Obtaining artifact-corrected signals in fiber photometry: Isosbestic signals, robust regression and dF/F calculations

doi:10.21203/rs.3.rs-3549461/v1

Download PDF

Article

Obtaining artifact-corrected signals in fiber photometry: Isosbestic signals, robust regression and dF/F calculations

https://doi.org/10.21203/rs.3.rs-3549461/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this older preprint version

Read the latest preprint version →

Fiber photometry is a powerful tool to measure a wide variety of dynamics from targeted cell populations and circuits in freely-behaving animals. However, measured biosensor signals are contaminated by various artifacts (photobleaching, movement-related, noise) that undermine analysis and interpretation. Here, we consider existing approaches for obtaining artifact-corrected neural dynamic signals from fiber photometry data. We show using real and simulated photometry data that a specific form of robust regression, iteratively reweighted least squares (IRLS), is preferable to ordinary least squares (OLS) regression for fitting isosbestic signals to experimental signals. We also demonstrate the efficacy of low-pass filtering signals and baseline-normalization via dF/F calculations. Considerations and recommendations for analyses, including methods for detrending and normalization are discussed.

Physical sciences/Optics and photonics/Optical techniques/Imaging and sensing

Physical sciences/Mathematics and computing/Statistics

Biological sciences/Neuroscience

Biological sciences/Biological techniques/Imaging/Fluorescence imaging

Biological sciences/Biological techniques/Sensors and probes/Fluorescent proteins

Fiber photometry is an increasingly popular technique in the neurosciences that combines genetically-encoded fluorescent biosensor and implanted optic fiber to measure cell-type and circuit-defined population dynamics in vivo ^[1,2]. The key concept underpinning this technique is that the amount of fluorescent light emitted by a biosensor in response to excitation light depends upon the presence of molecules it is designed to detect (e.g., calcium [Ca2+], specific neurotransmitters). Thus, biosensor fluorescence can be used as a proxy measure for biologically-relevant processes, like intracellular Ca2+ (proxy for neural activity) or neurotransmitter release within a brain region. The implanted optic fiber is simply the means by which excitation light is delivered and emission light is collected from a target region. This light is typically relayed between implanted fibers and the photometry system via a flexible fiber optic patchcord, allowing an animal to behave relatively freely during recordings.

An important caveat is that fluorescence signals obtained using fiber photometry contain artifacts that contaminate the signal, undermining detection and interpretation of signal changes. These artifacts include 1) system noise, 2) photobleaching of the biosensor, and 3) motion-related artifacts (Figure 1A). Noise is a feature of any measurement system. In the case of photometry, one source of noise is from the electrical components of the system. Given electrical noise is thought to be predominantly higher frequency than the kinetics of commonly used biosensors, this noise is usually attenuated by applying low-pass filters to recorded signals. Photobleaching occurs when a fluorophore is exposed to excitation light, causing it to output less fluorescence ^[3,4]. Given photometry involves exposing the biosensor to excitation light across a recording session, this results in signal decay across the session, introducing an artifactual reduction in signal. This is partially mitigated by using low-intensity excitation light, but some degree of signal decay is expected. Therefore, the effect of photobleaching has traditionally been dealt with through statistical procedures (discussed below). An especially problematic source of artifacts stems from animals’ movements. Movement generally causes the optic fiber patchcord to bend, reducing light transmission through the fiber ^[5,6] (Figure 1B). This effect of bending on measured signal is dynamic, non-linear and can correlate with an animal’s location and specific behaviors. This complicates any attribution of signal change to locations and behaviors, which are often the key dependent variable in photometry studies. While the physical recording set-up can be optimized to reduce the contribution of fiber bending and other movement artifacts to recorded signals, these are an insidious source of artifacts requiring direct addressal.

Artifact-correction via isosbestic control signals

An elegant solution to counteract these artifacts is to concurrently measure an isosbestic control signal and use this to remove the artifactual component of the neural dynamic (aka experimental) signal ^[1] (Figure 2A). The key principle underpinning this method is that the degree to which a biosensor’s fluorescence fluctuates with a neural dynamic depends on the wavelength of excitation light being delivered. For example, green fluorescence emitted by GCaMP is highly dependent on calcium binding when exposed to blue (~460-490nm) excitation light – it emits very little green light in a calcium-free environment and strong fluorescence in a calcium-saturated environment. However, when exposed to violet (~410nm) excitation light, the amount of green fluorescence emitted by GCaMP is relatively independent of calcium binding ^[1,7]. Therefore, any fluctuations in fluorescence obtained from isosbestic wavelength excitation are not due to the neural dynamic of interest and are instead attributable to artifacts. Given key artifacts affecting the isosbestic signal are the same artifacts affecting the experimental signal, the isosbestic signal can be used to remove the artifactual component of the experimental signal to obtain a truer readout of the dynamic being studied (Figure 2A). An alternative to measuring a biosensor’s isosbestic signal is to co-express a non-biosensor fluorophore (e.g., mCherry) to track artifacts. Although this is a less direct solution, it can be the best option in scenarios where a biosensor does not have an amenable isosbestic wavelength.

There are several ways the isosbestic signal has been used to correct the experimental signal across studies ^[8,9]. A common way is to fit the isosbestic signal to the experimental signal via linear regression (ordinary least-squares [OLS] regression) ^[1]. The obtained fitted isosbestic signal is then subtracted from the experimental signal and, when the original isosbestic dF/F equation is followed, used to divide that difference signal. This subtraction is intended to remove photobleaching decay and movement-related artifacts from the experimental signal, while dividing by the isosbestic baseline counteracts signal loss due to photobleaching.

A problem with this popular approach is that OLS regression assumes all datapoints in the experimental signal should be used to fit the isosbestic signal. This may be unsuitable for fiber photometry data because the experimental signal is, by design, expected to diverge from the isosbestic signal in a neural dynamic-related manner. These neural dynamic-related divergences function as outliers in the prediction, which have an outsized influence on OLS-based regression due to these “errors” being squared. This causes the isosbestic signal to be overfit to the neural dynamic component of the experimental signal while being underfit to the intended artifactual component (Figure 2B). This issue, in turn, negatively influences obtained dF/F and subsequent analyses.

One well understood solution to this influence of outliers on regression is a family of approaches known as robust regression. There are various forms of robust regression, but the one deemed most relevant for the current purpose is iteratively reweighted least-squares (IRLS) (Figure 2C). In IRLS, datapoints are weighted between 0 and 1. The initial regression gives all datapoints an equal weighting of 1 (equivalent to OLS). Weightings are then re-calculated based on the residual per datapoint and chosen weighting function; outliers are downweighted and another weighted regression is conducted to obtain a new fit. This is repeated iteratively, with each iteration recalculating weightings and the weighted fit. The key idea is that this process allows the isosbestic signal to be increasingly fit onto the components of the experimental signal that it strongly predicts (the putative artifactual component of the experimental signal), while disregarding parts of the experimental signal it does not predict well (the putative neural dynamic component). Conceptually, IRLS is preferable to OLS for the purpose of fitting an isosbestic signal to the experimental signal in fiber photometry.

In addition to this, researchers face other crucial decisions when obtaining “artifact-corrected dF/F” from photometry signals. This includes filtering to remove other unwanted noise components from recorded signals. A common choice is to low-pass filter signals to suppress high frequency noise (e.g., from electrical components) while preserving lower frequency signals attributable to biosensor kinetics. This is often done in real-time as recordings are made but can also be done within the analysis pipeline. Another consequential choice is the calculation in which the fitted isosbestic is used to artifact-correct the experimental signal. As the term dF/F implies, the dF/F calculation ([experimental signal - fitted isosbestic]/fitted isosbestic) ^[1] involves dividing by the isosbestic signal to baseline-normalize signals. This theoretically counteracts proportional loss of signal due to photobleaching across the session. However, not all studies reporting a “dF/F” divide by the changing isosbestic baseline. Instead, some procedures only subtract the fitted isosbestic from the experimental signal ^[8,9] and are thus better characterized as calculating a dF signal. Although these procedures typically include initial detrending and normalization of experimental and isosbestic signals to remove overall decay trends attributable to photobleaching, these approaches do not counteract putative attenuation of signal dynamics across the session. That is, these procedures address the change in signal baselines but do not address corresponding changes in signal scaling (i.e., its effect on event-related signals) across the session.

We empirically tested these different steps for artifact correction using simulated photometry data. We developed MATLAB scripts to simulate key components of photometry data and their influence on each other (all data and MATLAB code available at https://github.com/philjrdb/RegressionSim). Simulations were used because the true parameters (e.g., underlying neural dynamics) are known, unlike real photometry data where underlying neural dynamics are technically unknown.

The components of the simulation were, in order of their operation (Figure 3A):

A neural dynamic component composed of 100 identical 3sec event transients (waveform taken from ^[10]), which was tracked by experimental but not isosbestic signals.
A photobleaching component, modelled as a double-exponential decay function ^[4] that attenuated experimental and isosbestic signals.
A movement-related component that stochastically attenuated both experimental and isosbestic signals by a shared percentage.
A unique Gaussian noise component per experimental and isosbestic signal.

The order of these operations was designed to emulate artifact-related transformations of experimental and isosbestic signals across key junctures of the recording set-up (recording site → photodetector → signal processor). Neural and photobleaching components were fixed, whereas movement-related and noise components were stochastic and randomized between simulations. These simulations were used to assess the impact of 3 aforementioned choices made by researchers when obtaining an “artifact-corrected dF/F” from photometry signals: 1) low-pass filtering signals before isosbestic fitting, 2) applying OLS regression for isosbestic fitting, and 3) using the fitted isosbestic to perform baseline normalization across the session.

Specifically, we generated 10 simulations of experimental and isosbestic signals for a 20min recording session (10Hz sampling rates). These signals were then either low-pass filtered (3Hz) or not. Isosbestic signals (filtered or unfiltered) were then fit to corresponding experimental signals using OLS or IRLS (Figure 3B). IRLS regression was performed by the in-built MATLAB function robustfit, using Tukey’s bisquare as the weighting function. This function is favorable as it entirely ignores large outliers (putative neural dynamics), unlike other weighting functions that only attenuate the influence of large outliers, allowing IRLS to better concentrate the fit onto the artifactual baseline. The narrowness of Tukey’s bisquare is determined by a tuning constant. The default constant is 4.685, but two smaller constants (c = 3 vs 1.4) that make the weighting function narrower (i.e., more aggressive downweighting) were also tested. Larger constants were not tested as this shifts IRLS towards an OLS solution. Finally, fitted isosbestic signals were applied in dF/F ([experimental signal - fitted control]/fitted control) and dF (experimental signal - fitted control) calculations, reflecting the disparate choice to apply vs. not apply baseline normalization using the fitted isosbestic signal. In summary, there were 3 independent factors tested for each of the 10 simulations conducted: low-pass filter [3Hz vs none] x regression [OLS vs IRLS c=4.685 vs 3 vs 1.4] x baseline correction [dF vs dF/F].

The key dependent variable was how much extracted “artifact-corrected” signals (dF or dF/F) reflected the actual underlying neural dynamic signal. Given true and extracted signals are on different scales due to artifact attenuation factors as well as dF/F by definition being a fraction of dF, we normalized true and artifact-corrected signals by dividing whole-session signals by their sum squared deviation from 0 (i.e., they were z-scored relative to 0 [null-Z]; REF). The average absolute residual (i.e., difference) between normalized true and extracted signals were calculated for baseline (non-event) versus event periods (Figure 3C). Repeated measures ANOVA of residuals across simulations (n = 10) revealed each analysis factor had a significant effect on the accuracy of extracted signals for both baseline and event periods. These effects were essentially additive, so we discuss them in turn.

Applying a 3Hz low-pass filter to experimental and isosbestic signals resulted in a more accurate artifact-corrected signal (baseline periods: F(1,9) = 12137.4, p < .001; event periods: F(1,9) = 2665.0, p < .001) (Figure 3C). As theorized, IRLS regression produced a more accurate artifact-corrected signal than OLS regression, with this effect scaling with the tuning constant (baseline periods: F(3,27) = 37284.5, p < .001; event periods: F(3,27) = 8608.8, p < .001). That is, more aggressive downweighting reliably resulted in better accuracy across parameters (baseline periods: all t(9) > |14.9|, p < .001; event periods: all t(9) > |10.5|, p < .001[Bonferroni-adjusted]). Baseline normalization via a dF/F calculation was preferable to its omission via dF calculation (baseline periods: F(1,9) = 72.80, p < .001; event periods: F(1,9) = 308.4, p < .001).

Causes for these differences are clear when examining true vs extracted peri-event signals (Figure 3D). The true underlying event transient was a 3sec waveform that was identical in shape and amplitude across the 100 event trials, each flanked with a null baseline of 0. OLS vs IRLS regression and dF vs dF/F calculations had distinct impacts on the degree to which extracted peri-event signals reflected this.

dF vs dF/F calculations had a clear impact on event transient extraction over trials (Figure 3D). Despite the true transient not changing, there was an obvious trend across trials for dF-derived event signals. dF-based transients from initial trials were substantially larger than those from later trials, regardless of whether they were OLS- or IRLS-derived, due to signal attenuation from simulated photobleaching. By contrast, baseline normalization via the dF/F calculation mitigated this influence of photobleaching, largely restoring the stability of transients across the session.

OLS vs IRLS regression had an orthogonal effect on extracted peri-event signals, instead influencing the average waveform (Figure 3D). Mean peri-event signals across simulations (n = 10) were analyzed using waveform confidence intervals (95% t-test CI, 1/3sec temporal threshold ^[10]) against the null of 0, as commonly done in the literature ^[11-15]. Analysis of the true peri-event signal confirmed the 3sec waveform period as significantly deviating from 0, whereas pre- and post-event periods (null baseline) did not. When extracting OLS-derived “artifact-corrected” signals, mean waveforms (both dF and dF/F) were downshifted relative to the true signal (also evident in Figure 3B), resulting in a misidentification of when and how the peri-event waveform deviated from the null of 0. OLS-derived transients failed to identify the full 3sec extent of the positive waveform (Type 2 error) while pre-, late and post- event periods were misidentified as being below 0 (Type 1 error).

Taken at face value, these results would promote misconstrued inferences about peri-event neural dynamics. Alternatively, this reliable below 0 “inhibition” prior to the event of interest might be used as indication of an invalid baseline, which is often combatted by re-baselining each trial to a manually selected pre-event period. This extra arbitrary analysis step comes with the assumption that the designated pre-event period acts as a valid baseline, i.e., is free of neural dynamics that might systematically or unsystematically skew trial data (including a potentially real pre-event inhibition). This can be problematic for tasks in which animal behavior and potential anticipatory neural dynamics pervade the recording session (e.g., free operant tasks ^[12]), precluding a valid one-size-fits-all pre-event baseline. It is worth emphasizing that invalid pre-event baselining is equally liable to create the Type 1 and Type 2 errors observed here with OLS incorrectly baselining “artifact-corrected” signals.

IRLS-derived dF/F theoretically bypasses these issues by more directly fitting the isosbestic to the artifactual baseline of the experimental signal. Accordingly, the IRLS-derived transients (both dF and dF/F) accurately identified the 3secs transient from null pre/post periods without the need for re-baselining or trial-by-trial based corrections. This provides empirical support for the theoretical preferability of IRLS over OLS.

In summary, these simulations indicate low-pass filtering signals, applying IRLS regression to fit isosbestic to experimental signals, and using these signals within a baseline-normalizing dF/F calculation, produces better artifact-corrected neural dynamic signals than not low-pass filtering, using OLS regression, and not baseline-normalizing (i.e., only subtracting fitted isosbestic from experimental signals). Although these specific results are demonstrated with a particular set of simulation parameters, their logical basis applies across photometry scenarios, and we have found these results hold across simulation parameters. We have made a wide range of parameters accessible in the simulation scripts provided at https://github.com/philjrdb/RegressionSim; we welcome further tests and refinements.

To further demonstrate the ramifications of employing standard OLS versus robust IRLS regression, we applied these regression approaches to a subset of published photometry data ^[15]. In brief, the data used come from an experiment where GCaMP6f was selectively expressed in ventral tegmental area (VTA) dopamine (DA) neurons using transgenic TH-Cre rats. Calcium-dependent and -independent GCaMP fluorescence (465nm and 405nm excitation light, respectively) was recorded while rats made nosepoke responses for beer. Using exemplar data from a single subject during late acquisition phase (all data and MATLAB code available at https://github.com/philjrdb/RegressionSim), OLS and IRLS regression (Tukey’s bisquare, c = 1.4) were used to fit the isosbestic signal to the experimental signal.

In this dataset, OLS and IRLS regressions generated notably different fits (Figure 4A). The IRLS fit followed the baseline of the experimental signal more closely, whereas the OLS fit was shifted towards the large peaks in the experimental signal, which reliably occurred following nosepoke-elicited beer delivery. This difference impacted dF/F around beer deliveries (-3sec to 10sec) (Figure 4B). Without any other adjustments, the OLS-derived dF/F around beer deliveries was downshifted relative to IRLS-derived dF/F. For OLS-derived dF/F, 95% confidence intervals (t-based, 1/3sec temporal threshold ^[10]) indicated the entire 3sec lead-up to beer delivery as significantly below session baseline, while the excitatory component was only reliably detected 1sec after beer delivery. Taken at face value, this might be used to interpret pronounced reductions in VTA DA activity prior to nosepokes and delayed excitation to beer delivery (more strongly coinciding with beer consumption). Alternatively, this reliable “inhibition” prior to the event of interest might be used as indication of an invalid baseline, which could be combatted by re-baselining each trial to a manually selected pre-event period. However, as mentioned, this choice comes with the problematic assumption that the specified pre-event period acts as a valid baseline across trials.

IRLS appeared to bypass these issues without the need for additional, arbitrary analysis choices (Figure 4B). As with the simulation data, re-baselining was not necessary for IRLS-derived dF/F around beer deliveries; only a brief but significant reduction in dF/F is observed immediately before the nosepoke, and a robust excitatory transient is observed immediately following beer delivery. Although the advantage of IRLS over OLS for this dataset is conjecture, as the true calcium response of VTA DA neurons is technically unknown, these observations strongly align with findings from simulated data where true parameters were known. Thus, IRLS-based dF/F is demonstrably preferable to OLS-based dF/F with real photometry data.

Fiber photometry is a powerful neuroscientific tool, but experimental signals that track neural dynamics of interest are contaminated by photobleaching, movement-related and noise artifacts that undermine analysis and interpretation. Typically, noise is attenuated via low-pass filtering, while photobleaching and movement-related artifacts are addressed by using an isosbestic control signal within early steps of the analysis to obtain an artifact-corrected neural dynamic signal. Despite these solutions being identified in the early days of fiber photometry ^[1], approaches for extracting neural dynamics from experimental signals remain relatively unstandardized. There continues to be substantial variability in whether and how these analysis steps are applied across studies and analysis pipelines. We examined the impacts of low-pass filtering, baseline-normalization (dF/F vs dF), and types of regression (OLS vs IRLS) on artifact correction. As predicted, we found clear and specific advantages of low-pass filtering, IRLS and dF/F calculations over alternatives. We discuss these and other considerations below.

Low-pass filtering is a ubiquitous, relatively uncontroversial step in photometry analyses. It takes advantage of the fact that biosensors tend to operate on slower timescales (chronic to sub-second, but not microsecond) than major noise components, such as noise from electrical components. However, further signal processing, such as the use of bandstop filters, may be warranted if noise components below the ideal low-pass filter threshold are detected.

The common procedure of baseline-normalizing signals using the dF/F calculation as stipulated in Lerner et al. ^[1] was also shown to be more effective for counteracting artifacts, particularly photobleaching artifacts, than simply subtracting the fitted isosbestic from the experimental signal as done in some procedures. It is important to note that this within-session baseline-normalization is separate to, and is in fact undermined by, detrending and/or normalizing of experimental and isosbestic signals prior to the dF/F calculation ^[8,9]. That is, detrending and/or z-scoring experimental and isosbestic signals precludes their use in dF/F calculations as the original scale and all-positive values of recorded signals is necessary for appropriate baseline normalization. However, once dF/F is calculated, detrending and normalization steps for the artifact-corrected dF/F is possible (discussed below).

A key argument we make is that standard OLS regression is not ideal for fitting the isosbestic to the experimental signal because it overfits the isosbestic signal to the experimental signal’s neural dynamic component, while underfitting to the artifactual component it is meant to track. We show that robust regression, specifically IRLS regression, is conceptually and demonstrably better suited for this purpose. Interestingly, we found that lower tuning constants (i.e., more aggressive downweighting) produced better results across the constants tested. It is unclear whether this trend holds true for even lower tuning constants, or the degree to which this trend applies across scenarios. However, our findings suggest stronger over weaker applications of IRLS can be beneficial for extracting neural dynamics from experimental signals.

The key benefit of IRLS is that it better concentrates the fit of the isosbestic onto the artifactual component of the experimental signal, resulting in a more accurate extraction of the neural dynamic signal. By doing so, it also tracks the baseline of the neural dynamic signal more closely, reducing the need for problematic trial-by-trial re-baselining (discussed in previous sections). That said, local re-baselining may be warranted when longer-term fluctuations in neural dynamics skew the analysis of phasic signals. These baseline changes can be handled by a specified pre-event baseline period, albeit with the difficult assumption that this period serves as a valid baseline. Another option is to detrend these fluctuations. We recommend employing a smoothed moving median instead of a moving mean to avoid potential baselining issues that the latter is more susceptible to (overfitting to large phasic neural dynamics instead of the local baseline). Relatedly, if further normalization of the artifact-corrected dF/F is desired, we recommend using the null-Z, as employed in this study, to normalize dF/F instead of standard z-scoring. This is because a standard z-score normalizes signals relative to the overall mean, which may not reflect the IRLS dF/F baseline. In so doing, the overall mean would be made 0 as per OLS dF/F. The null-Z avoids this by normalizing the signal relative to 0 (i.e., the IRLS-derived baseline) instead of the overall mean.

Finally, a consideration not previously mentioned is the common step of excluding invalid recording periods, such as non-recording or patchcord disconnection periods. This logical step should be done as the first pre-processing step, as these invalid periods and transitions between valid and invalid recording periods can adversely affect signal processing and isosbestic regressions steps.

Taken together, we recommend measuring an isosbestic control signal within the same subject (preferably the isosbestic wavelength of the biosensor, otherwise a static fluorophore [e.g., mCherry]) and applying 4 analysis steps to better extract the true neural dynamic component of experimental biosensor signals:

Exclude invalid recording periods, such as logged disconnections, from further analysis.
Apply a low-pass filter to remove high-frequency noise components from experimental and isosbestic signals. The cut-off for this filter will depend on the kinetics of the biosensor and problematic noise components.
Fit the isosbestic signal to the experimental signal using iteratively reweighted least squares (IRLS) regression. We recommend using Tukey’s bisquare with a low tuning constant.
Calculate a baseline-normalized dF/F using the equation stated in ^[1].

These initial analysis steps are not exhaustive. There may be a need to apply other corrections and adjustments in some cases, as discussed above. However, these specific steps are beneficial over existing alternatives and should apply to most use-cases. Therefore, we argue these steps could serve as a standard component across photometry analysis pipelines.

Acknowledgements

The work described here was supported by grants from the Australian Research Council (PJRDB: DP220102317; GPM: DP220100040 and DP200102576). We thank Christopher Nolan for helpful comments on the manuscript.

Lerner, T. N. et al. Intact-Brain Analyses Reveal Distinct Information Carried by SNc Dopamine Subcircuits. Cell 162, 635-647 (2015). https://doi.org/10.1016/j.cell.2015.07.014
Gunaydin, L. A. et al. Natural neural projection dynamics underlying social behavior. Cell 157, 1535-1551 (2014). https://doi.org/10.1016/j.cell.2014.05.017
Xu, F., Shi, D.-Q., Lau, P.-M., Lin, M. Z. & Bi, G.-Q. Excitation wavelength optimization improves photostability of ASAP-family GEVIs. Molecular Brain 11, 32 (2018). https://doi.org/10.1186/s13041-018-0374-7
Zhang, Y. et al. Fast and sensitive GCaMP calcium indicators for imaging neural populations. Nature 615, 884-891 (2023). https://doi.org/10.1038/s41586-023-05828-9
Kallweit, J. et al. An Overview on Methods for Producing Side-Emitting Polymer Optical Fibers. Textiles 1, 337-360 (2021).
Lustermann, B. et al. Experimental determination and ray-tracing simulation of bending losses in melt-spun polymer optical fibres. Scientific Reports 10, 11885 (2020). https://doi.org/10.1038/s41598-020-68568-0
Barnett, L. M., Hughes, T. E. & Drobizhev, M. Deciphering the molecular mechanism responsible for GCaMP6m's Ca2+-dependent change in fluorescence. PLoS One 12, e0170934 (2017). https://doi.org/10.1371/journal.pone.0170934
Martianova, E., Aronson, S. & Proulx, C. D. Multi-Fiber Photometry to Record Neural Activity in Freely-Moving Animals. JoVE, e60278 (2019). https://doi.org/doi:10.3791/60278
Conlisk, D. et al. Integrating operant behavior and fiber photometry with the open-source python library Pyfiber. Scientific Reports 13, 16562 (2023). https://doi.org/10.1038/s41598-023-43565-1
Jean-Richard-dit-Bressel, P., Clifford, C. W. G. & McNally, G. P. Analyzing Event-Related Transients: Confidence Intervals, Permutation Tests, and Consecutive Thresholds. Frontiers in Molecular Neuroscience 13 (2020). https://doi.org/10.3389/fnmol.2020.00014
Jean-Richard-dit-Bressel, P., Clifford, C. W. G. & McNally, G. P. Analyzing Event-Related Transients: Confidence Intervals, Permutation Tests, and Consecutive Thresholds. Frontiers in Molecular Neuroscience 13, 1 (2020). https://doi.org/10.3389/fnmol.2020.00014
Jean-Richard-dit-Bressel, P., Tran, J., Didachos, A. & McNally, G. P. Instrumental aversion coding in the basolateral amygdala and its reversion by a benzodiazepine. Neuropsychopharmacology 47, 1199-1209 (2022). https://doi.org/10.1038/s41386-021-01176-2
Ghareh, H. et al. Role of anterior insula cortex in context-induced relapse of nicotine-seeking. eLife 11, e75609 (2022). https://doi.org/10.7554/eLife.75609
Crouse, R. B. et al. Acetylcholine is released in the basolateral amygdala in response to predictors of reward and enhances the learning of cue-reward contingency. eLife 9, e57335 (2020). https://doi.org/10.7554/eLife.57335
Liu, Y. et al. The Mesolimbic Dopamine Activity Signatures of Relapse to Alcohol-Seeking. J Neurosci 40, 6409-6427 (2020). https://doi.org/10.1523/JNEUROSCI.0724-20.2020

No competing interests reported.

Download PDF

Version 1

posted

You are reading this older preprint version

Read the latest preprint version →

Obtaining artifact-corrected signals in fiber photometry: Isosbestic signals, robust regression and dF/F calculations

Status:

Version 1

Abstract

Figures

Introduction

Comparing artifact correction options via photometry simulations

Comparing OLS versus IRLS regression on real photometry data

Considerations and recommendations

Declarations

Acknowledgements

References

Additional Declarations

Supplementary Files

Status:

Version 1