Strong activation of PV cells in V1 impairs discrimination performance

Strong activation of PV cells is frequently used as a tool to silence cortical activity and study the functional role of brain areas [6,19,29,30]. However, different strengths of PV cell activation in V1 may differentially affect behavior [7,21]. To examine how manipulating PV cell activity over a wide range of laser strengths affects perception, we used photostimulation of PV cells expressing Channelrhodopsin-2 (ChR2) in transgenic mice. We confirmed the average percentage of PV cells that were also positive for ChR2-EYFP was 95.83% (N = 168 cells) (S2A and S2B Fig), consistent with previous reports using transgenic mouse lines [14,31,32]. We estimated the boundaries of V1 and sealed all adjacent areas with black paint (see Methods; S3 Fig, example intrinsic imaging map in relation to the mask). We then centered an optic fiber above the craniotomy to deliver blue 473 nm light to V1 (Fig 1D). To exclude the possibility of behavioral effects arising due to nonspecific effects from the light itself [33], we used silencing laser powers in mice lacking ChR2 as a control (Fig 1F). At 2 mW/mm² laser power illumination, discrimination in control wild-type mice remained high (d’ 3.57 ± 0.28 to 3.31 ± 0.26, Wilcoxon signed-rank test, p = 0.297, N = 6, 7 sessions). Moreover, the laser light did not affect pupil size dynamics (S4A Fig). In contrast, discrimination in transgenic PV-ChR2 mice was fully abolished (Fig 1F; d’ 3.01 ± 0.16 to −0.25 ± 0.07, Wilcoxon signed-rank test, p = 0.002, N = 10, 10 sessions).

Optogenetically manipulating V1 activity could potentially induce behavioral responses by itself. To determine whether mice responded during PV cell activation in the absence of a visual stimulus, we activated PV cells at different intensity levels in trained mice presented with gray screen. PV cell activation alone lasting for 2 s did not induce licking behavior or changes in running for the duration of stimulation (S4B and S4C Fig).

PV cell activation during entire stimulus presentation impairs performance at all nonsilencing levels

Since PV cells are well situated to modify feedforward transmission [34], and have been linked to altering response gain and orientation selectivity, and form stimulus-selective ensembles with Pyr neurons [21,22,25,35], we hypothesized that low levels of activation could sharpen sensory representations to improve discrimination. We therefore tested PV cell activation at a range of nonsilencing laser powers (< 0.5 mW/mm², see Methods), while mice performed easy discriminations (Fig 2A, 90° angle difference between ‘go’ and ‘no-go’ stimuli). When light was delivered throughout the entire stimulus presentation, visual discrimination progressively decreased as a function of laser power. Performance declined from 2.86 ± 0.14 (d’, S5A Fig) to between 1.5 and 2 units of d’ lower at the four highest powers (Fig 2B, ‘entire’). As expected [6,7], the strongest PV cell activation at 0.43 mW/mm² reduced behavior to low performance levels (d’ 0.91 ± 0.19, S5A Fig), confirming that V1 activity is required for visual orientation discrimination. Interestingly, the effect was mostly mediated by a sharp decrease in licking to the rewarded stimulus (Fig 2C), reaching as low as 0.49 ± 0.08 (hit rate, S5B Fig). Despite a small significant reduction (Fig 2D), licking to the unrewarded stimulus remained relatively stable around 0.22 ± 0.04 (S5C Fig). Notably, individual mice showed variability in false alarm (FA) rate change across laser powers with some increasing and others reducing their licking rate (S5E Fig), suggesting that mice adopted different strategies when presented with the same sensory limitations.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Stimulating PV cells at nonsilencing levels during entire stimulus presentation or late phase V1 activity impairs discrimination performance, whereas stimulating during early phase V1 activity improves performance.

A) Example of grating stimuli separated by 90° for easy task (top), schematic of optogenetic manipulations activating only PV cells expressing ChR2 (middle), optogenetic stimulation timings for all four conditions aligned to stimulus onset (STIM ON) or offset (STIM OFF) (bottom). B) Normalized average discrimination performance, d’, C) normalized average probability of correct ‘go’ trials (hit rate), and D) normalized average probability of incorrect ‘no-go’ trials (false alarm rate), as a function of laser power for ‘entire’ window (black, N = 10 mice, 41 sessions) and ‘early’ window (red, N = 10 mice, 37 sessions) conditions. Data is normalized by subtracting the value of the 0 power level (see Methods). E–G) same as in B–D) but for ‘late1’ window (blue, N = 10 mice, 40 sessions) and ‘late2’ window (green, N = 8 mice, 32 sessions) conditions. All plots: shaded line is the 95th percentile of data shuffled 1,000 times, using bootstrapping with replacement. Horizontal bars with corresponding colors indicate significant deviation (p < 0.05) from no laser stimulation control condition. See also S5 Fig.

https://doi.org/10.1371/journal.pbio.3003518.g002

We did not observe an activation level of PV cells that significantly enhanced discrimination. Where previous findings suggested that improvements arise from the significant increase in hit rate [21], in our task, mice already showed a high hit rate (S5B Fig, above 0.95 for 0 mW/mm²). We therefore determined whether improvements in behavioral d’ were masked by a ‘ceiling effect’. We restricted the analysis to include only sessions with a hit rate lower than 0.95 for the baseline (no laser) condition and confirmed our initial findings with no behavioral improvements following PV cell activation (S5D Fig).

PV cell activation restricted to the early phase V1 activity improves performance at all nonsilencing levels

It has been previously suggested that the earliest visually evoked activity in V1 is sufficient to enable perceptual discriminations [20]. We therefore investigated if moderate PV cell activation restricted to this initial 120 ms following stimulus onset (Fig 2A, ‘early’), without interfering with sustained V1 activity that is modulated by recurrent and feedback signals, improved visual discrimination. Indeed, we observed a consistent significant improvement in behavioral d’ across all nonsilencing laser powers (Fig 2B), from 3.11 ± 0.12 up to 3.45 ± 0.16 (S5A Fig). These behavioral improvements were also present across individual mice (S5F Fig, in 8 of 10 mice), and resulted from a significant decrease in FA rate across all levels (Fig 2D), from 0.20 ± 0.03 down to 0.09 ± 0.02 (S5C Fig), and a stable hit rate (Fig 2C), with values remaining high in the ranges of 0.95 ± 0.02 and above (S5B Fig).

We next wondered whether these perceptual benefits would disappear at high levels of inhibition. Indeed, performance at silencing laser powers was impaired in two out of three mice (S5G Fig), and the third mouse delayed its response for the same duration as the laser stimulation (S5H Fig).

Undisturbed late phase V1 activity is required for optimal performance

Even at lower hierarchical levels such as V1, visual cortical neurons remain active after participating in the initial feedforward sweep of information following the stimulus onset [27]. While the early response window is thought to encode basic visual features, activity in the late phase response is more closely associated with the behavioral choice of the animal [27]. Since we found that moderate photostimulation restricted to the early phase V1 activity improved discrimination, we asked whether increasing PV cell-mediated inhibition during the sustained V1 activity window could also benefit performance, by targeting top-down influences or biasing the behavioral choice. To spare the initial feedforward sweep, we restricted light delivery to specific time windows, delayed by 120 ms or 180 ms from stimulus onset, and lasting until stimulus offset (Fig 1E, ‘late1’ and ‘late2’ respectively). This approach allowed us to simultaneously investigate the effects of varying the strength and timing of PV cell activation.

Irrespective of the delay with which PV cells were activated, we found that performance dropped gradually with increasing laser power (Fig 2E). Behavioral d’ for late1 condition decreased from 3.36 ± 0.12 in the absence of stimulation to reach values between 2.0 and 2.4 at the strongest stimulations (S5A Fig). On the other hand, d’ for late2 condition decreased less, from similar levels of 3.40 ± 0.13 to levels between 2.7 and 2.9 (S5A Fig). This divergence in performance was most evident at laser powers over 0.2 mW/mm². In both delayed conditions the hit rate was maintained at high levels (although slightly reduced, Fig 2F) over 0.90 ± 0.03 across all tested laser powers (S5B Fig), contrasting the entire window condition where hit rate was reduced. The reduced d’ was mainly explained by a change in FA rate (Fig 2G). The FA rate in the late1 condition gradually increased over laser powers to double from 0.16 ± 0.03 at baseline to 0.33 ± 0.05 at 0.43 mW/mm² (S5C Fig). The FA rate in the late2 condition was maintained at a low rate across all laser powers at around 0.17 ± 0.04 (S5C Fig). Interestingly, all mice either sustained or increased their FA rate (in contrast to the entire window condition where mice adopted variable strategies, S5E Fig).

The performance of mice was less impaired when they were given more time to perceive the stimulus without manipulating the activity of V1. A previous study has shown that the first visually evoked spikes in V1 are sufficient for making perceptual decisions, resulting in performance above chance levels [20]. Importantly, our results demonstrate that easy discriminations rely on unaltered late phase V1 activity to achieve maximum performance.

Difficult discriminations do not benefit from PV cell activation

After establishing the effects of varying the strength and timing of PV cell activation on easy discriminations, we next wanted to determine whether task difficulty influences the effects. While mice can discriminate differences as small as 9° [36], we wanted to have a condition where the discrimination was significantly more difficult but mice were still able to reliably discriminate. Behavioral measurements from 8 mice showed that stimuli separated by 15° were an appropriate choice where performance was lower but still above chance levels (Figs 3A and S6A). When analyzing the distribution of reaction times in the easy and difficult conditions, we observed that the distributions were distinct, consistent with previous reports [37]—difficult discriminations had a broader distribution with longer latencies (S6B Fig). We trained mice on the new stimuli until they achieved d’ > 1.5 (average performance during testing d’ > 2, S7A Fig). We found that the effects of the entire (Fig 3B–D) and both delayed (Fig 3E–3G) window conditions were similar to those reported in easy discriminations (Fig 2), with reductions in d’ of a comparable size.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Stimulating PV cells at nonsilencing levels does not improve performance of difficult visual discriminations.

A) Example of grating stimuli separated by 15° for difficult task (top), schematic of optogenetic manipulations activating only PV cells expressing ChR2 (middle), optogenetic stimulation timings for all four conditions aligned to stimulus onset (STIM ON) or offset (STIM OFF) (bottom). B) Normalized average discrimination performance, d’, C) normalized average rate of correct ‘go’ trials (hit rate), and D) normalized average rate of incorrect ‘no-go’ trials (false alarm rate), as a function of laser power for ‘entire’ window (black, N = 7, 24 sessions) and ‘early’ window (red, N = 8, 28 sessions) conditions. Data is normalized by subtracting the value of the 0 power level (see Methods). E–G) same as in B–D) but for ‘late1’ window (blue, N = 7, 25 sessions) and ‘late2’ window (green, N = 7, 25 sessions) conditions. All plots: shaded line is the 95th percentile of data shuffled 1,000 times, using bootstrapping with replacement. Horizontal bars with corresponding colors indicate significant deviation (p < 0.05) from no laser stimulation. See also S7 Fig.

https://doi.org/10.1371/journal.pbio.3003518.g003

After restricting laser stimulation to the early window, we saw no improvements in d’ across laser powers (Fig 3B). The hit rate significantly decreased with stronger PV cell activation at powers greater than 0.2 mW/mm² (Fig 3C), counteracting the beneficial effects of a slightly decreased FA rate (Fig 3D).

Since difficult discriminations require sustained V1 activity [20], we wondered whether longer PV cell stimulation would be beneficial. To test this, we introduced two new windows lasting from stimulus onset to either 180 ms or 340 ms. We found that the changes in d’, hit and FA rates were comparable to the standard early window (S8 Fig). Thus, we did not identify manipulations in the strength or timing of PV cell activation that improved performance in difficult discriminations.

To further understand the effects of moderate PV cell activation in V1 on behavior, we investigated the influence on various aspects of the decision-making (Fig 4). First, we investigated whether improvements in discrimination could be explained by a speed-accuracy trade-off strategy [38,39]. Second, we investigated how the laser illumination influenced the animal’s decision criterion [40]. Third, we investigated the effect of PV cell activation on the profile of licking, running and pupil size. Fourth, we investigated the effect of PV cell activation on different parameters of a simple model describing decision-making dynamics.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. PV cell activation selectively affects different decision-making components.

A) Response latency for time to first lick in laser off (0 mW/mm²) vs. laser on (0.04–0.43 mW/mm²) in the early window condition (easy discrimination) for ‘go’ (left panel) and ‘no-go’ (right panel) stimuli. Circle, median; line, 95% CI of bootstrapping. Asterisks indicate significant deviation (p < 0.05) from no laser stimulation control condition. B) Normalized decision criterion c across time windows in easy (left panel) and difficult (right panel) discriminations. Shaded line is the 95% CI of data shuffled 1,000 times, using bootstrapping with replacement. Horizontal bars with corresponding colors indicate significant deviation (p < 0.05) from no laser stimulation. Data is normalized by subtracting the value of the 0 power level (see Methods). C) Licking proportion in response to the presentation of the ‘go’ (solid line) and ‘no-go’ (dashed line) stimuli. Plots in C–E) show data for the entire window condition (easy discrimination), with colors ranging from blue to red corresponding to increasing laser powers, ranging from 0.04 to 0.43 mW/mm², as indicated in A). D) Time course of classification accuracy of running speed (left panel) and pupil size (right panel) linear decoders (probability of correctly identifying ‘go’ vs. ‘no-go’ trials), based on the cumulative evidence of running speed or pupil diameter for different PV cell activation levels. E) Normal cumulative distribution function (solid line) with fixed mean and standard deviation as a model fit for cumulative d’ (circles) across stimulus presentation (for visualization purposes, every second data point is shown). The number of mice and sessions is the same as in Figs 2 and 3.

https://doi.org/10.1371/journal.pbio.3003518.g004

Behavioral improvements are not due to speed-accuracy trade-off strategy

We asked whether the improvements from early window PV cell activation could result from mice adopting a speed-accuracy trade-off strategy, with slower responses reflecting more accumulated evidence and improved performance [38,39]. To quantify reaction times for each trial, we measured lick latencies following stimulus onset (Figs 4A and S9). Although some PV cell activation levels showed changes in the median time to first lick to the rewarded stimulus, delays were minor (20–40 ms, Fig 4A, left panel). Moreover, there were no significant delays in the responses to the unrewarded stimulus (Fig 4A, right panel), despite discrimination improvements arising from reductions in FA rate (Fig 2D). Thus, a speed-accuracy trade-off strategy did not explain the results.

PV cell activation affects decision criterion

In signal detection theory [40], the sensory process can be characterized by a sensitivity parameter, d’, and the decision process by a decision criterion parameter, c (see Methods). While d’ quantifies the separation of two distributions, the decision criterion c quantifies whether the animal is more liberal (negative criterion, animal more likely to respond) or more conservative (positive criterion, animal less likely to respond). We investigated whether the changes we observed in d’ were associated with changes in the decision criterion. Regardless of task difficulty, the criterion in the entire window condition shifted from negative to positive values, i.e., lick to nonlick trends (from −0.55 ± 0.08 at 0 mW/mm² to 0.61 ± 0.22 at 0.43 mW/mm² in the easy task and from −0.63 ± 0.10 at 0 mW/mm² to 0.30 ± 0.34 at 0.43 mW/mm² in the difficult task; Fig 4B, black). Smaller shifts in criterion were also observed in the early window condition (from −0.49 ± 0.07 at 0 mW/mm² to −0.15 ± 0.08 at 0.43 mW/mm² in the easy task and from −0.36 ± 0.11 at 0 mW/mm² to 0.10 ± 0.18 at 0.43 mW/mm² in the difficult task; Fig 4B, red). We noted that the criterion increased in particular for the entire window condition. We determined whether this was associated with the animal waiting longer with their response. However, a lick latency increase was seen for the go stimulus but not the no-go stimulus (S9A Fig). Furthermore, a significant change in lick latency was also observed in conditions without an increase in decision criterion (S9C and S9D Fig). Thus, with the increased strength of PV cell activation, mice adopted a more conservative strategy in these two conditions. Interestingly, despite interleaving laser powers within recording sessions, criterion values were different, suggesting mice rapidly adapted their behavior on a trial-by-trial basis [41]. Together, these results show that PV cell activation in V1 not only influences perceptual evidence but also the decision processes.

Similar PV cell activation effects on running speed and licking responses

We next investigated how PV cell activation in V1 influenced other behavioral readouts such as running speed and pupil size. After learning, running shows a stereotypical speed profile, with an early detection component after stimulus onset—a reduction in running speed for both the ‘go’ and ‘no-go’ stimulus, and a late component discriminating between the ‘go’ and ‘no-go’ stimulus (S1A Fig, right bottom panel). Pupil dynamics have been linked to decision-making performance in other tasks [42,43]. To compare the effects of PV cell activation on running and pupil size with the effects on the licking proportion during both stimulus conditions (Fig 4C) we used a linear decoder (see Methods). The decoder predicted based on the running speed and pupil size the identity of each trial (classifying it as a ‘go’ or ‘no-go’ trial), enabling comparison of classification accuracy and licking proportion. We found that the running speed decoder classified trials with accuracy above 80% in the absence of optogenetic PV cell manipulations (Fig 4D, left panel; here shown for the entire window condition). The profile of running speed resembled the profile of the licking proportions (Fig 4C), with a clear separation between laser conditions suggesting similar effects on both behavioral readouts. Interestingly, changes in running speed diverged earlier than the licking proportion responses. In contrast, in our task, pupil size was a poor predictor of stimulus identity, with decoder performance accuracies below 60%, and was not systematically modulated by the laser manipulations (Fig 4D, right panel).

In conclusion, running speed accurately predicts discrimination performance and is modulated by PV cell activation in V1, providing an early readout of the animal’s intention to respond to the stimulus by licking. In contrast, pupil dynamics are not a good predictor of behavioral performance.

PV cell activation modulates amplitude of evidence accumulation

We next determined which function described the dynamics of decision-making and the effects of PV cell activation. We quantified the decision-making dynamics with the cumulative d’ (quantifying the d’ up to a certain time, see Methods). We found that the normal cumulative distribution function provided good fits with three parameters (the mean describing the onset, the standard deviation quantifying the variability, and the amplitude describing the strength of evidence integration). We created three versions of the model. In model one, the mean, standard deviation, and amplitude were all free to vary across laser powers. In model two, the mean and standard deviation were fixed across laser power conditions. In model three, only the amplitude was fixed. Model two with fixed mean and standard deviation but allowing changes in the amplitude (Fig 4E, here shown for the entire window condition), significantly outperformed the other two models across all time windows and difficulty levels (median R² of model 2 = 0.995; 2.5th and 97.5th percentiles of model 1 = 0.984 and 0.987; 2.5th and 97.5th percentiles of model 3 = 0.705 and 0.748). In conclusion, the evidence accumulation was well described by a simple function. Interestingly, the onset and variability of evidence accumulation remained relatively consistent across all conditions, while the amount differed across laser powers.

Changes in neural selectivity reflect behavioral performance in easy but not difficult discriminations

Since improvements in easy discriminations were not due to mice adopting a speed-accuracy trade-off strategy, we wondered whether changes in performance were associated with changes in the selectivity of the V1 network to stimulus. To simultaneously record the activity of V1 neurons while activating PV cells, we repeated the same experiments in a different cohort of PV-Cre mice, where we expressed a calcium reporter GCaMP7s and a red-shifted opsin ChrimsonR in layer 2/3 of V1 (Fig 5A and 5B, see Methods). We confirmed the average percentage of PV cells that were also positive for ChrimsonR-tdTomato was 97.01% (N = 134) (S2C and S2D Fig). We confirmed that red light (639 nm) illumination increased PV cell responses and suppressed Pyr cell responses to the visual stimulus and the effects scaled with laser power (Fig 5C, example cells; S10C Fig, population average). Out of 1,156 Pyr cells, 965 of those cells were significantly modulated by the PV cell activation, accounting for 83.48%. As in the behavioral experiments, we moderately suppressed activity of Pyr cells (S10C Fig; approximately 50% suppression) [21].

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Changes in neural selectivity caused by PV cell activation reflect behavioral performance in easy but not difficult discriminations.

A) Optogenetic stimulation timings for entire and early window conditions aligned to stimulus onset (STIM ON) or offset (STIM OFF) (top), schematic of optogenetic method combined with 2-photon imaging (middle), schematic of laser activating only ChrimsonR-expressing PV cells while GCaMP7s reports the activity of all neurons (bottom). B) Example frames of in-vivo imaging plane showing GCaMP7s-expressing neurons when illuminated with 920 nm wavelength (left), and ChrimsonR-tdTomato-expressing neurons when illuminated with 1,040 nm wavelength (right). Arrow shows example PV cell. Scale bar, 50 µm. C) Average response of two example neurons to optogenetic laser illumination during the entire stimulus duration (0–2 s). The activity to the visual stimulus of the identified Pyr cell is suppressed, whereas the PV cell is increased. Shading, SEM. D) Median laser-induced changes in absolute selectivity for entire (left panel, N = 1,156 cells, 6 mice) and early (right panel, N = 754 cells, 6 mice) time window conditions, for easy task, significant at all laser levels (Wilcoxon signed-rank test, p < 0.002). Error bars, SEM. Laser levels in D–G) are calibrated for each mouse to cause moderate suppression in Pyr cell activity (see Methods) and range from 0.02 to 0.37 mW. E) is same as D) showing entire (left panel, N = 542 cells, 3 mice) and early (right panel, N = 561 cells, 3 mice) time window conditions but for difficult task (Wilcoxon signed-rank test, ‘entire’ laser level 1 = −0.03 ± 0.01, p = 0.017; ‘early’ laser level 2 = 0.01 ± 0.01, p = 0.037, all other laser levels n.s.). F) Stimulus selectivity index of the same cells as in D) in the no laser condition (0 mW, off) and across laser levels for easy task. Cells were ordered by their selectivity at each laser level. G) is same as F) but for difficult task. Color bar range indicates selectivity with positive values for cells preferring the rewarded ‘go’ stimulus and negative values for cells preferring the unrewarded ‘no-go’ stimulus. H) Histograms of data in F) to quantify the shift towards go or no-go preference in the easy task. Entire: Median ± SEM, laser off = 0.16 ± 0.02, laser on = 0.18 ± 0.02, Wilcoxon signed-rank test, p < 0.002; averaged across laser levels. I) Histograms of data in G) to quantify the shift towards go or no-go preference in the difficult task. Entire: Median ± SEM, laser off = 0.15 ± 0.02, laser on = 0.12 ± 0.01, Wilcoxon signed-rank test, p = 0.931; averaged across laser levels. Early: Median ± SEM, laser off = 0.15 ± 0.01, laser on = 0.10 ± 0.01, Wilcoxon signed-rank test, p < 0.002; averaged across laser levels.

https://doi.org/10.1371/journal.pbio.3003518.g005

Response selectivity in V1 has been previously linked to higher discrimination performance [7]. The selectivity is defined as the difference in the response to ‘go’ and ‘no-go’ stimulus divided by the pooled standard deviation, with positive values indicating a preference for the ‘go’ and negative values a preference for the ‘no-go’ stimulus (see Methods). We quantified the median of the stimulus selectivity and found that more cells preferred the go stimulus after learning (S10A Fig) [7,35]. We investigated the selectivity changes as a result of PV cell activation for each time window and laser power separately. When mice performed easy discriminations, the absolute selectivity across all laser levels decreased when activating PV cells during the entire stimulus presentation (Fig 5D, left panel; largest change for laser level 1 = −0.07 ± 0.01, p < 0.002, n = 1,156 cells, 6 mice) and increased when restricting activation to the early window only (Fig 5D, right panel; largest change for laser level 3 = 0.04 ± 0.01, p < 0.002, n = 754 cells, 6 mice), reflecting the previously reported behavioral changes (Fig 2B). We confirmed that the observed increases in selectivity were not caused by rebound activity after ending PV cell activation at 120 ms (S10B Fig). This is in agreement with previous reports showing that photoinhibition durations shorter than 500 ms, regardless of laser intensity, causes minimal rebound [44].

Furthermore, the increase in the absolute selectivity in the early window condition resulted from a shift toward the rewarded ‘go’ orientation (Fig 5F and 5H, Early: median ± SEM, laser off = 0.18 ± 0.02, laser on = 0.31 ± 0.02, Wilcoxon signed-rank test, p < 0.002, N = 754, 6 mice; averaged across laser levels; Note: positive values indicate go stimulus preference, negative values indicate no-go stimulus preference). A match between reductions in behavioral performance and neuronal selectivity was not observed in difficult discriminations. Most laser levels showed no significant changes in absolute selectivity (Fig 5E) and no consistent bias in ‘go’ or ‘no-go’ selectivity (Fig 5G and 5I). Thus, our results show that increased behavioral performance during PV cell activation in the early time window during easy discriminations is associated with increased response selectivity in V1.

Circuit model captures changes in neural selectivity with PV cell activation

How can activating PV cells at different times have such contrasting effects on neural selectivity? And how can brief PV cell activation lead to increased selectivity? To address these questions, we constructed a simplified circuit model that captures the minimal mechanisms required to explain our experimental observations. The model consists of two excitatory populations (Pyr neurons) that compete via lateral inhibition and receive additional inhibition from two PV cell populations, both of which are activated by laser stimulation (Fig 6A). Stimulating the circuit over 100 trials (50 go and 50 no-go) reproduced key features of the data. First, PV cell activation throughout the stimulus period led to a reduction in Pyr cell activity (S11E Fig) and a corresponding decrease in absolute selectivity (Fig 6B). Second, early PV cell activation induced a transient suppression, followed by a gradual rebound in activity (S11F Fig), accompanied by an increase in absolute selectivity (Fig 6B). Third, the model captured the experimentally observed asymmetry: the selectivity increase was primarily driven by enhanced selectivity in go-preferring cells (Fig 6C). Two elements of the model proved critical for these effects: competitive interactions among Pyr cells, and a nonlinear activation function in the Pyr cells. In the model, go-preferring Pyr cells are embedded in stronger competitive interactions than no-go-preferring cells, implemented by differences in their inhibitory connectivity. Depending on the strength of this competition, the circuit settles into different steady-state activity patterns, reflected in the more pronounced and persistent change in activity for go-preferring cells following early PV activation. Removing this competition reduced selectivity overall and abolished the selectivity enhancement seen with early PV cell activation (Fig 6D). Likewise, replacing the nonlinear activation function with a rectified linear one eliminated the selectivity increase (Fig 6E), highlighting the importance of nonlinear response properties in shaping the effect.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Increased selectivity requires competition and nonlinear activation function in pyramidal cells.

A) Illustration of the circuit model. Light gray triangles denote excitatory pyramidal populations, circles denote inhibitory populations. Excitatory populations inhibit each other via local interneurons (dark gray) and receive additional inhibition from PV cells (red). Top: Go-preferring cells circuit interactions. The left pyramidal population is a go-preferring population, receiving stronger input during ‘go’ trials. The dark gray connections are strong and hence there is strong lateral inhibition leading to competition in the circuit. Bottom: No-go-preferring cells circuit interactions. The left pyramidal population is a no-go-preferring population, receiving stronger input during ‘no-go’ trials. The inhibitory interactions between the pyramidal cells are less strong and hence there is less competition in the circuit. B) The difference in absolute selectivity between laser on and laser off conditions for ‘Entire’ and ‘Early’ PV cell activation for the experimental data (left panel) and the model (right panel). Same data as in Fig 5D but averaged across laser levels. C) The difference in selectivity between laser on and laser off conditions for go- (blue) and no-go-preferring (red) populations for ‘Entire’ and ‘Early’ PV cell activation for the experimental data (left panel) and the model (right panel). Same data as in Fig 5F but averaged across laser levels. Note that no-go selectivity was multiplied by −1 to indicate increased selectivity if above 0 and decreased selectivity if below 0. D) The difference in absolute selectivity between laser on and laser off conditions for ‘Entire’ and ‘Early’ PV cell activation in a model with weaker competition. E) The difference in absolute selectivity between laser on and laser off conditions for ‘Entire’ and ‘Early’ PV cell activation in a model with (rectified) linear activation function. Across all panels: Error bars, SEM.

https://doi.org/10.1371/journal.pbio.3003518.g006

The model was intentionally kept minimal to allow for mechanistic insight. Although simplified relative to actual cortical circuits, it demonstrates that the experimentally observed effects can be accounted for by basic circuit motifs. The conclusions were not strongly dependent on precise parameter choices (S11C and S11G Fig). Rather, the essential requirement was that certain qualitative connectivity constraints are met (e.g., strong competitive inhibition). This minimal modeling approach allowed us to isolate the key circuit ingredients responsible for the specificity of PV-mediated modulation.

Ethics statement

All experiments were performed at the University of Cambridge in accordance with the Animals (Scientific procedures) Act 1986 and with AWERB (Animal Welfare and Ethical Review Board at the University of Cambridge) and UK Home Office review and approval (PPL numbers PA41D42A4 and PP6681989).

Animals

Mouse lines were obtained from the Jackson Laboratory (PV-Cre, stock number 008069, and Ai32, stock number 012569), and inbred to maintain the original colony or crossed to create a PV-Cre::Ai32 colony [82]. Sixteen adult mice (11 males and 5 females) underwent surgeries at ages between 7–13 weeks. Animals were co-housed in a 12-hour (08:30–20:30) reversed light/dark cycle room and provided with cage enrichment and standard diet and water ad libitum. During behavioral testing, mice were food-deprived to maintain a minimum of 85% (but typically 90%) of their free-feeding body weight (between 3–5 g of standard food pellets per animal per day).

Surgical procedure

Mice were anesthetized with 2.5% Isoflurane (1%–2% for maintenance) with 2 L/min oxygen, administered with pre-operative analgesia of 5 mg/kg injectable Metacam, and positioned in a stereotactic apparatus. A subset of mice also received intraperitoneal injections with a 10 mg/kg dose of Dexamethasone the day before and day of surgery to prevent brain swelling. Body temperature was maintained at 36.5–37.0°C throughout the procedure using a heating pad. Ointment (Xailin Night) was applied to the eyes to protect them from drying out. The head was shaved, and a circular piece of scalp was removed to expose the skull which was then cleaned and dried. All mice were fitted with a headpost attached by dental cement (C&B Metabond) to allow head-fixation onto the experimental apparatus. A craniotomy was made over the left hemisphere with the center positioned 2.5 mm left of the midline and the posterior edge aligned with the anterior edge of the transverse sinus. Where viral injections were administered (PV-Cre mice only), Picospritzer (Parker Picospritzer III) was used to inject a total of 800 nL of a 2:18 mixture of AAV-syn-jGCaMP7s-WPRE (Addgene: 104487-AAV1) and AAV-syn-FLEX-rc[ChrimsonR-tdTomato] (Addgene: 62723-AAV5), respectively, in 2 injection sites (coordinates ML2.5 and ML2.3 corresponding to the monocular segment of the primary visual cortex [83]) of 200 nL each at depth of 200 μm and 400 μm from the brain surface. The craniotomy was sealed with a double-layer 4 mm-diameter cranial window (two no. 1 thickness windows glued together with index-matched adhesive Norland #71) to provide stable optical access for photostimulation during behavior [84]. A custom-designed 3D-printed cap was used to protect the cranial window. Mice were monitored for 5 days post-surgery with free access to food and water, with post-operative analgesia of 1.5 mg/ml oral Metacam given in the first 3 days. After full recovery, mice were food-deprived, and habituated to experimenter handling and head-fixation apparatus, after which training) commenced.

Behavioral setup

Mice were head-fixed onto a Styrofoam cylinder (Fig 1A). Running was measured with an incremental encoder attached to the cylinder shaft (Kübler). Behavioral task trials were initiated when running exceeded a threshold for a specified time (typically 1–2 s). An ITI—gray pre-stimulus illumination, was then presented ranging from 3 to 25 s to discourage timed licking (3–6 s for testing). To avoid excessive licking during training, the ITI was extended if mice licked during the gray screen, and up to 6 s timeout was given when mice licked during the unrewarded ‘no-go’ stimulus. A reward spout positioned near the snout of the mouse delivered strawberry-flavored soy milk via a pinch valve (NResearch), with licks being detected via a piezo disc sensor. Reward was delivered upon first lick only during a reward zone (1–2 s after the rewarded ‘go’ stimulus onset) or by an auto-reward trigger at 2 s if mice failed to lick. The stimuli consisted of static black and white gratings with a fixed spatial frequency (0.06 cpd) with 100% contrast and had either 135° or 225° orientation. Stimuli were presented pseudorandomly (no more than four of the same stimulus type in a row). Stimuli were presented with an equal probability of 0.5, except in rare cases where mice failed to show behavioral improvements and the ‘no-go’ stimulus was biased for brief periods with a 0.7 probability to discourage licking. The rewarded stimulus was counter-balanced across mice. Stimuli were presented on a 27“ DELL U2715H monitor positioned at 45° to the mouse body axis and 23 cm away from its right eye. Stimulus presentation, reward delivery, and laser stimulation were controlled by a custom-written Matlab script using the Psychophysics Toolbox [85].

Schematics in Figs 1–3 and 5 are adapted from SciDraw.io, from the following contributing authors: Emmett Thompson (3925987); Ethan Tyler and Lex Kravitz (3926057); Long Li (5496322); and Manish Kumar (4914800).

Experimental protocol

For 2–3 days, mice were habituated to the cylinder and milk rewards. For 2–3 days, mice learned to run on the cylinder (monitor off) uninterrupted and to lick the spout for rewards without visual stimulation. During discrimination training (1–3 weeks), mice learned to discriminate between ‘go’ and ‘no-go’ stimuli. Each training session lasted ~1 to 1.5 hours (usually between 100–400 trials). Mice were tested with optogenetics after achieving d’ > 2 on 3 consecutive sessions, except for two mice which were tested after 1 such session. This criterion was to prevent overtraining and ensure all four task outcomes (hit, miss, FAs, and CRs) were present for analysis. Overtraining will also result in a ‘ceiling effect’, possibly masking behavioral improvements resulting from PV cell activation. Three factors were tested: 1) strength and 2) timing of optogenetic stimulation, and 3) difficulty of discrimination. Each testing session had five interleaved laser powers (including 0 mW/mm² to establish baseline d’) and lasted until at least 400 trials were completed. Testing of timing was done in a pseudorandom order.

Optogenetic stimulation

A Stradus VersaLase laser was used to deliver 473 nm light via a 200 µm core, 0.39 NA optic fiber cable. Laser intensity was calibrated using a standard photodiode power sensor (S121C, Thorlabs). All laser powers are reported as mW/mm². A laser power was first identified to fully silence discrimination (Fig 1F). Lower test powers, including 0.04, 0.06, 0.08, 0.10, 0.20, 0.27, 0.35, and 0.43 mW/mm², were then selected to produce detectable changes in discrimination. The size and boundaries of V1 were estimated using intrinsic imaging data from 12 mice corresponding to previously reported boundaries [86]. Black paint was used to seal the window and restrict light delivery to V1. The paint was effective in blocking the laser, with an insignificant leakage of 0.02 µW at the highest testing powers. The optic fiber was centered over V1 at either a 10.7 mm or 17.7 mm distance from the window to deliver uniform illumination. Since the fiber was externally placed, side propagation through the brain to activate neighboring brain areas or to reach the retina to produce nonspecific effects is unlikely [87]. Additionally, to minimize laser light reaching the eyes, the optic fiber was shielded by a custom-designed 3D-printed cone. Light delivery lasted throughout the entire stimulus duration or for specific time windows (Fig 1E).

Behavioral and statistical analysis

Task performance was quantified with behavioral d’ [40], with higher values reflecting better performance. Behavioral d’ and decision criterion c were calculated by:

where Φ is the normal inverse cumulative distribution function, H is the rate of hit trials (licking for rewarded trials during the response window only) and F is the rate of false alarm trials (licking for unrewarded trials during entire stimulus presentation). While d’ captures perceptual sensitivity, the decision criterion c is a measure of the subject’s internal bias to respond; with negative values indicating high bias for ‘yes’ responses and positive values indicating high bias for ‘no’ responses. Lick latencies were determined in trials classified as hit or FA as the first lick detected in the stimulus presentation window (0–2 s). Where the data is normalized, this was done by subtracting the value of the control condition (0 mW laser power) to the laser condition.

No statistical methods were used to predetermine sample sizes, but the number of mice and sessions included in the behavioral experiment is comparable to previous studies [6,7,20,88]. All data are presented as mean ± SEM unless otherwise stated. To compare within-subject PV cell activation effects on d’, hit, FA, and criterion to the baseline condition (0 mW/mm²), we used bootstrap with replacement 1,000 times to generate 95% CI [89]. Significant effects were defined as nonoverlap between the no laser condition and the 95% CI for each laser power.

Eye tracking

Pupil positions were recorded with a Raspberry Pi camera (model 3B+ or 4), where infrared LEDs illuminated the eye. A small number of eye frames (<2%) were randomly selected from the whole recording of each session, and pupil extraction thresholds were manually adjusted. The ellipses to track pupil position and diameter were manually verified by inspecting a random subset of samples from each session. See previous studies [7,90] for more details on eye tracking.

Decoder analysis

A cumulative decoder was employed to quantify the accuracy with which running speed and pupil size could classify trials as either ‘go’ or ‘no-go’ at time t relative to stimulus onset [7]. The decoder received inputs of running speed or pupil size aligned to stimulus onset. Thirty trials were taken from each condition to make up the training data, which the decoder then used to construct a model of the running speed or pupil size response to the rewarded and unrewarded stimuli (calculating the mean response across trials for each condition). On the remaining trials, the log-likelihood at time t that trial k belongs to condition C (where C is the rewarded ‘go’, R or unrewarded ‘no-go’ stimulus, U) is proportional to:

where D indicates the running speed or pupil size. The trial was assigned as a rewarded stimulus response if L_R> L_U, otherwise classified as an unrewarded stimulus response. To obtain the cumulative likelihood L_c at each time point t, the summation only included time points from stimulus onset T_start to time t.

Modeling cumulative d-prime

The cumulative d’ was calculated to quantify the d’ up to a certain time point (by calculating the proportion of hits and false alarms at different timepoints relative to stimulus onset). A normal cumulative distribution function (CDF) was used to fit the experimentally observed cumulative d’ curves (Fig 4E):

where N is the normal cumulative distribution and the three parameters include a, the amplitude, µ the mean, and σ the standard deviation. Model 1 allowed all three parameters to be varied across laser levels, model 2 only allowed the amplitude to vary (while mean and standard deviation were fixed), and model 3 allowed the mean and standard deviation to be varied (while amplitude was fixed) across laser levels. We used half of the trials for the fitting and the other half of the trials for the testing, using R² to quantify the quality of the fits. The analysis showed that model 2 performed much better than model 3 and slightly better than model 1, indicating that a model with a mean and standard deviation fixed across laser levels, with an amplitude differing across levels, described the data well. Significance was determined by the overlap between the median R² of a model and the 95% confidence interval of the bootstrap distribution of the other models.

Immunohistochemistry and fixed tissue imaging

Immunohistochemistry and fixed tissue imaging were performed similarly to previous studies [35,91]. Mice were anaesthetized by an intraperitoneal injection of Pentobarbital and transcardially perfused with 15 ml PBS with heparin (102 mg/L) followed by 10 ml of 4% paraformaldehyde (PFA) in PIPES buffer (60 mM PIPES, 25 mM HEPES, 5 mM EGTA, and 1 mM MgCl2). Brains were extracted and postfixed in 4% PFA for 24 hours and replaced into PBS/Azide (0.02% NaN3). After embedding in 5% agarose, coronal sections were sliced from the primary visual cortex with 50 µm thickness using a vibratome (VT1000S, Leica Biosystems).

The brain slices were washed with PBS and incubated in 5% normal goat serum (G9023, Sigma) in PBS/Triton X-100/Azide (0.25% Triton X-100, 0.02% NaN3) for 2 hours at room temperature and then incubated overnight at 4°C in primary antibody solution (anti-PV, PV235, Swant, 1:500, in PBS/Triton X-100/Azide). Following primary antibody incubation, slices were washed three times in PBS for 5 min and incubated in secondary antibody solution (goat Alexa Fluor 647 anti-mouse, A21240, Invitrogen, 1:1000, in PBS/Triton X-100/Azide) for 2 hours at room temperature. Slices were then washed in PBS and mounted on glass slides (Menzel-Gläser) with mounting medium with DAPI stain (AB104139, Abcam).

Slices were imaged with a laser scanning confocal microscope (LSM 900, Carl Zeiss) using appropriate excitation and emission filters, a pinhole of 1 AU, and a 20x air objective. A series of z-stack images was obtained at a laser power and gain set to prevent signal saturation. All analysis was performed with Matlab.

Two-photon calcium imaging

Functional in-vivo calcium imaging was performed using a two-photon scanning microscope with a 3 mm working distance Nikon 16x 0.8 NA objective and Ti:Sapphire laser (Mai Tai HP Deep See, Spectra-Physics, < 70 fs pulse width, 80 MHz repetition rate) at 920 nm, wavelength for calcium indicator GCaMP7s excitation. A 12 kHz resonant scanner (Cambridge Technology) and an FPGA module (PXIe-7965R FlexRIO, National Instruments) imaged at a 30 Hz frame rate to acquire 512 x 512 pixels images covering a 500 x 500 μm field of view. Data was acquired using ScanImage 5.6. A National Instruments DAQ card was used to record triggers for synchronization of neural responses, stimulus presentation, optogenetic stimulation, eye camera videos, licking, and running. Laser power was set to 15–40 mW and was kept consistent throughout consecutive recording sessions. Where identification of tdTomato-tagged PV cells was required, the structural marker tdTomato was excited with a wavelength of 1,040 nm.

Imaging commenced no earlier than three weeks after surgery to ensure stable viral expression. A screening session for each mouse was conducted before data collection commenced. During the initial screening session, ThorCam Software (Thorlabs) was used to identify the injection site coordinates guided by blood vessel identification. Retinotopic and orientation mapping helped identify a responsive site for recording with RFs located near the center of the monitor, confirming that cells in the imaging field where responsive to the center of the screen (corresponding to the monocular visual field). A brief recording was made for anatomical identification of tdTomato-labeled PV cells, subsequently used for identifying PV cells. Recordings were made from layer 2/3 neurons at a depth of 150–250 μm below the pial surface. Mice were removed from the study if imaging was compromised due to bone regrowth or recording site becoming functionally unresponsive due to overexpression of GCaMP7s, if no single session of d’ > 1.5 was achieved after extended training, or if neuronal activity was fully silenced even at the lowest possible optogenetic laser power.

Simultaneous two-photon imaging and optogenetic stimulation

A Stradus VersaLase laser was used to deliver 639 nm light to the imaging site via a 200 µm core, 0.39 NA optic fiber cable coupled with a 20 mm Fiber Optic Cannula. The cannula was positioned at a 30° angle and 7 mm distance from the cranial window and held in place by a micromanipulator. Laser intensity was measured across several levels using a standard photodiode power sensor (S121C, Thorlabs), and calibrated for each animal. Selected test levels ranged between 0.02 and 0.37 mW and caused moderate suppression (~50% decrease in activity) of Pyr cell activity, consistent with previous reports [21]. ChrimsonR-tagged PV cells were stimulated by centering the laser beam over V1 to deliver a uniform illumination. A custom-made 3D-printed cone with attached Blackout Nylon Fabric (Thorlabs BK5) was carefully inserted around the objective and optic cannula to protect the PMTs from external light sources while also eliminating optogenetic laser light from illuminating the eyes.

To allow simultaneous two-photon imaging and optogenetic stimulation within the same neuronal population, the stimulus monitor and optogenetic laser were blanked during the linear phase of the resonant scanner such that the monitor was turned off and the laser was blocked out during the opening of the PMTs shutter, and vice versa.

Imaging data analysis

Data was pre-processed using Suite2p, where frames were registered to correct for brain motion and active neurons were automatically selected as regions of interest (ROIs) to extract calcium traces for further analysis [92]. After automated detection, all sessions underwent a manual inspection to ensure identified ROIs only included cell somas and excluded axons, synaptic boutons, or other artifacts. Pixels within each ROI were averaged to obtain raw fluorescence time series F(t). The raw fluorescence values for each ROI were then converted to Δ F/F. Neuronal activity was aligned to the onset of stimuli and Δ F/F was averaged over 0–1 s time window following stimulus onset (excluding the response window 1–2 s where reward is delivered) to estimate the response of a cell to oriented gratings. The Wilcoxon signed-rank test was used to determine if the response to the gratings (0–1 s) significantly decreased or increased relative to the pre-stimulus baseline activity (−0.5–0 s). To calculate the z-scored Δ F/F in the 0–1 s response window, we subtracted the pre-stimulus response and divided by the pre-stimulus standard deviation. Cells included in Fig 5B and 5C were cells where the z-scored Δ F/F was higher than 0.5.

Neural selectivity

A selectivity index (SI) was calculated for each cell as [7,35]:

where R_R is the averaged stimulus response (0–1 s) to the rewarded ‘go’ stimulus and R_U is the averaged stimulus response (0–1 s) to the unrewarded ‘no-go’ stimulus. The difference is divided by the pooled standard deviation of the responses (), where n_i is the number of trials in i for k conditions. A positive or negative value indicates cells with a preference for the rewarded or unrewarded stimulus, respectively. To calculate the median selectivity across cells, we computed the absolute of the selectivity of each cell. For each cell, the absolute selectivity during laser manipulations was subtracted from the absolute selectivity during the control condition (no laser).

Circuit model

We modeled circuit interactions of go- and no-go-preferring cells, where each circuit consists of two excitatory pyramidal cell populations, which inhibit each other via an intermediate inhibitory population; and two additional inhibitory populations, corresponding to (laser-activated) PV cell populations. The activity of the Pyr cell populations is described by their response , which evolves over time according to the following equation:

The activity of the inhibitory cell populations, which mediate the lateral inhibition is described by their response , which evolves over time according to the following equation:

where are the time constants of the excitatory and inhibitory populations, respectively. is the input to population. Each excitatory population receives a different random constant input . Population 1 receives a higher input during trials with its preferred stimulus, but not during trials with the nonpreferred stimulus. For the go circuit in go trials, this input is (0.5,0.1) and (0.1,0.1). For the go circuit in no-go trials, this input is (0.1,0.1) and (0.1,0.1). Similarly, for the no-go circuit in go trials this input is (0.1,0.1) and (0.1,0.1). For the no-go circuit in no-go trials, this input is (0.5,0.1) and (0.1,0.1). is the recurrent inhibitory input from the inhibitory population which is driven by the other Pyr cell population and is the effective synaptic weight from the inhibitory population to the Pyr cell population. In our model, go- and no-go-preferring cells are identical in all intrinsic properties. The only difference lies in their connectivity. Specifically, go-preferring cells are embedded in a circuit with stronger lateral competition compared to no-go-preferring cells. This is reflected by the different weights from the inhibitory to the excitatory population, which were as follows: , . is a PV cell population that targets population and is activated by the laser. is a sigmoid activation function:

is a rectified linear activation function:

Finally, is the weight from the excitatory to the inhibitory population. The remaining parameters were chosen as follows:

The activity of the PV cell populations depends on the laser condition:

The activity of each Pyr cell in each trial is initialized randomly, where we sample from a normal distribution and take the absolute value:

We simulated three different conditions (1,000 time steps each): (1) no PV cell activation, (2) entire PV cell activation, and (3) early PV cell activation. In the first condition, we simulated the network for 1,000 time steps without intervention. In the second condition, we simulated the laser activation by setting from time step 400 (after the network reached steady-state) till the end. In the third condition, we simulated the laser activation between time steps 400 and 500. We used a simulation timestep of = 0.05.

To calculate the selectivity in each circuit in the model, we subtracted the mean activity of the go or no-go population, respectively, in go/no-go trials from the mean activity of that population in no-go/go trials:

In Fig 6B, 6D, and 6E, we calculated the difference in absolute SI between laser on and laser off conditions. In Fig 6C, we calculated the difference in SI between laser on and laser off conditions.

We studied the impact of the competition controlled by the mutual inhibitory connections - by reducing them to 1.0. We studied the impact of the nonlinear activation function by replacing it with a rectified linear one: