Predicting anatomic therapeutic chemical classification codes using tiered learning.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Pharmacology

Background:

• Drug discovery faces low success rates and high costs, necessitating novel identification paradigms.
• The World Health Organization's Anatomical Therapeutic Chemical (ATC) Code System classifies compounds by therapeutic, pharmacological, and chemical properties.
• Predicting ATC codes aids in creating high-quality chemical libraries for drug screening and drug repositioning.

Purpose of the Study:

• To propose and evaluate a machine learning architecture, tiered learning, for predicting ATC codes.
• To leverage the hierarchical structure of ATC codes to improve prediction efficiency and accuracy.

Main Methods:

• Developed a tiered learning architecture that uses predictions from higher ATC code levels to inform lower-level predictions.
• Validated the approach using cross-validation and test sets on a diverse set of compounds.
• Compared tiered learning against established methods using various chemical descriptors, initialization techniques, and classification algorithms.

Main Results:

• Tiered learning achieved prediction accuracy comparable to or better than established methods.
• The architecture demonstrated generalizability, improving prediction rates for most machine learning algorithms.
• The approach effectively constrains the learning space for ATC code prediction based on higher-level classifications.

Conclusions:

• The hierarchical nature of ATC codes allows for a constrained search space, improving prediction accuracy.
• Tiered learning capitalizes on the characteristic distribution of ATC codes.
• This method offers a simplified and more accurate approach to ATC code prediction, benefiting drug discovery efforts.