ATC-NLSP

ATC-NLSP predicts World Health Organization Anatomical Therapeutic Chemical (ATC) classes for compounds using Network-based Label Space Partition (NLSP) to capture multilabel correlations among drug labels.

Key Features:

• Network-based Label Space Partition (NLSP): Uses a data-driven NLSP approach to partition the label space and model correlations among ATC labels within a multilabel learning framework.
• Similarity-based Features: Trained using chemical-chemical interactions and structural and fingerprint similarities of compounds to relate query compounds to known drugs.
• Community Detection Algorithms: Applies community detection to identify potentially intersecting label clusters in the label relation graph.
• Label Propagation: Employs label propagation on the label relation graph to detect community structures and propagate label information.
• Cluster-specific Predictors and Ensemble Labels: Trains predictors for each detected label cluster and combines their outputs as an ensemble to produce final multilabel ATC predictions.
• Multilabel Classification of WHO ATC Classes: Supports simultaneous prediction of multiple ATC classes per compound within the WHO ATC classification system.
• Evaluation by Jackknife Test: Experimental evaluation using the jackknife test on benchmark datasets reported an absolute true rate improvement from 0.6330 to 0.7497 compared to prior methods.
• Label-wise Analysis: Label-wise analyses indicate the multilabel learning approach outperforms single-label models.

Scientific Applications:

• ATC Classification: Predicts WHO Anatomical Therapeutic Chemical (ATC) classes for compounds to support drug classification tasks.
• Drug Discovery and Annotation: Facilitates annotation of compound therapeutic, pharmacological, and chemical properties in drug discovery and pharmacology research.

Methodology:

Network-based Label Space Partition (NLSP); similarity-based features including chemical-chemical interactions, structural and fingerprint similarities; community detection and label propagation on the label relation graph; training predictors for each label cluster and combining their outputs as an ensemble; evaluation by jackknife test on benchmark datasets.