Introduction

Under normal conditions, self-generated movements are coupled to sensory feedback in different modalities. Locomotion, for example, is coupled to proprioceptive feedback from muscle movements, somatosensory and auditory feedback from touching the ground, vestibular feedback from translational acceleration, and visual feedback in the form of expanding visual flow. This coupling is key to our brain’s ability to learn internal models of the world (Jordan and Rumelhart, 1992). Internal models are the transformations between cortical areas and can be used to predict the sensory consequences of movement. Hallucinations and delusions can be explained as a misconfiguration of these internal models: inaccurate predictions of the sensory consequences of movements can result in self-generated activity that is not correctly distinguished from externally driven activity (Fletcher and Frith, 2009; Frith, 2005).

Self-generated movements result in broad activation of most regions of the brain (Musall et al., 2019; Stringer et al., 2019), including some primary sensory areas like visual cortex (Keller et al., 2012; Saleem et al., 2013) where they have been shown to depend on visual context (Pakan et al., 2016). The framework of predictive processing postulates that in cortex, these movement related signals are internal model based predictions of the sensory consequences of movement that are compared to externally generated bottom-up input to compute prediction errors (Jiang and Rao, 2021; Keller and Mrsic-Flogel, 2018). Evidence for this interpretation has mainly come from the discovery of movement related prediction error responses in a variety of different cortical regions and species (Attinger et al., 2017; Audette et al., 2021; Ayaz et al., 2019; Eliades and Wang, 2008; Heindorf et al., 2018; Keller et al., 2012; Keller and Hahnloser, 2009; Stanley and Miall, 2007; Zmarz and Keller, 2016), and the observation that top-down inputs from motor areas of cortex appear to convey movement related predictions of sensory input to sensory areas of cortex (Leinweber et al., 2017; Schneider et al., 2018). The neurons on which the opposing bottom-up input and top-down prediction converge are referred to as prediction error neurons and are thought to reside primarily in layer 2/3 (L2/3) of cortex (Jordan and Keller, 2020). To complete the circuit, there also needs to be a population of neurons that integrate over prediction error responses. These are referred to as internal representation neurons. It is still unclear which cortical neuron type functions as internal representation neurons, but based on anatomical arguments it has been speculated that they can be found in deep cortical layers (Bastos et al., 2012).

In humans, externally driven and self-generated activity patterns can be experimentally separated at least partially, for example, by showing subjects a visual stimulus and later asking them to visualize the same stimulus (Le Bihan et al., 1993). This type of experiment is considerably more difficult to perform in animals. Using movement as a proxy for self-generated activity patterns, however, using a virtual environment, we can compare a condition in which self-generated and externally generated activity patterns are correlated to a condition in which they are not. In the visual system, we can use a virtual reality environment to experimentally couple or decouple locomotion from visual flow feedback. We will refer to the former as a closed loop condition and to the latter as an open loop condition. Assuming the predictive processing model is correct, we should find differences in the activation patterns in dorsal cortex in response to locomotion onsets as a function of whether they occurred with or without coupled visual flow feedback. Both prediction error neurons as well as internal representation neurons should respond differentially. Thus, we set out to investigate whether there are differences between closed and open loop activation patterns of mouse dorsal cortex, and whether these differences are neuron type specific. If hallucinations are indeed the result of aberrant activation of internal representations, possibly as a result of misconfigured internal models, antipsychotic drugs should alter the differential activation patterns in these neurons.

Results

Tlx3 positive layer 5 (L5) intratelencephalic (IT) neurons distinguished closed and open loop visuomotor coupling

To investigate neuron type specific differences in activity patterns at locomotion onsets with and without self-generated visual feedback in mouse dorsal cortex, we quantified locomotion onset activity using widefield calcium imaging (Wekselblatt et al., 2016). A GCaMP variant was expressed using either a retro-orbital injection of an AAV-PHP.eB vector (Chan et al., 2017), or a cross between a Cre line and the Ai148 GCaMP6 reporter line (Daigle et al., 2018) (see Methods; Figure 1A, Table S2). Mice expressed a GCaMP variant either brain wide (C57BL/6), or in genetically identified subsets of excitatory or inhibitory neurons using Cre driver lines. Please note, we use the term ‘brain wide’ here to describe an expression pattern that was determined by the intersection of the AAV-PHP.eB tropism and the expression pattern of EF1α or hSyn1 promoters but that was otherwise brain wide (see Discussion). We used Emx1-Cre mice to target expression to cortical excitatory neurons (Gorski et al., 2002), Cux2-CreERT2 mice to target expression to L2/3 and L4 layer excitatory neurons (Franco et al., 2012), Scnn1a-Cre mice to target expression to a subset of L4 excitatory neurons (Madisen et al., 2010), Tlx3-Cre mice to target expression to a subset of L5 IT neurons (Gerfen et al., 2013), and Ntsr1-Cre mice to target expression to a subset of L6 excitatory neurons (Gong et al., 2007). To target subpopulations of inhibitory neurons, we used PV-Cre (Hippenmeyer et al., 2005), VIP-Cre (Taniguchi et al., 2011), and SST-Cre mice (Taniguchi et al., 2011) (see Methods). Prior to the start of the imaging experiments, mice were implanted with a crystal skull cranial window (Kim et al., 2016). We used crystal skull preparations as we found that hemodynamic artifacts were smaller in this preparation compared to those observed in clear skull preparations (Figures S1A and S1B). For all experiments, mice were head fixed and free to locomote on a spherical treadmill (Figure 1B). Activity from dorsal cortex was imaged while mice were in a closed loop visuomotor condition, in which locomotion on the spherical treadmill controlled forward movement in a virtual corridor. Following this, mice were exposed to an open loop condition during which mice were presented with a replay of the visual flow they had self-generated in the preceding closed loop condition.

Figure 1.

We first compared locomotion onset responses in the closed loop condition to those in the open loop condition in mice that expressed GCaMP brain wide (C57BL/6). Consistent with previous findings (Musall et al., 2019; Stringer et al., 2019), we found that locomotion onset resulted in broad activation across most of dorsal cortex in both conditions (Figure 1C). However, comparing locomotion onsets in mice that expressed GCaMP6 only in Tlx3 positive L5 IT neurons, we found that the activation pattern was strikingly different between closed and open loop conditions (Figure 1D). While locomotion onset in the open loop condition resulted in an increase in activity across dorsal cortex similar to that observed in C57BL/6 mice with brain wide GCaMP expression, locomotion onset in the closed loop condition resulted in a very different activation pattern: In posterior parts of dorsal cortex, the activity decreased, while in anterior parts of the dorsal cortex, the activity increased. To quantify these activation patterns across mice, we selected six regions of interest: primary visual cortex (V1), antero-medial secondary visual cortex (V2am), retrosplenial cortex (RSC), primary motor cortex (M1), anterior cingulate cortex at the level of bregma (A24b), and secondary motor cortex (M2) (Figure 1E). In C57BL/6 mice that expressed GCaMP brain wide, closed loop locomotion onset resulted in similar activity patterns across these six regions (Figure 1F). This activity pattern was not markedly different from that observed during locomotion onsets in an open loop condition (Figure 1G). To quantify the contribution of visual flow onset to the closed loop locomotion onset activity, we measured responses to visual flow onsets that occurred in the absence of locomotion in the open loop condition (Figure 1H). In C57BL/6 mice with brain wide GCaMP expression, visual flow onsets resulted in responses that were small compared to those observed during locomotion onset, even in V1. Note that a visual flow onset is a relatively weak visual stimulus compared to a drifting grating stimulus. A grating stimulus is typically composed of a gray screen or a standing grating stimulus that suddenly transitions to a drifting grating, and thus the visual flow undergoes a very high acceleration. A visual flow onset in an open loop condition is the replay of previously self-generated visual flow and thus exhibits much lower visual flow acceleration. The absence of a strong visual flow onset response in C57BL/6 mice that expressed GCaMP brain wide is consistent with a high similarity of closed and open loop locomotion onset responses.

In mice that expressed GCaMP6 in Tlx3 positive L5 IT neurons, the closed loop locomotion onset patterns differed markedly when compared across the six regions of interest (Figure 1I). Moreover, these activation patterns were different from those observed during open loop locomotion onsets (Figure 1J). While these responses could partially be explained by hemodynamic occlusion, the time course of hemodynamic occlusion in Tlx3-Cre mice that expressed eGFP in L5 IT neurons was similar to that observed in C57BL/6 mice that expressed eGFP brain wide and mainly resulted in a decrease of fluorescence (Figures S1C and S1D). In Tlx3 positive L5 IT neurons, open loop visual flow onsets that occurred in the absence of locomotion resulted in quantifiable responses in V1 and V2am (Figure 1K), but the closed loop locomotion onset responses in V1 could not be explained by the sum of locomotion and visual flow onset responses (Figure S2). Most of the widefield signals we record here likely originate in compartments in superficial layers which are composed of axons and dendrites of local neurons, and long-range inputs (Allen et al., 2017) (see Discussion). To test how well the widefield signals recorded in V1 reflected local somatic activity in V1, we repeated the experiments using two-photon calcium imaging of somatic activity in Tlx3 positive L5 IT neurons. We found that while there were differences between the responses measured with widefield imaging and those measured at the soma, both for locomotion onset responses and visual responses, these differences did not change our conclusions (Figure S3).

We then repeated these experiments and analyses for the other excitatory and inhibitory cortical neuron types (Figure S4). For each neuron type, we quantified the similarity of the activity patterns associated with closed and open loop locomotion onsets by calculating the correlation coefficient between the activity traces in a window around locomotion onset (−5 s to +3 s) and averaging across regions of interest (see Methods). We found the highest similarity between closed and open loop locomotion onsets in C57BL/6 mice that expressed GCaMP brain wide and in mice that expressed GCaMP6 in PV positive interneurons, and the similarity was lowest in mice that expressed GCaMP6 in Tlx3 positive L5 IT and Ntsr1 positive L6 excitatory neurons (Figure 1L). Thus, excitatory neurons of deep cortical layers exhibited the strongest differences between closed and open loop locomotion related activation. The remarkable aspect of these findings is not that there are differences between closed and open loop locomotion onsets, particularly in visual regions of cortex one would perhaps expect to find differences, but that these differences are larger specifically in the deep cortical layers.

To quantify how the interaction between locomotion and visual flow feedback changes across dorsal cortex, we computed correlations between calcium activity and locomotion speed for each region of interest, for closed and open loop conditions separately. In C57BL/6 mice that expressed GCaMP brain wide, this correlation was comparable between closed and open loop conditions, was highest in posterior regions of dorsal cortex, and lower in more anterior regions (Figures 2A and 2B). In mice that expressed GCaMP6 in Tlx3 positive L5 IT neurons, the correlation between locomotion speed and calcium activity in the closed loop condition was negative in posterior regions of dorsal cortex and positive in anterior regions. Conversely, in open loop conditions, the correlation was positive throughout dorsal cortex (Figures 2C and 2D). To test whether it was the coupling of visual flow that resulted in a reduction of the correlation in the closed loop condition or whether the presence of an uncoupled visual stimulus resulted in an increase of correlation in the open loop condition, we also compared these correlations to those observed when mice were locomoting in relative darkness (see Methods). In C57BL/6 mice that expressed GCaMP brain wide, the correlation of calcium activity with locomotion in darkness was indistinguishable from that in closed or open loop (Figure 2B). In mice that expressed GCaMP6 in Tlx3 positive L5 IT neurons, an interesting pattern emerged. In visual regions of dorsal cortex, the correlation between locomotion speed and calcium activity in relative darkness resembled that observed in open loop conditions, while in frontal regions, it resembled that observed in closed loop conditions. In visual regions, the coupling of visual flow resulted in a reduction of correlation, while in frontal regions the presence of an uncoupled visual stimulus resulted in an increase of correlations. Thus, the difference between closed and open loop conditions likely had different origins in visual and frontal regions of dorsal cortex. Overall, we concluded that the coupling of locomotion and resulting visual flow feedback not only activated Tlx3 positive L5 IT neurons differentially but did so throughout dorsal cortex, well beyond visual regions.

Figure 2.

Visuomotor prediction errors originated in visual cortical regions and had opposing effects on Tlx3 positive L5 IT activity

A consequence of the natural coupling of locomotion and visual flow feedback is that movements are a good predictor of self-generated sensory feedback. In our closed loop experiments, forward locomotion was a strong predictor of backward visual flow. It has been speculated that such predictions are used to compute prediction errors in L2/3 of cortex (Jordan and Keller, 2020). Prediction errors come in two variants: Positive prediction errors that signal more visual flow than predicted and negative prediction errors that signal less visual flow than predicted (Keller and Mrsic-Flogel, 2018; Rao and Ballard, 1999). To probe for negative and positive prediction error signals, we quantified the responses to visuomotor mismatch events and to unexpected visual stimuli (representing less and more visual flow input than expected, respectively). Visuomotor mismatch was introduced by briefly (1 s) halting visual flow during self-generated feedback in the closed loop condition. Visual responses were elicited by presenting drifting grating stimuli to a passively observing mouse (see Methods). Please note, both types of prediction errors can be probed using a variety of stimuli, but the two we are using here, in our experience, are the most robust at eliciting strong responses in V1. In C57BL/6 mice that expressed GCaMP brain wide, both visuomotor mismatch and grating stimuli resulted in increases of activity that were strongest and appeared first in visual regions of dorsal cortex (Figures 3A-3C). In mice that expressed GCaMP6 in Tlx3 positive L5 IT neurons, visuomotor mismatch resulted in an increase of activity in V1, while grating onsets resulted in a decrease of activity in V1 (Figures 3D-3F). Thus, in dorsal cortex, both mismatch and drifting grating stimuli primarily activated visual cortex but resulted in opposing activation patterns in Tlx3 positive L5 IT neurons. An increase in activity to negative prediction errors and a decrease to positive prediction errors would be consistent with either a population of negative prediction error neurons or a type of internal representation neuron (Rao and Ballard, 1999). Given that L5 neurons in V1 do not exhibit the opposing influence of top-down and bottom-up inputs necessary to compute prediction errors (Jordan and Keller, 2020), we speculated that Tlx3 positive L5 IT neurons might function as a type of internal representation neuron.

Figure 3.

The antipsychotic drug clozapine alters visuomotor integration in L5

The activation of an internal representation is the brain’s best guess at what the stimuli in the environment are. We speculated that the activation of such an internal representation would be particularly susceptible to substances that are known to reduce illusory percepts, like antipsychotic drugs. We thus quantified the changes in dorsal cortex activity associated with a single intraperitoneal injection of an antipsychotic drug (clozapine). In C57BL/6 mice that expressed GCaMP brain wide, we found that closed and open loop locomotion onsets were almost unaffected by clozapine (Figures 4A and 4B), while there was a small increase in open loop visual flow onset responses (Figure 4C). In mice that expressed GCaMP6 in Tlx3 positive L5 IT neurons, however, clozapine fundamentally changed both closed and open loop locomotion onset responses. In V1, both types of locomotion onset now resulted in a massive increase in activity (Figures 4D and 4E). Conversely, open loop visual flow onset responses remained largely unchanged (Figure 4F). These effects could not be explained by clozapine induced changes in hemodynamic occlusion (Figure S5A) or locomotion behavior (Figures S5B and S5C). Clozapine also had an opposing effect on the average correlation of activity with locomotion speed in Tlx3 positive L5 IT neurons. While the average correlation between activity and locomotion speed was increased in C57BL/6 mice that expressed GCaMP brain wide, the same measure was decreased in mice that expressed GCaMP6 in Tlx3 positive L5 IT neurons (Figure S5D). Finally, we confirmed that this effect was also present when recording somatic activity in V1. Using two-photon imaging, after clozapine administration we observed an increase of somatic responses during locomotion onsets that was very similar to that observed with widefield imaging (Figures S5E and S5F). Thus, consistent with the speculation that Tlx3 positive L5 IT neurons might function as internal representation neurons, clozapine exhibited a substantially stronger influence on locomotion onset responses in Tlx3 positive L5 IT neurons than would be expected from the effect of clozapine on brain wide responses.

Figure 4.

Antipsychotic drugs decouple long-range cortico-cortical activity

L5 IT neurons are likely the primary source of long-range cortical communication (Chen et al., 2019; Harris et al., 2019; Leinweber et al., 2017). Given that clozapine increased locomotion onset responses in L5 IT neurons in V1, we wondered whether clozapine might more generally change the long-range influence between L5 IT neurons. To investigate this possibility, we computed the correlations of calcium activity between the six regions of interest across both hemispheres and all recording conditions. We did this first in C57BL/6 mice that expressed GCaMP brain wide before (Figure 5A) and after a single intraperitoneal injection of clozapine (Figure 5B). Overall, clozapine resulted in a relatively modest decrease in correlations. On repeating this experiment in mice that expressed GCaMP6 in Tlx3 positive L5 IT neurons, we found a larger decrease in correlations that appeared to be stronger for regions that were further apart (Figures 5C and 5D). To quantify these changes, we plotted the correlations as a function of linear distance between the regions as measured in a top-down view of dorsal cortex (Figure S6A). Note, this is only an approximation of the actual axonal path length between the regions. To visualize the distribution of correlations, we interpolated the data and represented them as heatmaps (Figure S6B). The resulting representation for the data from C57BL/6 mice that expressed GCaMP brain wide before and after clozapine injection is shown in Figures 6A and 6B, respectively. We then split the data into approximately equal portions of short-and long-range correlations using a cutoff of 0.9 times the bregma-lambda distance (approximately 3.8 mm). While both short-and long-range correlations were reduced, this reduction was not significant in either case (Figure 6C). We also found no evidence of a difference between short-and long-range correlation changes. In mice that expressed GCaMP6 in Tlx3 positive L5 IT neurons, however, we found a significant reduction in both short-and long-range correlations (Figures 6D and 6E), and both were more strongly reduced than in C57BL/6 mice that expressed GCaMP brain wide (C57BL/6 vs Tlx3-Cre x Ai148, short-range: p < 0.05; long-range: p < 10^-5; rank-sum test). Moreover, the reduction in activity correlations was stronger for long-range correlations than it was for short-range correlations (Figure 6F). To control for possible clozapine induced changes in hemodynamic occlusion, we repeated the experiment and analyses in Tlx3-Cre mice that expressed eGFP in L5 IT neurons and found that alterations in hemodynamic occlusion could not account for the decrease in correlations but resulted in a small increase of the same metric (Figure S6C-S6E). To further dissect the spatial scale at which this decorrelation occurs, we repeated the experiments in Tlx3-Cre x Ai148 mice that expressed GCaMP6 in L5 IT neurons, recording calcium activity using both a widefield and a two-photon microscope. We found that while the activity correlations measured with the widefield imaging decreased (Figure S7A-S7C), pairwise somatic activity correlations measured by two-photon imaging did not (Figure S7D). This is consistent with the idea that the decorrelation primarily occurs for long-range interactions. Finally, to confirm that the clozapine induced reduction in correlation was stronger in deep cortical layers, we repeated the experiment in a population of L2/3 and L4 excitatory neurons using Cux2-CreERT2 x Ai148 mice. We found that clozapine also decreased correlations of cortical activity in superficial excitatory neurons (Figure S8). However, this reduction was significantly weaker than the reduction we observed in mice that expressed GCaMP6 in Tlx3 positive L5 IT neurons (Cux2-CreERT2 x Ai148 vs Tlx3-Cre x Ai148, short-range: p < 0.005, long-range: p < 10^-8, rank-sum test). Thus, administration of the antipsychotic drug clozapine resulted in a decorrelation of activity across dorsal cortex that was stronger in L5 IT neurons (Tlx3) than it was in either upper layer excitatory neurons (Cux2) or in the brain wide average (C57BL/6).

Figure 5.

Figure 6.

To test whether the decorrelation of the activity of L5 IT neurons is specific to clozapine or might more generally be a functional signature of antipsychotic drugs, we repeated the experiments with two additional antipsychotic drugs, aripiprazole and haloperidol. We found that the principal effect of decorrelation of the activity patterns of Tlx3 positive L5 IT neurons was preserved with both drugs (Figures 7A-7F). These changes in correlation were absent upon injection of a psychostimulant, amphetamine (Figures 7G-7I). Thus, it is possible that a decrease in long-range correlations between cortical L5 IT neurons might be a common mechanism of action for antipsychotic drugs.

Figure 7.

Changes in the correlation between cortical regions could either be driven by changes in correlated external inputs or through changes in communication between cortical regions. If antipsychotic drugs primarily reduce correlations by reducing the strength of communication between cortical regions, independent of whether this communication is direct or indirect, we would expect to find changes in the way responses spread between cortical regions. We used visuomotor prediction error responses that originated in V1 (Figure 3) to quantify how the spread of these responses in Tlx3 positive L5 IT neurons was influenced by antipsychotic drugs. The computation of negative prediction errors during movement is thought to be driven by an excitatory prediction of visual flow mediated by long-range cortical input from L5 neurons in regions like A24b (Leinweber et al., 2017), as well as an inhibitory bottom-up visual input. By contrast, the computation of a positive prediction error during passive observation is thought to be driven by an excitatory bottom-up visual input and an inhibitory top-down input. Thus, we would primarily expect to find a reduction in mismatch responses in L5 IT neurons and a reduction in the spread of this signal to secondary visual regions. This is indeed what we observed: Mismatch responses were partially reduced in V1 and almost absent in V2am after injection of antipsychotic drugs (Figure 8A). While responses to onsets of drifting gratings were also reduced in V2am, they were overall less affected by the antipsychotic drugs than mismatch responses (Figure 8B). These results are consistent with the interpretation that antipsychotic drugs reduce lateral communication in L5. Whether this communication is mediated by direct long-range connections between L5 neurons, or indirectly via other cortical neurons or subcortical loops, remains to be tested.

Figure 8.

Discussion

When interpreting our results, it should be kept in mind that the source of the widefield imaging signal in these experiments is not entirely clear and is likely a combination of somatic, dendritic, and axonal signals. Given that the signal attenuation half depth in cortex is on the order of 100 µm (Aravanis et al., 2007), most signals likely come from the superficial 100 µm of tissue (Allen et al., 2017). In the case of the Tlx3 recordings, this would mean that we were recording a mixture of axonal and dendritic signals of the Tlx3 neurons that have their soma in L5 but have a dense net of axons and dendrites in L1. In the case of the brain wide expression of GCaMP using an AAV-PHP.eB in C57BL/6 mice, the signals we recorded likely also include contributions from thalamic axons. In addition to this, there is a hemodynamic contribution to the signal (Allen et al., 2017; Ma et al., 2016; Valley et al., 2020) that is the result of varying occlusion by blood vessels. We attempted to correct for the hemodynamic response using a 405 nm wavelength isosbestic illumination of GCaMP6 but could not recover the hemodynamic responses measured in eGFP experiments. Accurate correction likely requires calibration on eGFP measurements (Valley et al., 2020). Given that hemodynamic responses should not systematically differ between different mouse lines, and that in the crystal skull preparation the contribution of hemodynamic occlusion to increases in fluorescence was relatively small (Figure S1), these hemodynamic artefacts should not substantially alter our conclusions. Lastly, it should be kept in mind that responses in these widefield recordings inherently reflect population averages. Assuming there is some form of lateral inhibition among neurons, population averages need not reflect responses of neurons that are tuned to the specific stimulus presented. For example, the suppression to grating stimuli we observed in V1 Tlx3 positive L5 IT neurons could be the result of a subset of Tlx3 neurons that are strongly responsive to the specific grating presented suppressing the rest of the population of Tlx3 neurons. The sum of both could then result in a net decrease in calcium activity. Thus, an understanding of the mechanism that results in opposing population responses of Tlx3 positive L5 IT neurons to positive and negative prediction errors will require measurements with single neuron resolution.

In our experiments we have focused on Tlx3 positive L5 IT neurons, but given that L5 IT neurons are densely connected to L5 pyramidal tract neurons (Anderson et al., 2010; Kiritani et al., 2012) and to L6 neurons (Xu et al., 2016; Yamawaki and Shepherd, 2015), it is not unlikely that these neuron types would follow similar patterns of activity and exhibit similar susceptibility to the influence of antipsychotic drugs. In the comparison of closed and open loop locomotion onsets, we do indeed find that Ntsr1 positive L6 excitatory neurons exhibit similar levels of dissimilarity as those observed in Tlx3 positive L5 IT neurons (Figure 1L). It is possible that responses in the Ntsr1 recordings, however, were noisier than those in the Tlx3 recordings (Figure S4), likely because of the lower dendritic and axonal volume of Ntsr1 positive neurons in L1. Thus, while we have focused on the Tlx3 positive L5 IT neurons mainly for reasons of experimental convenience, it is possible that both the differential activation in closed and open loop locomotion onsets as well as the susceptibility of long-range correlations to antipsychotic drugs is a general feature of cortical layers 5 and 6.

It is still unclear what the relevant sites of action in the brain are of antipsychotic drugs. Antipsychotic drugs are often antagonists of the D2 receptors, which are most strongly expressed in striatum. Consequently, it is sometimes assumed that the relevant site of action of antipsychotics is the striatum. Consistent with this, an increase in striatal dopamine activates auditory cortex (Maia and Frank, 2017) and mediates hallucination-like perceptions in mice (Schmack et al., 2021). It is intriguing, however, to speculate that the effects we observed in L5 IT neurons, which provide a dominant input to striatum (Morita et al., 2019), may be an alternate site of action. While it is possible that the primary site of action of the antipsychotic drugs with respect to the L5 decorrelation phenotype we observed is striatum, this is unlikely. Changes to striato-thalamo-cortical loops could appear as changes in functional connectivity observed in cortex. However, feedback from striatum to cortex is exclusively mediated through thalamus, and neither L4, the primary thalamo-recipient layer in cortex, nor our brain wide recordings exhibited any of the effects we saw in L5 IT neurons. Thus, the effects we saw in L5 IT neurons are either not inherited from striatum or are mediated by a separate set of thalamic neurons via direct thalamic input to deep cortical layers (Constantinople and Bruno, 2013; Douglas and Martin, 1991). Thalamocortical collaterals in deep layers, however, come from thalamic neurons that project primarily to superficial layers (Oberlaender et al., 2012). Thus, we think it is unlikely that the observed effects in L5 IT neurons were inherited from striatum. Independent of whether the treatment-relevant site of action of antipsychotic drugs is indeed in L5 IT neurons, the long-range decorrelation of activity we observed after antipsychotic drug injection could be used as a basis for a functional screen for compounds with antipsychotic efficacy. Given that in the brain wide recordings the decorrelation of activity was weaker compared to that observed in Tlx3 positive L5 IT neurons, it might be possible to test for similar effects in human subjects using layer specific fMRI recordings (Haarsma et al., 2020).

A rather surprising implication of our results is that the dominant mode of coupling in cortex might be lateral, not vertical. Two findings would argue for this interpretation. First, while L5 was differentially activated across cortical regions during closed and open loop locomotion onsets, L2/3 exhibited rather limited differences. Second, antipsychotics decorrelated L5 activity but had no effect on L2/3. Both findings could be explained if the long-range lateral coupling within L5 is stronger than the local vertical coupling between L2/3 and L5.

One of the diseases treated with antipsychotic drugs is schizophrenia. It has been speculated that disrupted resting-state functional connectivity in brain networks involving cortex is central to the pathogenesis of schizophrenia (Li et al., 2017; Northoff and Duncan, 2016). Treatment-naïve schizophrenia patients predominantly exhibit increases in resting state functional connectivity between cortical regions, while medicated schizophrenia patients exhibit more balanced fractions of increased and decreased functional connections (Li et al., 2019). Auditory hallucinations, one of the hallmark symptoms of schizophrenia, are associated with increased activation of diverse cortical regions (Catafau et al., 1994; Dierks et al., 1999; Shergill et al., 2000). Thus, it is conceivable that antipsychotic drugs counteract this by reducing the strength of long-range cortical connectivity mediated by L5 IT neurons, independent of whether the increased activity in schizophrenic patients is indeed driven by aberrant increases in the strength of long-range cortical connections or is a consequence of an aberrant increase in the strength of striato-thalamo-cortical loops. However, a direct involvement of cortical circuits in the pathogenesis of schizophrenia is consistent with the glutamate hypothesis of schizophrenia. This hypothesis is based on the finding that NMDA receptor antagonists induce schizophrenia-like symptoms (Moghaddam and Javitt, 2012), and that disturbances of NMDA and AMPA receptor related gene expression are associated with schizophrenia (Harrison and Weinberger, 2005; Singh et al., 2020). In V1, it has been shown that NMDA receptors are necessary during visual development to establish normal prediction error signaling (Widmer and Keller, 2021). Thus, it is conceivable that a disruption of glutamatergic signaling in cortex that alters the way internal representations are activated by prediction errors is central to the etiology of schizophrenia. If so, we should find impairments in all domains of cortical function, including sensory processing. In the case of the sensory cortex, cortical dysfunctions should be apparent earlier in development, concurrent with the critical period for plasticity in sensory areas of cortex. While the core symptoms of schizophrenia typically appear in late adolescence, sensory processing is indeed impaired in schizophrenic patients (Javitt, 2009) and is apparent, for example, in the reduction of surround suppression in V1 (Yoon et al., 2009). Interestingly, some of these disturbances are detectable already during childhood as visuo-perceptual and reading anomalies in some subjects that later develop schizophrenia (Parellada et al., 2017). Positive symptoms of schizophrenia can be explained in a predictive processing framework of cortical function as a disturbance of a prediction error based update of internal representations (Fletcher and Frith, 2009; Frith, 2005). There are two types of prediction error neurons, positive and negative, and, given that cortex likely implements a non-hierarchical variant of predictive processing (Garner and Keller, 2022), two types of internal representation neurons. One set of internal representation neurons is excited by positive prediction errors and inhibited by negative prediction errors, and the other one is inhibited by positive prediction errors and excited by negative prediction errors (Keller and Mrsic-Flogel, 2018). Tlx3 positive L5 IT neurons exhibited opposing responses to positive and negative prediction errors (Figure 3), consistent with the responses of internal representation neurons. Given that the effects of antipsychotic drugs are predominantly observed in L5 IT neurons, and that anesthesia is thought to decouple L5 neuron soma from their apical tuft and thereby prevent conscious perception (Suzuki and Larkum, 2020), we speculate that Tlx3 positive L5 IT neurons may be one type of internal representation neuron and one of the functionally relevant sites of antipsychotic action.

Supplemental figures

Figure S1.

Figure S2.

Figure S3.

Figure S4.

Figure S5.

Figure S6.

Figure S7.

Figure S8.

Supplemental tables

Table S1.

Table S2.

Methods

Key Resource Table

Mice

All animal procedures were approved by and carried out in accordance with guidelines of the Veterinary Department of the Canton Basel-Stadt, Switzerland. The mice used for widefield imaging in this study were kept on a C57BL/6 background and were of the following genotype: 6 C57BL/6 (Charles River Laboratories), 4 Emx1-Cre (Gorski et al., 2002), 4 Cux2-CreERT2 (Franco et al., 2012), 7 Scnn1a-Cre (Madisen et al., 2010), 22 Tlx3-Cre (Gerfen et al., 2013), 3 Ntsr1-Cre (Gong et al., 2007), 2 PV-Cre (Hippenmeyer et al., 2005), 6 VIP-Cre (Taniguchi et al., 2011), and 5 SST-Cre (Taniguchi et al., 2011) mice. Ai148 ((Daigle et al., 2018), Jackson Laboratories, stock #030328) mice were used as breeders to drive GCaMP6f expression in Cre positive neurons. An additional 8 C57BL/6 and 7 Tlx3-Cre mice were used in experiments controlling for hemodynamic responses. To induce expression of Cre in the Cux2-CreERT2 line, mice were provided for at least 1 week with tamoxifen containing food (Envigo, 400 mg tamoxifen per kg of food) as their sole food source. Mice were group housed in a vivarium (light/dark cycle: 12/12 hours). Experimental mice were of either sex.

Surgery and virus injections

For all surgical procedures, mice were anesthetized using a mixture of fentanyl (0.05 mg/kg; Actavis), midazolam (5.0 mg/kg; Dormicum, Roche), and medetomidine (0.5 mg/kg; Domitor, Orion). In a subset of mice (6 C57BL/6, 4 Emx1-Cre, and 4 SST-Cre mice), we injected an AAV vector based on the PHP.eB capsid (Chan et al., 2017) retro-orbitally (6 μl each eye of at least 10¹³ GC/ml) to drive expression of GCaMP under either the EF1α or hSyn1 promoter pan-neuronally, or in a cell type specific manner using a double-floxed inverted open reading frame (DIO) construct. For hemodynamic response control experiments, we used an AAV-PHP.eB-EF1α-eGFP to pan-neuronally express eGFP (Figures S1A and S1B) or an AAV-PHP.eB-EF1α-DIO-eGFP to restrict eGFP expression to L5 IT neurons in Tlx3-Cre mice (Figures S1C and S1D, S5A, S6C-S6E). To improve optical access to the cortex, we implanted crystal skull cranial windows (Kim et al., 2016). Prior to removing the skull plate overlying dorsal cortex, we recorded the location of bregma relative to other landmarks on the skull. Superglue (Pattex) was used to glue the crystal skull in place in the craniotomy. In a subset of mice, for the comparison of hemodynamic responses (Figure S1A and S1B), we used a clear skull preparation (Guo et al., 2014). To do this, we attached the crystal skull coverslip directly onto the cleaned skull surface with a three-component polymer (C&B Metabond, Parkell). A custom-machined titanium head bar was attached to the skull using dental cement (Paladur, Heraeus). An epifluorescence overview image was taken to mark reference points on the dorsal cortical surface. Anesthesia was antagonized by an intraperitoneal injection of a mixture of flumazenil (0.5 mg/kg; Anexate, Roche) and atipamezole (2.5 mg/kg; Antisedan, Orion Pharma). For peri-and post-operative analgesia we injected buprenorphine (0.1 mg/kg; Reckitt Benckiser Healthcare (UK) Ltd.) and Metacam (5 mg/kg; Boehringer Ingelheim). Imaging commenced at the earliest 1 week after head bar implantation or 3 weeks after retro-orbital AAV injection. Clozapine (Novartis), aripiprazole (Otsuka Pharmaceutical), haloperidol (Janssen), or amphetamine (Hänseler) was intraperitoneally injected at 0.2 μg/g - 10 μg/g, 0.2 μg/g, 0.1 μg/g or 4 μg/g body weight, respectively.

Virtual reality setup and stimulus design

For all experiments we used a virtual reality setup as previously described (Leinweber et al., 2014). Mice were head-fixed and free to run on a spherical, air-supported Styrofoam ball. A virtual corridor was projected (using a Samsung SP-F10M projector) onto a toroidal screen positioned in front of the mouse covering a field of view of approximately 240 degrees horizontally and 100 degrees vertically. Recordings were blocked into 5-minute sessions. Experiments typically commenced with 4 sessions of a closed loop condition in which the rotation of the spherical treadmill was coupled to the movement in a virtual corridor. To introduce mismatches, we broke the coupling between locomotion and visual flow by briefly halting visual flow for 1 s at random times. Closed loop sessions were followed by 4 sessions of an open loop condition, in which rotation of the spherical treadmill and movement in the virtual corridor were decoupled. The visual stimulus consisted of a replay of the visual flow recorded in the previous closed loop condition. Following this we recorded 2 sessions of what we refer to as a dark condition in which the virtual reality was switched off. Please note, this was not complete darkness, primarily because the shielding of the blue excitation light was not light proof. Finally, we recorded activity in 2 sessions of a grating condition in which we presented drifting grating stimuli (8 directions of 6 s ± 2 s (mean ± SD) duration each with an inter-grating interval of 4.5 s ± 1.5 s (mean ± SD) gray screen, presented in randomized order). These sessions, totaling approximately 1 hour of recording time per day, were acquired for 3 consecutive days. For mice in which we went on to test the effect of drugs on dorsal cortex activity, the first post-drug injection data were acquired +1 h after drug injection (+30 mins in the case of amphetamine because of faster onset kinetics of the drug) and further data were collected at the +24 h and +48 h timepoints. Data in the manuscript are shown as the combined data over either all data collected before drug injection (labelled as naive) or after drug injection (labelled as clozapine, aripiprazole, haloperidol or amphetamine, respectively).

Data acquisition

All widefield imaging experiments were performed on a custom-built macroscope consisting of commercially available objectives mounted face-to-face (Nikon 85 mm/f1.8 sample side, Nikon 50 mm/f1.4 sensor side). We used a 470 nm LED (Thorlabs) powered by a custom-built LED driver for exciting GCaMP (or eGFP) fluorescence through an excitation filter (SP490, Thorlabs) reflected off a dichroic mirror (LP490, Thorlabs) placed in the parfocal plane of the objectives. Green fluorescence was collected through a 525/50 emission filter on an sCMOS camera (PCO edge 4.2). Apertures on objectives were usually kept maximally open and the current at the LED driver was used to adjust fluorescence intensity to a value that was kept below 25% of the maximum dynamic range of the sensor. In cases where this was not possible (e.g., transfection of eGFP yielded extremely bright fluorescence, often visible with the naked eye), objective apertures were gradually reduced to avoid overexposure of the sensor. LED illumination was adjusted with a collimator (Thorlabs SM2F32-A) to achieve homogenous illumination across the surface of the cranial window. The resulting Gaussian profile of the illumination cone was further trimmed with black tape on the sample side objective to avoid light shining directly into the mouse’s eye. An Arduino board (Arduino Mega 2560) was used to control LED onsets synced to the frame trigger signal of the camera. The duty cycle of the 470 nm LED was 90%. Raw images were acquired at full speed (100 Hz; except for eGFP controls, which were acquired at 50 Hz effective frame rate) and full dynamic range (16 bit) of the sensor. The raw images were cropped on-sensor and the resulting data was streamed to disk with custom written software in LabVIEW (National Instruments), resulting in an effective pixel size of 60 μm² at a standardized imaging resolution of 1108 pixels x 1220 pixels (1.35 MP).

Two-photon calcium imaging was performed using a custom-built microscope based on a Thorlabs B-scope (Leinweber et al., 2014). Illumination source was a tunable femtosecond laser (Coherent Chameleon) tuned to 930 nm. Emission light was band-pass filtered using a 525/50 filter for GCaMP and detected using a GaAsP photomultiplier (H7422, Hamamatsu). Photomultiplier signals were amplified (DHPCA-100, Femto), digitized (NI5772, National Instruments) at 800 MHz, and band-pass filtered around 80 MHz using a digital Fourier-transform filter implemented in custom-written software on an FPGA (NI5772, National Instruments). The scanning system of the microscope was based on a 12 kHz resonant scanner (Cambridge Technology). Images were acquired at a resolution of 750 pixels x 400 pixels at a 60 Hz frame rate and a piezo-electric linear actuator (P-726, Physik Instrumente) was used to move the objective (Nikon 16x, 0.8 NA) in steps of 15 µm between frames to acquire images at 4 different depths. This resulted in an effective frame rate of 15 Hz. The field of view was 350 µm x 350 µm.

Data processing

Off-line data processing and data analysis (section below) were done with custom-written MATLAB (2021a) scripts. For widefield macroscope imaging, raw movie data was manually registered across days by aligning subsequent mean projections of the data to the first recorded image sequence. We placed regions of interest (ROIs) relative to readily identifiable anatomical landmarks that had been previously noted during cranial window surgery (see above). This resulted in the selection of six 20 pixels x 20 pixels ROIs per hemisphere. Of those ROIs, we calculated the activity as the ΔF/F₀, wherein F₀ was the median fluorescence of the recording (approximately 30 000 frames in a 5 min recording). ΔF/F₀ was corrected for slow fluorescence drift caused by thermal brightening of the LED using 8^th percentile filtering with a 62.5 s moving window similar to what was described previously for two-photon imaging (Daniel A. Dombeck et al., 2007).

Raw two-photon imaging data were full-frame registered to correct for brain motion. Neurons were manually selected based on mean and maximum fluorescence images. Raw fluorescence traces were corrected for slow drift in fluorescence using an 8^th percentile filtering with a 15 s window (Daniel A Dombeck et al., 2007). ΔF/F traces were calculated as mean fluorescence in a selected region of every imaging frame and subsequently subtracted and normalized by the median fluorescence.

Data analysis

Locomotion and visual flow onsets were determined based on threshold crossing of locomotion or visual flow speed. To select only well isolated onsets, we used a speed threshold of 30 cm/s and excluded all onsets with locomotion or visual flow in the 3 s window preceding the onset. To increase the number of locomotion onsets that passed the speed threshold criterion for the eGFP imaging data (Figures S1, S5A), we excluded only those onsets if there was locomotion in the 1 s window preceding onset.

All stimulus response curves (Figures 1F-1L, 3B and 3C, 3E and 3F, 4, 8, S1-S4, S5A and S5B, S5E and S5F) were baseline subtracted. The baseline subtraction window for unpredictable stimuli (mismatches and grating onsets) was −200 ms to 0 ms before stimulus onset. For closed loop and open loop locomotion onsets, and open loop visual flow onsets, to accommodate calcium indicator offset dynamics and potential anticipatory activity, we placed the baseline subtraction windows for these onsets at −2900 ms to −2700 ms for GCaMP imaging, or −900 ms to −700 ms for eGFP imaging. Varying this window within the bounds of offset dynamics of the indicator and anticipatory activity onset did not change the conclusions drawn in this manuscript. Analyses in which image sequences were aligned to event onsets (Figures 1C-1D and 3) used the same parameters as above.

On rare occasions, we observed seizure-like activity after antipsychotic drug injection. These events resulted in atypically high levels of correlation, and to avoid them from influencing our conclusions we excluded these data from further analysis. We excluded sessions if they contained an event in which average activity of all 12 ROIs remained elevated above 30% ΔF/F for more than 10 s. This criterion led to the exclusion of 0.26% of data (10 out of 3801 of 5 min recording sessions).

To quantify the similarity of closed loop and open loop locomotion onset responses shown in Figure 1L we calculated the correlation coefficient between the averaged responses (onsets and mice, 1 s moving average) in a window −5 s to +3 s, for each ROI.

To calculate the maps that show the timing of activation of individual dorsal cortex regions (Figures 3B and 3C, 3E and 3F), we first aligned the craniotomies of all mice within one genotype to the first mouse’s average grating response map and then calculated an average dorsal cortex response time course. For each pixel of this image sequence, we then determined the time at which the pixel’s value crossed a threshold of the maximum value in the entire time course, divided by 6 standard deviations of the pixel’s baseline value (−200 ms to 0 ms before event onset, as stated above). The resulting maps were denoised using a Gaussian filter and we removed all pixel clusters that contained fewer than 500 connected pixels. For illustration purposes, these maps were overlaid over an example crystal skull craniotomy to produce the panels shown in Figure 3.

For the correlation analyses shown in Figures 5-7, S6-S8, we first calculated the correlation coefficient of neuronal activity (or fluorescence change, in the case of Tlx3-Cre mice that had been retro-orbitally injected with an AAV-PHP.eB-Ef1α-DIO-eGFP to express eGFP in L5 IT neurons) for all possible pairs of the 12 regions of interest. We then determined the distance between each pair of ROIs and normalized this value by the distance between bregma and lambda obtained from the images during initial surgery for each mouse (Figure S6A). We plotted the correlation against the distance for each pair of regions of interest and binned the data in a 40 x 40 grid and smoothed the resulting image with a gaussian filter (Figure S6B), to obtain the heatmaps shown in Figures 6A and 6B, 6D and 6E, 7A and 7B, 7D and 7E, 7G and 7H, S6B-S6D, S7A and S7B, S8C and S8D. Contour lines in the heatmaps were then drawn to 50% of the maximum value in the plot. The boxplot quantifications (Figures 6C, 6F, 7C, 7F, 7I, S6E, S7C, S8E) were calculated as the change in correlation coefficient of activity after and before treatment for all pairs of regions of interest, normalized to the value before treatment and split into short-and long-range using a threshold of 0.9 of the bregma-lambda distance for each mouse. For pairwise neuronal correlations that had been recorded with two-photon imaging (Figure S7D), we calculated the correlation coefficient separately for all pairs of neurons within each field of view. We noticed extreme synchrony of activity of L5 IT soma recorded in 1 field of view with two-photon imaging after clozapine injection and show, both distributions for all data, as well as the distribution when this 1 field of view is excluded.

For Figure S2, onset aligned data were first averaged for each mouse and then summed in the given combinations. The SEM was calculated over mice.

To obtain the differences in the correlation between activity and locomotion after and before clozapine injection (Figure S5D), we first averaged all correlation coefficients in the individual sessions of a condition (closed loop, open loop, or dark) to generate one correlation coefficient per mouse and ROI, for each condition. We then calculated and plotted the difference between these correlation values after and before clozapine injection.

Acknowledgements

We thank Tingjia Lu for production of the AAV vectors, Rebecca Jordan for performing the clear skull implants used in the hemodynamic response measurements, and members of the Keller lab for discussion and support. This project has received funding from the Swiss National Science Foundation, the Novartis Research Foundation, and the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No 865617).

Author contributions

MH performed the experiments and analyzed the data. All authors wrote the manuscript.

Conflict of interest

Part of the results presented herein have been included in patent application EP22153051.1.