Tox_(R)CNN: Deep learning-based nuclei profiling tool for drug toxicity screening.

Area of Science:

• Computational toxicology
• Drug discovery and development
• Cellular imaging and analysis

Background:

• Drug development is frequently hindered by toxicity, necessitating efficient prediction methods.
• Microscopy images of fluorescently labeled nuclei offer potential for toxicity prediction through pattern recognition.
• Deep learning, particularly convolutional neural networks (CNNs), excels at automated image analysis and complex pattern classification.

Purpose of the Study:

• To explore the utility of deep learning models for predicting drug-induced toxicity from microscopy images of cell nuclei.
• To develop and validate automated methods for toxicity prediction in drug screening.
• To assess the sensitivity, specificity, and transferability of the developed models.

Main Methods:

• Trained deep convolutional neural networks (CNNs) on DAPI-stained cell images pre-treated with various drugs.
• Employed different cropping strategies, focusing on nuclei-cropping for the Tox_CNN model.
• Implemented region-based CNNs (RCNN) for automated nuclei detection, classification, and per-cell toxicity prediction.

Main Results:

• The nuclei-cropping-based Tox_CNN model demonstrated superior performance in classifying cell health status.
• Tox_CNN enabled automated feature extraction, clustering compounds by their mechanism of action.
• The Tox_(R)CNN models showed higher sensitivity and broader specificity than existing toxicity assays, predicting toxicity across different drug mechanisms and cell assays.

Conclusions:

• The Tox_(R)CNN models provide robust, sensitive, and cost-effective tools for in vitro screening of drug-induced toxicity.
• These models can automate the extraction of predictive features from cell nucleus images.
• Adoption of these models can enhance compound prioritization in drug screening, increasing overall drug discovery efficiency.