(403e) Ffiber: Framework for Fluorescent Neuroimaging Based Experimental Routines

Background: The modern lab produces more data than it can analyze using traditional experimental techniques. For neuroimaging labs, terabytes of imaging data can easily be generated in the course of a single experiment. However, only a small portion of this data is accessible for extracting results based on existing methods. To enhance the portion of data that can be analyzed, many labs are turning to data science for high throughput processing and analysis. However, to synthesize highly skilled domain scientists with an effective data science strategy, a robust method for developing well-designed and data informed experimental pipelines must be followed. We developed a framework for fluorescent neuroimaging based experimental routines (FFIBER) that implements key practices of data science, software engineering, chemical engineering, and neuroimaging. The goal of this pipeline is to produce experimental workflows for high throughput neuroimage acquisition and analysis. For each experiment we follow a six-step process: (1) develop a data awareness, (2) design a data management plan, (3) determine an optimal experimental pipeline, (4) build out supporting data science infrastructure, (5) perform primary and supplemental imaging, and (6) produce interpretable visualizations of results. The objective of this project is to apply our FFIBER method during initial experimental stages and retroactively to generate, process, and analyze high volumes of neuroimaging data without an overall increase in labor-hours.

Methods: To develop a data awareness, we performed a data assessment of relevant metadata, data storage, and data relationships. When designing a data management plan, we identified the optimal data storage locations, file structure for raw data and results, and personnel responsibilities. To determine an optimal experimental pipeline, we performed software reviews, addressed labor-hour bottlenecks, and identified necessary programs, packages, and scripts. Then we built out the supporting data science infrastructure utilizing Python and Jupyter notebooks as the base script for pipelines, electronic laboratory notebook integration, and version control with the team through GitHub and Google Drive. After the pipeline was built, we completed the primary set of experimental imaging using confocal microscopy of brain slices and fed the data through the pipeline for analysis. We then completed supplemental imaging to experimentally support a representative set of results obtained through FFIBER. Finally, we wrote Python scripts to output consistent interpretable visualizations of our analyses and align these with other experimental results.

Results: We applied FFIBER to a data science experiment analyzing the phenotype of microglial cells in ex vivo brain slices to assess extent of glial cell activation. We were able to create and modify a pipeline during the initial stages of (1) an oxygen-glucose deprivation experiment and retroactively for (2) an inflammation-sensitized model using E. coli derived lipopolysaccharide, (3) and a neonatal hypoxia-ischemia (HI) model. A preliminary data assessment revealed a storage necessity of 5 GB per experiment with five individual file types from five separate personnel. We completed the data assessment and built a data management plan with university-supported Google Drive as our storage location, .tiff, .csv, and .ipynb as our file types, and established a slice-based, double-blind file structure. We wrote Python scripts automating image upload, cleaning and automatic feed to our analysis pipeline that reduced labor-hours from five hours per image set to less than ten minutes. Our software review led us to integrate a published cell morphometrics package, VAMPIRE, with our diff_register package for skeletonized morphometrics of microglia. Our image analysis pipeline developed with FFIBER split uploaded confocal images into fourths, segmented, skeletonized and identified the microglia in the images, and uploaded the images into our integrated morphometric analysis. FFIBER structured the development of our work pipeline that decreased work hours from a full day to less than an hour to analyze an entire experiment. Finally, our FFIBER structured pipeline enabled us to detect and quantify shape differences between non-treated and injured ex vivo slices – providing additional information for our initial oxygen-glucose deprivation experiment and retroactive insight to our inflammation sensitized and neonatal HI models. Our first application of FFIBER decreased our human labor time from > 24 hours per 5 GB image set to less than an hour, decreased our storage budget to $0, and produced information about cell morphometrics previously unavailable to our lab due to lack of analysis software and time constraints.

Conclusions: By applying FFIBER, we were able to create workflow pipelines both during initial stages of experimental design and retroactively in neuroimage sets already collected. We decreased labor hours for data sets from 24 hours to less than an hour through using Python scripts rather than manual work and a well-designed pipeline. Our framework creates a high throughput approach for experimental design in research dependent on neuroimaging by integrating traditional experimental methods with modern data science applications. We can further iterate the methodology to include metadatabases for easier data access and data connections, budget vs storage matrices for determining best data storage location, and templates for electronic lab notebooks, file structures, data relationships, and generalized processing pipelines. Additionally, while FFIBER was developed for fluorescent neuroimaging datasets, the current workflow is robust, and flexible enough for application to other image analysis techniques, other organs, and other fluorescent imaging methods.