The return of the BRAIN initiative notes

Figure 1: (A) Diagram showing implant of microendoscope (GRIN) into the rodent dorsomedial striatum to imaging and record neural activity with GCaMP (Ca²⁺ indicator) and a miniature microscope (Ghosh et al., 2011; Ziv et al., 2013). (B) Example raw Ca²⁺ imaging movie collected from miniature microscope along with the maximum projection from processed ΔF∕F version of the movie. The bright white spots in the ΔF∕F max projection frame are putative cells that need to be automatically identified and their frame-by-frame activity traces reconstructed. (C) Map of cell images colored by overall Ca²⁺ transient firing rate over entire imaging session. (D) Cell activity traces extracted with PCA-ICA. © Biafra Ahanonu 2017.

Find the posting on the BRAIN initiative notes webpage at Calcium imaging cell identification and fluorescence activity trace reconstruction, part 1.

Large-scale calcium (Ca²⁺) imaging, using the change in Ca²⁺ concentration inside neurons as a relatively direct measure of neural activity, has become an essential tool for neuroscientists to probe neural ensemble activity and identify coding principles the brain uses to process sensory information, store memories, and produce behaviors (Hamel et al., 2015). Originally, chemical fluorescent dyes such as OGB-1 allowed experimenters to measure Ca²⁺ concentration in neurons and thus neuronal activity (Paredes et al., 2008). These synthetic dyes allowed researchers to begin studying detailed intracellular Ca²⁺ dynamics along with activity of multiple neurons at once. However, the dyes had drawbacks: they did not allow long-term (weeks to months) imaging of neural activity, dye loading could be problematic in certain brain cell types (e.g. those with thick cell walls) or regions (deep brain structures), dyes would target non-specifically to various cellular compartments/organelles and would leak out of cells, and they could not be targeted in a genetically defined manner (though there is limited indication that this might be possible (Tour et al., 2007)). On the other hand, genetically encoded Ca²⁺ indicators (GECIs) have helped expand the types of experiments and questions that can be answered using Ca²⁺ imaging (Chen et al., 2013; Dana et al., 2018; Mank et al., 2008) by addressing many of the issues with the Ca²⁺ dyes: using adeno-associated and other viruses, they can be expressed chronically from weeks to years ((Ziv et al., 2013) and personal experiments/observations); specific promoters upstream of the GECI gene can lead to expression in specific cell types; tagging the GECI with certain amino acid sequences can enable targeting to specific regions of the cell (e.g. the axon terminal (Broussard et al., 2018; Dreosti et al., 2009)); and many other advantages. However, once experimenters had collected their Ca²⁺ imaging movies, they faced a formidable challenge in extracting both cell locations/shapes and their associated fluorescence activity traces from each of these movies.

One method of obtaining cell shapes and activity traces is via manual identification and selection of regions-of-interest (ROI) by humans and then taking the average pixel value within said ROI on each frame to reconstruct the activity trace for that cell (ROI) during the entire movie. The manual identification can take a lot of time, requires humans to browse through the entire movie frame-by-frame, and can lead to false negatives especially for low signal-to-noise ratio cells. The later is especially true for one-photon movies and modern GECIs where the baseline is too low to allow cells to be seen except for when they are active. Beyond missing cells, the ROI approach also tends to introduce crosstalk by contaminating each cell’s activity trace with fluorescent signals from nearby or overlapping cells (while cells themselves do not physically overlap, this occurs due to optical limitations and the fact that many Ca²⁺ imaging datasets are 2D, which leads to perceived spatial overlap when a 3D volume is flattened), background (e.g. neuropil from axons and dendrites), blood vessels, and other sources. Thus, a good Ca²⁺ imaging cell extraction algorithm needs to meet several criteria to allow biologists to obtain high-quality data from their movies:

• (1) be automated—reduces time needed to process data, especially as dataset sizes grow, and removes a degree of human introduced variability from manual ROI selection;
• (2) scale well with dataset sizes—this is not required, but preferred as large-scale Ca²⁺ imaging videos are becoming more widespread;
• (3) maintain high fidelity—finding all Ca²⁺ events associated with a particular cell;
• (4) have low crosstalk—minimize number of Ca²⁺ transients or other fluorescent signals in a cell’s activity trace that are not from that cell.

To solve the above problems and meet those criteria, many groups have developed and published new methods. An early popular method used Principal Component Analysis followed by Independent Component Analysis (PCA-ICA) (Mukamel et al., 2009); this was a great step forward and met several "good algorithm" criteria. However, it could fail in cases with a lot of correlated background noise or in regions of high cell density and activity. Over the next several years, additional techniques were published to tackle the problem and among these non-negative matrix factorization (NMF) (Pnevmatikakis et al., 2014; Maruyama et al., 2014) emerged as a promising candidate to meet many of the ”good algorithm” criteria. After publication, variants and enhancements of the method (CNMF, sc-CNMF, CNMF-E, etc.) sought to improve the cell activity traces by integrating Ca²⁺ dynamics into the model (constrained NMF—CNMF) or to compensate for issues in the original algorithm (e.g. handling of correlated background activity as seen in miniature microscope movies, see CNMF-E). In addition, there is an emerging focus on online (e.g. close as possible to real-time) processing of Ca²⁺ imaging data for closed-loop experiments and also methods to improve compression of data (Buchanan et al., 2018; Giovannucci et al., 2017).

Beyond those algorithms discussed above, a variety of other techniques have been released that either directly allow experimenters to simultaneously obtain cell images and activity traces or that focus on detecting cells or improving already acquired activity trace accuracy. For example, several publications have initial or limited methods for post-hoc automated removal of false positive signals (e.g. PCA-ICA, CNMF, etc. all output signal sources that upon manual inspection turn out to not be cells or other biologically relevant signals [e.g. dendrites]); look for this to be tackled to a greater degree going forward (I will have more on this in the future).

To help new experimenters gain a clearer picture of the landscape, see below table for a list of many different Ca²⁺ imaging cell extraction and activity trace reconstruction algorithms. This table will be updated to include any missing algorithms and further discussion about Ca²⁺ imaging analysis will take place in future posts.