A Neuron-Based Screening Platform for Optimizing Genetically-Encoded Calcium Indicators

Fluorescent protein-based sensors for detecting neuronal activity have been developed largely based on non-neuronal screening systems. However, the dynamics of neuronal state variables (e.g., voltage, calcium, etc.) are typically very rapid compared to those of non-excitable cells. We developed an electrical stimulation and fluorescence imaging platform based on dissociated rat primary neuronal cultures. We describe its use in testing genetically-encoded calcium indicators (GECIs). Efficient neuronal GECI expression was achieved using lentiviruses containing a neuronal-selective gene promoter. Action potentials (APs) and thus neuronal calcium levels were quantitatively controlled by electrical field stimulation, and fluorescence images were recorded. Images were segmented to extract fluorescence signals corresponding to individual GECI-expressing neurons, which improved sensitivity over full-field measurements. We demonstrate the superiority of screening GECIs in neurons compared with solution measurements. Neuronal screening was useful for efficient identification of variants with both improved response kinetics and high signal amplitudes. This platform can be used to screen many types of sensors with cellular resolution under realistic conditions where neuronal state variables are in relevant ranges with respect to timing and amplitude.


Introduction
Fluorescent protein-based sensors of neuronal activity are beginning to revolutionize neurophysiology [1]. Protein sensors enable in vivo recording of hundreds of neurons simultaneously [2] over chronic timescales [3][4][5]. Protein sensors also facilitate measurements of excitation in tiny neurons [6] and neuronal microcompartments, such as dendritic spines [7] and axonal terminals [8,9], which are inaccessible to electrophysiological methods. Protein sensors can be targeted to specific neuron types using gene regulatory elements [10]. They can thus be delivered to cells of interest in a non-invasive manner [11,12].
The optimization of sensors for neuronal activity would greatly benefit from testing in neuronal systems. Neurons [13,14] and neuronal microcompartments [15,16] have unusually small and fast calcium dynamics, which are difficult to model in non-excitable cells [17]. Previous efforts in engineering sensors have tested candidates in non-neuronal assays, including assays using purified proteins and tissue culture cells [5,[18][19][20]. Other efforts have tested sensors in lower throughput in vitro and in vivo systems, such as rat neuronal slice cultures, fly neurons, and fish neurons [7,[21][22][23][24].
Recently, a high-throughput stimulation and imaging system for neurons has been developed for drug screening using an indicator of synaptic function [25]. Here we combine a similar system with high-resolution imaging, suitable to assess the millisecond timescale dynamics of sensors in individual neurons. This allows us to screen sensors under physiological conditions. We tested this platform by assaying performance of GECI variants.

Indicator expression in neuronal cultures on screening platform
We developed a neuron-based platform to screen activity indicators ( Figure 1). We evaluated the capabilities of this novel screening platform by expressing and imaging of variants of the green fluorescent GECI, GCaMP3 [5]. GCaMP detects calcium increases via a calcium-dependent calmodulin and M13 peptide interaction that modulates the chromophore environment of a tethered GFP domain and augments fluorescence. Protein variants were made by mutating the GCaMP3 coding region using site-directed mutagenesis at selected positions in a lentiviral expression vector ( Figure 1B). The large size of the lentiviral vector and the presence of recombinogenic repeat sequences made whole plasmid PCRbased mutagenesis difficult. Mutagenesis was instead carried out by mutation of coding regions and subsequent sequence assembly with the lentiviral vector.
Individual variants were packaged in lentiviral particles in tissue culture cells. Dissociated neonatal rat hippocampal cells were then infected at 3 days in vitro in 24-well, glass-bottom plates for 18 hours, with unconcentrated viral particles. Glial proliferation was inhibited on day 4 by addition of the nucleoside analogue, AraC. Neuron-selective expression was achieved using a 476-bp human synapsin-1 promoter element [26]. Additionally, a nuclear localization signal (nls) tagged red fluorescent protein was co-expressed in neurons using an internal ribosome entry site (IRES)-nls-mCherry sequence ( Figure 1B) [27]. Red fluorescence was used both for image segmentation and measurement of relative GCaMP expression across cells. Hippocampal cultures included excitatory glutamatergic and inhibitory GABAergic neurons (data not shown).

Electrical field stimulation and fluorescence imaging
At 16-18 days in vitro, GCaMP expression reached sufficient levels for imaging, and neurons were capable of firing trains of APs [28]. GCaMP variants were tested on the platform ( Figure  1C) by measuring fluorescence changes in response to field stimulation of neuronal cultures. Growth medium was exchanged with imaging buffer. Neuronal culture wells were stimulated using a custom-built 24-well cap stimulator ( Figure  1D). Each well contained a pair of parallel platinum wires, and each pair could be independently controlled. Wells were illuminated by LEDs, and images were captured using an EMCCD camera. A motorized microscope stage allowed automated image acquisition from multiple wells in sequence ( Figure 1C).
Fields of view (800 µm x 800 µm) were located from a predetermined list of stage coordinates, corresponding to centers of each well. Imaging and stage movement were controlled by MetaMorph software (Molecular Devices). The variation in the glass bottom of wells ranged up to 500 µm in z position across a plate, which made well-by-well focusing necessary. An imaging-based autofocusing routine, based on computing nearest neighbor pixel contrast in the mCherry channel [29], was used. After focusing, reference mCherry, GCaMP, and bright field images were acquired.
A trigger (provided by MetaMorph) simultaneously initiated imaging and field stimulation. Electrical stimulation parameters were controlled as a slave by a separate custom software package (Ephus) [30]. Image timing data from the camera, temperature from thermocouples, LED light levels, stimulus current, and stimulus voltage were recorded at 10 kHz. Buffer temperature was ~30°C. Fluorescence image streams (35 frames/s, 250 frames) were captured using a cooled EMCCD camera controlled by MetaMorph.

Image segmentation and analysis
To assay GECI performance at cellular resolution, we identified regions of interest corresponding to nls-mCherry-and GCaMP-positive neuronal cell bodies using custom software. To define regions of interest, a raw mCherry image was lowpass filtered with a circular kernel roughly the size of a cell nucleus (Figure 2A,B). Putative locations of cells were determined as the local maxima of the filtered image whose intensity crossed an adaptively defined threshold. Based on these seed locations, a Voronoi diagram was drawn to cut the image into multiple sub-regions ( Figure 2C) [31]. Adaptive thresholding was performed on the GCaMP and mCherry image within each sub-region to define pixels belonging to the cytosolic and nuclear regions of interest, respectively ( Figure  2D). Averaged baseline GCaMP and mCherry fluorescence was measured within each cytosolic or nuclear region of interest, and cells were excluded if the average mCherry level did not reach a predefined threshold ( Figure 2E). Regions of interest that touched the boundary of the image were also excluded as a further quality control. Segmentation increased the sensitivity of the assay by narrowing regions of interest to exclude extraneous background ( Figure 2F). The median peak ∆F/F 0 (change in fluorescence normalized by initial fluorescence) response of GCaMP3 to 10 AP stimulus increased from 31.6 ± 4.1% to 70.7 ± 5.9% after segmentation (median ± s.e.m.; Student's paired t-test, p=1.6 x 10 -12 ). Responses were also less variable because segmentation excluded inactive pixels, which varied in number from well to well. The coefficient of variation decreased from 62.8% to 46.6% after segmentation.
Examples of evoked GCaMP3 ∆F/F 0 responses under optimized stimulation conditions and with glutamate and GABA receptor blockers in the imaging buffer (see further below) are shown in Figure 3A,B,C. Fluorescence traces from segmented neuronal cell bodies over time exhibited variable ∆F/F 0 responses across different regions of interest ( Figure 3D,E,F). Possible sources of neuron-to-neuron variability included neuronal subtype diversity, differential electrical stimulation, connectivity differences, and variations in segmentation accuracy.

Stimulus parameters for action potential generation
Stimulus parameters were optimized for AP generation using GCaMP3 by varying field pulse (FP) voltage, frequency, and width [9,32]. In terms of voltage, frequency, and pulse width, the average ∆F/F 0 response to 10 FP in segmented neuronal cell bodies reached an absolute maximum at 30 V, 83 Hz, and 500 µs, respectively ( Figure 4A,B,C). The response remained stable from the same well even after delivering 15 trains spaced ~1 min apart of 10 field pulses at 40 V, 83 Hz, and 1 ms pulse width ( Figure 4D). Additionally, the first and last replicate wells of GCaMP3-expressing neurons on plates responded within similar ranges, indicating that neurons remained healthy throughout the assay period ( Figure 4E). Based on these results, parameters were fixed at 40 V, 83 Hz, and 1 ms width to ensure suprathreshold stimulation. Calcium responses evoked by these stimulus conditions were dependent on stimulator evoked APs ( Figure 4F). The calcium response of neurons loaded with the synthetic calcium dye, Fluo4, was completely and reversibly abolished by blocking voltage-gated sodium channels with tetrodotoxin. To further confirm that APs were evoked, we also measured membrane potentials using the archaerhodopsin-3-based voltage sensor (ArchWT-GFP; Figure 4G,H) [33]. Red laser illumination (638 nm, ~500 W/cm 2 ; Figure 1C) and a high-speed EMCCD camera (500 frame/s) were used. With 40 V, 83 Hz, and 1 ms stimulus parameters, 10 FP evoked 10 AP ( Figure 4G,H, Figure S1). Subthreshold depolarizations were observed with submaximal field stimulation ( Figure S1). In screening, stimulus trains of 1, 2, 3, 5, 10, 20, 40, 80, and 160 AP were delivered with an intrastimulus interval of ~20 s. These trains were chosen to assess GCaMP performance over the full dynamic range of these cultured hippocampal neurons. Stimulus order was kept constant for screening.

Control of calcium response amplitude and variability with glutamate and GABA receptor blockers
Neuronal cultures formed functional networks. To control for network variability across wells, we sought to isolate individual neuronal responses from network influences by using neurotransmitter receptor inhibitors, including AMPA and NMDA glutamatergic receptors blockers CNQX and CPP (10 µM; Figure 5A). Without these antagonists, the variability and magnitude of responses were relatively large ( Figure 5A, inset), likely reflecting variable recurrent excitation in networks formed in culture. Consistent with the presence of inhibitory GABAergic neurons, addition of a GABA A receptor antagonist, gabazine (10 µM), disinhibited the GCaMP3 response, doubling the ∆F/F 0 for 10 AP stimulus ( Figure 5A). A slow component of the GCaMP3 response was blocked by MCPG (1 mM). This metabotropic glutamate receptor blocker can inhibit G qmediated activation of intracellular calcium release. CNQX, CPP, gabazine, and MCPG were thus used together in screening to dampen variability due to network effects and intracellular calcium release and increase calcium response amplitudes over those obtained with vehicle ( Figure 5B).

Neuron-based screening platform performance
We evaluated platform performance using the GCaMP3 construct. Nuclear mCherry fluorescence levels were correlated with GCaMP3 basal fluorescence ( Figure 6A; linear regression, R 2 =0.38, F(1,247)=153.1, p<2.2 x 10 -16 ). In screening, mCherry was thus used to normalize basal fluorescence values for GCaMP3 variants with variable F 0 related to mutations. The ∆F/F 0 response to 10 AP was weakly and inversely related to GCaMP expression levels, estimated by mCherry fluorescence ( Figure 6B; linear regression, R 2 =0.08, F(1,247)=21.52, p=5.68 x 10 -6 ). Higher expression levels might have resulted from a greater multiplicity of viral infection in wells with fewer neurons. Consistent with this notion, mCherry fluorescence was inversely correlated with the number of infected cells ( Figure 6C; linear regression, R 2 =0.07, F(1,247)=17.44, p=4.11 x 10 -5 ). Additionally, the GCaMP3 response of neurons was correlated with the number of infected neurons ( Figure 6D; linear regression, R 2 =0.31, F(1,247)=109.9, p<2.2 x 10 -16 ), perhaps reflecting less calcium buffering by GCaMP3 with lower expression in higher density wells. This relationship was used to correct for response variability (see below). Clumping of neurons in wells was associated with markedly increased ∆F/F 0 GCaMP3 responses in the low AP range ( Figure S2). In wells with clumped neurons, the GCaMP3 1 AP responses were often unusually high, and so these wells were excluded.
Without identifying underlying sources of variable responses, we normalized well-to-well differences by compensating based on infected cell density. A linear fit of the median GCaMP3 ∆F/F 0 responses in wells as a function of number of infected cells was calculated for control wells from all plates. For screening, compensated response values of GCaMP variants were obtained by dividing raw values by values on this GCaMP3 standard curve according to density for each well ( Figure 6D).

Platform detection sensitivity and throughput
We estimated the detection sensitivity of our assay for improvements over GCaMP3 performance by examining the distribution of GCaMP3 responses across wells. Using uncompensated data from 249 GCaMP3 wells, we selected random sets of 3, 4, 6, 8, 10, and 12 replicates 10 5 times to generate simulated distributions of GCaMP3 medians. The difference between the mean and the 99th percentile of these distributions normalized by the mean defined the detection sensitivity (for α=0.01). Figure 6E shows the percent changes that would be detectable by the assay as a function of replicate number. Increasing replicate number made the assay more sensitive. The assay was less sensitive with low AP stimulation. For example, with 4 replicates, 280%, 156%, 78%, and 20% ∆F/F 0 improvements over GCaMP3 could be detected for 1, 3, 10, and 160 AP stimulation, respectively (for α=0.01). Additionally, 23% decay time improvements over GCaMP3 could be detected for 10 AP stimulation.
The same analysis was carried out using cell density compensated data ( Figure 6F). Compensation improved sensitivity in the low AP range and with low replicates for ∆F/F 0 measurements. With 4 replicates, 256%, 139%, 69%, and 20% ∆F/F 0 improvements over GCaMP3 could be detected for 1, 3, 10, and 160 AP stimulation, respectively (for α=0.01). For decay time, 25% improvements over GCaMP3 could be detected for 10 AP stimulation.
We also evaluated the sensitivity of our assay by imaging calcium indicators previously tested in a variety of assays [5,20,34]. For small numbers of action potentials, the response amplitudes for Fluo4 and GCaMP5G [20] were larger than for GCaMP3 ( Figure 7A). Additionally, the fast rise time of Fluo4 was readily observed relative to GCaMP3 and GCaMP5G ( Figure 7B).

Comparison of GCaMP3 variant performance in neuronal culture versus solution
Our goal is to engineer protein indicators of neuronal function. How predictive are solution measurements with respect to calcium sensor sensitivity and kinetics in neurons? To address this question, we characterized mutated GCaMP3 variants in purified protein experiments and in the neuronal assay. Mutations in one of several GCaMP3 domains could affect peak fluorescence, F 0 , calcium affinity, and kinetics. We compared results from solution measurements with the neuronal assay ( Figure 8).
For purified protein, we measured the maximal fluorescence change (∆F sat /F apo ) and the apparent affinity for calcium (K d ) for several GCaMP3 variants. The peak response improvements for GCaMP5G were seen in both the neuronal and purified protein assays (Figure 8, Table S1). In general, larger

Discussion
Fluorescent protein sensors for neuronal activity are key reagents for dissecting neuronal circuit function [35][36][37]. Despite considerable effort, optimization of calcium indicator proteins has been slowed by development cycles that relied on testing in non-neuronal assays and then validating performance in neuronal systems. Here we developed a primary neuron-based screening platform to test sensors in the relevant context.
Other screening systems have been previously reported to image neuronal dynamics in response to trains of APs [25]. Our method is based on high-speed imaging with cellular resolution and is coupled to physiological stimulation of neural activity. Automated image segmentation and analysis allowed measurements from individual neurons, which increased the sensitivity of the assay. High-speed imaging allows optimization of sensors for speed. Similar image segmentation methods could be used to perform measurements on neuronal microcompartments, such as axons, presynaptic terminals, and dendritic spines. Simultaneous optimization of sensor sensitivity and kinetic parameters in neurons will be more efficient than solution-based methods for improving neuronal activity detectors. Because measurements in cultured neurons predict performance in vivo ( Figure 7C,D), neuronal screening will greatly shorten the optimization cycle. This assay system could be adapted for optimization of sensors of virtually any neuronal state variable (e.g. calcium concentration, membrane voltage, extracellular neurotransmitter levels, intracellular signaling pathways) [33,38]. It also could be used for the optimization of effector proteins that perturb neuronal activity [39,40]. Finally, the assay is also suitable for dissection of the genetic pathways involved in controlling calcium and other neuronal variables that can be imaged with fluorescence microscopy.

Lentiviral particle production and infection
A prolentiviral construct was combined with packaging and coat pseudotyping DNA constructs (pCAG-SIVgprre, pCAG4-RTR-SIV, pCMV-VSV-G) [41,43] and transfected into 32 x 10 6 HEK293T/17 cells (ATCC) cultured in DMEM and 10% fetal bovine serum in 10-cm plates. After 72 h, supernatant was collected (6 mL) and passed through a 0.45-µm filter. For each well of a 24-well plate, 0.5 mL of lentivirus was combined with 0.5 mL of conditioned growth medium and incubated for 18 h at 37°C. Medium was exchanged with 1 mL growth medium supplemented with 4 µM AraC to inhibit glial proliferation. Lentiviral particles were used in a biosafety level 2 laboratory. Sufficient virus (up to 4 mL) can also be produced in 6-well format (3.5-cm wells).

Image analysis
Analysis was implemented in MATLAB (release 2010a, MathWorks). The neuronal cell body segmentation strategy was to fit a circle of fixed radius around pixels to find centers of greatest mCherry fluorescence above a defined threshold. Center locations were used to construct Voronoi subregions [31]. Center locations were then used to draw circles of greatest GCaMP fluorescence above threshold within subregions. Circles were shrunk to exclude the lowest fluorescence quartile and holes were filled to define regions of interest. Regions of interest were excluded that contacted the image boundary. For GCaMP fluorescence transients, background subtraction was applied through subtraction of the mean of the lowest 5% intensity values. F 0 was defined as the mean of fluorescence 1 s prior to stimulus onset.

Calcium titrations of purified proteins
pRSET-A plasmids containing GCaMP variants were used to express protein in T7 Express E. coli cells (New England Biolabs) using 100 mL of ZYM-5052 autoinduction media [44] and ampicillin at 30°C for 48 h. Pelleted cells were lysed in B-PER (Thermo Scientific), 1 mg/mL lysozyme, and 15 U/mL DNase at 22°C for 30 min. After clearing by centrifugation, variants were purified using nickel-charged Profinity IMAC resin (Bio-Rad). Columns were washed with 20 mM Tris pH 8.0, 300 mM NaCl, 1 mM imidazole and then with 20 mM Tris pH 8.0, 500 mM NaCl, 10 mM imidazole. Variants were eluted in 20 mM Tris pH 8.0, 100 mM NaCl, 100 mM imidazole. Eluted protein concentrations ranged from 9-67 µM. Eleven-point calcium titrations were done using EGTA-buffered calcium solutions, similar to the protocol of the Calcium Calibration Buffer Kit #1 (Life Technologies). Green fluorescence intensities (excitation 485 nm, 5 nm bandpass; emission 510 nm, 5 nm bandpass) were measured using a Safire2 plate reader (Tecan). Apparent calcium affinty (K d ) was calculated from the inflection point of a sigmoidal fit to the fluorescence intensities.  Table S1. GCaMP3 variant calcium-induced fluorescence changes and calcium affinity measured from purified proteins and action potential-induced fluorescence changes and decay kinetics measured in neurons.

Supporting Information
(DOC)