A neural network based classification scheme for cytotoxicity predictions:Validation on 30,000 compounds

Area of Science:

• Computational chemistry
• Drug discovery
• Toxicology

Background:

• Eliminating cytotoxic compounds early in drug discovery significantly reduces research and development expenses.
• Predictive modeling for in vitro human cytotoxicity is crucial for efficient lead optimization.
• Current methods require extensive screening, incurring high costs and time investment.

Purpose of the Study:

• To develop and validate an artificial neural network (ANN) model for predicting in vitro human cytotoxicity.
• To utilize atomic fragmental descriptors for categorizing drug-like molecules based on toxicity.
• To enable early-stage filtering of potentially cytotoxic compounds in the drug development pipeline.

Main Methods:

• An artificial neural network was trained using atomic fragmental descriptors derived from the Atomic7 linear logP method.
• Cytotoxicity data from an in-house screening of 30,000 drug-like molecules were utilized.
• The model was trained on the most and least toxic compounds (12,998) and validated on the entire dataset.

Main Results:

• The ANN model demonstrated high accuracy in classifying compounds by their in vitro human cytotoxicity.
• Less than 5% of non-toxic and 9% of toxic compounds were misclassified by the trained neural network.
• The approach provides a reliable method for early identification of potential drug candidates.

Conclusions:

• The developed ANN approach effectively predicts in vitro human cytotoxicity using fragmental descriptors.
• This method offers a cost-effective strategy for filtering cytotoxic compounds during early drug discovery phases.
• Implementing this predictive model can streamline lead development and lead optimization processes, reducing R&D costs.