In silico methods for drug-target interaction prediction

Summary

Drug-target interaction (DTI) prediction is a crucial component of drug discovery. In recent years, in silico approaches have attracted attention for DTI prediction, primarily because of their potential to mitigate the high costs, low success rates, and extensive timelines of traditional drug development, while efficiently using the growing amount of available data. This review identifies four major factors that influence DTI predictions, highlights persistent challenges, and proposes insights and strategies from the perspectives of data, features, and experimental setups to address these challenges. Furthermore, it emphasizes the importance of refining established approaches—such as the “guilt-by-association” concept—to manage data sparsity, and integrating emerging technologies, including large language models and AlphaFold, to advance feature engineering. We hope that this work will provide valuable guidance and novel perspectives for advancing future research on DTI predictions.

In this review, Ru et al. present actionable strategies for in silico DTI/DTA prediction, covering dataset curation, true-negative construction, and rigorous cold-start evaluation. By integrating multimodal data, heterogeneous networks, LLMs, and AlphaFold, this work outlines a pathway toward more reliable, interpretable, and translational drug discovery pipelines.

Introduction

Drugs are essential for improving human health, extending life expectancy, and enhancing the quality of life.^1,2 The development of new drugs is fundamental to medical progress and has driven the discovery of innovative therapies, facilitating more effective disease management. Many once-untreatable conditions, such as specific cancers and infectious diseases, have become manageable through advances in drug discovery. Likewise, although chronic diseases such as diabetes, hypertension, and cardiovascular disorders remain incurable, drug research has substantially improved patient survival. Thus, drug development remains a cornerstone of medical advancement and a vital contributor to global health outcomes.³

Recent breakthroughs in technologies such as artificial intelligence (AI), gene editing, and high-throughput screening have accelerated the drug development process.^4,5,6,7 In early 2020, BenevolentAI leveraged its AI platform to identify baricitinib, a JAK inhibitor, as a potential therapeutic agent capable of inhibiting SARS-CoV-2 entry into host cells. This discovery led to the inclusion of baricitinib in multiple clinical trials for COVID-19 treatment and ultimately resulted in its emergency use authorization approval by the U.S. Food and Drug Administration (FDA).^8,9 Insilico Medicine leveraged its generative AI platform in 2019 to design and optimize a novel drug candidate for idiopathic pulmonary fibrosis within just 46 days. The compound entered clinical trials in 2022, marking the world’s first AI-designed drug candidate to reach the clinical stage.¹⁰ A recent report by the U.S. FDA noted the approval of 29 new drugs in the first 9 months of 2024, reflecting sustained momentum in U.S. drug development and the agency’s regulatory efficiency. Despite the increased number of approvals, long-standing challenges persist. The development of a new drug—from initial research to the market—typically requires approximately $2.3 billion and spans 10–15 years.^11,12,13 These high costs often pose entry barriers for small- and medium-sized pharmaceutical companies, while prolonged timelines compromise the industry’s capacity to rapidly respond to public health emergencies. Moreover, such extensive investments in time and cost do not necessarily correlate with improved success rates in drug development.¹⁴ Recent data indicate that the overall success rate fell to 6.3% by 2022, suggesting that over 90% of drug candidates ultimately fail to reach the market.¹⁵ These unsuccessful projects further increase research and development costs, underscoring the financial challenges inherent to drug discovery.

Drug development typically comprises five stages: discovery, preclinical research, clinical trials, regulatory approval, and post-marketing surveillance. Each stage has distinct objectives and challenges. Given the high costs, extended timelines, and considerable risks associated with this process, researchers have strived to enhance the efficiency and cost effectiveness across all stages. Drug-target interaction (DTI) prediction is a pivotal component of the discovery phase and is integral to advancing new drug development.¹⁶ Accurate target prediction and drug molecule optimization help mitigate the risk of clinical trial failures. Precise target identification minimizes the validation of ineffective drug-target pairs, allowing for more focused experimentation and efficient resource utilization. DTI prediction aids in identifying potential off-target effects, facilitating early detection of safety risks, and thereby improving drug safety.¹⁷ It is also valuable for identifying multi-target drugs that are promising for complex disease treatment. Consequently, DTI predictions have attracted considerable attention in drug research.

With the extensive growth in bioactivity data, compound libraries, and protein sequence data, in silico methods have become powerful tools for predicting DTIs.^18,19,20,21 These computational approaches enable the preliminary screening of thousands of compounds, notably reducing the reliance on labor-intensive experimental validations and accelerating the drug development pipeline. Recent work by Tanoli et al.,²² a large-scale literature analysis of 3,286 DTI prediction studies, emphasized the importance of combining computational strategies with experimental assays to ensure the biological and translational relevance of DTI models. Their review highlighted the lack of rigorous validation practices across the field. Building on this foundation, this review further summarizes four key factors that influence DTI prediction and systematically outlines the persistent challenges faced by current computational approaches. We propose targeted strategies and insights to address critical issues such as data sparsity and feature representation. We hope this work contributes to narrowing the significant gap between computational prediction and experimental validation and promotes the integration of DTI research into practical drug discovery pipelines.

Application of in silico methods in DTI prediction

Early in silico-based approaches

Early in silico methods for DTI prediction primarily focused on molecular docking and ligand-based virtual screening techniques.²³ Molecular docking, the earliest computational approach in this field, was introduced by Kuntz et al. in 1982.²⁴ This technique uses the three-dimensional (3D) structure of target proteins to position candidate drug molecules within the active sites, thereby simulating potential binding interactions.²⁵ Docking algorithms estimate binding free energies to predict the most favorable binding configuration between drugs and their targets.

Ligand-based virtual screening methods,²⁶ such as quantitative structure-activity relationship (QSAR)²⁷ and pharmacophore models,²⁸ predict new drug candidates by leveraging known bioactivity data. QSAR models establish mathematical correlations between the molecular structure and bioactivity. Pharmacophore models identify the spatial arrangements of functional groups that are essential for bioactivity. By capturing the shared characteristics among bioactive compounds, pharmacophore models facilitate the efficient virtual screening of compound libraries for structurally similar candidates.

Early in silico methods of DTI prediction have notable limitations. Molecular docking is highly dependent on the availability of protein 3D structures, which were scarce or difficult to obtain during the early years of its application.^29,30 Although homology modeling can approximate unknown structures, its accuracy declines significantly when sequence similarity between the template and target protein is low.³¹ Ligand-based methods also assume linear relationships between the chemical structure and biological activity, yet real-world molecular interactions are often complex and nonlinear, making such methods insufficient for capturing dynamic binding behaviors.

Moreover, these techniques rely heavily on known active compounds, limiting their potential to explore novel chemical spaces and reducing their general applicability in early-stage drug discovery.³²

In summary, the limitations of early in silico methods for DTI prediction, such as their dependency on 3D structural data, inadequacy in capturing complex structure-activity relationships, and difficulties in addressing data scarcity, have catalyzed the adoption and development of machine learning techniques in DTI prediction.

Machine learning-based methods

The advent of machine learning has led to substantial breakthroughs in DTI prediction. Machine learning has become a powerful tool for drug screening and target identification because it enables computational models to autonomously learn patterns and relationships from data. Yamanishi et al.³³ pioneered machine learning-based DTI prediction by constructing a dual-layer model that integrates chemical and genomic information.

Various machine learning algorithms and methods have been applied to DTI prediction, each contributing unique strategies and frameworks. Table 1 summarizes the milestones and representative studies in the field, and the following section highlights several influential studies and applications.

• •KronRLS: this approach integrates drug chemical structure similarity with the Smith-Waterman similarity scores of target sequences within a Kronecker regularized least-squares framework. This is the first study to formally define the DTI prediction problem as a regression task, laying the foundation for quantitative DTI prediction.³⁴
• •SimBoost: the first nonlinear approach for continuous DTI prediction, SimBoost introduces prediction intervals as a confidence measure and interpretable features derived from drug similarity matrices, protein similarity matrices, drug-target affinity matrices, and neighboring relationships.³⁵
• •DGraphDTA: DGraphDTA was the first method used to construct protein graphs based on protein contact maps by leveraging the spatial information inherent in protein structures. A protein contact map is a two-dimensional (2D) matrix that captures residual interactions within the 3D structure of a protein, which is essential for accurately predicting binding affinities.³⁶
• •MT-DTI: this model was the first to apply attention mechanisms to drug representation, addressing the limitations of convolutional neural network (CNN)-based methods in capturing associations between distant atoms, thereby improving the interpretability and predictive power of DTI models.³⁷
• •MVGCN: unlike most supervised learning approaches for DTI prediction, it introduces a multiview graph convolutional network (MVGCN) framework for link prediction within biomedical bipartite networks. By integrating similarity networks with bipartite networks, the MVGCN constructs a multiview heterogeneous network and uses self-supervised learning for the initial node embeddings.³⁸
• •DrugVQA: adapting concepts from visual question answering (VQA), DrugVQA frames the DPI task as a VQA problem. The protein’s distance map is used as an “image,” the drug’s SMILES string as a “question,” and the interaction prediction as the “answer,” providing an innovative perspective on DPI tasks.³⁹
• •DeepAffinity: this model captures nonlinear dependencies between protein residues and compound atoms through unsupervised pretraining. These “long-distance” dependencies are crucial for compound-protein interactions, as residues or atoms in proximity within the 3D space may participate jointly in molecular interactions.⁴⁰
• •BridgeDPI: while learning-based methods focus on individual DPIs, BridgeDPI introduces “guilt-by-association” principles to enhance network-level information, effectively combining network- and learning-based approaches to improve DTI prediction.⁴¹
• •DTINet: DTINet integrates data from diverse sources (e.g., drugs, proteins, diseases, and side effects) and learns low-dimensional representations of drugs and proteins to manage noise, incompleteness, and high-dimensional characteristics of large-scale biological data.⁴²
• •DeepICL: this model characterizes four protein-ligand interaction patterns (hydrophobic interactions, hydrogen bonds, salt bridges, and π–π stacking) and proposes a 3D molecular generation framework that incorporates interaction-aware features to advance structure-based drug design.⁴³
• •MMDG-DTI: leveraging pre-trained large language models (LLMs), MMDG-DTI captures generalized text features across biological vocabulary, demonstrating strong capabilities in handling unseen samples and extracting robust, discriminative features.⁴⁴
• •NerLTR-DTA: framing DTI prediction as a ranking task, NerLTR-DTA uses learning-to-rank (LTR) principles, creating unique queries for applications across multiple scenarios, including the discovery of novel drugs and targets.⁴⁵
• •BHCNS: this method improves prediction accuracy by identifying reliable negative samples and applying an inverse hypothesis in which proteins dissimilar to a known target are unlikely candidates for a compound’s interaction. This approach enhances the sample reliability and distinguishes BHCNS from conventional CPI prediction models.⁴⁶

Table 1.

Milestones and representative studies in the DTI and DTA prediction

	Drug feature	Protein feature	Explanation
AGraphDTA47	drug molecule graph features based on GNN	amino acid sequence features based on CNN;protein contact map features based on GCN	the protein graph features and amino acid sequence features, extracted via the neural network and having the same dimensions, are fused using element-wise matrix addition, thereby providing a more comprehensive representation of the complex information of the protein
BINDTI48	drug molecule graph features based on GCN	amino acid sequence features based on self-attention mechanism and convolution	the features of the drug and the target are fused through a bidirectional intention network, which combines the intention mechanism and the multi-head attention mechanism
BindingSiteDTI49	GNN-based fixed-scale substructure features of drug molecules	GNN-based multiscale substructure features of protein maps	explicitly extracts the multiscale substructure of the target and the fixed-scale substructure of the drug to facilitate the identification of structurally similar substructure markers and models hidden relationships at the substructure level to construct interactive features
BiComp-DTA50	extracting features from drug sequences based on separable CNN layers	extracting features from encoded protein sequences using fully connected neural networks	captures informative features from protein sequences, a unified measure, BiComp, is proposed that builds on alignment-free (i.e., Lempel-Ziv-Markov chain algorithm [LZMA]) and alignment-based (i.e., Smith-Waterman) similarity measures
CPInformer51	functional class fingerprint (FCFP) is fused with structural GCN features as the final representation of the compound	multi-scale protein features with different receptive fields extracted from different network layers	ProbSparse self-attention is applied to protein features under the guidance of compound features to eliminate information redundancy and improve the accuracy of CPInformer
CPGL52	representing drug molecule information based on graph attention network (GAT)	representing protein information based on long short-term memory neural network (LSTM)	LSTM is able to learn the relationship between words that are far apart in a sequence. This unique use of LSTM can better extract the spatial features of protein structure
DeepDTA53	learning representations from drug SMILES sequences based on CNN	learning representations from raw protein sequences based on CNN	first model for predicting DTA based on deep learning
DeepConv-DTI54	extracting drug fingerprint information based on fully connected layers	extracting local residue patterns of target protein sequences based on CNN	method for capturing local residue patterns using CNN successfully enriched the protein features in the original sequence
DataDTA55	fingerprint characterization of compound structure information based on algebraic graphs, extracting SMILE sequence features based on CNN	protein binding pocket descriptor, extracting protein sequence features based on CNN	a dual-interaction aggregation neural network strategy is developed to ensure effective learning of multi-scale interaction features
DeepGLSTM56	a graph convolutional network (GCN) module is introduced, using the power graph representation to process drug compounds	processing protein sequences using a bidirectional LSTM layer	method based on GCN and LSTM, capable of predicting the binding affinity values between FDA-approved drugs and SARS-CoV-2 viral proteins
DrugVQA39	extract SMILES features based on multi-head self-attention BiLSTM	feature extraction based on dynamic CNN protein distance map	interpretable model inspired by the visual question answering (VQA) paradigm, which can directly predict DPI based on protein distance maps and molecular SMILES
DeepAffinity40	extracting SMILES sequence features based on bidirectional recurrent neural network (RNN)	represent protein sequences with a new alphabet of structural and physicochemical properties and then extract features based on a bidirectional RNN	the performance on new protein categories with limited labeled data can be further improved through transfer learning. Additionally, we developed independent and joint attention mechanisms and embedded them into our model to enhance its interpretability
DGraphDTA36	extracting drug molecule graph features based on GNN	extracting protein contact graph features based on GNN	the proposed method is the first attempt to construct a protein graph based on the protein contact graph
DeepCDA57	encoding drug SMILEs based on CNN	encoding protein sequences based on LSTM	the adversarial domain adaptation method is used to learn the feature encoder network for the test domain to handle the different distributions of training and test domains
DeepEmbedding-DTI58	extracting drug molecule graph features using GNN with attention mechanism	the bidirectional encoder representation from Transformer (BERT) model learns embedding vectors from the text composed of these protein words	to improve the training efficiency, a Bidirectional Encoder Representation from the Transformer pre-training method is introduced to extract substructure features from protein sequences, and a local breadth-first search is introduced to learn subgraph information from molecular graphs
FusionDTA59	BiLSTM is used as a feature encoder for drug SMILES	BiLSTM is used as a feature encoder for protein sequences	to address the loss of implicit information, a novel multi-head linear attention mechanism is employed to replace the coarse pooling methods. FusionDTA is able to aggregate global information based on attention weights, rather than selecting the maximum information like max pooling
GraphCPI60	using GNNs to learn graph representations of compounds	building blocks using CNNs to learn low-dimensional vector representations of protein sequences	a framework that combines advanced graph neural representations of compounds with pre-trained embedding techniques for protein sequences is proposed. To the best of our knowledge, this study is the first to integrate local chemical context and topological structures to learn the interactions between compound-protein pairs
GraphDTA61	drug molecule graph representation learning based on GCN\GAT\GIN\GAT-GCN	protein sequence learning based on CNN	the first study to use GNN for DTA prediction
HyperAttentionDTI62	using CNN to learn the feature matrix of drugs	using CNN to learn the feature matrix of proteins	unlike previous attention-based models, our model infers an attention vector for each amino acid-atom pair. These attention vectors not only capture the interactions between amino acids and atoms but also control the representation of features across channels
IIFDTI63	GAT is used to extract independent features of drugs, which can capture the topological information between atoms	convolutional structures with different kernel scales are used to extract independent features of proteins	fuses the interaction and independent features between drugs and targets to predict DTI
MIDTI64	GCN is used as an encoder to learn drug embeddings from integrated drug similarity networks	GCN is also used as an encoder to learn target embeddings from integrated target similarity networks	a novel multi-view similarity network fusion strategy is proposed, which leverages a multi-view attention mechanism to integrate different similarity networks in an unsupervised manner, as long as the nodes and sizes of these networks are consistent
MGraphDTA65	extracting drug features based on multi-scale graph neural network	multi-scale convolutional neural network to extract target features	a multi-scale graph neural network and a novel visual explanation method called gradient-weighted affinity activation mapping are proposed for DTA prediction and interpretation
MCANet66	learning low-dimensional representation features from drug sequences based on CNN	learning low-dimensional representation features from protein sequences based on CNN	using a cross-attention mechanism to extract interaction features between drugs and proteins, the feature representation capability of drugs and proteins is enhanced. Additionally, the PolyLoss function is employed to mitigate the overfitting and class imbalance issues in drug-target datasets
MATT-DTI67	extracting sequence features based on CNN	extracting sequence features based on CNN	we propose a relation-aware self-attention module to model drugs from SMILES data while considering the correlations between atoms. The relative self-attention module enhances the relative positional information between atoms in the compound, while accounting for the relationships between all elements
MMDG-DTI44	a graph convolutional network is applied to extract the relationship between atoms	the local features of the protein are represented by a 3D-CNN to represent the spatial features of the binding site. The global features of the protein are represented by a one-dimensional convolutional neural network to represent the features of the amino acid sequence	we propose a multi-scale convolutional network that utilizes different types of convolutional networks to extract local and global features of proteins as well as topological features of compounds
MFR-DTA68	processing FCFP information based on BioMLP and extracting molecular structure features based on GNN	based on BioCNN processing sequence, amino acid embedding and word embedding information	BioMLP/CNN is the first module designed to extract individual features of biological sequence elements, simultaneously extracting both individual and relational features of elements in a sequence
Tsubaki et al.69	extracting compound subgraph features using graph neural networks	extract sequence features using convolutional neural networks	using neural attention mechanisms alleviates the issue of poor interpretability in the black-box nature of deep learning, allowing us to identify which subsequences in the protein are more important when predicting the interactions of drug compounds
TC-DTA70	extracting sequence features based on CNN	use the encoder module of Transformer to extract amino acid sequence features	the results of this study demonstrate the effectiveness of the Transformer encoder and CNN in extracting meaningful representations from sequences
TEFDTA71	extracting molecular features based on MACCS fingerprints and encoders in Transformer	extracting protein sequence features based on CNN	most methods have been primarily developed for predicting non-covalent binding affinity, and currently, there are no deep learning methods specifically designed for predicting covalent binding affinity. In this paper, we propose a new model for predicting both covalent (bonded) and non-covalent (non-bonded) binding affinities in drug-protein interactions
TransformerCPI72	solving the molecular representation problem based on GCN	the word2vec of the protein is obtained based on its sequence, and then the sequence feature vector of the protein is passed to the encoder to learn a more abstract representation of the protein	learns the required interaction features and reduces the risk of hidden ligand bias. By mapping attention weights onto protein sequences and compound atoms, we can explore the interpretability of the model, which helps us determine whether the predictions are reliable and physically meaningful
TransformerCPI2.073	extracting compound molecular graph features based on GCN	calculate protein sequence representation using the pre-trained protein language model TAPE-BERT	demonstrates that sequence-to-drug models can achieve virtual screening performance close to structure-based approaches (without relying on any prior knowledge of protein 3D structures) and also validates the feasibility of applying this concept to drug discovery

Representative machine learning and deep learning methods are listed together with their drug-side features, protein-side features, and explanatory notes. The “explanation” column highlights the distinctive aspects and meaningful contributions of each study, summarizing what makes the corresponding method noteworthy.

Major factors influencing DTI predictions

Four key factors influencing DTI prediction have been identified: problem formulation (binary classification or regression), data quality and quantity, feature engineering, and the experimental setup.

Problem formulation

Most studies have formulated DTI prediction as a binary classification task to determine whether a drug (typically, a small-molecule compound) interacts with a biological target (such as a protein, enzyme, or receptor).⁷⁴ The affinity between a drug and its target, which reflects the binding strength of the interaction, is a critical indicator of drug efficacy. Therefore, affinity prediction constitutes a specialized and precise subtask within DTI prediction.⁷⁵

Affinity prediction is generally formulated as a regression task and requires sophisticated models to accurately quantify the binding strength between drugs and targets. Although DTI and drug-target affinity (DTA) predictions focus on different aspects of drug development, they are complementary and critical. DTI prediction provides a foundation for drug discovery by identifying potential interactions, whereas DTA prediction provides detailed insights into optimizing drug properties. Both DTI and DTA prediction tasks have attracted increasing interest and considerable research investments. Figure 1 outlines the steps involved in DTI and DTA prediction using machine learning approaches (A) and illustrates the number of studies on DTI and DTA conducted in recent years (B).

Figure 1.

Data inputs, modeling workflow, and research trends in in silico DTI/DTA prediction

(A) Drug attributes, target attributes, and drug-target relationships provide the fundamental data basis for in silico prediction. In different studies, these data are represented in forms such as feature vectors, interaction networks, or affinity matrices and then transformed into inputs for model construction. The subsequent machine learning and deep learning workflow typically includes preprocessing, feature encoding, model building, training, and evaluation for two main tasks: drug-target interaction (DTI) classification and drug-target affinity (DTA) regression.

(B) Annual number of published studies on DTI and DTA prediction in recent years, showing the steady growth of research activity in both fields.

Data quality and quantity

Data quality and quantity are critical factors that influence the performance and generalization capabilities of a model. Models trained on unrepresentative data may perform adequately in specific contexts but struggle to generalize to novel situations. Therefore, high-quality datasets are essential for promoting model generalizability, reducing bias and error, enhancing robustness, and mitigating overfitting risks. For DTI prediction, various datasets have been developed to meet specific experimental requirements and align with algorithmic characteristics, such as the enzyme, ion channel, G protein-coupled receptor (GPCR), and nuclear receptor datasets for classification tasks³³ and the Davis,⁷⁶ KIBA,⁷⁷ and Metz⁷⁸ datasets for regression tasks.

Table 2 provides a summary of the commonly used datasets for DTI prediction. Based on these datasets, several key characteristics can be identified.

Table 2.

Commonly used datasets in DTI/DTA prediction

Task type	Datasets	Compound	Protein	Interaction
Classification	enzyme33,64,66,79	445	664	2,926
	ion channel33,64,66,79	210	204	1,476
	GPCR33,63,64,66,79	223	95	635
	nuclear receptor33,64,66,79	54	26	90
	DUD-E39,49,58,69,80	22,886	102	22,645
	human39,44,48,49,51,52,58,60,63,65,69,72	1,052	852	3,369
	C. elegans44,51,52,60,63,65,69,72	1,434	2,504	4,000
	BindingDB39,44,48,49,51,52,63,72,81	49,745	812	33,772
	DrugBank44,48,62,63,66	6,645	4,254	17,511
	BIOSNAP48,81	4,510	2,128	27,464
	ChEMBL73	69,616	3,348	117,513
	Luo et al.38,42,64	708	1,512	1,923
Regression	Davis44,47,53,56,57,65,66,68,70,71,81,82	68	442	30,056
	KIBA44,47,53,56,57,65,66,68,70,71,82	2,111	229	118,254
	Metz56,65,82	1,423	170	35,259
	BindingDB57,71	80,324	5,561	1,254,402
	ToxCast56,65	7,657	328	342,869
	STITCH56	724,471	15,258	1,244,420
	Sc-PDB68	6,326	4,782	16,034
	DTC56	5,983	118	67,894

Datasets are grouped by task type. For each, the numbers of unique compounds, proteins, and interaction pairs are provided.

(1)Clinical relevance and functional richness of target proteins: most datasets focus on proteins such as enzymes, ion channels, GPCRs, nuclear receptors, and protein kinases—which are among the most important and well-validated therapeutic targets in clinical pharmacology.⁷⁹ These target families are extensively annotated in public databases with respect to their structure, function, and mechanism of action. (2) Data integration from multiple trusted sources: many datasets are constructed by integrating records from publicly available databases such as BindingDB, DrugBank, ChEMBL, ToxCast, STITCH, and curated datasets like Human and C. elegans. In contrast, datasets like Davis and Metz are derived from high-throughput screening (HTS) assays conducted under consistent experimental protocols, which provide more accurate and quantitative interaction measures. (3) Species specificity and cross-species extension: the majority of datasets involve human protein targets, reflecting the primary focus of drug development. However, datasets such as C. elegans include protein targets from the model organism Caenorhabditis elegans, providing valuable opportunities to evaluate model generalizability across species. (4) Distinct naming with shared origins: some datasets have unique names but are still built upon publicly available resources. For instance, the Human and C. elegans datasets were curated by Liu et al.⁴⁶ based on data from DrugBank, Matador, and STITCH. Additionally, the DUD-E dataset includes a broader range of target classes, including GPCRs and ion channels, and is derived from sources such as ChEMBL and ZINC. (5) Regression datasets adapted for classification tasks: although originally designed for regression, several datasets can be converted into binary classification tasks by applying threshold-based labeling strategies. For example, in the BindingDB dataset, drug-target pairs are labeled positive if the reported ${IC}_{50}$ value is below 100 nM and as negative if the ${IC}_{50}$ value exceeds 10,000 nM.⁸³ Thresholds of 5.0 (${pK}_d$) for the Davis dataset and 12.1 (KIBA score) for the KIBA dataset are commonly used to binarize interaction strength.³⁵

Feature engineering

Feature engineering is the process of transforming raw data into a format suitable for machine learning models. This process encompasses feature extraction, optimization, and interaction, which collectively aid models in interpreting data structures, enhancing prediction accuracy, and reducing computational complexity.

A critical step in machine learning-based DTI prediction is to numerically encode compound and protein information.^84,85 Numerous studies have sought to boost model accuracy by characterizing compound and protein features from multiple perspectives. For example, compound properties can be represented using molecular fingerprints, general descriptors, molecular structures, and functional groups. Protein information can be described through sequence composition, amino acid physicochemical properties, and secondary structural features. These descriptors form handcrafted features that serve as inputs for classification or regression algorithms. For example, iDTI-ESBoost⁸⁶ combines structural and evolutionary features with an AdaBoost classifier for DTI prediction. RFDT,⁸⁷ a predictor based on the rotation forest algorithm, encodes protein sequences into position-specific scoring matrices and represents drugs using fingerprint feature vectors. Several specialized toolkits have been developed to process compound and protein features. Table 3 provides an overview of some widely used options. A concise overview of the tools listed in Table 3 is presented here for reference.

• •BioPython⁸⁸ is an open-source Python toolkit widely used for the processing and analysis of biological sequence data. It provides modules for reading and writing various sequence file formats, multiple sequence alignment, sequence transcription and translation, sequence comparison, structure parsing, and access to online biological databases.
• •PyMOL⁸⁹ is a powerful open-source molecular visualization tool supporting multiple structure file formats. It is widely applied in structural biology, drug design, and molecular modeling and is indispensable for protein structural feature extraction and mechanistic studies.
• •DSSP⁹⁰ is a classic tool for protein secondary structure analysis, automatically identifying and annotating secondary structure types for each amino acid residue based on 3D protein structures (usually from PDB files). It also computes hydrogen bonds, solvent accessibility, backbone torsion angles, and other structural features.
• •iFeature⁹¹ is a versatile Python-based tool for biological sequence feature extraction, covering 18 encoding schemes and capable of computing 53 types of feature descriptors. It offers both command-line and graphical interfaces and is commonly used for standardized preprocessing of high-throughput sequence data in machine learning tasks such as sequence classification, function prediction, and interaction analysis.
• •Pfeature⁹² provides six main modules (composition, binary profiles, evolutionary information, structural features, patterns, and model building), enabling the calculation of over 200,000 features for protein-level and residue-level annotations, as well as for predicting the functions of chemically modified peptides.
• •ProtDCal⁹³ is a Java-based program that calculates general numerical descriptors for protein sequences and 3D structures, covering features such as electronic interactions, van der Waals forces, torsion potentials, and topological indices related to folding rates.
• •modlAMP⁹⁴ is a Python package designed for antimicrobial peptide data analysis, providing tools for descriptor calculation, sequence retrieval from public or local databases, peptide design, classification, and visualization.
• •ProtParam⁹⁵ is an online protein sequence analysis tool developed by ExPASy (Swiss Institute of Bioinformatics), capable of computing physical and chemical properties such as molecular weight, theoretical isoelectric point (pI), amino acid composition, extinction coefficients, estimated half-life, instability index, aliphatic index, and hydropathicity.
• •RDKit⁹⁶ is an open-source cheminformatics toolkit developed in C++ with a complete Python interface, offering efficient and flexible functionalities for molecular modeling, structure parsing, fingerprint generation, and descriptor computation.
• •Open Babel⁹⁷ is an open-source chemical toolbox supporting multiple chemical data languages, offering functionalities such as file format conversion, conformer search, 2D depiction, filtering, and substructure/similarity searching.
• •OpenChem⁹⁸ is a deep learning toolkit based on PyTorch for computational chemistry and drug design, providing a flexible, modular framework that integrates molecular representations (e.g., SMILES and molecular graphs) with various neural network models for molecule-level machine learning tasks.
• •ChemPy⁹⁹ is an open-source chemistry calculation library written in Python, primarily used for handling idealized chemical reaction systems, quantitative chemical calculations, and solving basic chemical equations, including reaction kinetics and concentration modeling.
• •ChemAxon Marvin¹⁰⁰ is a professional molecular structure drawing and chemical information processing tool developed by ChemAxon, supporting molecule visualization, standardized input, reaction scheme editing, molecular property evaluation, and initial modeling preparations.
• •PaDEL-Descriptor¹⁰¹ is a Java-based chemical toolkit integrating the Chemistry Development Kit (CDK), capable of calculating 797 molecular descriptors (1D, 2D, and 3D) and 10 fingerprint types, including atom-type E-state descriptors, McGowan volume, molecular linear free energy relationship descriptors, and various binary fingerprints.
• •ChemAxon JChem¹⁰² is an enterprise-level cheminformatics suite developed by ChemAxon, designed for processing and analyzing large-scale molecular structure data, offering functionalities such as structure parsing, standardization, searching, fingerprint generation, and property calculations.
• •Pybel¹⁰³ is the Python wrapper for Open Babel, offering a simplified programming interface for reading, writing, converting, and analyzing molecular structures within the Python environment.
• •ChemDes¹⁰⁴ is an integrated platform for molecular descriptor and fingerprint calculation, combining tools such as Pybel, CDK, RDKit, BlueDesc, Chemopy, PaDEL, and jCompoundMapper. It can compute 3,679 molecular descriptors and 59 types of fingerprints and provides utilities for format conversion, MOPAC optimization, and fingerprint similarity calculations.
• •CDK¹⁰⁵ is a Java-based open-source cheminformatics library designed for small-molecule modeling and computation, offering functionalities such as molecular structure parsing, descriptor calculation, structure standardization, and cleaning. It serves as the core engine behind platforms like PaDEL-Descriptor, ChemDes, KNIME, and Weka.
• •DeepChem¹⁰⁶ is a Python library designed for machine learning and deep learning on molecular and quantum datasets, offering standardized models, datasets, and workflows for tasks in drug discovery, molecular modeling, bioactivity prediction, material science, and computational physics.

Table 3.

Toolkits for feature extraction of proteins and compounds

Object	Toolkit	Address
Protein	BioPython88	https://biopython.org/
	PyMOL89	https://pymol.org/
	DSSP90	https://swift.cmbi.umcn.nl/gv/dssp/
	iFeature91	https://github.com/Superzchen/iFeature
	Pfeature92	https://github.com/raghavagps/Pfeature
	ProtDCal93	http://bioinf.sce.carleton.ca/ProtDCal/
	ModlAMP94	https://modlamp.org/
	ProtParam95	https://web.expasy.org/protparam/
Compound	RDKit96	http://rdkit.org/
	Open Babel97	http://openbabel.org/wiki/Main_Page
	OpenChem98	https://mariewelt.github.io/OpenChem/html/index.html
	ChemPy99	https://chempy.readthedocs.io/en/latest/
	ChemAxon Marvin100	https://chemaxon.com/marvin
	PaDEL-Descriptor101	https://github.com/ecrl/padelpy http://www.yapcwsoft.com/dd/padeldescriptor/
	ChemAxon JChem102	https://chemaxon.com/jchem-engines
	Pybel103	https://pypi.org/project/pybel/
	ChemDes104	https://github.com/ifyoungnet/ChemDes
	CDK105	https://cdk.github.io/
	DeepChem106	https://deepchem.io/

Commonly used software libraries and platforms for computing descriptors are listed. Official URLs are provided for each toolkit.

Unlike handcrafted features, deep learning-based automatic features do not rely on extensive domain knowledge. They are adept at processing unstructured data and perform exceptionally well in complex tasks. Various deep learning methods, each with unique characteristics and advantages, have been applied to capture information from diverse perspectives. For example, graph attention networks¹⁰⁷ were designed to process graph-structured data, allowing models to learn the connection strengths (e.g., chemical bonds) between different nodes (e.g., atoms). Graph convolutional neural networks¹⁰⁸ effectively extract information from simplified graphical representations of protein structures, offering interpretability for the learned features. Long short-term memory¹⁰⁹ networks, specialized in sequential data, capture temporal information and long-term dependencies within input sequences.

In DTI prediction, many studies have attempted to probe deeper into data by employing feature interactions across multiple perspectives. ProtDec-LTR3.0¹¹⁰ applied feature mapping to integrate information from ACC and top-gram methods, whereas Zhang et al.¹¹¹ used cross-term feature mapping to handle input features in ligand-based virtual screening studies. PKRank¹¹² utilizes pairwise kernels for feature processing. Recently, attention mechanisms have been incorporated to enhance the feature interactions. DeepCDA⁵⁷ introduced a dual-attention mechanism that encodes interactions between protein sub-sequences and compound substructures, calculating the attention coefficients between each compound and protein substructure pair to represent the binding strength. BINDTI⁴⁸ integrates a bidirectional intent network with multihead attention to combine drug and target features, and IIFDTI⁶³ employs a bidirectional encoder-decoder framework to capture interaction features between drug and target substructures. In summary, intrinsic associations between features remain an important direction for further exploration in DTI research.

Experimental setup

In DTI prediction, two primary entities are considered: drugs and proteins. Their interactions extend beyond simple one-to-one relationships, making the distribution of these entities across the training and test sets a critical aspect in DTI research. Based on data distribution patterns and data availability, DTI prediction tasks can be categorized into warm- and cold-start scenarios.^34,45 Figure 2 shows the interactions between certain compounds and proteins. The left panel highlights the overlap of protein targets among five selected drugs (CHEMBL10903, CHEMBL17657, CHEMBL16882, CHEMBL17881, and CHEMBL10874). The Venn diagram indicates how many targets are either uniquely associated with or shared among these compounds. It is evident that some targets are commonly targeted by multiple drugs. The right panel provides a global view of the compound-protein interaction network. The network clearly demonstrates that most drugs are linked to multiple protein targets, and some targets are concurrently associated with several drugs. This further supports the complexity of DTI prediction tasks and underscores the necessity of modeling such multi-relational interactions.

Figure 2.

Overlap and global network of drug-target interactions

The left panel shows a Venn diagram of protein targets among five drugs (CHEMBL10903, CHEMBL17657, CHEMBL16882, CHEMBL17881, and CHEMBL10874). It highlights how many targets are unique to each drug and how many are shared (for example, CHEMBL10903 and CHEMBL17657 share several common targets, while CHEMBL17881 also overlaps with CHEMBL16882), illustrating that certain proteins are commonly targeted by multiple drugs. The right panel presents a global compound-protein interaction network, where nodes represent drugs or proteins and edges represent interactions. The network demonstrates that most drugs are linked to multiple targets, and some targets are concurrently associated with several drugs, underscoring the complexity of DTI prediction and the need to model multi-relational interactions.

In the warm-start scenario, both the training and test sets contain the same or highly similar drugs and targets, enabling models to achieve improved predictive accuracy. Investigating warm-start scenarios, commonly referred to as drug repurposing, allows researchers to maximize the utility of existing drugs and identify potential new indications. Consequently, warm-start settings are prevalent in DTI research.^40,45,53,61

The cold-start scenarios involve novel drugs, targets, or drug-target pairs in the test set that are absent during training. This lack of prior knowledge significantly increases the difficulty of predictive tasks under cold-start conditions. Research on cold-start scenarios is crucial for new drug development, as it supports the screening and prediction of new drug candidates in data-limited settings, offering innovative treatment possibilities for diseases with currently limited therapeutic options. Cold-start scenarios can be further divided into three categories:

• •Drug cold-start: new drugs appear in the test set, with no prior interaction data available for these drugs during training.
• •Target cold-start: new targets appear in the test set, with no interaction data for these targets present in the training set.
• •Drug-target cold-start: both drugs and targets in the test set are entirely absent from the training set, representing the most stringent cold-start setting.

To investigate the cold-start problem in a more rigorous manner, it is essential to adopt stricter and more refined data partitioning strategies. One effective approach is to construct scenarios where the similarity between drugs or targets in the training and testing sets is minimized, thereby simulating more realistic and challenging settings. This can be achieved by clustering drugs or proteins and ensuring that samples from the same cluster appear exclusively in either the training or testing set.

• •Protein clustering: proteins can be clustered based on sequence similarity using tools such as BLAST¹¹³ and MAFFT¹¹⁴ or based on structural similarity using tools like Foldseek.¹¹⁵
• •Drug clustering: drugs can be clustered based on molecular scaffolds,^116,117 shape similarity,¹¹⁸ or chemical fingerprint similarity.¹¹⁹

In drug development, addressing both the optimization of existing data through warm-start scenarios and the exploration of new drug potentials in data-scarce conditions via cold-start scenarios are essential for a comprehensive and effective DTI prediction strategy.

Conclusions and future perspectives

This review identified several key challenges in DTI prediction, including issues related to data sourcing, integration, and representation. Several targeted strategies have been proposed to address these challenges.

Challenges

Data sourcing issues

Various databases offering extensive data resources are currently available for DTI prediction. However, these databases present several limitations. The primary issue is data redundancy or duplication across databases, often compounded by inconsistencies in the data formats and standards. Differences in data sources, measurement methods, experimental conditions, and standardization protocols can lead to conflicting affinity values or interaction statuses for the same drug-target pair across databases. Moreover, many databases are not regularly updated, leading to the omission of recent experimental findings or newly identified DTIs. Furthermore, although effective DTI prediction involves a comprehensive range of data types (chemical, biological, genomic, transcriptomic, and clinical), most existing databases focus on a single data type, limiting their utility for multimodal analysis. This fragmented data landscape underscores the need for more integrated and comprehensive datasets to enhance the robustness and generalizability of DTI models.

Data integration challenges

Data sparsity remains a critical challenge in machine learning-based DTI prediction because confirmed DTIs are vastly outnumbered by unknown interactions, particularly for novel targets or rare compounds. Existing databases are often biased toward positive samples representing validated interactions, with a scarcity of clearly annotated negative samples. A common strategy for DTI prediction is to treat unknown interactions as negative samples. However, some unknown pairs may represent unvalidated positive interactions, complicating the ability of a model to distinguish between positive and negative samples accurately. Additionally, the considerable imbalance between known and unknown interactions exacerbates class imbalance in binary classification tasks. Consequently, the incorporation of true negative interactions has emerged as a pivotal area for enhancing model accuracy and robustness in DTI prediction.

Data representation issues

For DTI prediction, traditional machine learning approaches rely on handcrafted features that require deep expertise in drug chemistry, protein structure, and biological networks. Furthermore, handcrafted features often fail to capture the complex, nonlinear relationships intrinsic to DTIs, thereby limiting their applicability to novel molecular structures or less-characterized targets. These features are generally tailored to specific tasks or datasets, which restricts their generalizability and renders their optimization a demanding and time-consuming process.

Despite their proficiency in capturing complex nonlinear patterns and high-dimensional data, automatically learned features are frequently perceived as “black boxes” because of the opacity of their internal decision-making processes, which complicates their interpretation. Moreover, models that use these features typically require extensive hyperparameter tuning, which is an intricate and resource-intensive process requiring substantial experimentation and validation. Neural network architectures, such as CNNs, recurrent neural networks (RNNs), and graph neural networks (GNNs), have shown potential in feature representation for DTI prediction; however, each architecture has inherent limitations. CNNs are constrained in capturing global structural features and are typically unsuitable for molecular graphs that lack a defined Euclidean structure. Although RNNs are effective in processing sequential data, they struggle to retain their long-term dependencies. Despite their effectiveness in representing 2D atomic structures and their use in graph-based encoder-decoder frameworks, graph convolutional networks (GCNs) face challenges in assigning appropriate weights to critical neighboring nodes, exhibit reduced flexibility, and often converge at a slower rate.

Strategies (insights)

Ensuring high-quality data sources

To construct datasets that are comprehensive, representative, and enriched with samples, it is essential to collect data from databases that include diverse sources, measurement techniques, experimental conditions, and standardization protocols.^120,121 The following guidelines are recommended to ensure data reliability:

• (1)Prioritize high-quality databases: select databases with well-documented experimental conditions, precise measurement methodologies, and transparent data sources to ensure consistency and dependability. Table 4 provides a list of authoritative databases containing information on proteins, compounds, drug-target pairs, and associated biological entities.
• A concise overview of the tools listed in Table 4 is presented here for reference.

  • •
    UniProt (Universal Protein Resource)¹²² is one of the world’s most comprehensive and authoritative protein databases, encompassing over 120 million protein sequences with detailed functional annotations across all branches of life. It provides essential sequence, function, structure, evolutionary relationships, and literature information, serving as a foundational resource in proteomics, bioinformatics, and drug discovery.
  • •
    PDB (Protein Data Bank)¹²³ is the first open-access digital resource in biology and medicine and the most authoritative global repository for 3D structures of proteins and other biomacromolecules.
  • •
    NCBI Protein¹²⁴ is maintained by the U.S. National Center for Biotechnology Information (NCBI), integrating predicted and experimentally validated protein sequences from multiple sources, along with comprehensive sequence and functional annotations.
  • •
    InterPro¹²⁵ integrates predictive models from multiple databases (e.g., Gene3D, PANTHER, Pfam, and PROSITE) to annotate protein domains, families, and functional sites, supporting functional annotation and classification research.
  • •
    CDD (Conserved Domain Database)¹²⁶ maintained by NCBI focuses on the identification and annotation of functional conserved domains within protein sequences.
  • •
    BRENDA¹²⁷ is the most comprehensive enzyme information system, covering data on enzyme functions, catalytic reactions, physiological roles, substrates, inhibitors, tissue specificity, and disease associations.
  • •
    GO (Gene Ontology)¹²⁸ is the international standard for gene and protein functional annotation, organized into three core categories: biological processes, molecular functions, and cellular components.
  • •
    PubChem¹²⁹ is the world’s largest open-access chemical structure and bioactivity database, hosting detailed structure, property, and activity information on small molecules, drugs, intermediates, and natural products.
  • •
    ChEMBL¹³⁰ is a large open-access drug discovery database integrating bioactivity data of small molecules, approved drugs, and clinical candidates, widely used in pharmaceutical research.
  • •
    ChemDB¹³¹ is a small-molecule database constructed from multiple vendors and public resources, containing millions of real and virtual molecules for structure retrieval, property calculation, and target analysis.
  • •
    DrugBank¹³² is a comprehensive drug and drug-target knowledge base covering basic drug information, mechanisms of action, metabolic pathways, indications, and drug-drug and drug-food interactions, regarded as a “gold standard” resource in drug information.
  • •
    ZINC¹³³ is a freely accessible database for virtual screening, containing tens of billions of enumerated small molecules with atomically precise structures.
  • •
    ChemSpider¹³⁴ is a chemical database integrating hundreds of independent sources, offering detailed physical and chemical properties, molecular structures, spectra, synthesis routes, and safety information.
  • •
    DrugCentral¹³⁵ is an integrated platform compiling ingredient information for FDA-approved and other regulatory agency-approved drugs, including structure, bioactivity, pharmacology, indications, and regulatory status.
  • •
    Drugs@FDA¹³⁶ is the U.S. FDA’s official platform for accessing regulatory information on approved drug and biological products, covering listing information, indications, manufacturers, labeling, and clinical trial data.
  • •
    BindingDB¹³⁷ is a publicly accessible protein-small molecule interaction database, containing over one million experimentally measured binding data entries and links to resources like ZINC.
  • •
    KEGG¹³⁸ is a comprehensive database for genome and pathway information, systematically integrating genes, proteins, compounds, diseases, and drugs to assist in understanding biological functions and disease mechanisms.
  • •
    DGIdb¹³⁹ aggregates drug-gene interaction information from publications, databases, and online resources, standardizing and merging them into unified concept groups.
  • •
    ToxCast¹⁴⁰ is a vital database for toxicology research and drug safety evaluation, predicting the potential toxicity of chemicals through high-throughput in vitro screening and computational modeling.
  • •
    STITCH¹⁴¹ is a database of protein-chemical interactions, integrating experimental data, curated information, text-mined results, and predictions to support functional studies and drug discovery.
  • •
    SuperTarget¹⁴² is a comprehensive drug database integrating indications, adverse effects, metabolic pathways, target annotations, and GO terms, with high-quality manual curation for part of the content.
  • •
    CovalentInDB¹⁴³ is a specialized database focusing on covalent inhibitors and their targets, systematically curating known covalent drugs, small-molecule inhibitors, binding mechanisms, and clinical statuses.
  • •
    PGxDB¹⁴⁴ is a professional pharmacogenomics database supporting the analysis of drug-gene interaction data and promoting individualized therapy and translational medicine research.
  • •
    DTP (DrugTargetProfiler)¹⁴⁵ is an interactive network platform designed to enhance bioactivity modeling for multitarget anticancer compounds, particularly suited for precision oncology applications.
  • •
    GLASS¹⁴⁶ is a specialized database focused on GPCR-ligand interactions, annotating interactions involving natural ligands, small-molecule drugs, and endogenous signaling molecules.
  • •
    CTD (Comparative Toxicogenomics Database)¹⁴⁷ integrates information on interactions among chemicals, genes, and diseases, aiming to uncover the health impacts of environmental exposures.
  • •
    DTC¹⁴⁸ is a community-curated platform integrating and standardizing drug-target interaction data, including clinical development compounds, target-disease associations, and mutant protein targets.
  • •
    TTD (Therapeutic Target Database)¹⁴⁹ systematically catalogs descriptions of FDA-approved drugs, clinical drugs, investigational drugs, and their therapeutic targets to facilitate target-based drug research.
  • •
    PharmGKB¹⁵⁰ is the world’s leading pharmacogenomics knowledge base, systematically collecting information on gene variants, drug responses, and clinical phenotypes to support precision medicine.
  • •
    Reactome¹⁵¹ is a high-quality, manually curated database of human biological pathways, covering processes such as metabolism, signal transduction, the cell cycle, and immune responses.
  • •
    STRING¹⁵² is a high-quality platform for integrated protein-protein interaction data, combining experimental, predicted, text-mined, and curated database sources to build cross-species PPI networks.
  • •
    ConsensusPathDB¹⁵³ integrates multiple public resources to provide interaction data across human, mouse, and yeast, including protein-protein interactions, metabolic reactions, signaling pathways, and gene regulation networks.
  • •
    HPRD (Human Protein Reference Database)¹⁵⁴ compiles high-confidence human proteomic information, including protein phosphorylation modifications, subcellular localization, and functional annotations.
  • •
    BioCyc¹⁵⁵ is a portal integrating thousands of microbial genomes and their inferred metabolic pathways, supporting genome browsing, metabolic network modeling, and enrichment analysis.
  • •
    IntAct¹⁵⁶ is an open-access molecular interaction database focusing on high-quality experimental data integration, with all entries manually reviewed and standardized.
  • •
    BioGRID¹⁵⁷ is a comprehensive repository of molecular interaction data, providing protein and gene interaction information across multiple species, widely used in functional genomics and systems biology research.
  • •
    TDR Targets¹⁵⁸ is a target prioritization platform for neglected tropical diseases, systematically integrating pathogen genome data and druggability information to facilitate therapeutic target discovery.
• (2)Data preprocessing: outliers, duplicates, and noise should be removed to maintain data integrity. Unique identifiers (e.g., drug names, target names, and structural descriptors, such as SMILES and InChI) should be used to identify and eliminate redundant drug-target pairs across different databases.
• (3)Harmonizing inconsistent data: in cases where multiple databases report different affinity values for the same drug-target pair, researchers may adopt strategies such as calculating a consistency score, selecting data from the most credible source, or averaging the values to ensure a balanced representation.
• (4)Supplementing insufficient data: when existing data quality or coverage is inadequate, conducting additional experiments on critical drug-target pairs can provide valuable data points, thereby enhancing the overall quality and robustness of the dataset.

Table 4.

Databases of proteins, compounds, drug-target pairs, and various associations

Object	Database	Address
Protein	UniProt122	https://www.uniprot.org/
	Protein Data Bank123	https://www.rcsb.org/
	NCBI Protein Database124	https://www.ncbi.nlm.nih.gov/protein/
	InterPro125	https://www.ebi.ac.uk/interpro/
	CDD126	https://www.ncbi.nlm.nih.gov/cdd/
	BRENDA127	https://www.brenda-enzymes.org/
	GO128	https://geneontology.org/
Compound	PubChem129	https://pubchem.ncbi.nlm.nih.gov/
	ChEMBL130	https://www.ebi.ac.uk/chembl/
	ChemDB131	https://cdb.ics.uci.edu/
	DrugBank132	https://go.drugbank.com/
	ZINC133	https://zinc.docking.org/
	ChemSpider134	https://www.chemspider.com/
	DrugCentral135	https://drugcentral.org/
	Drugs@FDA136	https://www.accessdata.fda.gov/scripts/cder/daf/index.cfm
Drug-target	BindingDB137	https://www.bindingdb.org/
	KEGG138	https://www.kegg.jp/
	DGIdb139	https://www.dgidb.org/
	ToxCast140	https://www.epa.gov/comptox-tools/exploring-toxcast-data
	STITCH116	http://stitch.embl.de/
	SuperTarget142	http://insilico.charite.de/supertarget
	CovalentInDB143	http://cadd.zju.edu.cn/cidb/
	PGxDB144	https://pgx-db.org/
	DTP145	https://drugtargetprofiler.fimm.fi/
	GLASS146	https://zhanggroup.org/GLASS/
	CTD147	https://ctdbase.org/
	DTC148	https://drugtargetcommons.fimm.fi/
Other associations	TTD149	http://db.idrblab.net/ttd/
	PharmGKB150	https://www.pharmgkb.org/
	Reactome151	https://reactome.org/
	STRING152	https://string-db.org/
	Consensus PathDB153	http://cpdb.molgen.mpg.de/
	HPRD154	http://www.hprd.org/
	BioCyc155	https://biocyc.org/
	IntAct156	https://www.ebi.ac.uk/intact/home
	BioGRID157	https://thebiogrid.org/
	TDR Targets158	https://tdrtargets.org/

By adhering to these guidelines, researchers can ensure that their datasets support robust, reliable, and generalized DTI predictions.

Mitigating data sparsity

Several strategies based on current research, algorithmic innovations, and theoretical frameworks have been proposed to address the challenges posed by data sparsity in DTI prediction.

• (1)Guilt-by-association principle¹⁵⁹: this principle posits that structurally similar compounds are likely to interact with the same target and that target proteins with high homology (e.g., sequence or structural similarity) may share interactions with the same compounds. This approach has been integrated into DTI prediction tasks. For instance, NerLTRDTA⁴⁵ incorporates the properties of closely related neighbors into the profile of the target entity. BHCNS⁴⁶ generates reliable negative samples by assuming that proteins that differ from known targets of a compound are unlikely to interact with the compound.
• (2)Transfer learning¹⁶⁰: transfer learning addresses data scarcity by leveraging knowledge from related domains or tasks and enhancing model performance under data-limited conditions. In DTI prediction, related interaction networks (e.g., drug-disease, target-disease, or protein-protein interactions) offer valuable knowledge that can be transferred to improve prediction accuracy, especially for poorly studied drugs or targets.
• (3)Multitask learning¹⁶¹: when data for a specific target is sparse but the target belongs to a well-studied family, multitask learning can improve prediction by sharing data across related family members. This approach not only enhances DTI prediction but also improves generalizability by leveraging additional information from drug-disease associations.
• (4)Few-shot learning¹⁶²: few-shot learning enables models to generalize from minimally labeled data, achieving robust performance on new, unseen samples. Meta-learning,¹⁶³ a prominent few-shot approach, optimizes the learning process to allow rapid adaptation to novel tasks with limited data.
• (5)Data augmentation and active learning¹⁶⁴: active learning iteratively refines model training by selecting and labeling the most informative samples. Starting with a small, labeled dataset, the model selects high-value samples from an unlabeled pool using specific sampling strategies. These samples are then labeled and added to the training set, and the model is retrained in successive cycles until performance goals or budget constraints were met. This iterative process effectively mitigates data sparsity by maximizing the utility of available data.

Comprehensive and effective representation of drug and target data

Enhancing the representation of drug and target data is essential to advance DTI prediction accuracy and robustness. Figure 3 illustrates hierarchical relationships between data sources, data representation types, and handcrafted and automatically learned features, providing a framework for streamlining the effective representation of drug and protein information. Based on these insights, the following strategies are proposed.

• (1)Multi-view, multi-modality representation: drugs and targets can be represented using diverse data modalities, such as molecular sequences and graph structures.^165,166 Multi-modality data encapsulates diverse aspects of an entity and offers holistic representations. Each modality reflects distinct signal types, formats, and sources that may demonstrate complementary, correlated, or unique characteristics. Consequently, the integration of multi-modality and multi-view data substantially improves the accuracy, robustness, and reliability of the predictions.
• (2)Incorporation of heterogeneous network information: recent studies have emphasized the critical role of inter-entity relationships in elucidating biological functions. For example, analyzing drug-induced effects on microRNA expression, particularly in cancer progression, can inform drug mechanisms and guide novel therapeutic strategies.¹⁶⁷ Integrating such heterogeneous network information into DTI models can enhance biological interpretability and predictive performance.
• (3)Application of LLMs¹⁶⁸: unlabeled data serve as core resources for training LLMs. By leveraging semantic and structural information from massive unlabeled datasets, LLMs achieve cross-task generalization and extract sophisticated protein representations. Pre-training on large protein sequence datasets enables LLMs to capture complex biological relationships, offering promising avenues for improving DTI predictions.

Figure 3.

Hierarchical relationships between data sources, data representations, and features

Drug and target data, such as molecular structures, protein sequences, and interaction networks, can be represented in different ways and then processed to obtain features. Two complementary strategies are illustrated: handcrafted descriptors (e.g., fingerprints, physicochemical properties, and sequence composition) and automatically learned embeddings (e.g., CNN, RNN, GNN, and Transformer). Importantly, the same type of data may support both handcrafted and learned representations, highlighting the layered relationships between data sources, representation types, and feature extraction approaches.

Collectively, these strategies aim to improve data representation, thereby contributing to more accurate and interpretable DTI predictions.

Enhancing experimental data settings

Given that DTI prediction involves both drug and target entities, the risk of data overlap between training and test sets is substantial, potentially leading to overestimation of model performance. Rigorous experimental protocols are thus necessary.

• (1)Elimination of highly similar samples: prior to analysis, drugs or targets with high structural similarity—assessed using metrics such as Smith-Waterman scores¹⁶⁹ or Tanimoto similarity¹⁷⁰—should be removed to reduce redundancy and bias, ensuring a more representative dataset and improving experimental validity.
• (2)Cluster-based partitioning for enhanced data independence¹⁷¹: data partitioning based on the clustering of drugs or targets using molecular structures or protein sequences allows the creation of independent training and test sets with no overlap. This approach increases inter-cluster variance and simulates real-world scenarios involving novel drugs or targets more accurately. Consequently, this enhances the robustness and generalizability of the model when applied to unseen samples.

The influence of AlphaFold on DTI prediction

The 3D structure of target proteins is critical for drug design, particularly for evaluating binding sites and interactions. AlphaFold has revolutionized 3D structure prediction by offering reliable and high-precision structural data for DTI research.¹⁷² Using AlphaFold-predicted structures, researchers can identify binding sites with greater accuracy, significantly improving virtual screening processes. For instance, Johansson et al.¹⁷³ applied AlphaFold-based methodologies to refine peptide-protein docking, demonstrating its potential to enhance predictive performance. Integrating AlphaFold-generated structural data into DTI workflows holds great promise for enabling more precise interaction predictions and accelerating drug discovery.

Future perspectives

Contemporary computational approaches that encompass machine learning and deep learning methodologies have substantially enhanced the efficiency and accuracy of DTI predictions. These innovations have not only expedited drug discovery but also reduced development costs and contributed to multiple stages of the drug development pipeline. However, significant challenges remain, particularly in developing models that balance high predictive accuracy with interpretability and broad applicability.

Future research should prioritize the acquisition and integration of high-quality DTI datasets coupled with the development of advanced methods for representing drug and target features. Addressing persistent challenges, such as data scarcity, cold-start scenarios, and limitations in predictive precision, will be critical for further progress.

In summary, contemporary computational methods play a pivotal role in DTI prediction and hold immense potential for advancing drug discovery and development.

Acknowledgments

The work was supported by the National Natural Science Foundation of China (nos. 62450002, 32470693, and 62425107), Zhejiang Provincial Natural Science Foundation of China (no. LD24F020004), and the Municipal Government of Quzhou (no. 2024D001).

Author contributions

X.R. conceived the review framework, conducted the literature survey, synthesized and organized the content, drafted the manuscript, and prepared the figures and tables. Q.Z. provided overall supervision, conceptual guidance, and critical revisions to improve the clarity and rigor of the manuscript. W.H. and L.X. assisted in literature collection and provided helpful comments during manuscript preparation. All authors read and approved the final manuscript.

Declaration of interests

The authors declare no competing interests.