Resumen de Insights to the characterization of cell motility and intercellular communication through a bioimage analysis perspective
Metastasis, the main cause of cancer morbidity and mortality, refers to the process in which cells spread from the primary tumor to adjacent tissues, proliferate and invade healthy organs. Anticipating and treating metastasis remains one of the biggest clinical challenges. At the time that patients are diagnosed with cancer, the metastatic cascade has already been initiated by cancer cells. However, it remains unknown how to block this process without damaging healthy tissues. A common approach is to develop targeted treatments to impede cells from starting or continuing the invading excursion. Moreover, cells can undergo different migration modes and only a few of them encode a high motility phenotype, which are the ones leading the metastasis. Thus, understanding the mechanisms driving cancer cell migration is critical to characterize highly metastatic cells. This is key to develop new efficient treatments and improving precision medicine.
Due to the experimental complexity, most quantitative cell migration studies are performed on 2D cell culture experiments, while cell migration naturally occurs in 3D environments such as the embryonic mesenchyme or the Extracellular Matrix (ECM) stroma enclosing a tumor. Cells embedded in 3D substrates (e.g., Collagen type I matrices) behave very differently to those in 2D cultures in terms of their differentiation, genetic expression, migration mechanisms, and proliferation. Instead of the lamellipodia or filopodia usually observed in 2D substrates, cells migrating in 3D substrates form dendritic protrusions to anchor, exert forces, and propel. Consequently, there is a growing interest to study the role of cell protrusions in cell migration and mechanotransduction.
To analyze the physiological relation between cell dendritic protrusions and cell motility, migrating cells are tracked in time-lapse microscopy videos. Cell migration is known to be random when analyzed frame by frame. In 2D substrates, cells are known to follow a random walk distribution. Namely, in short time windows, the cell displacements are random with a null displacement average, but after a while, it is possible to distinguish a significant non-zero mean displacement. In 3D substrates, cells show a larger persistence than in 2D cultures and their migration does not follow a random walk distribution. However, when observed in short time windows, their movement remains random. Therefore, identifying patterns in cell migration can be only achieved if long enough videos (~12 hours) for each tracked cell are acquired.
Cell motility strongly depends on the tissue micro-environment: the presence of signaling proteins, ECM stiffness and fiber cross-linking, cell density, and molecular pathways among many others. Additionally, experimental setups can show a large inter-variability. Also, cells behave differently when they are isolated. For this reason, in biological research, statistical analysis should always be computed from independent replicates of the same experimental procedure. Otherwise, the findings could not be translated to the clinical setup. Likewise, cells undergoing mitosis or apoptosis show an altered motility pattern and should be discarded from the analysis. Hence, the experiments to conduct this kind of study usually follow a high throughput setup in which a considerable amount of long microscopy videos is acquired, and all mitotic or apoptotic cells can be discarded without compromising the size of the remaining dataset.
The analysis of time-lapse microscopy videos is a common practice to assess different biological conditions and it can be done either manually or automatically. However, the annotation of cells and cellular protrusions in high-throughput microscopy images is non-viable. Hence, the first contribution in this project is the implementation of a robust computational tool for the quantification of cellular protrusions. For that, we first build a consistent and heterogeneous Ground Truth image dataset. Then, we propose a workflow that combines deep Convolutional Neural Networks (CNN), existing tracking methodologies, and geometrical analysis of the cell morphology to automatically quantify cellular protrusions. With the proposed image processing pipeline, we can analyze the role of cellular protrusions in 3D cell migration. The computational development of the tools and the obtained results set the baseline to conduct future biological studies involving 3D cell migration and cell protrusions.
This thesis has been developed as part of a collaboration in which we have access to cancer cells treated with different drugs. In particular, when testing one of the image processing pipelines developed for cell and cell protrusion segmentation, we detected a statistical limitation of the classical Null Hypothesis Tests: whenever two groups of data with a large number of entries are compared, the obtained p-value will always be statistically significant. Here we find a second contribution for which we develop a new statistical method that models the p-value as a continuous function of the data size and characterizes the differences among the compared groups of data. Our method is based on a known exponential model of the p-value. Then, we state and prove the hypothesis that the p-value function depends on the differences among the distribution of the compared datasets. The methodology is tested on both simulated and experimental data. It proves to be robust to data variations and useful to analyze the information obtained from the high-throughput microscopy videos that need to be analyzed in this project.
A relevant process in cell migration is inter-cellular communication. Almost all cells release nano-scale (30-200 nm) particles called small Extracellular Vesicles (sEV) that are responsible for cell-to-cell information transfer. SEVs biomolecular cargo contains proteins, lipids, and genetic information (microRNAs) among others. Inter-cellular communication is present in every biological process (e.g. embryonic development, infective processes, neurodegenerative diseases, or immune responses). In cancer, sEVs participate in the formation of the pre-metastatic niche and in changing the tumor micro-environment. They are also involved in tumor expansion, proliferation, and metastasis. Thus, there is an increasing interest in revealing their function and studying a potential use in clinical applications. Because sEVs' shape can only be studied in the nano-scale, they are usually imaged with Transmission Electron Microscopy (TEM). However, the obtained TEM images are quite heterogeneous and full of artifacts due to the complexity of sEV's sample preparation and fixation for TEM image acquisition. Hence, the third contribution of this thesis is the deployment of an accurate method to automatically segment sEVs in TEM images. We implemented a CNN architecture that obtains accurate sEVs binary masks. To characterize the size and shape of each sEV, it is important to detect each individual sEV in the image. Under the premise that sEVs are roundish, we can use the Radon Transform to split clustered sEVs.
This thesis focuses on developing bioimage analysis tools to contribute to the study of cancer cell biology, and in particular 3D cancer cell migration and cell communication. We propose a number of cutting-edge bioimage processing techniques with potential usability in clinical research. Nevertheless, if one thinks about the adaptation of this kind of work to different experiments, one will rapidly notice an important limitation: the flow between the computational development and the standardization of the method in biological research remains a challenge for life-scientists. In other words, the analysis implemented relies on the work of an image processing developer or the purchase of private image processing software. Aware of the latter, there is an increasing interest in the life-sciences community to develop user-friendly bioimage processing open-source tools. In line with these initiatives, the last contribution faced here deals with the dissemination of deep learning solutions among final non-expert users. To reach our goal it is essential to work in two directions: (1) The development of a user-friendly open-source environment and (2) The democratization of DL in bioimage analysis by establishing general standards for DL image processing and releasing basic educational material. Those are the two main contributions made by the deepImageJ environment to the life-sciences community. Namely, it provides an open-source tool to deploy DL models in ImageJ, one of the most widely used and well-known ecosystems for bioimage analysis.
