Superhuman cell death detection with biomarker-optimized neural networks.

Area of Science:

• Neuroscience
• Computational Biology
• Biomedical Imaging

Background:

• Longitudinal live microscopy of neurons is crucial for understanding neurodegenerative diseases.
• High-throughput microscopy requires extensive human annotation, limiting scalability and speed.
• Existing methods struggle to capture subtle cellular events indicative of disease progression.

Purpose of the Study:

• To address the bottleneck of human annotation in high-throughput live microscopy.
• To develop an automated method for detecting cell death using artificial intelligence.
• To identify novel morphological indicators of cell death through interpretable AI models.

Main Methods:

• Development of biomarker-optimized convolutional neural networks (BO-CNNs), a type of interpretable computer vision model.
• Training BO-CNNs directly on biosensor activity data from live neuron microscopy.
• Evaluating BO-CNN performance in detecting cell death compared to human annotators.

Main Results:

• BO-CNNs accurately detected cell death with superhuman speed and accuracy.
• The models learned to identify subcellular morphology associated with cell vitality without explicit supervision.
• A novel intranuclear morphology signal linked to cell death was discovered by BO-CNNs.

Conclusions:

• BO-CNNs offer a broadly applicable and essential tool for analyzing live microscopy data.
• Automated analysis using BO-CNNs significantly enhances the interpretation of high-throughput experiments.
• This approach accelerates neurodegenerative disease research by overcoming annotation limitations.