Drug–Target Classifiers: Advances & Applications

Updated 11 August 2025

Drug–target classifiers are computational models that predict and rank interactions between drug molecules and proteins, playing a key role in virtual screening and repurposing.
They integrate methodologies such as multi-label classification, deep learning, and graph neural networks to handle data sparsity, extreme class imbalance, and ligand promiscuity.
Ensemble methods and multi-modal feature integration, including chemical, biological, and network descriptors, enhance prediction accuracy and model interpretability.

Drug–target classifiers are computational models designed to predict and rank potential interactions between drug-like molecules and biological targets, most commonly proteins. These classifiers underpin numerous stages of drug discovery—from virtual screening and target fishing to in silico repurposing—by efficiently mining high-dimensional chemical and biological data to identify compounds with specific activities or elucidate the target profile of novel or existing drugs. The rapid development of machine learning and, more recently, deep learning has led to increasingly sophisticated classifiers that address key challenges such as extreme class imbalance, data sparsity, the need for interpretability, and the biological reality of ligand promiscuity.

1. Classification Paradigms and Model Frameworks

Drug–target classifiers have evolved from basic supervised learning paradigms—where drug–target interaction prediction is treated as a two-class classification problem—into complex systems that reflect the many-to-many nature of molecular biology.

Single-label vs. Multi-label Classification: Early computational approaches generally assumed a “single target paradigm,” assigning each ligand to one protein target (single-label, multi-class). However, many compounds are promiscuous, binding multiple targets. Recognizing this, more recent models adopt a multi-label multi-class formulation. For example, Naive Bayes models were extended into multi-label multi-class form (MMM), using binary relevance transformations to allow simultaneous assignment of multiple targets to one ligand. Statistical significance testing (McNemar test) has demonstrated that MMM approaches typically generalize better to real ligand promiscuity than their single-target (SMM) counterparts, with higher recall rates and overall statistically superior performance despite a modest reduction in precision (Afzal et al., 2014).
Binary and Regression-Based Settings: While binary “interacts/does-not-interact” classifiers remain prevalent—particularly in virtual screening—there is a strong trend toward modeling binding affinity as a continuous variable. End-to-end deep learning models such as DeepDTA (Öztürk et al., 2018) and ResDTA (Ghosh et al., 2023) use regression objectives to directly predict binding affinity, enabling fine-grained ranking of potential drug–target pairs for lead prioritization.
Local, Global, and Hybrid Graph Approaches: Some frameworks, such as bipartite local models (BLM) and enhanced variants (BLMN), build separate local classifiers for each drug and target, integrating similarity-based information and addressing the cold-start problem via neighbor-based label inference (Mei et al., 2015). Others leverage global information (e.g., bipartite graph models, knowledge graph embeddings, and multi-level graph neural networks) to jointly encode network topology and molecular structure, addressing both transductive and inductive learning requirements (Ye et al., 2021, Song et al., 15 Jul 2025).

The selection and engineering of feature representations is central to the success of drug–target classifiers.

Chemical Descriptors:
- Bit-string fingerprints (e.g., Atom Pairs, CARHART Atom Pairs, Fragment Pairs, Pharmacophore Fingerprints) and continuous variables (e.g., Burden Numbers) are widely used, particularly in rare-event classification tasks (Tomal et al., 2013).
- Extended Connectivity Fingerprints (ECFP), MACCS keys, Estate fingerprints, and CDK descriptors offer alternative encodings of molecular structure and are often evaluated and compared for predictive accuracy (Liyaqat et al., 2022).
Protein/Target Descriptors:
- Protein Sequence Composition (PSC) vectors, Pseudo Amino Acid Composition (PseAAC), and advanced representations derived from protein LLMs (PLMs) are increasingly employed to capture sequence-based or structural features (Liyaqat et al., 2022, Bal et al., 2023).
- Contact map features, derived either from predictive models or docking simulations, are used to encode intra-protein or protein–ligand distance relationships as inductive biases in deep models (Bal et al., 2023).
Graph-Based and Multi-Scale Features:
- Molecular graphs (atoms as nodes, bonds as edges) processed by Graph Neural Networks (GNNs) or Graph Isomorphism Networks (GIN) enable detailed, permutation-invariant encoding of chemical structure (Liu et al., 16 Apr 2024, Song et al., 15 Jul 2025).
- Hierarchical representations incorporating information from atoms, motifs (chemical substructures/fragments), and full molecules further enrich drug encoding, as in HiGraphDTI (Liu et al., 16 Apr 2024).
Functional and Multi-Modal Features:
- Integration of gene expression/chemical perturbation data (e.g., from L1000 profiles) merges structural and functional perspectives, allowing the classifier to consider the phenotypic consequences of compound administration (Debnath et al., 3 Nov 2024).

3. Ensemble Methods, Model Selection, and Imbalanced Data

Handling high-dimensional data and the rare-class problem—where only a small fraction of compounds are active against a target—requires specialized strategies.

Phalanx-Based Ensembles: The Ensemble of Phalanxes (EPX) framework divides variables into synergistic groups (“phalanxes”), trains base classifiers on each group, and ensembles their probability outputs. This divide-and-conquer approach addresses dimensionality and class imbalance by adaptively grouping variables that “work well together,” as guided by average precision (AveP) and merging criteria (Tomal et al., 2013).
Ensemble and Adversarial Training: Multi-view and adversarial learning architectures combine descriptive (e.g., ECFP8) and differentiable (e.g., GraphConv) representations under a joint loss, e.g., incorporating GAN-style objectives to improve output sharpness and distributional robustness (IVPGAN) (Agyemang et al., 2019). These techniques have demonstrated improved performance, particularly in challenging cold-start (new drug/target) scenarios.
Rare Class Optimization: Evaluation metrics for rare-class discovery emphasize early ranking performance and recall. Principal metrics include the hit curve (cumulative actives vs. candidate rank), AveP, initial enhancement (IE; enrichment in early selection), ROC-AUC, and Precision-Recall AUC. Balanced cross-validation and undersampling or oversampling (via fuzzy-rough or synthetic sample methods) are used to mitigate training bias (Islam et al., 2020).

4. Deep and Graph-Based Models in Drug–Target Prediction

Deep neural architectures and GNNs have become dominant in modern drug–target classification strategies.

Sequence-to-Affinity Deep Networks: Architectures such as DeepDTA (Öztürk et al., 2018) and ResDTA (Ghosh et al., 2023) use parallel CNN streams to encode drug SMILES and protein sequences, sometimes augmented by residual connections or multi-stream feature fusion to preserve informative context and optimize regression performance for affinity values.
Attention and Transformer Models: Self-attention Transformers (e.g., Molecule Transformer in MT-DTI (Shin et al., 2019)) and cross-attention modules enable the modeling of complex, long-range dependencies both within molecules and between drug and protein contexts. Pretraining large Transformer models on molecular corpora and transferring to DTI tasks yields improvements in precision-recall performance and practical ranking (Shin et al., 2019).
Hierarchical Graph Learning: Hierarchical graph representation models integrate molecular graphs at multiple abstraction levels (atom, motif, molecule) and couple them with specialized attention and fusion modules for protein features. Message broadcasting bridges global (affinity network) and local (molecule/target) contexts (Liu et al., 16 Apr 2024, Chu et al., 2022). Graph-in-Graph (GiG) frameworks enable inductive (structure-based) and transductive (network-based) learning synergy (Song et al., 15 Jul 2025).
Energy-Based and Generative Models: Energy-based generative frameworks (e.g., TagMol) optimize a conditional energy function to generate ligands with high predicted binding affinity to a target, employing GCN/GAT architectures for discriminators and incorporating contrastive learning to bypass computational intractabilities in partition functions. These approaches support both ligand generation and classification (Li et al., 2022).

5. Integration of Networks, Knowledge Graphs, and Tensor Methods

Drug–target classifiers increasingly leverage the broader biological or chemogenomics context by integrating network and knowledge-graph-derived features.

Tensor Factorization with Knowledge Graphs: Incorporating gene/protein representations learned from knowledge graphs (e.g., Hetionet) into tensor factorization frameworks leads to performance gains in predicting clinically successful drug–target–disease triplets, as demonstrated by improvements in AUC and AUPR over matrix and tensor factorization baselines (Ye et al., 2021). The integration of multi-relational embeddings and external side information is critical for addressing data sparsity and enriching biological context.
Causal Intervention and Confidence Calibration: Causal intervention methods perturb KGE model embeddings and assess score sequence robustness via ranking consistency metrics. Calibrated confidence measures (e.g., P_ci(S) averaged over interventions) significantly improve calibration error and the accuracy of high-confidence predictions, allowing more reliable prioritization in drug discovery pipelines (Ye et al., 2023).

6. Interpretability, Practical Applications, and Limitations

Interpretability and practical deployment are increasingly emphasized in classifier research and application.

Model Interpretability: Hierarchical attention mechanisms and motif-level representations enable models to pinpoint critical chemical substructures and protein sequence regions that contribute to interaction, supporting mechanistic hypotheses, rational design, and repositioning strategies (Liu et al., 16 Apr 2024).
Multi-Modal and Functional Integration: Combining structural data with functional perturbation features (e.g., RNA-seq, L1000) enhances prediction accuracy and biological plausibility, at the cost of increased data dependency and potential data loss in current benchmarks (Debnath et al., 3 Nov 2024).
Scalability and Efficiency: Modern deep learning frameworks (e.g., PADME) are designed to scale linearly with data set size, unlike quadratic kernel-based methods. End-to-end or modular architectures (e.g., Barlow Twins + GBM in BarlowDTI) further reduce computational burden and promote accessibility via web interfaces (Schuh et al., 31 Jul 2024).
Challenges: Availability and completeness of functional data (gene expression) remains a bottleneck in candidate coverage for multi-modal frameworks. Ensuring chemical and semantic validity in generative or grammar-based models is also a challenge, requiring ongoing advances in parsing, data augmentation, and outlier detection.

7. Conclusion

Drug–target classifiers have rapidly transitioned from empirical descriptor-driven classifiers and naive Bayes formulations to complex, multi-level, multi-modal architectures integrating graph representation learning, sequence models, adversarial and self-supervised learning, and knowledge graph embeddings. The field has advanced metrics and evaluation standards to recognize the rare, multi-label, and context-rich nature of real-world pharmacology. Methodological innovations focusing on scalable learning, robust feature representation, data fusion, and confidence calibration are driving continual improvements in accuracy, interpretability, and translational value. Ongoing work on expanding data modalities, integrating additional biological context, and refining scalability and cold-start generalization will further enhance the impact of these classifiers in drug discovery and personalized medicine.