The interview is regarding his research about Accurate automated segmentation of autophagic bodies in yeast vacuoles using cellpose 2.0
I interviewed Dr Steven Backues, who is in the department of Chemistry at Eastern Michigan University.
This is his email for any other questions: sbackues@emich.edu
This is the google doc link of the entire transcript of the interview, in the google doc link -- Research transcript - Google Docs
Understanding the interview:
The research focuses on studying autophagy in yeast, a simple eukaryotic model organism, due to its cellular similarity to humans and ease of experimentation. Autophagosomes form through lipid transport from the ER and the attachment of the protein ATG8 to membrane lipids, facilitated by enzymes ATG7, ATG3, and ATG10. The team used CellPose 2.0, an AI-based image segmentation tool, to analyze electron microscopy images of autophagic bodies. Training the AI involved challenges like overtraining and difficulty distinguishing ambiguous images, but the model performed comparably to human segmenters. The model is specific to yeast vacuole images and not directly applicable to mammalian cells, which have lysosomes instead of vacuoles. Nitrogen starvation triggers autophagy via the TOR kinase pathway by signaling amino acid scarcity. Dr Backus said that the researchers are currently using the model to analyze mutants affecting autophagosome size and number, providing insights into protein functions in autophagy regulation. They also developed custom scripts to automate measurements post-segmentation. While the model is well-validated for their specific use, broader validation is limited, and integration with other computational tools is possible but not extensively pursued. Overall, the project demonstrates the utility of AI in automating complex image analysis in cell biology research.
Yeast is a simple eukaryote that shares many cellular features with humans, such as having a nucleus and performing autophagy. It is easier to work with because millions of yeast cells can be grown overnight, unlike mice or humans. Yeast has long been a standard model system for studying cellular processes due to its simplicity and ease of experimentation.
Autophagosome formation involves membrane-bound structures formed mainly by lipid transport proteins that channel lipids from the endoplasmic reticulum into the autophagosome membrane. A key protein, ATG8, is covalently linked to the head of a lipid, attaching it to the membrane, which is important for autophagosome growth. Enzymes such as ATG7, ATG3, and ATG10 facilitate this attachment process.
Initially, training improves the model's accuracy, but after a certain point, overtraining occurs where the model becomes too focused on the fine details of the training images, which are not relevant to new images. This reduces its ability to generalize to new data, causing performance on new images to plateau or decline.
The main challenges included finding existing programs insufficient for segmenting the challenging gray-on-gray images of autophagic bodies. Attempts with free and proprietary solutions failed, and a government-funded project to develop a new program was not funded. The breakthrough came with CellPose 2.0, which was effective despite the difficulty of the images.
Automated segmentation is broadly beneficial for microscopy image analysis, including human cells. However, the specific model trained for yeast autophagic bodies is not directly applicable to mammalian cells because mammalian autophagy structures differ significantly. Researchers would need to train their own models for human cell images.
Nitrogen starvation, specifically the lack of amino acids, triggers autophagy in yeast. The TOR kinase pathway senses amino acid levels; when amino acids are scarce, TOR activity decreases, leading to activation of autophagy proteins. This allows the cell to recycle nutrients by breaking down cytoplasmic components in the vacuole.
The next steps involve using the model to analyze data from mutants affecting autophagosome size and number, which serves as ongoing validation. The model is considered reasonably well validated for the specific use case, and further validation will come from continued sensible results in research applications.
The surprising result was that the AI model performed nearly as well as human segmenters, sometimes even better in ambiguous situations. However, it tended to slightly overestimate the number of autophagic bodies, especially in cells with few or no bodies, due to lack of training on empty images.
Measuring AVs helps determine how mutations in proteins like ATG3 affect autophagy by influencing the number and size of autophagosomes. This provides insights into the protein's role in regulating autophagosome initiation and growth, which affects the overall autophagy process.
Yes, CellPose 2.0 can be combined with custom scripts and other computational tools. For example, the researchers wrote custom Python scripts to measure segmented bodies and automate processing of multiple replicates. CellPose 2.0 supports both graphical and code-based interfaces, facilitating integration with other tools.