Periodic Reporting for period 1 - Visual Proteomics (Biomarker discovery by AI-guided, image based single-cell isolation proteomics)
Reporting period: 2019-04-01 to 2021-03-31
Proteins are the functional building blocks of life that determine health and disease states and constitute important drug targets. If a person gets sick and needs to see a doctor, it is because of a protein imbalance in our cells. The doctor then looks for irregularities in certain proteins to find out if there is inflammation in our body. In cancer for example, when chemotherapy does not work, it is because some of the cancer cells show changes in their protein repertoire, which can make them chemoresistant. The analysis of all proteins in a diseased tissue, the so-called proteome, is therefore of enormous value to understand the disease on a molecular level and to identify therapeutic vulnerabilities. The most widely used method to comprehensively study all the proteins in a biological system is mass spectrometry (MS) based proteomics. Using MS based proteomics to dissect a chemoresistant tumor and compare the sick tissue with healthy tissue from the same patient allows us to identify the proteins, which are most critical to the disease. However, the disease related proteome is complex, and analytical challenges exist due to the presence of different cell types and cellular states that can differentially promote disease progression or drug resistance. Classic approaches only provide average proteome descriptions of the disease as no methods exist to dissect and analyze the tissue with respect to cellular subsets.
The goal of the fellowship was therefore to develop an innovative MS based method that would address these limitations to obtain fine-resolved molecular maps of the disease related proteome. To achieve this, we married high-resolution microscopy with artificial intelligence guided image analysis and ultra-high sensitivity MS based proteomics. For the first time, this new ‘visual proteomics’ concept combines the visual dimension with the molecular phenotype and is generically applicable to cell cultures and patient biobank specimens. Our method represents an exciting new tool for biomedical research for biomarker discovery and next-level molecular disease profiling on the protein level.
The goal of the fellowship was therefore to develop an innovative MS based method that would address these limitations to obtain fine-resolved molecular maps of the disease related proteome. To achieve this, we married high-resolution microscopy with artificial intelligence guided image analysis and ultra-high sensitivity MS based proteomics. For the first time, this new ‘visual proteomics’ concept combines the visual dimension with the molecular phenotype and is generically applicable to cell cultures and patient biobank specimens. Our method represents an exciting new tool for biomedical research for biomarker discovery and next-level molecular disease profiling on the protein level.
The new method allowed us to uncover the protein landscape of cancer cells with unprecedented detail, so that we can identify where the problem is. But also to connect the protein profiles with the visual appearance of the cell. Artificial intelligence helps to detect hidden but informative patterns too subtle or complex for humans to discern by themselves and the underlying proteomic signatures identified will help to better understand the disease mechanism. It is a completely new tool that could help clinicians with diagnosis and treatment strategies in the future. While the doctor might look at 10-20 proteins, our approach can analyse thousands of proteins at the same time and identify their common visual features. In this way, clinicians can connect the morphological features of the cells that are visual through the microscope with their molecular profile, giving them a much deeper insight to what is actually happening below the surface of the cells. This knowledge can be of great help when it comes to giving the right diagnosis or finding the best treatment plan.
To prove that our pipeline works, we applied it to tissue samples of melanoma (skin cancer) and acinic cell carcinoma (salivary gland cancer) retrieved from Danish biobanks. Using artificial intelligence, the method automatically divided the tumor cells into subgroups based on visual features such as shape, protein localisation and more. The cells were then transferred to a laser microdissection microscope, which cuts out the cells individually and shoots them into a collection well, so the different subgroups are put together. By using a completely new, ultra-high sensitivity proteomics machine, we could then describe the protein landscape of these subgroups with unprecedented depth, precision and accuracy. Next, we compared the protein landscape of the tumor cells with the protein landscape of healthy tissue from the same patient to find out where the disease-causing protein imbalance is – which could be a possible drug target. Interestingly, we could identify prognostic biomarkers that were only expressed in specific regions in the cancer tissue. The spatial information retrieved from the microscopic read-out was key for the interpretation of these results.
Our results were presented on international conferences and released on a preprint server in order for it to be available to everyone/the public/colleagues as fast as possible (https://www.biorxiv.org/content/10.1101/2021.01.25.427969v1). The study is still undergoing peer review and has not yet been published yet by a scientific journal.
To prove that our pipeline works, we applied it to tissue samples of melanoma (skin cancer) and acinic cell carcinoma (salivary gland cancer) retrieved from Danish biobanks. Using artificial intelligence, the method automatically divided the tumor cells into subgroups based on visual features such as shape, protein localisation and more. The cells were then transferred to a laser microdissection microscope, which cuts out the cells individually and shoots them into a collection well, so the different subgroups are put together. By using a completely new, ultra-high sensitivity proteomics machine, we could then describe the protein landscape of these subgroups with unprecedented depth, precision and accuracy. Next, we compared the protein landscape of the tumor cells with the protein landscape of healthy tissue from the same patient to find out where the disease-causing protein imbalance is – which could be a possible drug target. Interestingly, we could identify prognostic biomarkers that were only expressed in specific regions in the cancer tissue. The spatial information retrieved from the microscopic read-out was key for the interpretation of these results.
Our results were presented on international conferences and released on a preprint server in order for it to be available to everyone/the public/colleagues as fast as possible (https://www.biorxiv.org/content/10.1101/2021.01.25.427969v1). The study is still undergoing peer review and has not yet been published yet by a scientific journal.
I received world-class training in state-of-the-art high-resolution microscopy and machine learning based image analysis software. This training aspect allowed me to build a novel discovery proteomics pipeline, which provides the basis for my future scientific career as independent research group leader In Germany after successful completion of the project. The concept is highly innovative and the resulting data provides unprecedented insights into protein dynamics in human health and disease states. In summary, we have successfully merged the best from two worlds, microscopy and proteomics, and developed a highly sensitive and multimodal method that provides much more accurate and biologically meaningful data for biomedical research.