Protein constructs and membrane fractionation
All experiments were performed with an Arch-eGFP fusion. A lentiviral backbone plasmid encoding Arch-eGFP (FCK:Arch-EGFP, accession number BAA09452) was used to clone the Arch gene into the pet28b vector using the restriction sites EcoRI and NcoI. The D95N mutation was created separately in the pet28b and FCK backbones, using the QuikChangeII kit (Agilent) and the same DNA primers for both backbones (Supplementary Table 2
Arch and its D95N mutant were expressed in E. coli
, following Ref. 31
. Briefly, E. coli
(strain BL21, pet28b plasmid) was grown in 1 L of LB with 100 µg mL−1
kanamycin, to an O.D. 600 of 0.4 at 37 °C. All-trans
retinal (5 µM) and inducer (IPTG 0.5 mM) were added and cells were grown for an additional 3.5 hours in the dark. Cells were harvested by centrifugation and resuspended in 50 mM Tris, 2 mM MgCl2
at pH 7.3 and lysed with a tip sonicator for 5 minutes. The lysate was centrifuged and the pellet was resuspended in PBS supplemented with 1.5% dodecyl maltoside (DM). The mixture was homogenized with a glass and teflon Potter Elvehjem homogenizer and centrifuged again. The solubilized protein in the supernatant was used for experiments.
Spectroscopic characterization of Arch and Arch(D95N)
The absorption spectra of fractionated E. coli
membranes containing Arch and Arch(D95N) were determined using an Ocean Optics USB4000 spectrometer with a DT-MINI-2-GS light source (Supplementary Fig. 1
). The peak extinction coefficients of microbial rhodopsins vary across rhodopsin types from 48,000 to 63,000 M−1
Due to the high homology between Arch and bacteriorhodopsin (BR), we used the BR extinction coefficient, 63,000 M−1
, for Arch. The differing wavelengths of maximum absorption of Arch (558 nm) and Arch(D95N) (585 nm) led to different extinction coefficients at 633 nm, as shown in . For Arch, 633 nm was in the tail of the absorption while for Arch(D95N) 633 nm lay half way down the shoulder. The relative extinction coefficients of Arch and Arch(D95N) at 633 nm are independent of our choice to use BR as the reference for the peak extinction coefficient. Absorption spectra for Arch and Arch(D95N) were measured as a function of pH between pH 6 and 11.
The fluorescence emission spectra of Arch and Arch(D95N) were determined using illumination with a 100 mW, 532 nm laser (Dragon Lasers, 532GLM100) or a 25 mW, 633 nm HeNe laser (Spectra-Physics) (Supplementary Fig. 1
). Scattered laser light was blocked with a 532 nm Raman notch filter (Omega Optical, XR03) or a 710/100 emission filter (Chroma), and fluorescence was collected perpendicular to the illumination with a 1,000 micron fiber, connected to an Ocean Optics QE65000 spectrometer. Spectra were integrated for 2 seconds. Arch and Arch(D95N) both had emission maxima at 687 nm. We do not know why the two proteins have such different peak absorption wavelengths but the same peak emission wavelength.
The fluorescence quantum yields of Arch and Arch(D95N) were determined by comparing the integrated emission intensity to emission of a sample of the dye Alexa 647 (Invitrogen). Briefly, the concentrations of micromolar solutions of dye and protein were determined using a visible absorption spectrum. We used the extinction coefficients of 270,000 M−1
for Alexa 647 and 63,000 M−1
for Arch and Arch(D95N), assuming that these microbial rhodopsins have the same extinction coefficient as bacteriorhodopsin. The dye solution was then diluted 1:1,000 to yield a solution with comparable fluorescence emission to Arch. The fluorescence emission spectra of dye and protein samples were measured with 633 nm excitation. The quantum yield was then determined by the formula
is the integrated fluorescence from 660 to 760 nm, ε is the extinction coefficient at 633 nm and c
is the concentration.
Relative photostability of Arch and eGFP
To perform a direct comparison of photostability of Arch and eGFP we studied the photobleaching of the Arch-eGFP fusion. This strategy guaranteed a 1:1 stoichiometry of the two fluorophores, simplifying the analysis. The experiments were performed on permeabilized cells, in the microscope, with video recording as the cells photobleached. We first recorded a movie of photobleaching of Arch under 640 nm illumination; then on the same field of view we recorded photobleaching of eGFP under 488 nm illumination, with illumination intensity adjusted to yield approximately the same initial count rate as for Arch. Fluorescence background levels were obtained from nearby protein-free regions of each movie and were subtracted from the intensity of the protein-containing regions. The area under each photobleaching timetrace was calculated, yielding an estimate of the total number of detected photons from each fluorophore. The eGFP emission (λmax = 509 nm) and the Arch emission (λmax = 687 nm) were collected through different emission filters, so the raw counts were corrected for the transmission spectra of the filters and the wavelength-dependent quantum yield of the EMCCD camera. The result was that the relative number of photons emitted prior to photobleaching for eGFP:Arch was 3.9:1, and for eGFP: Arch(D95N) this ratio was 10:1.
HEK cell culture
HEK-293 cells were grown at 37 °C, 5% CO2, in DMEM supplemented with 10% FBS and penicillin-streptomycin. Plasmids were transfected using Lipofectamine and PLUS reagent (Invitrogen) following the manufacturer’s instructions, and assayed 48–72 hours later. The day before recording, cells were re-plated onto glass-bottom dishes (MatTek) at a density of ~5,000 cells cm−2.
The concentration of endogenous retinal in the HEK cells was not known, so the cells were supplemented with retinal by diluting stock retinal solutions (40 mM, DMSO) in growth medium to a final concentration of 5 µM, and then placing the cells back in the incubator for 1 – 3 hours. All imaging and electrophysiology were performed in Tyrode buffer (containing, in mM: 125 NaCl, 2 KCl, 3 CaCl2, 1 MgCl2, 10 HEPES, 30 glucose pH 7.3, and adjusted to 305–310 mOsm with sucrose). Only HEK cells having reversal potentials between −10 and −40 mV were included in the analysis.
Simultaneous fluorescence and whole-cell patch clamp recordings were acquired on a home-built, inverted epifluorescence microscope, operated at room temperature. Here we summarize the design considerations; a detailed specification is given in Supplementary Fig. 2
. A key challenge was to collect fluorescence with high efficiency, while also achieving a large enough field of view to image an entire neuron and its processes. Typically, microscope objectives offer a tradeoff between magnification and light-gathering capacity (numerical aperture), which we sought to avoid. Additionally, we wanted the ability to change magnification while maintaining a patch on a single cell. The vibrations associated with switching objectives—particularly water or oil immersion objectives—are incompatible with simultaneous patch clamp. Finally, we wanted the capability to split the field of view into two wavelength bands, and to change magnification without changing the registration of the two halves of the image.
To achieve these goals simultaneously, we designed our microscope around a 60× NA 1.45 oil immersion objective (Olympus 1-U2B616), with variable zoom camera lenses to change illumination area and magnification. The magnification was continuously variable between 10 × and 66 ×, without touching the objective. The microscope readily converted between single-band and dual-band imaging, with only minor realignment.
On an upright electrophysiology setup retrofitted with a laser and EMCCD camera, a dipping objective (Olympus LUMPlanFl – 40 × W/IR; NA 0.8) collected enough light to record voltage-dependent fluorescence of HEK cells. However, recording of action potentials with high signal-to-noise ratio required a high NA objective (e.g. Olympus 1-U2B893 60 × Water NA 1.2; or 1-U2B616 60 × Oil NA 1.45).
Filamented glass micropipettes (WPI) were pulled to a tip resistance of 3–10 MΩ fire polished, and filled with internal solution (containing, in mM: 125 Potassium gluconate, 8 NaCl, 0.6 MgCl2, 0.1 CaCl2, 1 EGTA, 10 HEPES, 4 Mg-ATP, 0.4 Na-GTP, pH 7.3; adjusted to 295 mOsm with sucrose). The micropipettes were positioned with a Burleigh PCS 5000 micromanipulator. Whole-cell, voltage clamp recordings were acquired using an AxoPatch 200B amplifier (Molecular Devices), filtered at 2 kHz with the internal Bessel filter, and digitized with a National Instruments PCIE-6323 acquisition board at 10 kHz. Ambient 60 Hz noise was removed using a HumBug Noise Eliminator (AutoMate Scientific). For experiments requiring rapid modulation of transmembrane potential, series resistance and whole-cell capacitance were predicted to 95% and corrected to ~50%. Electrical stimuli were generated using the PCIE-6323 acquisition board and sent to the AxoPatch, which then applied these signals in either constant current or constant voltage mode.
Measurements of photocurrents were performed on HEK cells held in voltage clamp at 0 mV while being exposed to brief (200 ms) pulses of illumination at 640 nm at an intensity of 1,800 W cm−2.
All experiments were performed at 24 °C.
Ramp and step-response of Arch and Arch(D95N)
To measure fluorescence as a function of membrane potential, a triangle wave was applied, with amplitude from −150 mV to +150 mV and period 12 s, with video recording at 100 ms per frame. A pixel weight matrix was calculated according to Eq. 2
(below) and applied to the movie images to generate a fluorescence number for each frame. These fluorescence values were divided by their minimum value (at V
= −150 mV). The result is plotted as a function of V
in – . This procedure preferentially weighted data from pixels at the cell membrane, but did not entail any background subtraction. Comparable results were obtained by manually selecting pixels corresponding to a region of plasma membrane, and plotting their intensity as a function of V
, without background subtraction. Background subtraction from the raw fluorescence would have yielded considerably larger values of ΔF/F.
The step response was measured in a similar manner, except that test waveforms consisted of a series of voltage pulses, from −70 mV to +30 mV with duration 300 ms and period 1 s. Cells were subjected to 20 repetitions of the waveform, and the fluorescence response was averaged over all iterations.
Frequency-dependent response functions of Arch and Arch(D95N)
Test waveforms consisted of a concatenated series of sine waves, each of duration 2 s, amplitude 100 mV, zero mean, and frequencies uniformly spaced on a logarithmic scale between 1 Hz and 1 kHz (31 frequencies total). The waveforms were discretized at 10 kHz and applied to the cell, while fluorescence movies were acquired at a frame rate of 2 kHz.
The model parameters for extracting FL
) were calculated from the fluorescence response to low frequency voltages. These parameters were then used to calculate an estimated voltage at all frequencies.
The applied voltage was downsampled to 2 kHz to mimic the response of a voltage indicator with instantaneous response. For each applied frequency, the Fourier transform of FL
) was calculated and divided by the Fourier transform of the downsampled V
). The amplitude of this ratio determined the response sensitivity. It was crucial to properly compensate pipette resistance and cell membrane capacitance to obtain accurate response spectra. Control experiments on cells expressing membrane-bound GFP showed no voltage-dependent fluorescence.
The power spectrum of FL
) under constant V
= 0 was also measured to enable calculations of signal-to-noise ratio for any applied V
Estimates of membrane potentials from fluorescence images
A common practice in characterizing fluorescent voltage indicators is to report a value of ΔF/F per 100 mV of membrane potential. We feel that this parameter is of limited use, for several reasons. First, the value of ΔF/F is highly sensitive to the method of background subtraction, particularly for indicators in which F approaches zero at some voltage. Second, ΔF/F contains no information about signal-to-noise ratio, which depends on absolute fluorescence levels, background, and membrane targeting of the indicator. Third, the ratio ΔF/F contains no information about the temporal stability of the fluorescence. Fluctuations may arise due to intracellular transport, photobleaching, or other photophysics. Fourth, definitions of absolute fluorescence are often subjective, depending on a user-defined Region of Interest which might or might not include pieces of internal membranes or other cells.
We therefore sought a measure of the performance of a voltage indicator which reported the information content of the fluorescence signal. We sought an algorithm to infer membrane potential from a series of fluorescence images. We used the accuracy with which the estimated membrane potential matched the true membrane potential (as reported by patch clamp recording) as a measure of indicator performance. The algorithm described below is implemented in our analysis (Supplementary Software
The estimated membrane potential, FL
), was determined from the fluorescence in two steps. First we trained a model relating membrane potential to fluorescence at each pixel. We used the highly simplified model that the fluorescence signal, Si
), at pixel i
and time t
, is given by:
are position-dependent but time-independent constants, the membrane potential V
) is time-dependent but position independent, and εi
) is spatially and temporally uncorrelated Gaussian white noise with pixel-dependent variance:
indicates an average over time.
This model neglects nonlinearity in the fluorescence response to voltage, finite response time of the protein to a change in voltage, photobleaching, cell-motion or stage drift, and the fact that if εi
) is dominated by shot-noise then its variance should be proportional to Si
), and its distribution should be Poisson, not Gaussian. Despite these simplifications, the model of Eq. 1
provided good estimates of membrane potential when calibrated from the same dataset to which it was applied.
The pixel-specific parameters in Eq. 1
are determined by a least-squares procedure, as follows. We define the deviations from the mean fluorescence and mean voltage by
Then the estimate for the slope i
, and the offset is:
A pixel-by-pixel estimate of the voltage is formed from:
The accuracy of this estimate is measured by
A maximum likelihood weight matrix is defined by:
This weight matrix favors pixels whose fluorescence is an accurate estimator of voltage in the training set.
To estimate the membrane potential, the pixel-by-pixel estimates are combined according to:
Within the approximations underlying Eq. 1
, Eq. 3
is the maximum likelihood estimate of V
In cases where the membrane potential is not known, one can replace V(t) by the total intensity of the entire image I(t), provided that there is only a single cell with varying membrane potential within the image. In this case, the algorithm preferentially weights pixels whose intensity co-varies with the mean intensity. Such pixels are associated with the membrane. This modified procedure yields an estimate of the underlying intensity variations in the membrane. The output resembles the true membrane potential, apart from an unknown offset and scale factor. A key feature of this modified procedure is that it enables spike identification without a patch pipette.
On a video record of 30,000 frames taken (e.g. 30 s of data at 1,000 frames s−1), the training phase of the algorithm took approximately 3 min to run on a desktop PC. Application of the weighting coefficients to incoming video data could be performed in close to real time. Small shifts in the field of view due to stage drift or bumps of the apparatus are compensated by using image registration techniques to translate the pixel weight map. Large changes in focus or movement to a new field of view required re-training of the algorithm.
Molecular biology and virus production
Plasmids encoding Arch-EGFP (FCK:Arch-EGFP) were either used directly for experiments in HEK cells, or first used to produce VSVg-pseudotyped virus according to published methods16
. For pseudotyping, HEK-293 cells were co-transfected with pDelta 8.74, VSVg, and either of the Arch backbone plasmids using Lipofectamine and PLUS reagent (Invitrogen). Viral supernatants were collected 48 hours later and filtered using a 0.45 µm membrane. The virus medium was used to infect neurons without further concentration.
The D95N mutation was introduced using the QuickChange kit (Stratagene), according to the manufacturer’s instructions using the same primers as the E. coli plasmid.
Neuronal cell culture
E18 rat hippocampi (BrainBits) were mechanically dissociated in the presence of 1 mg mL−1 papain (Worthington) before plating at 5,000 – 30,000 cells per dish on poly-L-lysine and Matrigel-coated (BD Biosciences) glass-bottom dishes. At this density synaptic inputs did not generate spontaneous firing. Cells were incubated in N+ medium (100 mL Neurobasal medium, 2 mL B27 supplement, 0.5 mM glutamine, 25 µM glutamate, penicillin-streptomycin) for 3 hours. An additional 300 µL virus medium was added to the cells and incubated overnight, then brought to a final volume of 2 mL N+ medium. After two days, cells were fed with 1.5 mL N+ medium. Cells were fed with 1 mL N+ medium without glutamate at 4 DIV, and fed 1 mL every 3–4 days after. Cells were allowed to grow until 10–14 DIV. Cells were supplemented with retinal by diluting stock retinal solutions (40 mM, DMSO) in growth medium to a final concentration of 5 µM, and then placing the cells back in the incubator for 1 – 3 hours, after which they were used for experiments.
Whole-cell current clamp recordings were obtained from mature neurons under the same conditions used for HEK cells recordings. Series resistance and pipette capacitance were corrected. Only neurons having resting potentials between −50 and −70 mV were used in the analysis.
We used a simple spike identification algorithm that could be applied either to electrically recorded V
) or to optically determined
). The input trace was convolved with a reference spike. Sections of the convolved waveform that crossed a user-defined threshold were identified as putative spikes. Multiple spikes that fell within 10 ms (a consequence of noise-induced glitches near threshold) were clustered and identified as one.