Geometric deep learning and multiple-instance learning for 3D cell-shape profiling.

Area of Science:

• Computational biology
• Cell biology
• Biophysics

Background:

• Cellular 3D morphology reflects cell state and function.
• Understanding morphology-phenotype relationships is crucial for biological insights.

Purpose of the Study:

• To develop a deep learning pipeline (MorphoMIL) for profiling 3D cell and nuclear shapes.
• To analyze morphological signatures and understand cell states under various perturbations.
• To apply MorphoMIL for drug discovery and dissecting phenotypic heterogeneity.

Main Methods:

• Utilized geometric deep learning and attention-based multiple-instance learning.
• Processed 3D point-cloud data for morphological signature extraction.
• Applied the pipeline to over 95,000 melanoma cells with chemical and genetic perturbations.

Main Results:

• MorphoMIL accurately predicted drug perturbations and cell states.
• Identified subtle morphological changes linked to perturbations.
• Revealed key shapes correlating with signaling activity and cell-state heterogeneity.

Conclusions:

• MorphoMIL provides scalable, high-throughput morphological profiling.
• The framework offers interpretable insights into cell biology and drug responses.
• Demonstrated superior performance and generalization across datasets for drug discovery.