data to achieve stable performance.

Peng et al. (Liu et al. 2024) presented NFSADTI, which integrates neural feature selection and attention for drug and protein feature extraction. NFSADTI relies on predefined features and random split evaluation, limiting generalisation to unseen drugs or targets and lacking explicit modelling of molecular interactions. NFSADTI leverages feature selection and attention mechanisms, but still depends on the quality of initial feature representations, limiting its ability to capture complex and fine‐grained interaction patterns; additionally, feature selection can lead to potential information loss, and the model may require substantial data for stable performance. Nguyen et al. (Wen et al. 2026) proposed LGABAN, which utilises a graph attention network that learns discriminative molecular representations by emphasising chemically important atoms and bonds, outperforming traditional descriptors and static GNN approaches. LGABAN employs graph attention and bilinear aggregation but still depends on the quality of initial feature representations, limiting its ability to capture fine‐grained and global interaction patterns and reliance on large‐scale data for stable performance.

Despite substantial progress in DTI prediction, ranging from sequence‐based deep learning models such as DeepConv‐DTI to attention and transformer‐based frameworks like MolTrans, BINDTI and CAT‐DTI, several critical challenges persist in learning robust and informative representations. Models like DeepConvDTI concentrate on feature extraction but miss out on refining features, hence the inclusion of noisy and less significant features. The methods are efficient in identifying the presence of local sequence patterns. Nonetheless, they have no mechanisms for modelling context and do not offer ways of selecting or denoising features. Models that apply transformer architectures like MolTrans perform well in modelling long‐range dependence and dependency contexts. Nevertheless, there is no guarantee that these models will consider noise‐aware feature extraction during the modelling process, hence the inclusion of redundant and irrelevant features due to the lack of an effective refinement method. Attention‐based architectures, such as BINDTI and CAT‐DTI, offer more efficient interaction modelling techniques based on cross‐modal attention techniques. However, these models lack strong feature selection and refinement mechanisms, thereby posing challenges when denoising noise in feature extraction. Recent models like NFSA‐DTI and LGABAN integrate feature selections and attention methods to solve some of the existing problems. However, despite these efforts, there are still difficulties associated with effective task‐specific feature refinement and multi‐level contextual feature learning. Addressing this gap is essential for developing robust, noise‐resistant and contextually rich representations for accurate drug target interaction prediction.

2.2. Drug Repurposing

Machine learning techniques have been extensively employed to computationally identify DTIs. Traditional drug repurposing approaches include similarity‐based methods that leverage chemical or biological resemblance, feature‐based machine learning models that use handcrafted descriptors with classical classifiers, and network‐based techniques that predict new associations by modelling drug target disease relationships as heterogeneous networks. Yang et al. (2014) designed a conditional random field model and trained it using a stochastic gradient contrastive divergence algorithm that identifies drug target associations by integrating genomic, chemical and pharmacological data for drug repurposing, for predicting antibacterial drugs by assessing functional similarities of compounds in infectious diseases. This approach relies on integrating heterogeneous handcrafted features from multiple data sources, making it sensitive to data quality and missing information, while lacking automated feature learning and refinement, which can limit its ability to capture complex and high‐level interaction patterns.

Deep learning techniques have become increasingly important for DTI prediction due to their ability to overcome several limitations of traditional ML approaches. Unlike ML models, which rely heavily on handcrafted drug and protein features, DL models are capable of learning multi‐level, task‐oriented representations directly from raw inputs such as SMILES strings, molecular graphs and protein sequences, thereby reducing dependence on manual feature engineering. Drug repurposing methods can be categorised into DNN‐based approaches that learn nonlinear relationships from integrated features, CNN‐based models that extract local patterns from drug representations, GNN‐based frameworks that capture relational dependencies using graph structures, and attention‐based methods that highlight biologically important drug and target features to enhance repurposability.

Monteiro et al. (2021) proposed a CNN‐based model to derive evolutionary information from custom protein sequences and drugs, which are integrated with a FCNN for DTI prediction, demonstrating improved performance over traditional descriptor‐based methods in drug repurposing. This approach relies on handcrafted and modality‐specific features with conventional ML/DL models, limiting its ability to capture deep and long‐range interaction patterns, while lacking feature refinement and being sensitive to data quality.

Iwata et al. (Berenstein et al. 2016) designed a DL model that utilised bioactive drug molecules and potential targets within the proteomes of pathogens. The work utilised molecular descriptors and depends on predefined features that fail to generalise to novel drugs. Furthermore, geometrical and physicochemical descriptors demand extensive computational resources, making them less scalable (Roy 2004). This method relies on similarity‐based and manually engineered features, limiting its ability to learn deep, non‐linear interaction patterns and making performance highly dependent on the quality and completeness of similarity information.

Mullins (2022) reviewed dedicated in silico virtual screening platforms for drug repurposing, covering disease areas from cancer, tuberculosis, to COVID‐19, and demonstrated that the combination of molecular docking with machine learning improves the identification of repurposing candidates. However, the platform‐level approach heavily depends on high quality three‐dimensional structural data for docking simulations, which limits its use in the absence of experimental protein structures, and does not make use of sequence‐based contextual representations for feature learning. Huang et al. (2024) proposed a foundation model for clinician‐centred drug repurposing by integrating heterogeneous biomedical knowledge graphs and deep learning to predict novel drug indications at scale. Although this approach works well for indication‐level repurposing, it is disease‐drug association focused, not modelling direct drug target interaction, limiting its ability to capture fine‐grained molecular binding patterns between specific drug compounds and protein targets.

Zhu et al. (2021) validated neuroinflammation through JAK2/STAT3 pathway validated JAK family as a target for repurposing of CNS‐active compounds. However, this work is experimental and does not provide a computational screening framework. This work relies solely on wet‐lab validation and does not provide a generalisable predictive model for identification of novel JAK‐targeting candidates from existing drug libraries. Wang et al. (2022) proposed a drug repurposing approach based on a high‐throughput screening and identified the FGFR inhibitor AZD4547 as a potent inhibitor of necroptosis by targeting the kinase RIPK1. This study further highlights that already existing kinase inhibitors can also be directed against off‐target kinases of independent therapeutic relevance. This approach requires experimental high‐throughput screening for candidate confirmation, a resource intensive process that does not scale to genome‐wide drug target pairs enumeration limiting its utility as a stand‐alone computational repurposing strategy.

Bansod et al. (2021) introduced pharmacological inhibition of DDR1 and DDR2 using imatinib suppressed TGF‐β1/Smad‐mediated pancreatic fibrosis in experimental chronic pancreatitis, establishing dual DDR1/DDR2 inhibition as a viable anti‐fibrotic repurposing strategy. This study, however, only utilises a single known DDR inhibitor, and there is no use of a predictive computational model to screen for novel DDR targeting compounds from broader drug libraries, which restricts the scope of candidate identification on pre‐selected molecules. Wang et al. (2025) proposed DTIBFAI which integrates sequence encoders based on ChemBERTa and BioBERT with molecular fingerprints and dipeptide composition features for DTI prediction and repurposing. A key limitation of DTIBFAI is that the extracted features are directly forwarded for interaction prediction without an intermediate feature refinement process. This preserves redundant or noisy information and limit the discriminative capability of the learned representations.

Even with these advancements, many drug repurposing methods still depend on descriptor‐based features or custom‐trained models for extracting drug and protein representations. This creates practical challenges, as feature extraction often needs to be repeated for different datasets, increasing computational cost and limiting scalability. As a result, these approaches tend to perform well only on specific datasets and struggle to generalise to new or unseen drug target pairs. Many existing models rely on sequence or similarity‐based features that often contain irrelevant or overlapping information, which can degrade predictive performance. The refinement of features is an important process in improving the quality of learned representations through noise reduction, redundancy removal and accentuation of the most relevant signals present within a diverse collection of biological information. Feature refinement helps models learn to recognise only those patterns that are of biological importance and helps capture drug target associations effectively and accurately. All of these issues collectively point towards a need for scalable and generalisable methods for modelling both chemical and biological context for DTIs.

The main contributions of this study are as follows:

• Variable‐length drug and protein sequences are encoded into fixed‐length contextual representations using BERT models.

• A novel ResNet‐based subnetwork is proposed for feature refinement to preserve key biochemical characteristics of drugs and proteins.

• A deep feedforward neural network is proposed to predict drug target interactions using refined drug and protein representations.

• The proposed framework was extensively evaluated across four benchmark DTI datasets with systematic comparison to state‐of‐the‐art methods.

• Paired t‐tests confirm significant improvements of the proposed framework over state‐of‐the‐art models.

• To demonstrate the practical applicability, the proposed framework was applied to screen previously unobserved drug target pairs. Candidate interactions were ranked and the top‐ranked pairs were identified as potential repurposing candidates. The results demonstrate the framework's ability to uncover biologically relevant drug target associations.

The rest of the article is arranged as follows: Section 3 details the datasets used, which consist of DTI and drug repurposing datasets. Section 4 elaborates on the methodology employed in this study, highlighting feature extraction, feature refinement using subnets and drug target interaction prediction. Section 5 highlights the experimental setup and discusses model performance in detail in the results section, including statistical significance analysis, followed by drug repurposing validation. Finally, Section 6 summarises the study and highlights possible future work.

3. Dataset Description

To evaluate the proposed model, we utilise two sets of datasets: one for DTI and another for drug repurposing.

3.1. Drug Target Interaction Dataset

Four publicly available benchmark datasets, namely DrugBank (Wishart et al. 2006), BindingDB (Gilson et al. 2016), Caenorhabditis elegans (Burns et al. 2015) and GPCR (Zhang et al. 2024), are used in this study. Each dataset consists of drug and protein sequence pairs along with interaction information. Table 1 presents the statistics of the four benchmark DTI datasets.

TABLE 1.

Characteristics of the benchmark datasets used for drug target interaction prediction.

Dataset	Drug compound count (Min/Max Len)	Target count (Min/Max Len)	DTI pair count (Pos/Neg)
DrugBank (Wishart et al. 2006)	6645 (11/1997)	4254 (5/1695)	35022 (17501/17511)
C. elegans (Burns et al. 2015)	1767 (54/13100)	1876 (1/252)	7786 (3893/3893)
BindingDB (Gilson et al. 2016)	43155 (93/4684)	757 (6/225)	50149 (28237/21912)
GPCR (Zhang et al. 2024)	5359 (98/1212)	346 (17/429)	13806 (7237/6569)

3.2. Drug Repurposing Dataset

The drug repurposing dataset consists of molecular descriptors of proteins and drugs, along with interaction information. The benchmark DTI dataset was compiled from DrugBank, Yamanishi et al., BioLiP and BindingDB, resulting in balanced positive and negative interaction pairs after redundancy removal and sequence similarity filtering (< 1% above 0.85). Protein sequences were obtained from UniProt, while drug structures were represented using canonical SMILES from PubChem and related databases. Molecular representations were constructed using CTD descriptors for proteins and a charge bioactivity threshold of 10 nM to capture weak interactions and structural properties. The detailed description about the dataset is available in Tables S1 and S2. Protein sequences were obtained from UniProt, while drug structures were represented using canonical SMILES from PubChem and related databases. Molecular representations were constructed using CTD descriptors for proteins and a charge bioactivity threshold of 10 nM to capture weak interactions and structural properties. The detailed description of this dataset, about protein and drug, is available in the Supporting Information under the section “Drug Repurposing Dataset Description”.

Unlike conventional DTI datasets that primarily focus on learning interactions for specific targets, this dataset adopts a polypharmacological framework in which a single drug interacts with multiple protein targets across diverse biological pathways. This includes the complex human diseases, particularly cancer, neurodegenerative disorders, inflammatory diseases and metabolic syndromes, where therapeutic outcomes depend on modulating interconnected molecular networks rather than a single target.

4. Methodology

The proposed model comprises two major components. The first component focuses on representation learning, where pre‐trained ProtBERT (Elnaggar et al. 2021) and ChemBERT (Chithrananda et al. 2020) models are employed to derive contextual embeddings of protein and drug sequences, respectively. These embeddings are subsequently refined through the proposed dedicated subnetwork to enhance discriminative feature refinement. The second component consists of the proposed deep feed‐forward neural network (DFFNN), which performs drug target interaction classification based on the concatenated refined representations. A detailed description of the implementation is provided in the subsequent subsections, along with a comprehensive architectural visualisation in Figure 1.

FIGURE 1.

The proposed model framework consisting of ProtBERT and ChemBERT along with feature subnets with DFFNN classifiers.

4.1. Representation Learning

4.1.1. Feature Extraction

Bidirectional Encoder Representations from Transformers (BERT) is a transformer‐based model built from stacked encoder blocks with residual connections, normalisation and multi‐head self‐attention, generating contextual embeddings for input sequences. It relies on positional encoding to capture the order of elements, self‐attention to model relationships across the sequence, and multi‐head attention to learn diverse features in parallel. These components enable BERT to effectively understand long‐range dependencies and contextual relationships in sequence data. BERT is trained using masked language modelling, in which randomly selected tokens in a sequence are masked and the model learns to predict them by leveraging the contextual information from surrounding tokens. BERT has two variants, namely, the large model and the base model. We used the ProtBERT (Elnaggar et al. 2021) large model for protein sequences to extract contextual embeddings and the ChemBERT (Chithrananda et al. 2020) base model for SMILES of drug compounds. The pre‐trained ProtBERT and ChemBERT models effectively capture bidirectional contextual information covering long‐range dependencies in the sequences. The attention mechanism assigns feature importance, preserving contextual information within the sequences.

ProtBERT is a large variant model of BERT consisting of 420 million parameters that were trained over 210 million proteins from the UniProtKB/UniRef90 database (Suzek et al. 2007). ProtBERT consists of sequences of 24 encoder layers, employed with 16 attention heads in each encoder layer, and each head has a dk/v (key/value dimension) of 64. The ProtBERT large architecture features a model dimension of 1024, enabling deep contextual learning of protein sequences through the multi‐head attention mechanism across sequences of encoder layers to capture long‐range dependencies. ProtBERT generates variable‐length protein sequences of input into fixed contextual embeddings of length 1024.

The ChemBERT base model consists of approximately 44 million parameters, trained on the 110 million ZINC database, and enables learning rich molecular representations from SMILES strings. The basic ChemBERT model has 12‐headed attention in each of the 12 layers in the encoder. The ChemBERT base architecture features a feed‐forward dimension of 3072 and a model dimension of 768. Pre‐training ChemBERT on diverse chemical‐generating activities are robust in detecting and analysing chemical information. ChemBERT generates fixed‐size 768‐dimensional structural information from the variable size representations of compounds.

4.1.2. Feature Refinement Using ResNet‐Based Subnetworks

Both protein and drug representations are processed through a novel, proposed dedicated ResNet‐based subnetwork for modality‐specific refinement. The proposed protein subnet maps high‐dimensional ProtBERT contextual representations through two dense layers, progressively reducing dimensionality from 1024 to 256, with batch normalisation and a projection‐based residual connection to preserve representational integrity while enabling efficient feature refinement. Similarly, the proposed ChemBERT subnet maps chemical contextual representations through two dense layers, progressively reducing dimensionality to 128, with batch normalisation and dropout to enhance generalisation while supporting effective feature refinement. All dense layers in the proposed subnets adopt He‐normal initialisation to maintain stable activation dynamics during training, consistently applied across both ProtBERT and ChemBERT subnets to facilitate faster convergence and improved gradient stability.

The proposed protein subnet processes high dimensional ProtBERT embeddings through two successive dense blocks. The first block projects the input to 512 units, followed by batch normalisation to improve training stability and reduce overfitting, while the second block further reduces the dimensionality to 256 units. A projection‐based residual connection aligns the 512‐dimensional intermediate output to 256 dimensions before being added to the second block output.

Similarly, the proposed ChemBERT subnet begins with a 256‐unit dense layer coupled with batch normalisation and dropout to capture informative sequence‐level chemical features, again normalised and activated, with a projection‐based residual connection mapping the 128‐unit features into the 128‐dimensional space to enrich semantic drug representations.

A projection layer is a linear transformation placed at the subnet to convert high‐dimensional input features into a compact latent space. In DTI models, it serves to align heterogeneous features such as protein sequences and drug representations, reduce dimensionality, stabilise optimisation, and control model complexity before deeper nonlinear feature learning.

The raw input feature vector of a subnet is:

$$x\in {ℝ}_{\mathit{\text{in}}}^d$$ (1)

where x_p represents the protein feature embedding and x_d represents the drug feature embedding.

The projection layer applies a linear transformation:

$$\begin{array}{c}z=\mathit{Px}+b\end{array}$$ (2)

where P ∈ ℝ^(d_proj×d_in) is the learnable projection matrix, b ∈ ℝ^d_proj is the bias term, and z ∈ ℝ^d_proj represents the projected feature vector.

The protein representation from subnet 1 (256 units) and the chemical representation from subnet 2 (128 units) are concatenated to form a unified 384‐dimensional fused embedding. This fusion enables the downstream DFFNN to learn complex interactions and relationships between drugs and their targets, ultimately improving DTI prediction accuracy.

4.2. Drug Target Interaction Classification

The refined feature sets of protein and drug subnets were fed into the proposed DFFNN for classification. The proposed DFFNN classifier model consists of an input layer, three hidden layers and an output layer. The model has three hidden layers with 256, 128 and 64 neurons, each using ReLU activation. The model is trained for up to 1500 epochs with a batch size of 128. Class weighting is applied to handle class imbalances during training.

The ReLU activation function is:

$$\begin{array}{c}f\left(x\right)=\max \left(0,x\right)\end{array}$$ (3)

here x is the input to the activation function: if x > 0, then f(x) = x (linear behaviour for positive values); if x ≤ 0, then f(x) = 0 (zeroes out negative values).

The model is compiled using the Adam optimiser, which adapts learning rates for efficient convergence, and sparse categorical cross‐entropy as the loss function suited for integer‐labelled classification tasks.

The sparse categorical cross‐entropy loss function is:

$$\begin{array}{c}\text{Loss}\left(H\right)=-\left(\frac{1}{N}\right)\Sigma y_i\log \left({ŷ}_i\right)+\left(\frac{\lambda }{2}\right){‖H‖}^{22}\end{array}$$ (4)

where λ is the regularisation coefficient, y_i is the real label, ŷ_i is the predicted label, and ‖H‖² ₂ is the L2 norm of the model parameters.

The final output layer contains 2 neurons with a softmax activation function:

$$\begin{array}{c}\sigma \left(z_i\right)=\frac{e_i^z}{\Sigma e_j^z}\end{array}$$ (5)

where zᵢ represents the input to the i‐th neuron, eᶻⁱ denotes the exponential function applied to zᵢ, N is the total number of neurons in the output layer, and ∑_j₌₁ᴺ eᶻʲ represents the sum of the exponentials of all output neuron inputs. This normalisation ensures that the output values are in the range of 0 to 1 and that their sum equals 1, allowing them to be interpreted as probabilities.

Pre‐trained BERT models are employed to capture contextual relationships within drug and protein sequences for drug target interaction prediction. The extracted representations are subsequently processed through dedicated subnetworks and then forwarded to the proposed deep feed‐forward neural network (DFFNN) for final classification. Accordingly, the proposed framework is hereafter referred to as PCBERT‐DFFNN‐DTI throughout this article.

5. Experiments and Results

5.1. Experimental Setup

Feature extraction was performed on an NVIDIA DGX A100 AI server equipped with eight NVIDIA A100 Tensor Core GPUs, 320 GB of GPU memory, and 1 TB of system memory RAM. The training and evaluation of the proposed DFFNN model were conducted on the Google Cloud Platform. The DFFNN architecture was implemented using TensorFlow within a Python 3.11.0 environment. Detailed descriptions of the evaluation metrics are provided in the Supporting Information under the section “Evaluation Metrics.” Each dataset was divided into 80% training, 10% validation and 10% testing for model development and evaluation. Additionally, 5‐fold cross‐validation was performed to assess the robustness and generalisation ability of the proposed framework.

5.2. Ablation Study on the Feature Refinement Module

To evaluate the contribution of the proposed feature refinement module, an ablation study was conducted by removing the ResNet‐based feature refinement subnetworks and directly utilising the pretrained ProtBERT and ChemBERT embeddings for DTI prediction. This experiment was designed to quantify the impact of feature refinement on the overall predictive performance of the PCBERT‐DFFNN‐DTI framework. Table 2 presents the performance of the model without the feature refinement module across the four benchmark datasets. The results indicate a consistent decline in performance compared with the complete PCBERT‐DFFNN‐DTI framework. Specifically, the ablated model achieved accuracies of 0.968, 0.828, 0.946 and 0.848 on the C. elegans, GPCR, BindingDB and DrugBank datasets, respectively. Similar reductions were observed for Precision, Recall, F1‐score, AUROC and AUPRC metrics.

TABLE 2.

Ablation study of the ResNet‐based FeatureRefinement module.

Dataset	Model	Accuracy	F1‐Score	AUROC	AUPRC
C. elegans	Without refinement	0.968	0.968	0.986	0.988
	PCBERT‐DFFNN‐DTI	0.971	0.970	0.991	0.993
GPCR	Without refinement	0.828	0.839	0.895	0.889
	PCBERT‐DFFNN‐DTI	0.831	0.840	0.907	0.898
BindingDB	Without refinement	0.946	0.952	0.981	0.981
	PCBERT‐DFFNN‐DTI	0.947	0.953	0.980	0.977
DrugBank	Without refinement	0.848	0.850	0.918	0.912
	PCBERT‐DFFNN‐DTI	0.859	0.861	0.916	0.907

Note: Bold values indicate the best performance.

The ablation study confirms the effectiveness of the feature refinement module in PCBERT‐DFFNN‐DTI. The complete model achieved higher performance on C. elegans with an accuracy of 0.971 and AUROC of 0.991, compared to 0.968 and 0.986 without refinement. Similar improvements were observed on GPCR, where accuracy increased to 0.831 and AUROC to 0.907, and on DrugBank, where accuracy and F1‐score improved to 0.859 and 0.861, respectively. On BindingDB, the full model maintained comparable performance to the ablated version. Overall, these results demonstrate that the refinement module improves the quality of fused drug–protein representations, resulting in more accurate and robust drug–target interaction predictions.

5.3. Results and Interpretation

To evaluate the proposed PCBERT‐DFFNN‐DTI model for DTI prediction, we compared its performance with state‐of‐the‐art (SOTA) models across two experimental settings, E1 and E2. In E1, the dataset was split into training, validation and test sets at 8:1:1 using a fixed random seed of 42, whereas E2 employed a 5‐fold cross‐validation strategy.

5.3.1. Result Analysis on E1

The performance of the proposed PCBERT‐DFFNN‐DTI model was assessed on four benchmark DTI datasets, namely DrugBank, C. elegans , GPCR and BindingDB. The results are tabulated in Table 2. The experimental setup follows the identical setup used in the IIFDTI published work (Cheng et al. 2022). The results of SOTA on all benchmark datasets are obtained from the IIFDTI published work, with (—) indicating unavailable results. From Table 3, the proposed PCBERT‐DFFNN‐DTI model consistently achieved superior performance across all evaluation metrics across three benchmark datasets, as clearly illustrated in Figure 2 with respect to AUC, AUPRC and F1‐score performance metrics.

TABLE 3.

Comparative performance analysis of the proposed PCBERT‐DFFNN‐DTI model against state‐of‐the‐art models.

Dataset	Model (Year)	AUC	AUPRC	Prec.	Recall	Accu.	F1
DrugBank	DeepDTA (Ozturk et al. 2018)	0.871	0.870	0.786	0.798	—	0.792
	DeepConv‐DTI (Lee et al. 2019)	0.836	0.831	0.736	0.767	—	0.751
	CPI‐GNN (Tsubaki et al. 2019)	0.695	0.734	—	—	0.659	0.609
	TransformerCPI (Chen et al. 2020)	0.865	0.868	0.774	0.821	—	0.796
	MolTrans (Huang et al. 2021)	0.862	0.862	0.809	0.767	—	0.787
	IIFDTI (Cheng et al. 2022)	0.923	0.930	0.854	0.860	—	0.857
	PCBERT‐DFFNN‐DTI (This Study)	0.916	0.907	0.847	0.875	0.859	0.861
C. elegans	DeepDTA (Ozturk et al. 2018)	0.983	0.984	0.970	0.960	—	0.965
	DeepConv‐DTI (Lee et al. 2019)	0.983	0.985	0.954	0.936	—	0.945
	TransformerCPI (Chen et al. 2020)	0.984	0.983	0.949	0.948	—	0.948
	MolTrans (Huang et al. 2021)	0.982	0.982	0.971	0.963	—	0.967
	IIFDTI (Cheng et al. 2022)	0.991	0.992	0.954	0.971	—	0.962
	PCBERT‐DFFNN‐DTI (This Study)	0.991	0.993	0.982	0.959	0.971	0.970
BindingDB	DeepDTA (Ozturk et al. 2018)	0.934	0.934	0.858	0.860	—	0.859
	DeepConv‐DTI (Lee et al. 2019)	0.922	0.921	0.835	0.846	—	0.840
	CPI‐GNN (Tsubaki et al. 2019)	0.608	0.517	—	—	0.597	0.373
	TransformerCPI (Chen et al. 2020)	0.933	0.934	0.840	0.891	0.811	0.865
	MolTrans (Huang et al. 2021)	0.899	0.897	0.826	0.768	0.864	0.796
	IIFDTI (Cheng et al. 2022)	0.944	0.945	0.879	0.873	—	0.876
	CAT‐DTI (Zeng et al. 2024)	0.960	0.947	—	0.884	0.896	—
	DrugBAN (Bai et al. 2023)	0.960	0.948	—	0.900	0.904	—
	NFSA‐DTI (Liu et al. 2024)	0.965	0.957	—	0.908	0.907	—
	LGABAN (Wen et al. 2026)	0.966	0.953	—	0.902	0.910	—
	PCBERT‐DFFNN‐DTI (This Study)	0.980	0.977	0.950	0.956	0.947	0.953
GPCR	DeepDTA (Ozturk et al. 2018)	0.776	0.762	0.713	0.712	—	0.712
	DeepConv‐DTI (Lee et al. 2019)	0.752	0.685	0.695	0.713	—	0.704
	TransformerCPI (Chen et al. 2020)	0.842	0.837	0.755	0.796	—	0.775
	MolTrans (Huang et al. 2021)	0.807	0.788	0.699	0.762	—	0.729
	IIFDTI (Cheng et al. 2022)	0.845	0.842	0.766	0.783	—	0.774
	PCBERT‐DFFNN‐DTI (This Study)	0.907	0.898	0.840	0.839	0.831	0.840

Note: Bold values indicate the best performance. Underlined F1 scores are computed from reported precision and recall.

FIGURE 2.

Performance comparison of the proposed PCBERT‐DFFNN‐DTI model with SOTA models across four datasets in the E1 setting.

From Table 3, it is evident that the proposed PCBERTDFFNN‐DTI model demonstrates competitive and often superior performance across multiple evaluation metrics compared to existing methods. Although the proposed model reports slightly lower AUC and AUPRC on the DrugBank dataset compared to certain SOTA methods, the proposed model reported an F1 score of 0.861, which outperforms the best model from the literature, IIFDTI, and earlier deep learning approaches such as DeepDTA, DeepConv‐DTI, MolTrans and TransformerCPI. While IIFDTI reports a slightly higher precision, the proposed model achieves a better balance between precision and recall, resulting in stable and reliable classification performance on this large and complex dataset.

On the C. elegans dataset, the proposed PCBERTDFFNN‐DTI model achieved an AUC of 0.991, an AUPRC of 0.993, and the highest F1 score of 0.970 among all SOTA methods. These results indicate a clear improvement over MolTrans and IIFDTI, demonstrating the effectiveness of the proposed architecture even on relatively smaller datasets. The strong generalisation capability suggests that the learned representations are robust and well‐suited to datasets with limited sample sizes.

The proposed model reported an accuracy of 0.980, an AUC of 0.947, and an AUPRC of 0.953, indicating strong overall discrimination and robust performance on the large‐scale and highly imbalanced DTI data. Among the SOTA methods, LGABAN reports the highest performance with an accuracy of 0.910, AUC of 0.966 and AUPRC of 0.953. Compared to these, PCBERT‐DFFNNDTI improves AUC by 1.4%, AUPRC by 2.4% and accuracy by 3.7%, demonstrating consistent gains over the baselines.

Finally, the proposed model significantly outperforms all existing SOTA models across all reported metrics for the GPCR dataset. It achieved an AUC of 0.907, AUPRC of 0.898 and an F1 score of 0.840, representing a substantial improvement over the previous best‐performing model, IIFDTI and TransformerCPI. This notable performance gain highlights the model's ability to effectively capture complex DTI patterns in GPCR.

5.3.2. Discussion

The proposed PCBERT‐DFFNN‐DTI model performed effectively on benchmark datasets such as C. elegans , BindingDB and GPCR, but shows comparatively lower performance on the DrugBank dataset in terms of AUC, AUPRC and precision. This performance gap should not be interpreted as a limitation of the model, but rather as a consequence of the intrinsic complexity of DrugBank and the trade‐offs involved in modelling such a diverse interaction space. In particular, the model must balance between learning generalised representations and capturing rare, dataset‐specific patterns, which becomes significantly more challenging in highly heterogeneous settings like DrugBank. DrugBank is characterised by an extremely large and heterogeneous chemical and biological space, with 6645 drugs (sequence length range: 11–1997) and 4254 protein targets (length range: 5–1695). This wide variability leads to a highly non‐uniform feature distribution, where a few drugs and proteins are well represented while many remain rare or unique. As a result, under a fixed 8:1:1 split (E1), the test set is more likely to contain underrepresented or unseen entities, giving rise to severe cold‐start scenarios. In such cases, the model has no prior knowledge of these drugs or proteins during training, limiting its ability to learn meaningful interaction patterns and reducing generalisation performance. The cold‐start problem is further influenced by the heterogeneous nature of DrugBank, where proteins span diverse functional classes with minimal overlap. Since DTI models rely on shared statistical and relational patterns between known drug target pairs, the absence of sufficient interaction neighbours for rare entities results in sparse and less informative embeddings. Consequently, predictions for unseen pairs often depend on superficial similarities rather than robust learned relationships. In contrast, datasets like BindingDB are characterised by greater redundancy and a more condensed interaction space, with similar drugs and proteins appearing repeatedly across samples. This redundancy reduces the impact of cold‐start issues, since unseen pairs often have closely similar examples in the training set, allowing the model to generalise more effectively.

5.3.3. Result Analysis on E2

Table 4 presents the performance comparison of PCBERT‐DFFNN‐DTI and five baseline models using 5‐fold cross‐validation. The proposed model achieved the highest AUC, AUPRC, accuracy and F1 score on both BindingDB and DrugBank, outperforming the baselines substantially, as illustrated in Figure 3.

TABLE 4.

Performance comparison of various SOTA models on BindingDB and DrugBank datasets under the E2 setting.

Dataset	Model	AUC	AUPRC	Precision	Recall	Accuracy	F1 score
DrugBank	CPI‐GNN (Tsubaki et al. 2019)	0.724 ± 0.080	0.767 ± 0.062	—	—	0.678 ± 0.061	0.650 ± 0.077
	TransformerCPI (Chen et al. 2020)	0.872 ± 0.004	0.877 ± 0.007	—	—	0.792 ± 0.007	0.799 ± 0.009
	MolTrans (Huang et al. 2021)	0.870 ± 0.008	0.879 ± 0.009	—	—	0.789 ± 0.010	0.798 ± 0.010
	IIFDTI (Cheng et al. 2022)	0.874 ± 0.003	0.876 ± 0.006	—	—	0.783 ± 0.012	0.806 ± 0.009
	BINDTI (Peng et al. 2024)	0.888 ± 0.006	0.888 ± 0.010	—	—	0.822 ± 0.007	0.825 ± 0.004
	PCBERT‐DFFNN‐DTI (This Study)	0.918 ± 0.005	0.920 ± 0.004	0.830 ± 0.010	0.862 ± 0.006	0.846 ± 0.005	0.846 ± 0.005
BindingDB	CPI‐GNN (Tsubaki et al. 2019)	0.680 ± 0.150	0.618 ± 0.176	—	—	0.647 ± 0.130	0.503 ± 0.215
	TransformerCPI (Chen et al. 2020)	0.940 ± 0.003	0.924 ± 0.004	—	—	0.874 ± 0.007	0.854 ± 0.008
	MolTrans (Huang et al. 2021)	0.941 ± 0.005	0.923 ± 0.010	—	—	0.876 ± 0.011	0.856 ± 0.015
	IIFDTI (Cheng et al. 2022)	0.927 ± 0.007	0.905 ± 0.010	—	—	0.836 ± 0.012	0.826 ± 0.011
	BINDTI (Peng et al. 2024)	0.965 ± 0.004	0.951 ± 0.004	—	—	0.913 ± 0.008	0.914 ± 0.006
	PCBERT‐DFFNN‐DTI (This Study)	0.985 ± 0.001	0.985 ± 0.001	0.953 ± 0.005	0.961 ± 0.002	0.951 ± 0.002	0.957 ± 0.002

Note: Values are mean (±std). Bold values indicate the best performance.

FIGURE 3.

Performance comparison of the proposed PCBERT‐DFFNN‐DTI model with SOTA models under the E2 setting.

From Table 4, it is evident that the proposed PCBERTDFFNN‐DTI model consistently achieved the best results across all evaluation metrics on both datasets. On the DrugBank dataset, PCBERT‐DFFNN‐DTI reached the highest AUC of 0.918 ± 0.005 and AUPRC of 0.920 ± 0.004. This performance surpasses the strongest baseline, BINDTI, which reports an AUC of 0.888 ± 0.006 and an AUPRC of 0.888 ± 0.010. The proposed model also shows better classification performance with an accuracy of 0.846 ± 0.005 and an F1 score of 0.846 ± 0.005. This indicates a clear improvement over all baseline models, including transformer‐based approaches like MolTrans and IIFDTI. These results show that the PCBERT‐DFFNNDTI model offers a better balance between precision and recall, which leads to more reliable predictions in cross validation settings. The performance gains are clearly illustrated in Figure 3.

On the BindingDB dataset, the proposed PCBERTDFFNN‐DTI model shows an even more pronounced improvement over existing method. The proposed model reported the highest AUC of 0.985 ± 0.001 and AUPRC of 0.985 ± 0.001, significantly surpassing the best‐performing baseline, BINDTI, which reports an AUC of 0.965 ± 0.004 and AUPRC of 0.951 ± 0.004. In addition, PCBERTDFFNN‐DTI attains an accuracy of 0.951 ± 0.002 and an F1 score of 0.957 ± 0.002, highlighting its near‐perfect classification capability on this large and highly imbalanced dataset. The substantial improvements across all metrics, as visualised in Figure 3, demonstrate the robustness of the proposed PCBERT‐DFFNN‐DTI model.

5.3.4. Discussion

The results presented in Tables 3 and 4 demonstrate that the proposed PCBERT‐DFFNN‐DTI model consistently outperforms existing state‐of‐the‐art (SOTA) methods under both E1 and E2 evaluation settings. The proposed framework achieves strong and stable performance across the DrugBank and BindingDB datasets, indicating its effectiveness in learning informative drug and protein representations for DTI prediction. The consistent improvements observed across all evaluation metrics further highlight the robustness and generalisation capability of the proposed architecture. Under the E2 setting, which employs 5‐fold cross‐validation, the evaluation becomes more reliable by reducing the dependence of the results on a single train–test partition. By averaging performance across multiple folds, this protocol provides a more comprehensive assessment of model behaviour under different data distributions and minimises the influence of partition‐specific variations. Consequently, the reported performance better reflects the model's overall predictive capability and stability.

For DrugBank, which contains a large and highly heterogeneous interaction space, different folds exhibit varying levels of complexity due to the diversity of drugs, proteins and interaction patterns. Therefore, averaging the results across all folds provides a more balanced and reliable estimate of the model's generalisation performance. In contrast, BindingDB contains a relatively denser interaction space with substantial similarity among many drug target pairs. As a result, the model demonstrates consistently strong performance across different folds, indicating stable learning behaviour and good predictive reliability. Overall, the E2 evaluation protocol offers a rigorous assessment of the model's generalisation capability by examining performance across multiple train–test partitions. The superior results achieved by PCBERT‐DFFNN‐DTI under this setting demonstrate its robustness, stability and effectiveness for large‐scale drug target interaction prediction.

While the proposed PCBERT‐DFFNN‐DTI framework achieved strong performance across multiple benchmark datasets and an independent external dataset, the current study did not include dedicated unseen‐drug or unseen‐target evaluation protocols. Therefore, the reported results should be interpreted as evidence of predictive generalisation under standard experimental settings rather than robustness to strict cold‐start scenarios. Future work will investigate the performance of the proposed framework under unseen‐drug, unseen‐target and unseen‐drug target split settings to further assess its transferability and practical applicability in real‐world drug discovery tasks.

5.4. Statistical Significance Analysis

The proposed PCBERT‐DFFNN‐DTI model demonstrated superior performance over SOTA methods on DrugBank, BindingDB and GPCR, and achieved comparable improved results on the C. elegans dataset, indicating strong generalisation capability. To statistically validate these observed improvements, a paired t‐test was performed by comparing the F1 scores obtained under identical experimental settings. The analysis evaluated the paired differences in F1 score between the proposed model and each SOTA model across the benchmark datasets. Specifically, statistical tests for DeepDTA (Ozturk et al. 2018) and DeepConv‐DTI (Lee et al. 2019) were conducted using E1 settings based on the results reported in Table 2. In contrast, tests for CPI‐GNN (Tsubaki et al. 2019), TransformerCPI (Chen et al. 2020), MolTrans (Huang et al. 2021), and IIFDTI (Cheng et al. 2022) were performed under both E1 and E2 settings, leveraging the single‐run and repeated‐run evaluations presented in Tables 3 and 4. The comparison with BINDTI (Peng et al. 2024) was carried out exclusively under E2 settings as reported in Table 4.

The null hypothesis states that there is no statistically significant difference between the F1 scores achieved by the proposed PCBERT‐DFFNN‐DTI model and those obtained by SOTA DTI prediction methods. All statistical tests were conducted at a 5% significance level. As shown in Table 5, all comparisons yielded p‐values below 0.05, leading to the rejection of the null hypothesis in every case. These findings confirm that the observed improvements in F1 score are statistically significant and demonstrate the reliability and effectiveness of the proposed approach over existing SOTA models.

TABLE 5.

Paired t‐test comparison (F1 score).

	Setting	p	Significance
DeepDTA (Ozturk et al. 2018)	E1	0.0143	Yes
DeepConv‐DTI (Lee et al. 2019)	E1	0.0082	Yes
CPI‐GNN (Tsubaki et al. 2019)	E1/E2	1.8e−4	Yes
TransformerCPI (Chen et al. 2020)	E1/E2	1.1e−4	Yes
MolTrans (Huang et al. 2021)	E1/E2	3.0e−4	Yes
IIFDTI (Cheng et al. 2022)	E1/E2	2.5e−4	Yes
BINDTI (Peng et al. 2024)	E2	6.0e−5	Yes

5.5. Case Study: Drug Repurposing

An independent drug repurposing dataset with predefined training and testing splits was used to evaluate the generalisability of the proposed PCBERT‐DFFNN‐DTI framework. Unlike the benchmark DTI datasets employed in the previous experiments, this dataset provides multiple predefined feature representations, including (i) molecular descriptors (MD), (ii) sequence‐based chemical and protein features and (iii) hybrid feature combinations. Consequently, minor architectural adaptations were implemented to accommodate these alternative input modalities.

For experiments based on molecular descriptors (MD), the ProtBERT and ChemBERT encoding modules were bypassed because the input data consisted of precomputed numerical descriptor vectors rather than raw protein sequences and SMILES strings. The descriptor vectors were directly supplied to the downstream prediction network, where feature refinement and classification were performed by the DFFNN module. For sequence‐based experiments, the complete PCBERT‐DFFNN‐DTI architecture was employed, with protein sequences and SMILES strings processed through the ProtBERT and ChemBERT encoders to generate contextual embeddings. For the hybrid configuration, the descriptor‐based features and sequence‐derived representations provided by the benchmark dataset were concatenated into a unified feature vector and subsequently fed into the DFFNN classifier. Therefore, the independent dataset evaluation served as an external validation study under different feature representations while maintaining the same prediction framework.

This dataset facilitates in silico drug discovery using a polypharmacological framework, in which a single drug molecule can interact with multiple off‐target proteins. This aspect is particularly important for complex, multifactorial diseases such as cancer, neurodegenerative disorders and metabolic syndromes, where disease progression is driven by multiple interconnected biological pathways rather than alterations in a single molecular target. By capturing diverse drug target interaction patterns, the dataset enables computational models to learn cross‐target generalisation and infer underlying polypharmacological relationships, thereby supporting the identification of potential drug repurposing candidates. Comparative performance across the MD, sequence‐based, and hybrid feature configurations is summarised in Table 6.

TABLE 6.

Drug repurposing dataset results using different feature sets.

Features	Model	AUROC	AUPRC	Prec.	Spec.	Rec.	Acc.	F1
MD‐based	MD + SVM (Simon and Bankapur 2023)	0.915	0.918	—	0.983	0.720	0.883	0.824
	MD + RF (Simon and Bankapur 2023)	0.988	0.982	—	0.989	0.819	0.925	0.892
	MD + FCNN (Monteiro et al. 2021)	0.955	0.949	—	0.963	0.827	0.911	0.876
	MD + SVM (Monteiro et al. 2021)	0.854	0.810	0.905	—	0.870	0.890	0.887
	MD + RF (Monteiro et al. 2021)	0.881	0.912	0.940	—	0.920	0.930	0.930
	PCBERT‐DFFNN‐DTI (This Study)	0.989	0.989	0.989	0.989	0.989	0.989	0.989
Seq‐based	CNN + SVM (Monteiro et al. 2021)	0.965	0.958	—	0.993	0.765	0.906	0.861
	CNN + RF (Monteiro et al. 2021)	0.988	0.983	—	0.992	0.833	0.931	0.902
	CNN + FCNN (Monteiro et al. 2021)	0.966	0.960	—	0.961	0.861	0.923	0.895
	T5 + SVM (Simon and Bankapur 2024)	—	—	0.920	—	0.880	0.910	0.900
	T5 + RF (Simon and Bankapur 2024)	—	—	0.930	—	0.930	0.930	0.930
	PCBERT‐DFFNN‐DTI (This Study)	0.997	0.997	0.997	0.997	0.997	0.997	0.997
Hybrid	CNN + FCNN+AE (Monteiro et al. 2021)	0.972	0.966	—	0.948	0.880	0.922	0.896

Note: The symbol (—) indicates unavailable results. Bold values indicate the best performance.

The proposed model PCBERT‐DFFNN‐DTI outperformed the best model from SOTA RF‐based model by about 6% in F1 score when assessed using molecular descriptors, achieving superior performance across all metrics. Notably, PCBERT‐DFFNN‐DTI consistently outperformed the strongest sequence‐based baseline (T5 + RF) by 6.7% when utilising only sequence derived features (F1 = 0.997). BERT's bidirectional contextual representation learning and efficient fine‐tuning capabilities, which allow for richer feature extraction straight from raw sequences, are responsible for this improvement.

To assess the practical implications of this capability, a large‐scale repurposing analysis was performed for Gabapentin. Using the same polypharmacology‐driven repurposing framework, the trained model screened previously unseen drug target combinations and identified several high‐confidence targets spanning neurological, inflammatory, oncological and metabolic pathways as summarised in Table 7. Many predicted targets, including GLRA1, HRH1, CXCR1 and adrenergic receptors, are closely associated with pain modulation, neuroinflammation and sensory signalling, supporting Gabapentin's established therapeutic role in neuropathic disorders. In addition, the model identified potential interactions with FGFR2, FGFRL1, IL2RA, CYP19A1 and GnRHR, suggesting opportunities in oncology and cancer‐supportive care. Notably, the prediction of CYP19A1 is consistent with reported clinical use of Gabapentin for managing aromatase inhibitor–associated symptoms in breast cancer patients. As shown in Table 7, the analysis also highlighted less‐explored targets, including TASK‐1, TASK‐2 and ABCB6, indicating potential applications in pulmonary hypertension, renal disorders, and mitochondrial dysfunction. Although these predictions require experimental validation, they demonstrate the model's ability to uncover previously unrecognised therapeutic relationships.

TABLE 7.

Top 10 predicted Gabapentin targets identified by PCBERT‐DFFNN‐DTI.

Target	Therapeutic area	Biological relevance
GLRA1 (Glycine receptor alpha‐1)	Neurology	Inhibitory neurotransmission and pain processing
HRH1 (Histamine H1 receptor)	Neurology/Dermatology	Sensory signalling and itch modulation
CXCR1 (IL‐8 receptor A)	Neuroinflammation	Inflammatory signalling and immune response
PTGS1 (COX‐1)	Inflammation	Prostaglandin‐mediated inflammatory pathways
FGFR2	Oncology	Tumour growth and growth‐factor signalling
FGFRL1	Fibrosis/Angiogenesis	Tissue remodelling and angiogenic regulation
IL2RA (CD25)	Immunology/Oncology	T‐cell activation and immune regulation
CYP19A1 (Aromatase)	Breast cancer	Oestrogen biosynthesis and hormonal regulation
GnRHR	Endocrinology/Oncology	Hormone‐dependent signalling pathways
KCNK3 (TASK‐1)	Pulmonary hypertension	Pulmonary vascular homeostasis

Overall, the Gabapentin case study provides a practical example of how the superior cross‐target generalisation achieved by PCBERT‐DFFNN‐DTI can be translated into actionable drug repurposing hypotheses. The diverse target profile predicted for a single approved drug highlights the effectiveness of the proposed framework in exploiting polypharmacological patterns for in silico drug discovery.

6. Conclusion

This study demonstrates the potential of artificial intelligence in accelerating drug discovery and repurposing through accurate drug target interaction (DTI) prediction. The proposed PCBERT‐DFFNN‐DTI framework leverages large‐scale pretraining on protein and chemical sequences to capture complex contextual and long‐range dependencies. Dedicated subnetworks further refine modality‐specific embeddings, preserving intrinsic biochemical characteristics while reducing redundancy and noise prior to classification. The proposed model was extensively evaluated against SOTA models under two experimental settings (E1 and E2) across four benchmark datasets: DrugBank, C. elegans , BindingDB and GPCR. The results show consistent performance improvements across multiple evaluation metrics. Notable gains were observed on challenging datasets such as GPCR and BindingDB, while competitive or superior performance on C. elegans and DrugBank highlights strong generalisation capability. Statistical validation using paired t‐tests at a 5% significance level confirmed that the observed F1‐score improvements over SOTA methods are statistically significant, reinforcing the reliability and robustness of the proposed framework. To further assess generalisability, the model was evaluated on an independent drug repurposing dataset. This study demonstrates that leveraging a multi‐source drug repurposing dataset enables effective in silico polypharmacological modelling of drug target interactions. The proposed approach achieved substantial improvements over hybrid and descriptor‐based models, even without relying on handcrafted molecular descriptors, demonstrating scalability and practical applicability in real‐world drug repurposing scenarios. By capturing complex cross‐target relationships, the proposed approach supports network‐level pharmacological inference and facilitates the identification of promising drug repositioning candidates with improved therapeutic efficacy, particularly for multifactorial diseases. Overall, PCBERT‐DFFNN‐DTI establishes a robust, statistically validated and generalisable framework for DTI prediction and drug repurposing.

Author Contributions

Sanjay Bankapur: investigation, supervision, resources. Essmily Simon: conceptualization, investigation, validation, writing – original draft.

Funding

The Google LLC Research Credits Program provided partial funding (Award GCP19980904).

Ethics Statement

No animals or humans were involved in this research.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Table S1: Number of positive and negative DTIs for drug repurposing.

Table S2: Final training and testing sets after elimination.

Table S3: Tuned hyperparameters of DFFNN.

Data Availability Statement

All data used are publicly available benchmark datasets; sources are cited in Section 3.