Deep learning framework for quantifying self-organization in Myxococcus xanthus.

Area of Science:

• Microbiology
• Developmental Biology
• Computational Biology

Background:

• Myxococcus xanthus undergoes multicellular development under starvation, forming fruiting bodies and differentiating cells.
• Understanding gene pathways requires quantifying subtle phenotypic changes, which is currently challenging.
• Existing methods lack reliable quantification of phenotype similarity for systematic pathway mapping.

Purpose of the Study:

• To develop a deep learning framework for quantifying multicellular phenotypes and self-organization dynamics in Myxococcus xanthus.
• To enable systematic mapping of gene pathways by providing reliable methods for phenotype similarity quantification.
• To explore the predictability of developmental trajectories from early-stage phenotypic signatures.

Main Methods:

• Applied a deep learning system integrating ResNet, StyleGAN2, and a Variational Autoencoder with a Siamese network.
• Trained the model on high-resolution microscopy images of 292 genetically distinct Myxococcus xanthus strains.
• Encoded images into 13-dimensional feature vectors capturing aggregation patterns and dynamics.

Main Results:

• The deep learning model accurately quantified phenotype patterns and self-organization dynamics.
• Feature vectors effectively captured variations in aggregation across time and strains.
• The model's feature space was interpretable, correlating with biological features like aggregate number and size.
• Developmental phenotypes and aggregation fate were predictable from early-stage images, across genetic and environmental variations.

Conclusions:

• Machine learning can reveal hidden aspects of complex multicellular dynamics.
• The developed methods allow for phenotypic analysis without manual annotation.
• Early-stage phenotypic signatures contain critical information about bacterial developmental trajectories.