NegStacking

NegStacking improves prediction of drug-target interactions (DTIs) by using a stacking ensemble that leverages multiple negative sample sets, feature subspacing, hyperparameter perturbation, and a logistic regression meta-learner to mitigate class imbalance.

Key Features:

• Stacking ensemble: Combines an ensemble of weak learners with a logistic regression meta-learner to produce final DTI predictions.
• Negative sample generation: Generates multiple distinct sets of negative samples via sampling techniques to exploit the large pool of non-interacting pairs.
• Weak learners trained on varied negatives: Trains each weak learner on a unique negative sample set paired with a consistent set of positive samples to increase learner diversity.
• Feature subspacing: Applies feature subspacing during training to promote diversity among weak learners.
• Hyperparameter perturbation: Introduces hyperparameter perturbation across weak learners to further diversify the ensemble.
• Class imbalance mitigation: Specifically addresses imbalance between known interacting pairs and non-interacting pairs in DTI prediction.
• Empirical performance: Experimental evaluations demonstrated improved accuracy and reliability in predicting new drug-target interactions compared to other approaches.

Scientific Applications:

• DTI prediction: Predicts potential interactions between drugs and biological targets for drug discovery studies.
• Novel target identification: Supports identification of novel therapeutic targets by prioritizing likely drug-target pairs.
• Drug development optimization: Assists in optimizing candidate selection and prioritization during early-stage drug development.
• Bioinformatics and pharmacology research: Serves as a computational approach for studies requiring large-scale DTI screening and imbalance handling.

Methodology:

NegStacking generates multiple negative sample sets via sampling techniques, trains an ensemble of weak learners each on a unique negative set with a consistent positive set using feature subspacing and hyperparameter perturbation, and combines learner outputs with a logistic regression meta-learner.