A noise-robust classification method for cryo-ET subtomograms with out-of-distribution detection

Area of Science:

• Structural biology
• Biophysics
• Computational biology

Background:

• Cryogenic electron tomography (cryo-ET) provides high-resolution 3D reconstructions of biological samples.
• Accurate subtomogram classification is essential for structural analysis in cryo-ET.
• Existing classification methods face challenges with out-of-distribution (OOD) data, leading to errors.

Purpose of the Study:

• To develop a unified framework for subtomogram classification that incorporates OOD detection.
• To enhance the accuracy and robustness of subtomogram classification methods.
• To distinguish between known (in-distribution, ID) and unknown (OOD) data classes.

Main Methods:

• A noise-robust classification method using a 3D discrete wavelet transform encoder.
• A Mahalanobis distance-based OOD detector tailored for 3D subtomograms.
• An adaptive classifier designed for datasets of varying scales.

Main Results:

• The proposed framework successfully distinguishes between ID and OOD data.
• The noise-robust method enhances subtomogram classification accuracy.
• The approach effectively models features and improves OOD detection capabilities.

Conclusions:

• The unified framework offers a significant improvement for subtomogram classification in cryo-ET.
• The integration of OOD detection addresses a critical limitation in current methods.
• This work provides a more reliable tool for structural analysis of biological macromolecules.