PECAN Predicts Patterns of Cancer Cell Cytostatic Activity of Natural Products Using Deep Learning.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Oncology

Background:

• Machine learning accelerates drug discovery by linking compound structure to biological function.
• Current models often provide binary activity predictions for limited drug targets, especially in anticancer research.
• A more nuanced approach is needed to predict drug efficacy across diverse cancer types.

Purpose of the Study:

• To introduce PECAN (Prediction Engine for the Cytostatic Activity of Natural product-like compounds), a novel feed-forward neural network.
• To enable simultaneous classification of antiproliferative activity for compounds against 59 cancer cell lines.
• To predict not just the presence but also the degree of biological activity.

Main Methods:

• Development of a feed-forward neural network architecture.
• Training and validation using National Cancer Institute (NCI) data.
• Evaluation of classification accuracy and analysis of compound structural features.

Main Results:

• PECAN achieved 60.1% accuracy in a six-way classification task.
• The model demonstrated high performance with a "within-one" measure accuracy of 93.0%.
• Evidence suggests PECAN utilizes meaningful structural features for its predictions.

Conclusions:

• PECAN offers a significant advancement in predicting anticancer drug activity.
• The model's ability to provide graded activity predictions enhances its utility in drug discovery.
• PECAN's reliance on structural features indicates its potential for identifying novel therapeutic compounds.