DOC: For non-neuronal cells, should the deconvolved output be disregarded?

[RandallJEllis]

Issue with current documentation:

My group is imaging GCaMP6f in cultured cancer cells. We have used Suite2p to analyze the data, but we are wondering, since they are not neurons, if we should use the deconvolved output (spks.npy), or the cell fluorescence (F.npy). Any clarification on this would be greatly appreciated.

Idea or request for content:

No response

[carsen-stringer]

maybe not? the calcium events in neurons are caused by spikes and that's what we try to extract with deconvolution since the gcamp decay timescale is much longer than those events, but I don't know what happens in cancer cells. cool to hear it's working in that cell type though!

[RandallJEllis]

Thank you @carsen-stringer! I'm trying to figure out which parameters would be worth changing to adapt the pipeline to cancer cells. I'm thinking win_baseline and sig_baseline because the activity is much slower than action potentials, but if any other parameters come to mind, that would be very helpful.

[RandallJEllis]

Trying to simplify this question: Is there anything about the oasis_trace function that is neuron-specific and disqualifies it from being used in other cell types?

1. I see that Reference 26 in the Suite2p paper is to a paper that uses non-negative deconvolution in xenopus laevis embryos, so I'm guessing oasis_trace is cell-type agnostic, unless there are specific details about the exponential kernel that make it neuron-specific?
2. In the review paper about deconvolution, there is this quote:

An incorrect interpretation of deconvolved spike trains is that they represent the “firing rates” or “probabilities” of neural spiking; this is not true. Instead, they represent a time-localized and noisy estimate of the calcium influx into the cell. The calcium influx events may vary in size from spike to spike (Figure 4e and f). Therefore, a deconvolution algorithm which estimates event amplitudes cannot perfectly reconstruct a spike train even when the exact spike times are known from simultaneous electrophysiology [47] (Figure 4d and g). This variability in spike-evoked amplitudes in turn places an upper bound on blind spike detection accuracy. The upper bound is saturated when performance is evaluated in bins of 320 ms or larger (Figure 4g). In smaller bins, the deconvolution does not saturate the upper performance bound (Figure 4g), which appears to be due to variability in the timing of the deconvolved spike times on the order of ≈100 ms (Figure 4h).

I don't see the 100ms-320ms window hard-coded anywhere in the Suite2p codebase, so I'm guessing this assumption wasn't used?

3. Does the frame rate impact how the exponential kernel acts based on assumed event length? As in, if you have a high frame rate for slower events that would occur in non-neural cells, could this make the kernel act in inappropriate ways during deconvolution?

Any clarification on this is tremendously appreciated.