Source-linked AI summary
Recent Advances in Deep Learning-Based Drug-Target Binding Affinity Prediction
Jafin Khan, Md Hossain Shuvo
TL;DR
Accurate drug–target binding affinity prediction remains an open problem in computational drug discovery. This paper reviews deep learning methods, benchmark datasets, representations, and evaluation practices, finding strong standard-benchmark performance but reduced generalization in cold-start settings and important evaluation limitations.
Problem
Drug–target binding affinity prediction remains difficult despite recent advances, while experimental affinity measurement is expensive and time-intensive.
Method
The paper conducts a comprehensive review and comparative analysis of recent deep learning DTA methods, datasets, representations, architectures, and evaluation methodologies.
Results
Most methods report CI values of approximately 0.89–0.91 on Davis and KIBA, but reported performance varies across datasets and evaluation settings.
Takeaways & Limitations
Future DTA research should use more diverse datasets, consistent evaluation protocols, stronger cold-start evaluation, and combined molecular-graph and pretrained-protein representations.
Takeaways & Limitations
Most methods are evaluated on limited benchmark datasets, and substantial performance degradation occurs under cold-start settings.
Abstract
from arXiv · showhide
Computational approaches to drug discovery involve multiple sub-problems, and among them, drug-target binding affinity prediction plays an important role. Despite recent advances, accurately predicting binding affinity remains an open research area. The major objective of our paper is to perform a comprehensive review and comparative analysis of recent machine learning methods for drug-target binding affinity prediction, with a focus on identifying strengths, limitations, and research gaps. We review representative recent deep learning approaches that use common benchmark datasets and evaluation metrics, covering a range of neural network architectures and representation strategies. In addition, we analyze seven widely used benchmark datasets and commonly adopted evaluation metrics for drug-target binding affinity prediction. Our analysis indicates that although many methods report strong performance on standard benchmarks, their effectiveness is often influenced by dataset bias and limited evaluation settings. Furthermore, most methods exhibit reduced performance in cold-start scenarios, highlighting challenges in generalization. We identify several limitations of current approaches, including dataset imbalance, the lack of standardized evaluation, limited real-world applicability, and challenges in cold-start scenarios. We also discuss future research directions, including better dataset design, more robust evaluation methods, improved handling of cold-start problems, and the integration of multimodal representations.
1 Introduction
Drug–target binding affinity prediction supports early drug discovery, but current deep learning methods remain difficult to compare and generalize reliably. This review examines representations, architectures, datasets, evaluation practices, limitations, and future directions.
- Motivation: Drug discovery commonly uses computational approaches because experimental development is costly, time-consuming, and requires evaluating many candidate compounds.Developing a new therapeutic drug typically requires billions of dollars and more than a decade.
- Motivation: DTA prediction estimates how strongly a drug molecule interacts with a protein target, a property that influences drug effectiveness and candidate identification.Binding occurs at a protein binding site, and stronger affinity generally indicates a more stable and potentially effective interaction.
- Recent approaches: Deep learning approaches use diverse architectures and representations, including CNNs, GNNs, RNNs, transformers, molecular graphs, sequences, embeddings, and multimodal features.Multimodal approaches combine sequence, structural, and spatial information to represent complex drug–target interactions.
- Research challenges: Benchmark performance can be affected by dataset bias, data leakage, limited evaluation protocols, and inconsistent dataset characteristics.Datasets vary in size, quality, molecular representation, and experimental conditions, complicating fair comparison and cross-dataset generalization.
- Review scope: The paper systematically compares recent methods across representation strategies, model architectures, benchmark datasets, and evaluation methodologies.It also identifies current limitations and promising future research directions.
2 Methods
DTA prediction generally transforms drug and protein information into features, learns their relationships with deep learning models, and outputs a binding-affinity value. The methods section focuses on how input and output representations and their characteristics influence model performance.
- Input representations: Drug molecules are represented with SMILES strings and proteins with sequence or structural information before feature learning.These inputs provide the drug and protein information used by the prediction framework.
- Modeling: Deep learning models process drug and protein representations to learn useful patterns and relationships between them.The framework treats representation learning as the central modeling stage.
- Prediction: The model predicts binding affinity as a regression output.The pipeline produces a numerical binding-affinity value after representation learning.
- Scope: The methods discussion examines datasets, input and output representations, and how their characteristics influence model performance.This scope connects dataset properties and representation choices to downstream prediction behavior.
2.1 Datasets
The review analyzes seven commonly used DTA benchmark datasets and the representations and affinity measurements used to construct them. Affinity values include several experimental measures, with Kd and Ki often transformed logarithmically so higher values indicate stronger binding.
- Representations: Drug molecules are commonly represented as SMILES strings, while target proteins are represented as amino acid sequences in FASTA format.These representations provide compact molecular connectivity and protein primary-structure information.
- Affinity labels: Benchmark labels include Kd, pKd, Ki, IC50, EC50, and KIBA scores that quantify drug–target interaction strength.KIBA combines Kd, Ki, and IC50 measurements into one binding score.
- Benchmark datasets: The collected benchmarks include Davis, KIBA, PDBbind General, PDBbind Refined, CASF-2016, Metz, and BindingDB.The review compares their numbers of drugs, protein targets, and affinity ranges.
- Preprocessing: The preprocessing retains targets with at least 10 amino acid residues and drugs with at least 5 atoms.These thresholds are applied across the collected datasets.
- KIBA: KIBA contains 243,251 interaction records, 52,498 unique ligands, and 467 protein targets after invalid-score removal.Its scores range from approximately −3.10 to 17.80.
2.1.4 Davis
The reviewed datasets differ substantially in scale, molecular and protein sequence lengths, and affinity distributions. These differences include structural information in PDBbind, extensive interaction coverage in BindingDB, and distinct ranges and distributions across Davis, KIBA, Metz, and other benchmarks.
- Davis: Davis contains 379 protein targets, 68 compounds, and 30,056 interaction records, with pKd values ranging from 5.0 to 10.80.It is smaller than KIBA and BindingDB and provides experimentally measured kinase inhibition data.
- PDBbind: PDBbind v2020 contains 19,443 protein–ligand complexes and adds experimentally resolved three-dimensional structural coordinates.This distinguishes it from primarily sequence-based datasets such as KIBA and Davis.
- PDBbind refined set: The PDBbind refined set contains 5,316 protein–ligand complexes involving 5,316 protein targets and 4,203 ligands.It is selected using stricter criteria for crystallographic resolution, structural completeness, and experimental reliability.
- BindingDB: BindingDB contains 4,649 protein targets, 252,224 drug molecules, and 598,426 interaction records across multiple affinity measurement types.Its broad drug and target coverage creates substantial dataset diversity.
- Distribution statistics: Drug SMILES lengths vary widely, with BindingDB reaching nearly 2,000 characters and Davis and Metz showing shorter compounds.Average lengths remain below maxima, indicating a small number of long-compound outliers.
- Distribution statistics: Protein sequence lengths also vary, exceeding 7,000 amino acids in BindingDB and 4,000 residues in PDBbind, while Davis, Metz, and CASF-2016 are shorter.Affinity distributions differ as well: Davis and Metz are comparatively narrow, whereas KIBA and several other datasets are broader.
2.2 Target Sequence Similarity
Benchmark datasets show relatively low pairwise sequence similarity among their protein targets, with average identities ranging from approximately 0.173 to 0.214.
- Average pairwise sequence identity among benchmark-dataset target proteins ranges from approximately 0.173 to 0.214.
2.3 Input and Output Representations
DTA methods use varied drug and protein representations, combining sequence, graph, structural, biochemical, and pretrained features before predicting affinity. Most methods formulate prediction as regression, while some apply dataset-specific classification thresholds.
- Drug Representations: Atom type is the most frequently used drug feature, followed by atom degree, aromaticity, hydrogen count, and SMILES encoding.These features are commonly computed from molecular graphs derived from SMILES using cheminformatics toolkits such as RDKit.
- Drug Representations: Graph-based drug representations model atoms as nodes and chemical bonds as edges, with node and edge descriptors encoding molecular structure.
- Target Representations: Protein targets use sequence-based, structure-based, or hybrid representations, with amino acid encoding the most widely adopted feature.Other common target representations include contact maps, sequence embeddings, physicochemical properties, and secondary structure information.
- Target Representations: Recent methods increasingly use pretrained protein embeddings that capture contextual and evolutionary information from large-scale sequence databases.These embeddings often improve predictive performance compared with handcrafted or task-specific sequence encodings.
- Target Representations: Hybrid approaches combine sequential, structural, spatial, and biochemical information to represent complex drug–target interactions.
- Output Representations: Most methods treat affinity prediction as regression, while some convert it into classification using thresholds such as 7 for Davis and 12.1 for KIBA.Regression preserves more detailed information about interaction strength, whereas classification simplifies the prediction task.
3 Network architectures
Recent DTA studies span graph, convolutional, recurrent, hybrid, and transformer architectures, with GNNs frequently used because molecular compounds naturally form graphs. Evaluation relies on multiple metrics and validation strategies.
- Network architectures: Recent DTA studies employ GNNs, CNNs, LSTMs, and transformer-based models across their neural network architectures.
- Network architectures: GNN-based methods are frequently used because drug compounds can naturally be represented as molecular graphs.GNNs learn molecular representations through message passing and neighborhood aggregation.
- Network architectures: GCN models learn local graph structures through graph convolutions, whereas GAT models use attention to weight neighboring nodes during aggregation.
- Network architectures: Hybrid architectures combine GNNs with CNNs or LSTMs to integrate graph-based molecular representations with sequence-based learning.
- Network architectures: Transformer-based models use self-attention to capture long-range dependencies and global contextual relationships in molecular and protein representations.GEFormerDTA is an example of this architecture in DTA prediction.
- Evaluation metrics: Existing methods are evaluated with varied metrics and validation strategies to measure agreement between predicted and experimentally observed affinity values.
4 Results
The review compiles reported DTA prediction results across Davis and KIBA, comparing metrics, methods, and neural architectures. Reported performance is generally strong, with GNN-based and hybrid methods often showing consistent results, though performance varies across datasets.
- Reported performance overview: Reported results were collected from published DTA methods using widely used datasets and evaluation metrics, without accounting for differences in train–test splits.The collection is intended to show development, cross-dataset variation, and overall performance trends.
- Performance distribution: Most methods report CI values between approximately 0.89 and 0.91 on Davis and KIBA, while reported MSE is concentrated around 0.18–0.23 on Davis and 0.12–0.15 on KIBA.Figure 4 summarizes reported CI, MSE, and R2 distributions; the supplied passage gives the CI and MSE ranges.
- Comparative analysis: CM-DTA achieved the highest reported Davis CI at 0.929, while CSAN-BiLSTM-Att achieved the highest reported KIBA CI at 0.971.These are reported benchmark values from Figure 5A.
- Cross-dataset variation: Several methods show larger CI or MSE differences between Davis and KIBA, indicating less consistent or more dataset-dependent performance.Examples include CSAN-BiLSTM-Att, SimCNN-DTA, GEFormerDTA, DeepMHADTA, DeepCDA, BiComp-DTA, and ImageDTA.
- Comparative analysis: GEFormerDTA achieved the lowest reported Davis MSE at 0.060, while CSAN-BiLSTM-Att achieved the lowest reported KIBA MSE at 0.014.These are reported benchmark values from Figure 5B.
- Network architectures: GNN-based methods are often among the highest-performing approaches and generally show strong, consistent performance across Davis and KIBA.Hybrid architectures are also competitive, whereas CNN-based methods show greater variability across datasets.
5 Challenges
The review identifies challenges involving benchmark diversity, cold-start evaluation, pretrained protein language model integration, and interpretability. These constraints limit how reliably current DTA methods can be compared, generalized, and biologically validated.
- Limited benchmark dataset diversity: Evaluating DTA methods mainly on Davis and KIBA limits evidence about performance across the broader range of available benchmark datasets.The datasets differ in SMILES lengths, protein sequence lengths, and binding-affinity distributions.
- Limited benchmark dataset diversity: Larger and more diverse datasets integrating multiple sources are proposed to improve model robustness and generalization.This recommendation follows the reported diversity of dataset characteristics.
- Limited evaluation under cold-start settings: Most methods use conventional random train–test splits and do not evaluate unseen drugs, unseen targets, or unseen drug–target pairs.Only BiComp-DTA, AttentionMGT-DTA, and PocketDTA among the reviewed representative methods explicitly reported cold-start evaluations.
- Limited evaluation under cold-start settings: Under reported cold-start evaluations, CI decreased and MSE increased substantially for BiComp-DTA, AttentionMGT-DTA, and PocketDTA.Reported CI decreases ranged from 10.5% to 31.2%, while MSE increases ranged from 152.3% to 249.3% across the cited examples.
- Limited integration of pretrained protein language models: Only a few representative methods use pretrained protein language models, although PocketDTA, 3DProtDTA, and GEFormerDTA achieve competitive performance.The review suggests improving integration of structural and sequence-based representations.
- Limited model interpretability: Most methods rely on benchmark cross-validation, provide limited independent-dataset validation, and rarely offer biological interpretations of predictions.The review connects improved interpretability and broader validation with increased reliability of DTA prediction methods.
6 Conclusion
The paper reviews recent deep learning-based DTA methods and finds generally consistent performance for GNN-based models across common benchmarks. It also identifies limited benchmark coverage and substantial cold-start degradation, motivating broader datasets, standardized evaluation, and improved representations.
- GNN-based methods generally demonstrate consistent performance across the Davis and KIBA datasets.
- Recent methods increasingly combine graph-based drug representations with pretrained protein language models.
- Limited evaluation on benchmark datasets, particularly Davis and KIBA, makes robustness and generalization difficult to conclude.
- Most existing methods show substantial performance degradation in cold-start settings involving unseen drugs and targets.
- Future work should develop diverse datasets, consistent protocols, stronger cold-start handling, and better integration of molecular graphs with pretrained protein language models.
Statements and Declarations
The paper reports programmatic and computational support, declares no competing interests, and states that an ethics declaration is not applicable.
- The work was supported by the RISE Undergraduate Research Program at PVAMU and computational resources from the ACCESS program.
- The authors declare no competing financial or non-financial interests.
- The ethics declaration is stated as not applicable.