MultiCell: geometric learning in multicellular development

Area of Science:

• Developmental Biology
• Computational Biology
• Biophysics

Background:

• Understanding cell self-organization during embryogenesis is crucial but challenging.
• Predicting individual cell behavior in living tissues over time remains a significant hurdle.
• Existing methods struggle to capture complex intercellular interactions.

Purpose of the Study:

• To present MultiCell, a novel geometric deep learning method for analyzing multicellular dynamics.
• To represent complex multicellular data using a unified graph structure.
• To predict single-cell behaviors during developmental processes with high resolution.

Main Methods:

• Developed MultiCell, a geometric deep learning framework.
• Utilized a unified graph data structure to represent cellular interactions and cell junction networks.
• Applied the method to four-dimensional morphological sequence alignment.
• Employed neural activation maps and model ablation studies.

Main Results:

• MultiCell accurately captures convoluted interactions among cells.
• Achieved interpretable 4D morphological sequence alignment.
• Successfully predicted single-cell behaviors during Drosophila embryogenesis at single-cell resolution.
• Demonstrated the essential roles of cell geometry and junction networks in predicting cell behaviors.

Conclusions:

• MultiCell provides a data-driven approach for quantitative studies of dynamic multicellular processes.
• The method enables precise prediction of cell behaviors during development.
• This work offers a pathway toward a unified morphodynamic atlas with single-cell precision.