Learning Drug Functions from Chemical Structures with Convolutional Neural Networks and Random Forests

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Drug discovery is a costly and time-consuming process, with billions spent annually on empirical chemical testing.
• Predicting drug efficacy in silico could accelerate lead prioritization and reduce development expenses.
• Current methods often rely on transcriptomic data, which is limited to a subset of molecules.

Purpose of the Study:

• To determine if drug function, specifically MeSH therapeutic use classes, can be predicted solely from chemical structure.
• To compare the predictive power of chemical structure-based methods against transcriptomic data-based methods.
• To assess the potential of these models for predicting drug side effects and identifying repurposing opportunities.

Main Methods:

• Utilized two chemical-structure-derived classification methods: convolutional neural networks with chemical images and random forests with molecular fingerprints.
• Compared the performance of these methods against prior predictions based on drug-induced transcriptomic changes.
• Employed a multilabel classification strategy to evaluate model effectiveness.

Main Results:

• Chemical structure-based methods significantly outperformed previous predictions using transcriptomic data.
• The structure of a chemical contains as much information about its therapeutic use as its cellular transcriptional response.
• Achieved high predictive accuracy ranging from 83% to 88% by leveraging larger training datasets based on chemical structure.

Conclusions:

• Chemical structure is a powerful predictor of drug therapeutic use, offering a more data-rich approach than transcriptomic responses.
• The developed in silico models enhance drug discovery efficiency by improving lead prioritization and reducing costs.
• This strategy shows promise for predicting side effects and discovering novel drug repurposing applications.