New multiscale algorithm reconstructs protein structures from noisy cryo-EM images with SOTA accuracy
Using explicit protein backbone priors, this method beats standard approaches on noise and contrast.
A team from KTH Royal Institute of Technology (David Y. W. Thong, Ozan Öktem, Joakim Andén) has developed a novel multiscale algorithm for directly recovering atomic protein structures from single-particle cryo-EM images. The method goes beyond traditional approaches by explicitly representing the protein backbone in terms of bonds, torsion angles, and bond angles. This provides rich prior information that makes the reconstruction robust to high noise, low contrast, and even misspecifications in the transmission electron microscope (TEM) image formation model. The algorithm operates by first estimating large-scale structural features, then progressively refining to atomic-level detail, which helps avoid convergence to poor local minima.
When tested on three protein cryo-EM datasets generated using an electron microscope digital twin, the multiscale algorithm consistently outperformed standard methods. It achieved lower root-mean-square deviation (RMSD) and higher template modelling (TM) scores relative to the ground truth. The multiscale approach prioritized larger-scale structures early in the reconstruction, reducing the risk of getting stuck in suboptimal solutions. This work, submitted to the Journal of Structural Biology, represents a significant step toward automated, high-accuracy protein structure determination from noisy experimental data, with potential applications in drug discovery and structural biology.
- Explicit representation of protein backbone (bonds, torsion angles, bond angles) provides rich prior information for reconstruction.
- Achieves state-of-the-art accuracy on high-noise, low-contrast cryo-EM data; robust to misspecifications in the TEM image formation model.
- Multiscale approach prioritizes larger-scale structures first, reducing convergence to bad local minima and improving RMSD and TM scores.
Why It Matters
Enables more accurate protein structure determination from noisy cryo-EM data, accelerating drug discovery and structural biology.