Quantifying Cell Activity

Quantifying Cell Activity#

The gold-standard formula for measuring cellular activity is “Delta F over F₀”, or the change in fluorescence intensity normalized by the baseline activity:

\[\Delta F/F_0 = \frac{F - F_0}{F_0}\]

Here, F₀ a user-defined baseline that may be static (e.g., median over all frames) or dynamically estimated using a rolling window.

This guide assumes you’ve read this scientifica article by Dr Peter Rupprecht at the University of Zurich.

Blog-Post Takeaways

  • \(\Delta F/F\) reflects intracellular calcium levels, but often in a nonlinear way

  • \(\Delta F/F\) was initially introduced for organic dye calcium indicators in the 1930s

  • GCaMP behaves differently due to its low baseline brightness and its nonlinearity

  • There is no single recipe on how to compute \(\Delta F/F\)

  • Computation of \(\Delta F/F\) must be adapted to cell types, activity patterns, and noise

  • Interpretation of \(\Delta F/F\) requires knowledge about indicators, cell types, and confounds

Cell Activity (DF/F)

ΔF/F traces for 30 cells, sorted by similarity using the Rastermap algorithm.#

Baseline F₀ Strategies#

Choosing the correct baseline strategy depends on many factors:

  • Cell type

  • Brain location

  • Frame-rate

  • Virus Kinetics

  • Scientific question

This table summarizes Calcium Activity detection strategies used in embryoinic development reviewed by Huang et al. [3].

Method

How It’s Performed

Pros

Cons

Percentile Baseline Subtraction

Use a moving window to define F₀ as low percentile (e.g., 10–20th) of the trace.

Adaptive baseline; handles long-term drift.

Choice of window size/percentile affects sensitivity.

Z-Score Thresholding

Normalize trace by mean/SD, define events above N standard deviations.

Removes baseline drift; good for noisy data.

Sensitive to noise if SD is low; assumes Gaussian distribution.

dF/F Thresholding

Compute (F − F₀)/F₀ and define a fixed or adaptive threshold for events.

Widely used; compatible with GCaMP, Fluo dyes.

Sensitive to F₀ definition; arbitrary thresholds can bias results.

Standard Deviation (SD) Masking

Define active frames/regions where ΔF exceeds N×SD of baseline.

Objective thresholding for event detection.

Threshold choice heavily affects results.

Image Subtraction (Frame-to-Frame)

Compute ΔF = Fₙ − Fₙ₋₁ or F − background to detect sudden changes.

Simple, fast; used for wave detection.

Sensitive to noise; misses gradual changes.
Comparison of dF/F baseline strategies

Comparison of different ΔF/F₀ baseline correction methods across selected cells.#

Resulting DFF values from baseline strategies

The resulting ΔF/F₀ traces can look different depending on the chosen baseline strategy.#

Resulting DFF values from baseline strategies with Events

The resulting ΔF/F₀ traces can look different depending on the chosen baseline strategy.#

Standardized Noise Levels#

\[\nu = \frac{\mathrm{median}_t\left( \left| \frac{\Delta F_t}{F_0} - \frac{\Delta F_{t+1}}{F_0} \right| \right)}{\sqrt{f_r}}\]

  • Compare shot-noise levels across datasets, recordings, experiments

  • Takes advantage of the slow calcium signal, which is typically similar between adjacent frames

  • Median to exclude outlies stemming from fast onset

  • Norm by sqrt(framerate) makes metric comparable across datasets with different framerates

  • Shot-noise decreases with #samples with a square root dependency

The resulting units, % for dF/F, divided by the square root of seconds, is strange.

This metric should be used relative to other datasets.

High vs Low Noise Levels

Top N neurons with highest and lowest standardized noise levels.#

Pipelines#

There are several pipelines for users to choose from when deciding how to process calcium imaging datasets. Often times, the outputs of these pipelines are in units that are not well documented. Even worse, the inputs to downstream pipelines pipelines require data to have particular units.

CaImAn#

See: detrend_df_f

CaImAn computes ΔF/F₀ using a running low-percentile baseline.

By default, it uses the 8th percentile over a 500 frame window. The idea is to track the lower envelope of the signal to get F₀ without being biased by transients.

Neuropil/background: CaImAn handles this as part of its CNMF model [5].

Background and neuropil are explicitly separated into distinct spatial/temporal components, so the output traces are background subtracted during this factorization (as was the issue in the above example).

There is a strong argument to be made that a matrix factorization CNMF is not complex enough to model the true background and neuropil.

CaImAn default DF/F strategy

CaImAn uses a default baseline of the lower 8th percentile and a moving window of 200 frames.#

Giovannucci et al. [2]

Suite2p#

Suite2p does not output traces in ΔF/F₀ format directly.

Instead, it gives you the raw trace and the neuropil, along with spike estimates if you ran deconvolution. The neuropil represents fluorescence from the surrounding non-somatic tissue. As an optional step, many experimentors apply a fixed subtraction:

# F is an [n_neurons x time] array of raw signal
# Fneu is an [n_neurons x time] array of neuropil
F_corrected = F - 0.7 * Fneu

The 0.7 is an empirically chosen scalar to account for the partial contamination. Essentially you subtract 70% of the signal contained in the surrounding neuropil.

To compute ΔF/F₀, you divide trace, be that neuropil-corrected or not, by an F₀ you calculate yourself.

Internally, when running spike deconvolution, suite2p will convert your data into DF/F0 internally. For this, the default F₀ estimate comes from a “maximin” filter: smooth the trace with a Gaussian (default ~10 s), take a rolling min, then a max over a 60 s window.

_images/dff_oasis.png

Raw, neuropil, ΔF/F₀ and resulting deconvolved spikes as output by Suite2p#

Pachitariu et al. [6]

EXTRACT#

EXTRACT outputs raw fluorescence signals without built-in ΔF/F₀ calculation. You compute it yourself using something like a low-percentile (e.g. 10%) as F₀. Most people use a global or sliding percentile window.

Neuropil: Handled implicitly. The algorithm uses robust factorization to ignore background and neuropil. There’s no explicit subtraction or coefficient to tune. It isolates only what fits a consistent spatial footprint and suppresses outliers by design.

Inan et al. [4], Dinç et al. [1]

Pipeline

F₀ Method

ΔF/F₀

Neuropil

CaImAn

8th percentile, 500-frame window

Yes, in pipeline

Modeled via CNMF, no manual subtraction

Suite2p

Maximin (default) or 8th percentile

No, user divides post hoc

0.7 × Fneu subtracted before baseline

EXTRACT

User-defined (e.g. 10th percentile)

No, user computes

Implicitly handled via robust model

Calculating ΔF/F#

The most important consideration you must consider is how you calculate your baseline activity, which depends on your experimental question.

Example ΔF/F trace baseline comparisons

Example of ΔF/F trace showing different baseline choices. Adapted from Fig. 1.#

  1. Spike inference from mouse spinal cord calcium imaging data (Rupprecht et al. [7])

  • DF/F = Lower 10th percentile

  • GCaMP6s

  1. Spike inference on GCaMP8 indicators Rupprecht et al. [8]

  • GCaMP8, GCaMP7f

  1. GCaMP6 calibration study Huang et al. [3]

  • ΔF/F = (F − F₀,local)/F₀,global, where F₀,local is the mean fluorescence over 100 ms before the first AP, and F₀,global is the minimum F₀,local across trials

  • GCaMP6f

Indicator Dynamics#

GCaMP Variant

Optimal Tau (s)

Notes / Sources

GCaMP6f (fast)

~0.5–0.7 s

Suite2p: ~0.7 s (docs). OASIS/CNMF: ~0.5–0.7 s (PMC). CaImAn: ~0.4 s (docs).

GCaMP6m (medium)

~1.0–1.25 s

Suite2p: ~1.0 s. OASIS: ~1.25 s (PMC). CaImAn: ~1.0 s (docs).

GCaMP6s (slow)

~1.5–2.0 s

Suite2p: 1.25–1.5 s. OASIS/Suite2p: ~2.0 s (PMC). CaImAn: ~1.5–2.0 s.

GCaMP7f (fast)

~0.5 s

Spike inference tuning (bioRxiv). Similar to GCaMP6f.

GCaMP7m (medium)

~1.0 s (est.)

Estimated by analogy to GCaMP6m. No default in tools.

GCaMP7s (slow)

~1.0–1.5 s

In vivo half-decay ~0.7 s (eLife). Tau ≈ 1.0 s.

GCaMP8f (fast)

~0.3 s

OASIS fine-tuning (bioRxiv). Fastest decay; tenfold faster than 6f/7f (Nature).

GCaMP8m (medium)

~0.3 s

Slightly slower than 8f, still ~0.3 s (bioRxiv; Nature).

GCaMP8s (slow)

~0.7 s

Spike inference optimal tau ~0.7 s (bioRxiv). Faster than 6s (Nature).

References#

[1]

Fatih Dinç, Hakan Inan, Oscar Hernandez, Claudia Schmuckermair, Omer Hazon, Tugce Tasci, Biafra O. Ahanonu, Yanping Zhang, Jérôme Lecoq, Simon Haziza, Mark J. Wagner, Murat A. Erdogdu, and Mark J. Schnitzer. Fast, scalable, and statistically robust cell extraction from large-scale neural calcium imaging datasets. bioRxiv, 2024. URL: https://www.biorxiv.org/content/early/2024/08/17/2021.03.24.436279, arXiv:https://www.biorxiv.org/content/early/2024/08/17/2021.03.24.436279.full.pdf, doi:10.1101/2021.03.24.436279.

[2]

Andrea Giovannucci, Johannes Friedrich, Pat Gunn, Jérémie Kalfon, Brandon L Brown, Sue Ann Koay, Jiannis Taxidis, Farzaneh Najafi, Jeffrey L Gauthier, Pengcheng Zhou, Baljit S Khakh, David W Tank, Dmitri B Chklovskii, and Eftychios A Pnevmatikakis. Caiman an open source tool for scalable calcium imaging data analysis. eLife, 8:e38173, jan 2019. URL: https://doi.org/10.7554/eLife.38173, doi:10.7554/eLife.38173.

[3] (1,2)

Lawrence Huang, Peter Ledochowitsch, Ulf Knoblich, Jérôme Lecoq, Gabe J Murphy, R Clay Reid, Saskia EJ de Vries, Christof Koch, Hongkui Zeng, Michael A Buice, Jack Waters, and Lu Li. Relationship between simultaneously recorded spiking activity and fluorescence signal in gcamp6 transgenic mice. eLife, 10:e51675, mar 2021. doi:10.7554/eLife.51675.

[4]

Hakan Inan, Murat A Erdogdu, and Mark Schnitzer. Robust estimation of neural signals in calcium imaging. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. URL: https://proceedings.neurips.cc/paper_files/paper/2017/file/e449b9317dad920c0dd5ad0a2a2d5e49-Paper.pdf.

[5]

Haifeng Liu, Zhaohui Wu, Xuelong Li, Deng Cai, and Thomas S. Huang. Constrained nonnegative matrix factorization for image representation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(7):1299–1311, 2012. doi:10.1109/TPAMI.2011.217.

[6]

Marius Pachitariu, Carsen Stringer, Mario Dipoppa, Sylvia Schröder, L. Federico Rossi, Henry Dalgleish, Matteo Carandini, and Kenneth D. Harris. Suite2p: beyond 10,000 neurons with standard two-photon microscopy. bioRxiv, 2017. URL: https://www.biorxiv.org/content/early/2017/07/20/061507, arXiv:https://www.biorxiv.org/content/early/2017/07/20/061507.full.pdf, doi:10.1101/061507.

[7]

Peter Rupprecht, Wei Fan, Steve J. Sullivan, Fritjof Helmchen, and Andrei D. Sdrulla. Spike inference from mouse spinal cord calcium imaging data. bioRxiv, 2024. URL: https://www.biorxiv.org/content/early/2024/07/22/2024.07.17.603957, arXiv:https://www.biorxiv.org/content/early/2024/07/22/2024.07.17.603957.full.pdf, doi:10.1101/2024.07.17.603957.

[8]

Peter Rupprecht, Márton Rózsa, Xusheng Fang, Karel Svoboda, and Fritjof Helmchen. Spike inference from calcium imaging data acquired with gcamp8 indicators. bioRxiv, 2025. URL: https://www.biorxiv.org/content/early/2025/05/26/2025.03.03.641129, arXiv:https://www.biorxiv.org/content/early/2025/05/26/2025.03.03.641129.full.pdf, doi:10.1101/2025.03.03.641129.

Contents