Utilizing high throughput screening data for predictive toxicology models: protocols and application to MLSCN assays.

Area of Science:

• Toxicology
• Cheminformatics
• Computational Biology

Background:

• Computational toxicology presents an alternative to traditional experimental methods.
• The NIH Molecular Libraries Screening Center Network (MLSCN) provides large, public screening datasets.
• Predictive models for cell toxicity can aid in compound evaluation and elimination.

Purpose of the Study:

• To develop computational models for predicting cell toxicity using MLSCN cell proliferation data.
• To address the challenges posed by imbalanced datasets in toxicity prediction.
• To explore the correlation between in vitro cell proliferation and in vivo animal toxicity.

Main Methods:

• Development of random forest ensemble models using MLSCN cell proliferation datasets.
• Implementation of protocols to handle imbalanced data.
• Comparison of in vitro proliferation data with animal acute toxicity data.
• Development of a visualization technique for dataset comparison.

Main Results:

• Achieved 70-85% correct classification rates on prediction sets for cell toxicity.
• Observed a significant drop in accuracy when models were applied to in vivo data.
• Demonstrated a correlation between MLSCN cell proliferation results and animal acute toxicity data.
• Presented a visualization tool for assessing the reliability of predicting new datasets.

Conclusions:

• Computational models can predict cell toxicity with moderate accuracy using MLSCN data.
• Extrapolation of in vitro cell proliferation models to in vivo toxicity requires caution.
• Further research is needed to improve the prediction of animal toxicity from cell-based assays.