Challenge: The transformative power of AI, particularly deep learning approaches, in biomedical image analysis is undeniable. However, most of these methods require large amounts of publicly available training data. Particularly in the field of cryo-electron tomography (cryo-ET), the lack of such data represents a critical roadblock, hindering the development and application of deep learning methods for cryo-ET data analysis. As a result, cryo-ET data annotation is currently slow and laborious, often performed by hand, hampering progress in unlocking the full potential of this powerful tool that is revolutionizing our understanding of the intricate world of cells.

Solution: To meet this challenge, the groups of former Helmholtz Pioneer Campus PI Ben Engel (now Biozentrum Basel) and Helmholtz AI PI Tingying Peng joined forces to develop and implement data-efficient AI-based solutions for analyzing membranes requiring minimal amounts of training data. Their PhD student Lorenz Lamm’s latest tool, called MemBrain v2, was published as a preprint in bioRxiv and the code was made publicly available via Github. One notable component of MemBrain v2 is MemBrain-seg: a program for generalizable membrane segmentation, delivering excellent results in a matter of minutes without the need to re-train the model with user data. This success is due to robust design choices during training, such as specialized data augmentations and loss functions. The team also gathered a comprehensive training dataset with multiple membrane variations by carefully annotating many tomograms, as well as collaborating with several cryo-ET expert groups and standardizing the training data.

Proof of concept: MemBrain-Seg was adapted by CZII scientist Utz Ermel to segment over 13,000 tomograms deposited on CZII’s CryoET Data Portal, which aims to provide access to high-quality, standardized data for biologists and developers to train and build new AI models. The segmentation was completed in just a few days, highlighting the value and power of this newly developed tool. The created segmentations can now be visualized directly on the data portal without the need to download and process the data locally.

Significance: This collaborative effort, aimed at providing open-source tools and shared data, demonstrates the power of scientific teamwork to unlock the full potential of cryo-ET analysis. The steady progress of the MemBrain algorithm, which first facilitated the detection of membrane proteins, then the segmentation of membranes, showcases the global contribution and impact of interdisciplinary work performed at the Helmholtz PioneerCampus, HelmholtzMunich and Helmholtz_AI.