Integrative approach for predicting drug-target interactions via matrix factorization and broad learning systems.

Area of Science:

• Computational chemistry
• Bioinformatics
• Drug discovery

Background:

• Experimental screening of drug-target interactions (DTIs) is time-consuming and costly.
• Existing computational methods for DTI prediction often neglect interaction representations, leading to impaired performance.
• Data sparsity and incompleteness are significant challenges in DTI prediction.

Purpose of the Study:

• To propose an integrative approach, ConvBLS-DTI, for accurate DTI prediction.
• To capture comprehensive drug-target representations and simplify network structure.
• To reduce the impact of data sparsity and incompleteness in DTI prediction.

Main Methods:

• Weighted K-nearest known neighbors (WKNKN) for preprocessing unknown drug-target pairs.
• Neighborhood regularized logistic matrix factorization (NRLMF) for feature extraction.
• A broad learning network with a convolutional neural network (CNN) for DTI prediction.

Main Results:

• ConvBLS-DTI demonstrated superior performance compared to mainstream methods on benchmark datasets.
• The model achieved improved prediction accuracy on the area under the receiver operating characteristic curve (AUC) and precision-recall curve (PRC).
• The proposed method effectively handles data sparsity and incompleteness.

Conclusions:

• ConvBLS-DTI offers an effective and efficient computational approach for DTI prediction.
• The integrative method enhances prediction accuracy by considering comprehensive drug-target representations.
• This approach has the potential to accelerate the drug discovery process.