I'm a 5th year graduate student in the Brock Lab at UT Austin. I combine single-cell transcriptomics, imaging, and deep learning to better understand phenotype composition, stability, and change in in-vitro cancer cell lines.
Email: tyler_jost@utexas.edu
LinkedIn: www.linkedin.com/in/tyler-jost
If you are interested in trained models, imaging data, or more details, please reach out!
Using an eco-evolutionary approach to cancer is of key importance when understanding cancer. However, it's uncommon even for in-vitro work to take into account distinctive phenotypes. Often, these phenotypes are only discovered after endpoint assays such as single-cell RNA sequencing. To address this, I used the Earth Mover's Distance (EMD) to quantify how separated distinctive phenotypes were using only one gene. Crucially, EMD has properties that make it able to identify genes that would be otherwised overlooked with other methods because EMD is sensitive to outlier populations. I was able to find genes that provided excellent separation identified phenotypes through gene expression, and my coworker isolated them through flow activated cell sorting. I then validated this separation using 3'-Tag RNA seq.
UMAP and gene expression histograms for a top MDA-MB-231 cell line marker, ESAM.
Link: TBD
Cell morphology work often focuses on shape and texture, and typically identification is not done on subpopulations within a cell line but instead on other features such as metastatic variability, gene deletions, etc. We wanted to know if cells with transcriptomic changes within a cell line could also be identified using basic phase contrast imaging. To do so, we used our isolated populations found from the statistical approach mentioned above. We labeled each subpopulations and used a convolutional neural network to identify subpopulation. We found that, despite being relatively unperturbed, intra cell line populations are identifiable through computer vision. Additionally, we found that including neighborhood interactions instead of just shape and texture (possible primarily due to our use of deep learning) allowed us to improve our classification abilities.
Link: TBD
Raman spectroscopy is an imaging modality that already has multiple industrial uses for identification of chemicals. It has also been demonstrated to be a reliable tool for label-free identification of biological matter such as pathogenic bacteria. We questioned whether it would be useful for identifying cells with minimal RNA expression differences that would normally be identifiable only through sequencing methods. While this project never reached publication status, it was an interesting venture into 3D convolutional neural networks. Previous work only used the maximum intensity image, however we hypothesized that using all of the spectra within a cell could provide a better input for a deep neural network. We maxed out at 80% accuracy, which is likely lower than what it can do, but gathering enough data to satisfy a deep neural network is one of the most difficult challenges that needs to be addressed on an experimental level.
Raman spectroscopy representations for a singular cell. On the left is the projection image, but the right contains all of the information for each pixel on the left. Incorporating this information could give greater insight into subtle changes in cell state.
In the field of mathematical biology, there is a strong need for the reuse of data. Machine learning has largely standardized a lot of this, and it's easy for me to try out a new network structure, etc. on MNIST, CIFAR-10, and more by just downloading the dataset. A lot of this lack of reuse could be solved by a central repository. But doing so requires something standardized and fast. To accomplish this, I designed a MySQL database as well as a frontend to access this data. This was a great way to become familiar with MySQL, specifically building out the database in a way that made sense for all types of experiments. For example, within our machine alone you can vary fluorescence, the way the amount of cells is reported, the level of depth the image was taken, and more. But after it was built, accessing the data was much more enjoyable than flat CSVs.
The entity relationship diagram for the database.