Unsupervised Identification of Protein Compositions and Conformations via Implicit Content-Transformation...

Area of Science:

• Structural biology
• Machine learning
• Biophysics

Background:

• Identifying protein composition and conformation from microscopic images is difficult.
• Existing methods struggle with the complex, unparameterized nature of protein conformation.

Purpose of the Study:

• To develop a novel method for accurately identifying protein composition and conformation from 3D microscopic images.
• To overcome limitations of existing disentanglement methods that require explicit transformation parameterization.

Main Methods:

• Proposed DualContrast, a contrastive learning-based method for implicit disentanglement.
• Modeled protein composition as content and conformation as transformation.
• Generated positive and negative pairs in data and latent spaces for content and transformation.

Main Results:

• DualContrast successfully disentangles protein composition and conformation.
• Demonstrated superior performance compared to existing self-supervised methods.
• Achieved the first fully unsupervised identification of protein composition and conformation in 3D microscopic images.

Conclusions:

• DualContrast offers a powerful new approach for analyzing protein structures in microscopy.
• The method advances unsupervised learning in structural biology.
• Enables accurate identification of protein mixtures without prior structural knowledge.