SelfAlign: Achieving Subtomogram Alignment with Self-Supervised Deep Learning

Cryo-Electron Tomography (Cryo-ET) and subtomogram averaging (STA) have been instrumental in advancing the analysis of high-resolution structural biology, enabling detailed insights into macromolecular complexes. However, due to limitations in sample thickness and electronic metrology, there are inherent issues with missing wedge artifacts and low signal-to-noise ratio in Cryo-ET. Researchers use STA to align and average subtomograms to address these two issues. Traditional STA methods, reliant on cross-correlation, are computationally expensive and not scalable for large datasets. The emerging method of using deep learning for STA has low accuracy and unstable performance at low signal-to-noise ratios. To address these issues, we proposed SelfAlign, a self-supervised deep learning approach for subtomogram alignment. To improve alignment accuracy, we introduce a rotation and translation method effectively reducing translation errors. Further, we present a self-labeling mechanism optimized for end-to-end processes,thereby abolishing the need for manual labeling. Additionally, we design a concise and efficient loss function to uphold stable training in scenarios with low signal-to-noise ratios. We demonstrate the efficacy of SelfAlign using four datasets, showcasing its superior performance in terms of alignment accuracy compared to existing methods. SelfAlign offers a robust and scalable solution for subtomogram analysis.