Supplementary Figure 4: Simulations of SLAP activity reconstruction.

We evaluated solver performance by reconstructing simulated data under a variety of conditions. The measurement parameters in these simulations were chosen to match our 1016 Hz imaging experiments. To assess sensitivity of source recovery to solver/simulation parameters, we modified each parameter while holding the others constant. In panels A,B, and H, the simulated sample consisted of numerous square segments (10*10 pixels) placed randomly (uniformly chosen) within a 500 pixel diameter. In the remaining panels, the simulated sample was generated from a segmented yGluSnFR-labeled volume from one of our experiments. Peak ΔF/F₀ for each segment was uniformly distributed in the interval [0.5, 5]. We performed 500 iterations of the dynamic Richardson Lucy algorithm, or normal Richardson Lucy deconvolution for comparison, where stated. To evaluate performance, we calculated the Pearson correlation between the estimated ΔF/F₀ trace and its ground-truth counterpart over time for each segment, averaged over segments. a. Effect of sample brightness on imaging performance. The expected Poisson-distributed photon rates per segment per frame was varied, in the range 0.1-10⁵. Typical values of this parameter in the recorded datasets in this work are around 100. Recovery improves at higher photon rates. b. Effect of number of sources on imaging performance. The number of sources was varied, in the range 10-1000, in both the generation of ground truth data and the segmentation algorithm. Increasing the number of sources decreases solver performance. c. Effect of incorrectly-estimated decay time constant. The ground truth activity was generated using a time constant of 100 frames. The time-constant of source recovery was varied in the range 0-1000. The solver performs best when the correct time constant is used. The solver is insensitive to underestimation of the time-constant, as this can be compensated with additional spikes (w) but does reduce the denoising benefit of the dynamics model. A time constant of 0 corresponds to normal Richardson-Lucy iterations. Substantial overestimation of the time constant degrades performance of the solver. d. Effect of incorrect reference image. We evaluated the performance of the solver in the case where ground truth sources were omitted from the segmentation. We added 0-50% additional unsegmented sources, while quantifying reconstruction performance over the segmented sources. This is meant to simulate fluorescence activity within the ‘ON’ region of the SLM not reflected in the reference image, a possibility in sensors with very low baseline fluorescence. Increasing unsegmented activity degrades the performance of the solver. e. Effect of alignment errors. We shifted the reference image provided to the solver by 0-10 pixels, after registration. Better alignment improves the reconstructions. f. Effect of segmentation spatial scale. We performed source recovery using finer or coarser segmentations than those used to generate the data. Segmentation differing from the ground truth degrade performance, which can be partially compensated for by using a finer segmentation scale. g. Estimation of pairwise correlations. In each simulation, we generated 3000 frames of correlated random spike amplitudes (w) using vines and the extended onion method⁴¹, to which we applied the dynamics model to produce ground truth ΔF/F₀ timeseries with source-to-source correlations in the range [−0.6 0.6]. Measurements were simulated from these ground truth traces and reconstructed. Plotted are pairwise correlation coefficients in the ground truth vs. reconstructed data using (left) normal RL iterations or (right) Dynamic RL iterations. h. SLAP performance with different numbers of tomographic angles. We simulated the performance of SLAP source recovery with different numbers of equally-spaced line angles, at two different sample densities. At the higher sample density (500 sources), increasing the number of line angles from four to 8 improved performance, equivalent to approximately doubling the photon rate. At the lower sample density (100 sources), the corresponding improvement was negligible. Parameters settings for the above experiments: # random instances: A,G:5 B,C,D,E,H:40; F:10. # frames per instance: A,B,E: 350 ; B,C,F,H:500 G: 3000. # segments: A,C,E,G: 500 ; Photons per frame per segment: B-F: 100. τ: 100