A Multimodal Sequence-to-Sequence Model for Automatic Assignment of ATC Codes in Drug Discovery and Repurposing

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• The Anatomical Therapeutic Chemical (ATC) system classifies drugs by therapeutic use.
• Accurate ATC code prediction is vital for drug repurposing and clinical trials.
• Challenges include ATC code hierarchy, polypharmacology, and data limitations.

Purpose of the Study:

• To develop a novel multimodal generative approach for predicting ATC codes.
• To leverage complementary molecular information from different representations.
• To address the multilabel and data scarcity issues in ATC code prediction.

Main Methods:

• A sequence-to-sequence architecture utilizing multimodal generative AI.
• Employing two distinct molecular representations: SMILES strings and molecular descriptors.
• Developing a method to determine the optimal stopping point for generating ATC labels.

Main Results:

• The multimodal approach demonstrated superior performance compared to baseline methods.
• Improved accuracy in predicting ATC codes for both new drugs and in drug repurposing scenarios.
• Validation of the hypothesis that combined molecular representations enhance prediction accuracy.

Conclusions:

• The proposed multimodal generative approach offers a significant advancement in automated ATC code prediction.
• This method effectively handles the complexities of drug classification, including polypharmacology and data imbalance.
• The publicly available code and datasets facilitate further research and application in drug discovery.