A recent study presented a novel computational method for inferring drug-target interactions, which exploits the primary protein sequence and molecular fingerprints of drug compounds, and evaluated the model’s predictive performance by conducting 5-fold Cross-Validations.
Inferring drug-target interactions
The identification of interacting drug-target pairs is of significant importance in drug discovery. In the past, the FDA declared that the demand for new drugs is hard to meet due to the adverse clinical outcomes of some new drug candidates. Researchers have previously used clinical experiments as the main approach for inferring drug target interactions. However, these traditional experiments are expensive and time consuming. Therefore, novel computer-aided drug development (CADD) methods need to be advanced in order to avoid these drawbacks.
Existing computational methods for inferring drug-target interactions
In recent years, researchers have made progress in predicting drug-target interactions by combining traditional computing methods and bioinformatics. The most common approaches are based on molecular docking and pharmacophores. Researchers use the molecular docking simulation to detect the optimal binding position between drug molecules and their targets based on matching energy. This method also requires knowledge of the complete three-dimensional (3D) substructures of the proteins, but these substructures are often hard to explore and determine by Nuclear Magnetic Resonance (NMR), electron microscopy, and X-ray crystallography.
Pharmacophores are a characteristic element of drug-active molecules, which play a pivotal role in the prediction of drug target interactions. Researchers suggest that the pharmacophore method can effectively infer drug target interactions for multi-target drugs and can therefore reduce the blindness of screening. However, a major drawback of this method is that when the conformation of the drug molecule changes, it will not match the existing pharmacophore model. Therefore, the formation of the pharmacophore model is still not comprehensive for further bioassays. At the same time, this method does not take 3D structures of targets into account, which reduces the accuracy of the pharmacophore model.
Subsequently, it is important to develop more robust and universal methods for inferring drug-target interactions, without the need for knowledge of the ligand or 3D structure.
A new method for inferring drug target interactions
In this study, the researchers present a novel ensemble learning-based method, which exploits the protein primary sequence and molecular fingerprints of drug compounds. More specifically the model integrates a pyramid histogram of oriented gradients (PHOG), position-specific scoring matrix (PSSM) and the rotation forest (RF) classifiers, for identifying drug-target interactions.
In practice, the protein primary sequences are first converted into PSSMs to describe the potential biological evolutionary information. Next, the researchers model employs PHOG to mine the representative features of PSSM from multiple pyramid levels, and the complete descriptors of drug-target pairs are generated by combining the molecular substructure fingerprints and PHOG features. The final step involves the researchers feeding the complete descriptors into the RF classifier to ensure effective prediction.
To verify the reliability further, the researchers applied 5-fold cross validations on enzyme, ion channel and nuclear benchmarking datasets. The results from this assessment showed that the model worked well on all of the four benchmark datasets. The model achieved mean accuracies of 88.96%, 86.37%, 82.88% and 76.92% on the enzyme, ion channel, G protein-coupled receptor (GPCR) and nuclear receptor data sets. This demonstrates that the PHOG features can trace the local characteristics and assist the model to improve accuracy when compared with other computational models for inferring drug-target interaction.
Summary
This study combined the PHOG, PSSM and RF descriptors into a novel computational model for inferring drug-target interactions and conducted a series of experiments to prove the reliability and accuracy of their proposed model. The model produced impressive mean accuracies on four benchmarking datasets, thereby demonstrating that it is able to effectively trace local characteristics. The researchers hope to go on to experiment with more existing models to further validate the accuracy and feasibility of their prediction methods.
Image credit: jigsawstocker – FreePik
