Adversarial Erasing Enhanced Multiple Instance Learning (siMILe): Discriminative Identification of Oligomeric Protein...

Area of Science:

• Cellular and Molecular Biology
• Biophysics
• Machine Learning in Biology

Background:

• Single-molecule localization microscopy (SMLM) enables nanoscale imaging of cellular protein structures.
• Analyzing structural variability in 3D SMLM data across different cell conditions remains a challenge.
• Interpretable discovery of subcellular structures requires methods that can handle complex, high-dimensional data.

Purpose of the Study:

• To develop a machine learning method for identifying condition-specific changes in protein assemblies from SMLM data.
• To enable interpretable subcellular discovery by analyzing shape and network features of protein structures.
• To overcome the limitations of current methods in capturing structural variability across diverse cellular contexts.

Main Methods:

• Implemented siMILe, a weakly supervised multiple instance learning method for SMLM data analysis.
• Utilized shape and network features for identifying protein assembly changes without requiring structure-level supervision.
• Enhanced structure classification using adversarial erasing and a symmetric classifier for improved instance selection.

Main Results:

• Validated siMILe on caveolin-1 (Cav1) labeled PC3 cells, differentiating caveolae structures based on cavin-1 expression.
• Demonstrated siMILe's ability to distinguish between caveolae, noncaveolar scaffolds, and 8S complexes based on cavin-1 association.
• Successfully applied siMILe to simulated SMLM data and clathrin-coated pit data to detect inhibitor-induced structural variations.

Conclusions:

• siMILe effectively identifies differential molecular structures in distinct cell conditions using SMLM data.
• The method provides interpretable insights into protein assembly dynamics and interactions.
• siMILe enhances the SuperResNET SMLM software platform, expanding its capabilities for comparative structural analysis.