Self-organization of modular activity in immature cortical networks

During development, cortical activity is organized into distributed modular patterns that are a precursor of the mature columnar functional architecture. Theoretically, such structured neural activity can emerge dynamically from local synaptic interactions through a recurrent network with effective local excitation with lateral inhibition (LE/LI) connectivity. Utilizing simultaneous widefield calcium imaging and optogenetics in juvenile ferret cortex prior to eye opening, we directly test several critical predictions of an LE/LI mechanism. We show that cortical networks transform uniform stimulations into diverse modular patterns exhibiting a characteristic spatial wavelength. Moreover, patterned optogenetic stimulation matching this wavelength selectively biases evoked activity patterns, while stimulation with varying wavelengths transforms activity towards this characteristic wavelength, revealing a dynamic compromise between input drive and the network’s intrinsic tendency to organize activity. Furthermore, the structure of early spontaneous cortical activity – which is reflected in the developing representations of visual orientation – strongly overlaps that of uniform opto-evoked activity, suggesting a common underlying mechanism as a basis for the formation of orderly columnar maps underlying sensory representations in the brain.

(Mean (+/-SEM) of n=40 trials for each experiment at varying power densities (n=48 experiments).Pearson's r=0.615, p<0.001) (g) The wavelength of opto-evoked events is invariant to power density (Pearson's r=-0.015,p=0.923) and resides in a narrow band consistent with the wavelength of spontaneous activity (blue, mean +/-std.dev.across animals).Average wavelength (+/-SEM) across modular trials (trials with modularity >2 standard deviations over mean at baseline in 1 second window before stimulus onset) for each power density tested.For (d-g) data is pooled across 7 animals.Similarity of spontaneous and opto-evoked correlation networks is high for all seed points.(c) Mean similarity of spontaneous vs opto-evoked networks is equivalent to that for spontaneous activity with itself (purple: spontaneous vs spontaneous, mean (+/-SEM)) and is significantly greater than surrogate data (Mean opto vs spontaneous similarity: 0.512 (+/-0.028);opto vs surrogate similarity: 0.018 (+/-0.006),p =0.016, WSR, n=7 animals).

Fig. S3 .
Fig. S3.Opto-evoked modularity increases nonlinearly with stimulus power, while the wavelength remains constant.(a-c) Uniform opto-evoked responses for low (0.5 mW/mm 2 , a), intermediate (1 mW/mm 2 , b) and high (10 mW/mm 2 , c) stimulus power.(Left) example traces of mean ∆F/F over ROI for 2 trials shown to right.(Right) Example opto-evoked activity at stimulus offset.Quantification of modular amplitude at the characteristic wavelength (F1), see Methods) and modularity (Mod.) are given for each event.(d) Modular amplitude of activity switches from a distribution peaked at low amplitude responses for low power densities to a distribution peaked at high amplitude responses at high power densities, consistent with the presence of a dynamic instability for sufficiently strong input drive.Density plot across animals, normalized within column.(e) Modularity shows non-linear increase with stimulus power.(f) Amplitude of overall activity (mean ΔF/F across field of view) increases with increasing power.(Mean(+/-SEM) of n=40 trials for each experiment at varying power densities (n=48 experiments).Pearson's r=0.615, p<0.001) (g) The wavelength of opto-evoked events is invariant to power density (Pearson's r=-0.015,p=0.923) and resides in a narrow band consistent with the wavelength of spontaneous activity (blue, mean +/-std.dev.across animals).Average wavelength (+/-SEM) across modular trials (trials with modularity >2 standard deviations over mean at baseline in 1 second window before stimulus onset) for each power density tested.For (d-g) data is pooled across 7 animals.

Fig. S4 .
Fig. S4.Uniform opto-evoked activity has similar wavelength and modularity to spontaneous activity.(a) Experimental schematic.Spontaneous activity is highly modular, consistent with prior work.(b) Distribution of wavelength across spontaneous events is narrow, consistent with a characteristic wavelength (transparent lines=distribution in individual animals, opaque line=group mean).(c-e) Opto-evoked activity patterns and spontaneous activity share similar spatial wavelength (c, p= 0.742, Wilcoxon signed-rank test (WSR)), narrowness of wavelength distribution across patterns (d, p= 0.383, WSR), and degree of modularity (e, p= 0.547, WSR).(Open circles=average across n=8 individual animals, horizontal line=group mean +/-SEM.)

Fig
Fig. S6.Spatially structured opto-evoked activity without spatial filtering.(a) As in Fig. 3a, three example optogenetic stimulation patterns.(b-c) Opto-evoked ΔF/F fluorescence activity without spatial filtering, for individual events (b) and mean responses across trials (c).(d-f) same as a-c, except for optogenetic stimuli of differing wavelengths, corresponding to data shown in Fig 4d-f.

Fig
Fig. S7.Luminance evoked visual responses are highly reliable but poorly explain variation in opto-evoked activity.(a) Experimental schematic.(b) Full-screen increases (ON) and decreases (OFF) in luminance drive reliable visually evoked responses through the closed eyelid.(c) Mean ON (left) and OFF (right) responses are modular.(d) Difference map showing ON -OFF selectivity.(e) Trial-to-trial correlation matrix comparing opto-evoked activity and visuallyevoked ON and OFF responses.Note reliability within and selectivity across ON and OFF responses, together with weaker and variable similarity to opto-evoked responses.(f-g) Visuallyevoked ON/OFF responses are less variable and lower dimensional than opto-evoked responses, requiring fewer principal components to explain the majority of variance in the data (f) and having significantly lower dimensionality (g).Circles indicate dimensionality of individual animals calculated from n=40 trials, lines indicate group mean (+/-SEM), n=7 animals.p=0.016,WSR test.(h) Projecting opto-evoked, spontaneous, and cross-validated visual events into the principal component space of visually-evoked activity, the first two PCs explain the majority of the variance of visual activity, while the majority of variance is concentrated on higher PCs for opto and spontaneous events.Inset: enlarged view of first 10 PCs.(i) The PCs of visually-evoked activity that explain the majority of variance of this activity poorly explain the variance in optoevoked activity, and are indistinguishable from surrogate events.Black dashed line indicates unity.

Fig. S8 .
Fig. S8.Structured optogenetic stimulation drives more reliable cortical responses than uniform stimulation.(a) Trial-to-trial correlations between individual trials are greater when comparing trials driven by the same input pattern (within pattern, circles) versus trials driven by non-matching input stimuli (across pattern, squares).Additionally, within pattern trial-to-trial correlations are also greater than both trial-shuffled responses (x's) and responses to uniform optogenetic stimulation (Blue line, mean trial-to-trial correlation, +/-SEM).Data shown for 1 representative experiment.(b) Across 6 animals, structured optogenetic stimulation (within pattern -WP) drives more reliable responses than trial shuffled (TS) and uniform optogenetic stimulation.Within pattern data are averaged across all 3 stimuli (shown as mean +/-average SEM).Crosses indicate mean (+/-SEM) for each group.Within pattern vs trial shuffled: p=0.016; within pattern vs uniform stimulation: p=0.016; trial shuffle vs uniform stimulation: p=0.578,WSR, n=7 animals.(c) For only pixels with large changes in activity (|z| > 2), average similarity of individual opto-evoked trials to their respective stimulus patterns, compared to trial shuffled similarity.Magnitude of similarity is comparable to when all pixels within FOV are included in calculating similarity (Fig 3f).Offset solid black circle and thick grey x indicate mean (+/-SEM) for group, p=0.016,WSR, n=7 animals.

Fig. S9 .
Fig. S9.Pharmacological silencing of feedforward LGN inputs to cortex.(a) Experimental schematic.Feed-forward activity is blocked via muscimol injection to LGN.Silencing is confirmed by measuring responses to full-screen luminance changes.(b) Mean response in V1 to luminance changes prior to (left) and following (right) muscimol injection confirms silencing of feed-forward projections.Black trace indicates stimulus triggered mean response, grey indicates individual trials (n trials=40).Peaks in activity following LGN muscimol are poorly timed to visual stimulus, and likely reflect spontaneous cortical activity.

Fig. S10 .
Fig. S10.Uniform opto-evoked events reveal large-scale correlated networks that are highly similar to spontaneous networks.(a) Spatial correlations across spontaneous (left) and opto-evoked (right) events show highly similar patterns for corresponding seed points.(b)Similarity of spontaneous and opto-evoked correlation networks is high for all seed points.(c) Mean similarity of spontaneous vs opto-evoked networks is equivalent to that for spontaneous activity with itself (purple: spontaneous vs spontaneous, mean (+/-SEM)) and is significantly greater than surrogate data (Mean opto vs spontaneous similarity: 0.512 (+/-0.028);opto vs surrogate similarity: 0.018 (+/-0.006),p =0.016, WSR, n=7 animals).

Fig. S11 .
Fig. S11.Overlap of spatially structured optogenetic stimuli to spontaneous activity.(a) Additional example of opto-evoked event (left, green), its maximally correlated spontaneous event (middle, black), and the overlap in active domains on top of the stimulus input pattern (right, grey).Data for stimulus pattern shown in Fig 3. (b) Maximum correlation to spontaneous activity for each individual trial.Opto (green) correlations from a single example stimulus pattern shown in (a), Null (black) correlations from random event-matched spontaneous vs spontaneous subsamples (1 example from n=500 simulations, see Methods.p=0.005,Wilcoxon rank-sum test).(c) Average maximum correlation across trials for each stimulus pattern (n=21 stimuli, pooled across animals.p<0.002 for 19 of 21 stimulus patterns, bootstrap test, see Methods.Closed circles=stimuli with p<0.002, open circles=n.s).(d-e) Computational modeling predictions (h=0.4),showing that structured inputs can drive patterns with novel components that are not entirely explained by spontaneous events.(d) Projections of activity patterns driven by spatially structured inputs (Opto, green) onto control activity patterns driven by non-structured noise (Spont, black).Surrogate events (yellow) for in silico activity patterns generated the same way as in vivo activity patterns (see Methods), so as to maintain statistics of image but randomize spatial pattern.Cross validated control projections done by randomly subsampling event-matched spontaneous patterns and projecting them onto a separate event-matched size spontaneous test distribution, as in Fig 6g.(e) Total variance explained by spontaneous PCs, for both control and spatially structured driven datasets (summed across first 40 components).Spontaneous PCs explain a significantly lesser proportion of the variance in the Opto data, indicating that spatially structured stimuli can drive outputs with novel components.n=10 stimulus patterns, p=0.0019,WSR.

Fig. S12 .
Fig. S12.Response to optogenetic stimulation across imaging field-of-view.(a) Average response to full-field optogenetic stimulation shows broad opto-response across millimeters of FOV.(Scale bar=1 mm).(b) Testing responsiveness to optogenetic stimulation with a 4 x 4 grid stimulus.Each grid square is sequentially stimulated (1 sec duration, 11 trials).(c) Stimulus triggered average response for pixels within stimulus ROI, showing robust and reliable response to optogenetic stimulation across the FOV (grey=individual trials, black=mean (+/-SEM)).ROIs on edge of FOV are poorly driven, due to only partial coverage with stimulus ROI.(d) Mean response to optogenetic stimulation across trials.Dashed red=opto-stimulus ROI, yellow=imaging FOV.

Fig. S13 .
Fig. S13.Model predictions in a network with homogeneous connectivity.The predictions in Fig. 1c-e are not dependent on heterogeneity of network connections and are largely unchanged in a model with homogenous local connectivity (compare Fig.S4).(a-e) As in Fig.S4a, c, e, f, g, respectively, but for homogenous (i.e.symmetric in rotation and translation, see Methods) cortical interaction.(e) Top, mean of n=40 trials for n=10 stimulus input patterns, p=0.002.Bottom, trial-averaged response to n=10 stimulus input patterns, p=0.002,WSR.(f-h) As in Fig.4a-c, but for homogenous connectivity.