modified gaussian mixture model (gmm)  (MathWorks Inc)

Journal: bioRxiv

Article Title: Spike inference from calcium imaging data acquired with GCaMP8 indicators

doi: 10.1101/2025.03.03.641129

Figure Lengend Snippet: a , Examples of isolated action potentials with the corresponding ΔF/F trace (gray) and the inferred spike rate (blue) for an increasing standardized noise level ν . A low-noise population recording corresponds to ν ≈ 2, a higher-noise population recording to ν ≈ 8. Arrows indicate the time point when the true isolated action potential occurred (red arrow if not detected). Single action potentials are accurately inferred for low noise levels for GCaMP8m/s. b , Top left: Illustration of the integral of inferred action potentials (APs) in the shaded time window around the true action potential. Other panels: Number of detected action potentials (APs) for a true isolated single AP, plotted as a distribution (kernel density estimate). Shades of gray indicate the different noise levels (lowest noise level, dark grey; highest noise level, light grey). c , Fraction of APs correctly detected as a single AP (spike count >0.5 and <1.5) across datasets and noise levels. d , Fraction of APs correctly detected from noise (spike count >0.5) across datasets and noise levels. e , Percent of APs per neuron correctly detected as a single AP as defined in (c), averaged across noise levels ν = 2-4 for robustness. f , Percent of APs per neuron detected as defined in (d), averaged across noise levels ν = 2-4. g , Example of a calcium imaging recording (GCaMP8s ground truth) together with spike rates inferred using GC8s-tuned CASCADE. Isolated minimal events are visually detectable (circles). h , Histogram of inferred spike numbers for isolated events, together with a Gaussian mixture model (GMM) fit to the underlying data. The unitary amplitude, defined as the mean of the first Gaussian component, is used for auto-calibration. i , Auto-calibration with unitary amplitudes derived from GMM fits decreases the overestimation of the spike rate for isolated spike events and brings back outliers. (But see also ).

Article Snippet: A modified Gaussian mixture model (GMM) in MATLAB was then used to perform a quantal analysis from the histogram to infer the unitary amplitude.

Techniques: Isolation, Imaging, Derivative Assay

Journal: bioRxiv

Article Title: Spike inference from calcium imaging data acquired with GCaMP8 indicators

doi: 10.1101/2025.03.03.641129

Figure Lengend Snippet: To evaluate the transfer function of various CASCADE models and their nonlinearity, a synthetic ground truth was generated. The ground truth spike patterns consisted of experimentally recorded action potentials of excitatory neurons from refs. , , , (n = 237 neurons with variable recording duration, each represented by a gray line; 102,093 action potentials, 32.7 hours of recordings) to ensure naturalistic spike patterns. These experimentally obtained spike patterns were convolved with a double exponential kernel (rise time 5 ms, decay time 500 ms, peak amplitude 40% ΔF/F), with Gaussian noise added to reach a standardized noise level of “8”. Spike inference algorithms were applied to this entire linear synthetic dataset. From the inferred spike rates and the synthetic ground truth, transfer functions as in were retrieved, allowing to judge the linearity of the models trained on specific datasets. The Default CASCADE model exhibited the highest nonlinearity (black average transfer curve) compared to the weighted linear fit of the inferred spike rates (red dashed line); the sublinearity of the average transfer function at high firing rates reflects the supralinearity of the training data. CASCADE trained with GCaMP8 ground truth yielded a relatively linear transfer function, with mild but clear signs of saturation for models fine-tuned for GCaMP8f and GCaMP8s. The transfer function obtained from a model trained with GCaMP8m (GC8m-tuned CASCADE) data was the most linear. Since the models reflect the nonlinearities of their training data, these analyses indicate that GCaMP8, and in particular GCaMP8m, are distinctly more linear calcium indicators than for example GCaMP6 (which is the basis of most of the training data for Default CASCADE) across the firing rate regime that is typically covered by experimentally obtained ground truth recordings. The bottom right panel shows quantified deviations from linearity as introduced in for the different CASCADE models. Quantifications were pooled across multiple simulations with variable setting of the kernel parameters to emulate GCaMP8f, GCaMP8m, GCaMP8s, GCaMP7f and GCaMP6 with the kernel rise times (2, 3, 4, 11 and 50 ms, respectively), decay times (45, 80, 190, 100 and 300 ms) and peak amplitudes (55, 71, 81, 70 and 25 % ΔF/F). Deviations from linearity were reduced compared to Default CASCADE by 30% (GC8-trained), 19% (GC8f-trained), 60% (GC8m-trained) and 48% (GC8s-trained). Therefore, the GC8m-trained model exhibited the lowest deviation from linearity, suggesting the highest linearity of the GC8m indicator used for its training. All comparisons p ≪ 10 −10 , Wilcoxon signed-rank test, compared across n = 1422 instances of simulated neurons.

Article Snippet: A modified Gaussian mixture model (GMM) in MATLAB was then used to perform a quantal analysis from the histogram to infer the unitary amplitude.

Techniques: Generated

Journal: bioRxiv

Article Title: Spike inference from calcium imaging data acquired with GCaMP8 indicators

doi: 10.1101/2025.03.03.641129

Figure Lengend Snippet:

Article Snippet: A modified Gaussian mixture model (GMM) in MATLAB was then used to perform a quantal analysis from the histogram to infer the unitary amplitude.

Techniques: Isolation