TL;DR
This paper introduces DAWN, a novel computational method for predicting drug-target interactions (DTIs). By combining drug substructure fingerprints, discrete wavelet transform (DWT) features from target protein sequences, and network topology information, DAWN achieves competitive prediction accuracy, particularly on balanced datasets.
What Problem Does This Paper Address?
Discovering new drug-target interactions (DTIs) is a cornerstone of drug discovery, but experimental methods are time-consuming and expensive. While computational approaches exist, there's a continuous need for more accurate and efficient methods to predict potential DTIs, especially given the vast number of chemical compounds and limited known interactions. Existing methods often fall into categories like molecular docking, matrix-based approaches, or feature vector-based methods, each with its own limitations. Molecular docking is computationally intensive, while matrix-based methods can struggle with high dimensionality and may not fully capture the physical/chemical properties of target proteins. This paper aims to address these limitations by proposing a feature-vector-based approach that integrates diverse data sources for improved DTI prediction.
Methods
The DAWN method employs a multi-faceted approach to feature extraction and prediction:
- Drug Representation: Drugs are encoded using MACCS fingerprints, which represent the presence or absence of common chemical substructures.
- Target Representation: The core novelty for target representation lies in the application of Discrete Wavelet Transform (DWT) to protein sequences. Protein sequences are first translated into numerical sequences based on six physicochemical properties of amino acids (hydrophobicity, volume, polarity, polarizability, solvent-accessible surface area, and net charge). DWT is then applied to these sequences to extract features that capture both low-frequency (approximated) and high-frequency (detailed) components, effectively compressing and denoising the signal. The paper specifically mentions using Nanni et al.'s method for DWT feature extraction, which involves calculating maximum, minimum, mean, and standard deviation values at different decomposition scales.
- Network Features: Information about existing drug-target interaction networks is incorporated. This is done by creating a binary vector representing known and unknown targets for a given drug, and vice-versa for a target.
- Feature Concatenation and Selection: The extracted drug, target (DWT), and network features are concatenated and normalized to form a comprehensive feature vector. Feature selection (using SVM-RFE+CBR) is then applied to identify the most important features, aiming to improve generalization and reduce overfitting.
- Classification: Finally, a Support Vector Machine (SVM) classifier, specifically using a Radial Basis Function (RBF) kernel, is trained on these feature vectors to predict DTIs.
Datasets
The study utilizes four benchmark datasets commonly used in DTI prediction research:
- Enzyme
- Ion Channel
- Guanosine-binding protein coupled receptor (GPCR)
- Nuclear Receptor
These datasets were sourced from the Kyoto Encyclopedia of Genes and Genomes (KEGG) database and other related databases. The paper distinguishes between balanced datasets (where positive and negative interaction pairs are roughly equal) and imbalanced datasets (where known interactions are far fewer than potential non-interactions).
Key Insights and Ideas
- DWT as a powerful feature extractor for protein sequences: The paper highlights that DWT can effectively capture complex patterns within protein sequences by analyzing them as signals, offering a novel way to represent target proteins beyond traditional sequence descriptors.
- Synergy of diverse features: Combining drug chemical features (MACCS), target biological features (DWT), and network topology information leads to improved prediction performance compared to using subsets of these features.
- Feature selection is crucial: Applying feature selection significantly enhances the prediction accuracy, demonstrating its importance in building robust DTI models.
- Network topology adds value: Incorporating network information provides a valuable context for DTI prediction, complementing molecular and sequence-based features.
Key Results
- Balanced Datasets: DAWN achieved competitive prediction performance on balanced datasets. For instance, the Area Under the ROC Curve (AUC) for DAWN with feature selection (DWT + NET + MACCS (FS)) reached 0.980 for Enzyme and 0.931 for Nuclear Receptor.
- Imbalanced Datasets: On imbalanced datasets, DAWN demonstrated strong performance, achieving average AUCs of 0.981 (Enzyme), 0.990 (Ion Channel), 0.952 (GPCR), and 0.906 (Nuclear Receptor). The Area Under the Precision-Recall Curve (AUPR) values were 0.895 (Enzyme), 0.921 (Ion Channel), 0.786 (GPCR), and 0.603 (Nuclear Receptor).
- Comparison with existing methods: DAWN showed comparable or superior performance to several state-of-the-art methods, including BLM, RLS, BGL, NetLapRLS, BLM-NII, CMF, KBMF2K, and NRLMF, across various datasets.
Limitations and Future Work
- Imbalanced Data: The authors acknowledge that the imbalanced nature of DTI prediction remains a challenge, and SVM, while effective, can struggle with highly imbalanced datasets. This is reflected in the lower AUPR for the Nuclear Receptor dataset.
- Dataset Size: The performance on the smaller Nuclear Receptor dataset might be affected by its size.
- Future Directions: The paper suggests that further improvements could be achieved by:
- Exploring more refined representations of structural and physicochemical properties.
- Investigating alternative machine learning models that might be better suited for imbalanced data.
- Developing strategies for sparse representation and gradient boosting decision trees.
- Building classification models using bootstrap sampling and sub-classifier weighting.
Summary updated