Lorenz Lamm, a PhD student at Helmholtz AI, has developed MemBrain, a Deep Learning-based programme that analyzes 3D tomograms to study membranes and its embedded proteins.
MemBrain: Detection of membrane-bound proteins in high-resolution 3D microscopy
Cryo-electron tomography (Cryo-ET) is a high-resolution imaging technique that can visualize the interior of single cells in 3D and at sub-nanometer resolution. It is used to study the structure, organization, and interactions between different compartments within cells, like membranes or proteins. However, the analysis of these 3D tomograms is time-consuming. One particularly tedious task is the examination of membranes and their embedded proteins. The new software “MemBrain” is focused on facilitating this. We asked Lorenz Lamm, who developed MemBrain during his PhD in Tingying Peng’s lab at Helmholtz AI and in Ben Engel’s lab at Helmholtz Pioneer Campus (now at Biozentrum Basel), to tell us more about it.
How would you explain what MemBrain is and how it works?
MemBrain is a Deep Learning-based program for the analysis of cryo-electron tomograms. It is specialized for the localization of membrane-bound proteins, which enables the study of interactions between these proteins within their native membrane environment. The idea behind MemBrain is to reduce the complexity of this detection task: We incorporate membrane segmentations as additional information in the program. This allows us to search only in relevant areas of the tomogram and to make our neural network robust against varying membrane orientations.
Figure: Image showing MemBrain's step-by-step pipeline for automatic detection of membrane-bound proteins in cryo-electron tomograms.
What are the challenges for the annotation of these membrane proteins?
An obvious challenge is that manual annotation in 3D is variable and error-prone. Luckily, there are tools such as Membranograms that facilitate manual picking. The main challenge, however, is the imaging data itself. Cryo-electron tomograms suffer from a low signal-to-noise ratio and other artifacts, which severely affects the visibility of certain structures. These limitations make it time-consuming to manually annotate proteins and difficult to automate this task.
How does MemBrain differ from other solutions in the field?
MemBrain is the first Deep Learning-based program that is specialized in localizing membrane proteins. There are other DL-approaches, like DeepFinder, for the general detection of particles, but their training typically relies on large annotated tomogram regions. MemBrain is very label-efficient compared to them: a single annotated membrane can be enough to train your model! Existing non-DL-based approaches are difficult to tune for membrane proteins or require further processing steps.
Why is your research important?
The importance of studying membrane proteins is immense — about one-third of human proteins are associated with membranes. Ben Engel’s lab is interested in plant cells, where photosynthesis is orchestrated by a variety of membrane proteins. In the current fast-progressing climate change, gaining a deeper understanding of photosynthesis is now more important than ever.
What are the remaining challenges for the detection of membrane proteins?
MemBrain is able to detect certain membrane proteins, once it is trained properly. Still, a few challenges remain: 1. Our provided pretrained model works only for certain protein shapes. We will try to gather a collection of different models, specialized for various proteins. 2. MemBrain requires membrane segmentations as input. This is another task that we are trying to automate, as currently, manual corrections are still necessary.
Cryo-ET is a rapidly growing field with many open challenges, and I’m excited to keep developing solutions for them.
Picture: Lorenz Lamm, PhD at Tingying Peng’s lab.
Lorenz’s interest in cryo-ET and membrane proteins was sparked when he wrote his Master’s thesis in a collaboration with Tingying Peng’s lab at Helmholtz AI, and Ben Engel’s lab at the Helmholtz Pioneer Campus (now at Biozentrum Basel). This collaboration persists, as Lorenz is now a second year PhD student shared between these two labs. MemBrain, is available as a preprint, on GitHub, and has been submitted for peer-review.
- Lorenz Lamm, Ricardo D. Righetto, Wojciech Wietrzynski, Matthias Pöge, Antonio Martinez-Sanchez, Tingying Peng, Benjamin D. Engel. “MemBrain: A Deep Learning-aided Pipeline for Automated Detection of Membrane Proteins in Cryo-electron Tomograms.” bioRxiv, 2022.03.01.480844; doi: https://doi.org/10.1101/2022.03.01.480844
- GitHub code & instructions
- Twitter thread