Towards molecular structure discovery from cryo-ET density volumes via modelling auxiliary semantic prototypes

Area of Science:

• Structural biology
• Computational biology
• Machine learning

Background:

• Cryo-electron tomography (cryo-ET) aims to reveal novel biological structures.
• General Class Discovery (GCD) methods identify new classes using labeled base classes for pseudo-labeling unlabeled data.
• Existing 2D GCD methods face challenges in 3D applications due to model bias and limited feature transferability.

Purpose of the Study:

• To extend General Class Discovery (GCD) to 3D cryo-electron tomography (cryo-ET) data.
• To develop a novel approach for identifying unknown structures in cryo-ET.
• To overcome limitations of traditional methods in handling 3D structural data.

Main Methods:

• Utilized a pretrained 2D transformer with weight inflation for 3D adaptation.
• Integrated CLIP (Contrastive Language-Image Pre-training) for textual information incorporation.
• Employed a graph convolutional network with CLIP's text encoder and a decoupled prototypical network for semantic distance distributions.

Main Results:

• The proposed method successfully adapted 2D models for 3D cryo-ET data.
• Semantic distance distributions were devised for effective representation of unlabeled samples.
• Empirical results demonstrated the method's capability in discovering novel structures within cryo-ET datasets.

Conclusions:

• The developed approach effectively bridges the gap between 2D GCD and 3D cryo-ET challenges.
• This method offers a novel pathway for exploring and discovering previously unknown structures in cryo-ET.
• The integration of vision-language models enhances the potential of GCD in structural biology.