Robustness of signal detection in cryo-electron microscopy via a bi-objective-function approach

Wei Li Wang, Zhou Yu, Luis R. Castillo-Menendez, Joseph Sodroski, Youdong MaoBMC Bioinformatics 2019 · DOI

TL;DR

This paper introduces a "bi-objective function" (BOF) approach for detecting weak signals in cryo-electron microscopy (cryo-EM) data, combining Fast Local Correlation (FLC) for initial particle picking and Maximum Likelihood Estimation (MLE) for verification. This method proves robust in identifying particles even at very low signal-to-noise ratios (SNRs), crucial for high-resolution structural determination.

What Problem Does This Paper Address?

Cryo-electron microscopy (cryo-EM) is a powerful technique for determining the 3D structures of biological molecules. However, a major hurdle is accurately identifying and selecting the individual particle images (micrographs) that contain the molecule of interest from noisy experimental data. These micrographs often have very low signal-to-noise ratios (SNRs), sometimes as low as 0.005-0.05, due to factors like electron scattering, imaging noise, and the weak-phase approximation. Manual selection is time-consuming and subjective. Automated methods exist, typically involving two steps: particle picking (finding potential particle locations) and particle verification (confirming they are indeed particles). While Fast Local Correlation (FLC) is efficient for picking, it's prone to false positives due to noise and "reference bias." Maximum Likelihood Estimation (MLE) is better at verification and less susceptible to reference bias, but computationally more intensive. This paper addresses the gap of how to optimally combine these two complementary approaches to achieve robust and automated signal detection, especially in challenging low-SNR conditions.

Methods

The core of this work is the Bi-Objective Function (BOF) approach, which integrates two distinct objective functions:

  • Particle Picking (Objective Function A): Fast Local Correlation (FLC)
    • This is a standard template-matching technique. A template image (e.g., a simulated particle view or a Gaussian circle) is slid across the micrograph, and the cross-correlation is calculated. Peaks in the correlation map indicate potential particle locations.
  • The paper uses SPIDER software for FLC-based particle picking.
  • Particle Verification (Objective Function B): Maximum Likelihood Estimation (MLE)
    • MLE is used to align the picked particles and verify their authenticity. It estimates the underlying signal by maximizing the probability of observing the data.
  • The paper highlights the importance of the starting reference for MLE. They explore using:
  • A randomly chosen noise image.
  • A Gaussian circle (which provides no prior structural information).
  • An unaligned average of the FLC-picked particles (which can introduce reference bias).

The BOF strategy involves using FLC to pick candidate particles and then employing MLE with different starting references to verify these candidates. The key innovation lies in systematically evaluating how the choice of MLE starting reference (especially the non-informative Gaussian circle) impacts the suppression of reference bias introduced by FLC, particularly at low SNRs.

Datasets

The study utilizes both simulated and experimental datasets:

  • Simulated Datasets:
    • Pure Noise Micrographs: 200 simulated micrographs with only Gaussian noise were generated to test the baseline performance and reference bias.
  • Simulated Influenza A Virus HA Trimer Micrographs: 120 noiseless micrographs were generated and then corrupted with Gaussian noise to create datasets with a range of SNRs (from 0.1 down to 0.0005). These were used to systematically evaluate the BOF's performance across different SNR levels.
  • Simulated Glucose Isomerase Micrographs: Not explicitly detailed as a separate dataset, but the principles tested on HA trimers are applied to this complex.
  • Experimental Datasets:
    • Experimental Ice Noise Micrographs: 218 micrographs of pure ice noise were collected from a cryo-EM session without any protein sample. This served as a control for experimental noise behavior.
  • Experimental Glucose Isomerase Complex: A dataset of the 173-kDa glucose isomerase complex was used. This dataset was particularly challenging due to low defocus, resulting in low single-particle SNRs (approximately 0.005–0.01).

Key Insights and Ideas

  • Complementarity of FLC and MLE: FLC is fast for picking, but prone to noise and bias. MLE is slower but more robust for verification. Combining them leverages their strengths.
  • Critical SNR Threshold: The BOF approach starts to lose its ability to reliably select and verify particles at SNRs around 0.005. This is a crucial finding for understanding the limits of automated detection.
  • Gaussian Circle as a Neutral Reference: Using a Gaussian circle as the starting reference for MLE significantly suppresses the reference bias that can be inherited from the FLC picking step. This is a key strategy for robust verification.
  • Reference Bias Transfer: Reference bias from FLC does not necessarily transfer to MLE when initialized with a neutral reference like a Gaussian circle. This is a major advantage for detecting weak signals.
  • Template Choice Matters for Picking, Less for Verification (at higher SNRs): While the choice of FLC template can affect particle centering and false positive rates, the MLE verification step, especially with a good reference, can overcome template-induced biases when the signal is strong enough.
  • Experimental Noise Behaves Like Gaussian Noise: The study found that experimental cryo-EM noise from amorphous ice behaves similarly to simulated Gaussian noise, validating the use of simulated data for testing.

Key Results

  • FLC Performance: FLC could successfully pick particles from simulated micrographs with SNRs as low as 0.002-0.005.
  • False Positive Rate: At SNRs of 0.005 and above, the false-positive rate for FLC picking was below 2.5%, increasing to ~7% at SNR 0.002.
  • BOF Robustness: For SNRs above 0.002, the BOF approach consistently recovered the correct particle structure (HA trimer) regardless of the FLC template or MLE starting reference (except for the unaligned average).
  • Suppression of Reference Bias: When a Gaussian circle was used as the MLE starting reference, the Fourier Ring Correlation (FRC) analysis showed a significant reduction in correlation between the class averages and the FLC picking template, indicating effective removal of bias. This was particularly evident at SNRs of 0.005 and above.
  • Experimental Validation: The BOF approach successfully generated class averages corresponding to the glucose isomerase complex from real cryo-EM data, even with low SNRs, especially when using a Gaussian circle for MLE initialization.

Limitations and Future Work

  • SNR Limit: The BOF procedure became inefficient at verifying signals at SNRs of 0.001 and below, resulting in blank or blurred class averages.
  • Template Choice for Picking: While a Gaussian circle is good for initial picking, it might miss particles with unique shapes. The authors suggest using a low-pass filtered model of the particle or exploring deep learning approaches for more complex particle shapes.
  • False Positives: FLC can still pick false positives (e.g., ice contamination), which may require manual curation or more advanced machine learning methods for removal.
  • Future Work: The authors suggest that replacing FLC with deep convolutional neural networks or other manifold learning-based methods could further improve the BOF framework's performance. They also note that their study was limited to the FLC-MLE combination, and other objective function pairs could be explored.

Summary updated

BibTeX
@article{wang2019robustness,
  author = {Wei Li Wang and Zhou Yu and Luis R. Castillo-Menendez and Joseph Sodroski and Youdong Mao},
  title = {{Robustness of signal detection in cryo-electron microscopy via a bi-objective-function approach}},
  journal = {BMC Bioinformatics},
  year = {2019},
  doi = {10.1186/s12859-019-2714-8},
}

Loading related content...

Figure