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

Wenjia Meng, Xueshi Yu, Tingting Zhang, Renmin Han

TL;DR

This paper introduces a novel framework for classifying cryo-electron tomography (cryo-ET) subtomograms that robustly handles noisy data and, crucially, identifies and discards "out-of-distribution" (OOD) subtomograms. This OOD detection prevents misclassification of unknown structures, leading to more accurate downstream analysis and reliable 3D reconstructions.

What Problem Does This Paper Address?

Cryo-ET is a powerful technique for visualizing biological molecules at near-atomic resolution. A key step in analyzing cryo-ET data is subtomogram classification, where individual 3D reconstructions (subtomograms) of molecules are grouped based on their type. However, existing classification methods struggle when faced with subtomograms that don't belong to any of the known classes (OOD data). These OOD subtomograms are often misclassified as one of the known classes, leading to errors in subsequent steps like alignment and averaging. This ultimately results in incorrect 3D structural models, hindering biological discovery. The paper aims to solve this by developing a classification method that can reliably distinguish between known (in-distribution, ID) and unknown (OOD) subtomograms, and then accurately classify only the ID ones.

Methods

The core of the proposed solution is a unified framework that integrates OOD detection with subtomogram classification. The framework has three main components:

  • 3D DWT-based Encoder: This component uses a 3D discrete wavelet transform (DWT) to process subtomograms. DWT is effective at decomposing signals into different frequency components, allowing the encoder to filter out high-frequency noise and extract more robust features from the noisy cryo-ET data. This is achieved through multiple "wavelet blocks" that iteratively apply DWT, convolution, batch normalization, and ReLU activation.
  • Mahalanobis Distance-based OOD Detector: To identify OOD data, the paper employs a Mahalanobis distance metric. This metric is well-suited for high-dimensional data and captures the intrinsic correlations between features by considering the covariance of the data. By calculating the Mahalanobis distance of a subtomogram's features to the distribution of known classes, the detector can effectively determine if it's an OOD sample. A confidence score is derived from these distances, and a threshold is used to classify samples as ID or OOD.
  • Adaptive Classifier: For the ID subtomograms, an adaptive classifier is used. This classifier is designed to handle datasets of varying scales. It incorporates an adaptive inner product operation between classifier weights and feature vectors, making it adaptable to both small real-world datasets and large simulated ones.

The overall workflow involves passing a subtomogram through the encoder, then the OOD detector. If identified as OOD, it's discarded; if ID, it proceeds to the adaptive classifier for label prediction.

Datasets

The study utilizes three datasets for evaluation:

  • SHREC19: A dataset containing 12 macromolecule classes and 20,785 samples.
  • SHREC21: A dataset with 13 classes and 16,291 samples.
  • Real Dataset: A dataset from the CryoET Data Portal, comprising 12 classes and 1,857 samples.

For training and testing, tomograms were segmented into 32x32x32 subtomograms. The paper also constructed several "ID&OOD" test sets by combining data from different sources to simulate real-world scenarios where unknown classes might be present.

Key Insights and Ideas

  • OOD detection is critical for cryo-ET: The paper strongly emphasizes that simply classifying all data, even if it's unknown, leads to significant downstream errors. Discarding OOD data is a crucial step for reliable structural analysis.
  • Wavelet transforms are powerful for noisy cryo-ET data: The use of 3D DWT effectively denoises the subtomograms, leading to more discriminative features and improved classification accuracy.
  • Mahalanobis distance offers robust OOD detection: This metric proves more reliable than simpler methods for identifying unknown subtomograms in the context of cryo-ET data.
  • Adaptability is key for real-world applications: The adaptive classifier ensures the method performs well across datasets of different sizes and characteristics.

Key Results

The proposed method demonstrates significant improvements over existing approaches:

  • Enhanced Classification Accuracy: When integrated with existing classification methods (like CB3D, DSRF3D-v2, RB3D, CFN), the proposed OOD detection framework consistently improved accuracy, often by substantial margins (e.g., CB3D accuracy increased by over 26% on one dataset).
  • Superior OOD Detection: The method achieved near-perfect OOD detection, with an FPR95 of 0.00% and AUROC of 99.99% on some datasets, significantly outperforming other OOD detectors like KNN and VIM.
  • Robustness: The noise-robust encoder and adaptive classifier contribute to high performance across various datasets, including real-world cryo-ET data.
  • Visualization: t-SNE visualizations show that the proposed method creates a more compact and well-separated embedding space for ID features, with OOD features clearly distinct, unlike other methods where OOD and ID features overlap.

Limitations and Future Work

The authors acknowledge that while their method excels at distinguishing structures with clear differences, classifying subtomograms with minor structural variations (e.g., different conformations of the same molecule) remains a challenge. They suggest that extending their method to handle such subtle differences will be a key focus for future research.

Summary updated

BibTeX
@article{meng2025noise,
  author = {Wenjia Meng and Xueshi Yu and Tingting Zhang and Renmin Han},
  title = {{A noise-robust classification method for cryo-ET subtomograms with out-of-distribution detection}},
  journal = {Bioinformatics},
  year = {2025},
  doi = {10.1093/bioinformatics/btaf274},
}
Figure