Deep learning models link local cellular features with whole-animal growth dynamics in zebrafish

Area of Science:

• Developmental biology
• Computational biology
• Machine learning

Background:

• Cellular behavior and organism growth are interconnected.
• The relationship between microscopic cell features and macroscopic body size is not fully understood.
• Investigating cellular-level predictors of organismal growth is crucial.

Purpose of the Study:

• To determine if local cellular features correlate with macroscopic animal body size.
• To develop a deep learning model for predicting organism size from cellular images.
• To identify specific cellular features that predict body size.

Main Methods:

• Utilized 722 zebrafish larva skin cell images and size data.
• Employed machine learning, specifically a Vision Transformer (ViT) model.
• Implemented random cropping and voting strategies for prediction.
• Applied gradient-weighted class activation mapping (Grad-CAM) for feature identification.

Main Results:

• The Vision Transformer model achieved high predictive performance (F-score of 0.91).
• Predicting individual zebrafish size was possible with as few as 27 skin cells (0.01 mm²).
• Identified key cellular features influencing size prediction using Grad-CAM.

Conclusions:

• Macroscopic organismic information can be predicted from microscopic cellular data.
• Deep learning offers a powerful approach to de-encrypt organismal traits from cellular snapshots.
• This study provides a proof-of-concept for linking cellular-scale data to whole-organism characteristics.