Visualizing the Interactions between Hundreds of Millions of Cells

Introduction:: Studying cellular systems in silico enables indispensable insight into disease states like angiogenesis and sickle-cell anemia. Current models tasked with studying these phenomena can capture hundreds of millions of cells at the cost of generating petabytes worth of data. At such a scale, data analysis becomes intractable, diminishing the usefulness of these models. Traditional post hoc methods require simulation data to be placed in persistent storage¹ until loading later on for subsequent processing. This practice can’t be achieved at such a scale without using undesirable practices necessary to reduce data volume². In situ or in place, data analysis circumvents these issues by working with data at the time of conception. The objective of this work is to create in situ-based frameworks that permit researchers to visualize and interrogate models that can produce up to hundreds of millions of cells.

Materials and Methods:: To enable in situ visualization of cells, we construct a pipeline within the computational fluid dynamics solver HARVEY. We utilize the open-source data visualization software ParaView and SENSEI in situ interface³ to facilitate the visualization of our simulation and movement of data to ParaView, respectively (Figure 1). With a proper pipeline, we employ an in-line framework where data is passed directly from the producers (HARVEY) to our pipeline. We assess the ability to visualize large-scale cellular data by capturing the flow through 100,000,000 red blood cells. To further gauge the utility of our pipeline, we compare the difference in the data storage required for our image-based analysis routine and the more traditional routine of checkpointing a simulation across several geometries. We utilize increasing-sized cubes filled with red blood cells as inputs, ranging from side lengths of 700 µm to 5,200 µm (with a cube with a side length of 700 denoted by the label C700).

Results, Conclusions, and Discussions:: We demonstrated the utility of image-based analysis with a large-scale cellular simulation, with Figure 2 showcasing the results of using our visualization pipeline. By working with data in memory, we could visualize the exterior of our run to track the simulation state while capturing a separate interior view to enable simultaneous coverage of regions of interest. In addition, we evaluated the size of complete checkpoints and images generated from our pipeline for increasingly large cubic inputs. We observed data reduction at magnitudes up to 10⁶ with the most extensive geometry, demonstrating the massive storage savings possible. Table 1 outlines the storage necessary for both analysis routines across a set of runs, considering that the last two data points (C4000, C5200) for the checkpointing routine are estimations due to storage concerns. With the first experiment, we were able to showcase the utility that is possible through in situ-based data analysis. The run containing 100,000,000 cells would’ve required terabytes of data per timestep using post hoc methods, while our image-based methodology takes up just megabytes. Working with the data directly after formulation means we inherently resolve any concerns over data integrity. The flexibility we demonstrated with capturing multiple views further outlines our pipeline's advantages. For cell tracking especially, monitoring the overall simulation state while also honing in on areas of concern can be invaluable. We further articulated attainable storage savings in the second set of experiments. As we increased the overall volume of our simulations, we saw an increase in the disparity between storage requirements between analysis routines. As model size continues to grow to enable continued exploration of complex real-life systems, it’ll become more advantageous to deploy lightweight analysis routines. We have shown that our in situ framework facilitates the visualization and interrogation of petabytes worth of cellular data, allowing researchers to fully utilize cellular models without compromising data integrity. In doing so, we are establishing frameworks for the continued use of in situ-based analysis as the scale of models continues to outpace I/O capabilities.


Acknowledgements (Optional): : An award of computer time was provided by the INCITE program. This research also used resources of the Oak Ridge Leadership Computing Facility, which is a DOE Office of Science User Facility supported under Contract DE-AC05-00OR22725.

References (Optional): :
¹Ahrens, James, et al. IEEE (2014) ²Baker, Allison, et al. Computer Graphics Forum (2019) Ayachit, Utkarsh, et al ³Ayachit, Utkarsh, et al. IEEE (2016).