Recent years have seen a surge in neuronal ensemble recording technology, making it feasible to simultaneously record from multiple brain areas hundreds to thousands of cells in behaving animals (Laubach et al., 1999; Gutierrez et al., 2005; Lebedev and Nicolelis, 2006; Wiest et al., 2007; Long and Carmena, 2011; Aggarwal et al., 2013; Schwarz et al., 2014). Studies have investigated how this massive amount of data relate to the functioning of neural circuits (Stevenson and Kording, 2011) and, ultimately, to animal behavior (Mohammed et al., 2016), but many open questions still remain. Following Hebb's description of the synaptic activity (Hebb, 1949), the neuronal assembly (NA) hypothesis propose that subgroups of co-activated cells constitute the fundamental unit of information processing in the brain (Nicolelis et al., 1997; Harris, 2005; Buzsáki, 2010). Importantly, subgroups may encompass distant neurons, distributed in different regions, possibly forming complex neuronal circuits through indirect connections or due to a common input (Buzsáki, 2010). The existence of NAs is widely discussed (Gerstein et al., 1989; Buzsáki, 2006; Picado-Muino et al., 2013; Russo and Durstewitz, 2017), as it correlates with several neurological phenomena (Sakurai, 1998; Fries, 2005; Buehlmann and Deco, 2008; Engel et al., 2013). However, despite the mounting experimental evidence supporting NAs and distributed brain interactions (Bower et al., 2015; Valdez et al., 2015; Carrillo-Reid et al., 2016; Mishra et al., 2016), their experimental verification is still a grand challenge in neuroscience.

The Hebbian theory sought to reconcile single-neuron dynamics to complex psychological phenomena, proposing that: (1) assembly neurons fire synchronously and as a consequence reinforce their synaptic efficiency; (2) one neuron may participate simultaneously in multiple NAs; (3) information coding is distributed; (4) NAs dynamically form and reconfigure; (5) NAs activate simultaneously, and a reduced number of NA cells can trigger the activity of the entire NA; and (6) active NAs can activate other NAs and evolve in phase-sequences, possibly decoupled from sensory or internal events, providing the basis for complex cognitive processing (Sakurai, 1998). However, the temporal dynamics of the formation and dissolution of assemblies, which prevent the brain from converging on undesirable states such as a global synchronization, was not fully theorized by Hebb (Buzsáki, 2006). The original proposal has thus been extended to contemplate not only neurons with synaptic connections, but all cells presenting above chance interactions. In this way, the term NA has no absolute definition with respect to neuronal activity and can be associated with varying degrees of temporal precision, scale and internal structure (Russo and Durstewitz, 2017). Here, we will consider that NAs are formed by neurons that jointly increase their average firing rates for some period (Lopes-dos Santos et al., 2011, 2013). In any case, there is growing agreement that NAs will be better understood by inspecting their causal link with motor commands and other behavioral features (Buzsáki, 2010, and references therein). In particular, characterizing NAs within behavioral tasks is relevant for assessing the distributed role of cortical regions in distinct cognitive processes (Gutierrez et al., 2005; Hu et al., 2005; Carrillo-Reid et al., 2016; Dejean et al., 2016), for the understanding of the neural code (Winters and Reid, 2010; Russo and Durstewitz, 2017), and for the design of new brain-machine interfaces (Wessberg et al., 2000; Hochberg et al., 2006; Ifft et al., 2013; Shokur et al., 2013; Ramakrishnan et al., 2015).

In the context of studying distributed neural interactions and NA dynamics, active tactile discrimination tasks with rats are elucidative: it is a consolidated animal model (Bush et al., 2016), involving tactile stimuli and higher-order cognitive processes such as sensory-motor integration, attention and decision making, hence engaging several cortical areas in the processing of information (Stoeckel et al., 2003; Cao et al., 2008). These task aspects are not considered in passive or anesthetized scenarios (Szwed et al., 2003; Krupa et al., 2004). During active tactile discrimination, the primary somatosensory cortex (S1) activity is directly linked to vibrissae stimulation (Simons, 1978), and recent studies focus on establishing the relationship between S1 modulations and other cortical areas. For example, Pais-Vieira et al. (2013) showed that the anticipatory modulation of S1 is dependent on the primary motor cortex (M1). Zagha et al. (2015), in a sensory detection task, found competing ensembles of neurons in M1, whose opposing spiking patterns were related to mapping of sensory stimulus to motor commands.

In particular, there is a growing interest on understanding how multisensory information processing might relate to animal performance, i.e., how information integrated from multiple sensing modalities encodes behavior information (Driver and Spence, 2000; Sheppard et al., 2013; Semprini et al., 2016; Bieler et al., 2017a,b). There is growing evidence supporting interactions between S1 and the primary visual cortex (V1): for instance, Zangaladze et al. (1999) used transcranial magnetic stimulation (TMS) to disrupt the activity in V1 and verified an interference in the tactile discrimination used for spatial orientation during the free exploration of novel objects. Vasconcelos et al. (2011) recorded S1 and V1 simultaneously during rat blind exploration of distinct objects in a dark environment and reported a similar amount of task information in the neural response of both areas, which points to multimodal neural interactions. Moreover, Sieben et al. (2013) noticed that visual stimuli induced neuronal oscillations and modulated the power of S1 activity in a task where animals received simultaneous light flash and whisker deflection; and Sieben et al. (2015) reported a reduction on direct connections between S1 and V1 when animals received less stimulation in neonatal development, further supporting the hypothesis that information processing in primary cortical areas is distributed. Finally, Bieler et al. (2017a) reported higher phase-coupling when animals received bimodal (visual-tactile) stimulation.

In addition, experiments highlight the role of the prefrontal cortex (PFC) and the posterior parietal cortex (PPC) in tactile decision-making, considering that these areas are anatomically connected to a variety of sensory regions (Reep et al., 1994; Miller and Cohen, 2001; Behrmann et al., 2004; Murray et al., 2012; Licata et al., 2017). PPC engages in cognitive processing of attention (Behrmann et al., 2004; Stilla et al., 2007) and is involved in the spacial representation of the environment (Whitlock et al., 2008) as well as in the association between tactile and visual information in rats (Winters and Reid, 2010). The PFC, in turn, is associated to planning (Tanji and Hoshi, 2008; Martinet et al., 2011), working memory (Funahashi and Kubota, 1994; Riley and Constantinidis, 2015) and reasoning (Miller and Cohen, 2001) and in somatosensory information coding (Murray et al., 2012). In a model of PFC columnar organization, Martinet et al. (2011) showed that PFC single-cell activation is involved in spatial navigation planning, whilst Reid et al. (2013) reported impaired object recognition when PFC was lesioned.

In this work, we test the hypothesis that neurons distributed in long-range fronto-parieto-occipital networks participate in different NAs and are involved in active tactile discrimination. For that, we investigate the dynamic formation of NAs within and among the primary somatosensory (S1), primary visual (V1), prefrontal (PFC), and posterior parietal cortices (PPC), simultaneously recorded from seven Long-Evans rats during an active tactile discrimination task. Simultaneous recordings in the four aforementioned areas are, to the best of our knowledge, original in the literature, and can contribute to shed light on neuronal population interactions within the same animal engaged in a cognitive task. First, we show that neuronal activity in all four regions is modulated by tactile discrimination and reward collection. Then, using a neuronal decoder, we show that the majority of single neurons carry little information about the stimulus and reward, but when the whole population is considered, average decoding performance resembles or surpasses that obtained by animals in the task. Finally, we identify the dynamic formation of NAs during tactile discrimination, and show that assembly neurons contain more information about animal behavior than neurons that do not participate in the assembly. These findings further support that distributed neural interactions encode active tactile discrimination.

The number of recorded units in each cortical area and the number of successful trials from each animal's dataset are presented in Table 2. Table 3 include task performance. Animal reaction times and movement times over the trial period, is presented in Figure 1B considering all rats, and in Figure 1C for each rat.

There is growing evidence that neural interactions are distributed (Nicolelis et al., 1997; Harris, 2005; Buzsáki, 2010). Thus, the study of neuronal assemblies (NAs) is relevant as it furthers the understanding of functional interactions and information processing in the brain. The characterization of NAs is still debatable (Lopes-dos Santos et al., 2013; Litwin-Kumar and Doiron, 2014; Russo and Durstewitz, 2017), though numerous studies have provided insights on the link between neural population dynamics and animal behavior (Terada et al., 2013; Siegle et al., 2014; Romano et al., 2015; Campagner et al., 2016; Dejean et al., 2016). In this work, we recorded 553 neurons from PFC, PPC, S1 and V1 of rats engaged in an active tactile discrimination task. Importantly, our analyses exploit simultaneous recordings from areas involved in primary sensory processing (S1 and V1) and higher-order cognition (PFC and PPC). Our results reveal task-related distributed neural activity and the formation and dissolution of NAs, which are suggestive of distributed neural interactions. We show that neural population responses, and in particular NA neurons, contain more tactile information than random groups of neurons. There are, however, single neurons scattered among the recorded regions with information content comparable to that of groups.

The first set of analysis, PSTH (Figure 3) and MFR (Figure 4), suggests that all of the selected cortical areas may be enrolled in the discrimination of tactile information: the overall neuronal activity presented significant firing modulations in the vicinity of the tactile discrimination period. Changes in neuronal activity are more prominently observed shortly before the nose poke (NP), when vibrissae stimulation and anticipatory effects are in place, and after the nose poke, which may be either due to the decision making process and reward collection, or to sensory-motor integration (Zangaladze et al., 1999; Krupa et al., 2004; Vasconcelos et al., 2011; Sieben et al., 2013; Namboodiri et al., 2015). The origins of anticipatory modulations were investigated by Pais-Vieira et al. (2013) in thalamo-cortical loops during the same tactile discrimination task. These authors indicate that anticipatory activity depends on top-down effects generated in part by motor gating. Therefore, despite the stereotypical animal behavior observed in this task, brain dynamics may be sensitive to task performance levels but also to the different behavioral strategies developed by animals (Nienborg et al., 2012; Sachidhanandam et al., 2013; Manita et al., 2015; Siegel et al., 2015; Yang et al., 2015). Further work should explore the relationship between neural dynamics and task performance.

The decoding accuracy of animal response from their neuronal firing modulations (Figure 5) increased above chance levels from the moment whiskers contacted the variable length bars to the collection of reward. In PFC and PPC, decoding accuracy peaked shortly after tactile stimulus, which is in agreement with the roles of these cortical areas in attention, planning and reasoning (Miller and Cohen, 2001; Behrmann et al., 2004; Stilla et al., 2007). For the primary cortical areas, the decoding accuracy for V1 peaked before S1, and reached higher values (mean of 85% vs. mean of 75%), but this difference may be attributed to the different number of neurons recorded in each area (Figure S6). In this task, sustained decoding accuracies after the tactile stimulation have been noted before in S1 and V1 (Krupa et al., 2004; Pantoja et al., 2007; Wiest et al., 2010; Vasconcelos et al., 2011), and here we provide the first report that the same occurs in PFC and PPC. Previously, Gardner et al. (2007) has reported anticipatory activity in PPC preceding the tactile responses during active touch in monkeys. From Figure S3, we observe markedly differences in spike rate between the wide and narrow scenarios from NP to reward, which explains the decoding profile. This sustained firing rate differences observed in all regions, but in particular in PFC, may be linked to the temporal evolution of the decision-making process required for tactile discrimination (Romo et al., 1999; Hernandez et al., 2002; Brody et al., 2003; Sreenivasan et al., 2014). Finally, the task studied comprises several aspects of cognitive processing, such as attention, decision making and sensory discrimination, thus further work is needed to clarify the neural information dynamics between primary sensory and associative areas in the context of this task.

Grand populational averages of MFR may mask the functional interactions in neural subgroups. The NAs characterization, performed by PCA associated with ICA, identified groups of neurons with joint activations beyond chance levels (Figure 6). From this characterization, we inferred a functional relationship map between the brain regions studied, before and during tactile stimulation (Figure 6B). Before NP, we noticed that the majority of functional network connections were restricted within single areas. At NP, functional connections emerged. The NAs comprising neurons from S1 and V1, found in 6 out of 7 animals, is in line with the findings of Vasconcelos et al. (2011), who reported distributed information processing in cortical primary sensory areas. PPC engaged in more NAs comparing the period prior NP and during NP, further supporting its role in multisensory integration (Winters and Reid, 2010; Harvey et al., 2012; Raposo et al., 2014). PFC established functional relationships with S1 in all of the recorded animals and with V1 in 6 out of 7 animals. Considering decision making tasks, this supports the observed central role of PFC in communicating with primary areas, engaged either in information coding (Murray et al., 2012) or planning and reasoning (Miller and Cohen, 2001). In addition, PFC is spatially distant from the other aforementioned cortical regions and their engagement in NAs during the discrimination task is an additional indicative that information processing is distributed in the brain (Nicolelis et al., 1995; Nicolelis, 2005; Winters and Reid, 2010; Guo et al., 2014). Finally, our results are aligned with the asynchronous convergence hypothesis (Nicolelis et al., 1995; Nicolelis, 2005), i.e., that active tactile discrimination results from the dynamic interplay of multiple descending, ascending, and local afferents that converge asynchronously on neurons located at each stage of the trigeminal pathway.

When assessing distributed interactions and the formation of NAs, the information content present in the neural population in comparison to that present in single-neurons provides further insights on the relationship between neural dynamics and behavior. As noted by others (Krupa et al., 2004; Pantoja et al., 2007; Wiest et al., 2010), the average single-neuron is minimally informative about aperture width whilst population decoding approximates or even exceeds animal performance (see Figure 5, Figures S4, S5, and Table 3). NA neurons contained more information about reward side than random groups of neurons, as demonstrated in Figure 6. However, there are highly informative single-neurons (Figure 5 and Figure S4). Bearing in mind subsampling distortions, we note that even if a single neuron presents a high decoding performance, Hebb emphasized that assemblies would evolve over time as phase sequences and these would facilitate cognitive acts, including motor behavior. Thus, according to Hebb, phase sequences would carry more information than neuronal firing rates (Almeida-Filho et al., 2014). Also, Buzsáki (2010) remarks in his comprehensive review of NA activity that determining the size of a NA is challenging and there is ample evidence that the contribution of single neurons to a given NA is skewed, i.e., a small number of assembly members may be as informative (from the decoder point of view) as a much greater number of active cells. However, one single neuron may not suffice to discharge downstream neurons - a key theorized role for NAs.

In addition to S1, we recorded from primary visual cortex because of previous reports on its involvement in cross-modal responses during tactile tasks (Zangaladze et al., 1999; Sadato et al., 2004; Vasconcelos et al., 2011). To prevent animals from using visual or auditory cues to perform the task, the behavioral apparatus was located inside a sound-attenuating and light-proof isolation box, previously described by Krupa et al. (2001). In the same task described here, the authors have shown that animal performance drops to chance when whiskers are cut. Besides, we only used infrared light during our recording sessions because rats do not have red nor infrared vision (Szél and Röhlich, 1992; Jacobs et al., 2001). Also, using the same behavioral apparatus, Vasconcelos et al. (2011) have shown that V1 neurons are recruited for tactile processing in the absence of visible light. In our work, modulations in firing rate are observed in V1 seconds before the rat samples the bars with whiskers (Figures 3, 4), but differences in firing rate due to narrow or wide apertures (Figure S3) only appear after tactile stimulation. James et al. (2006) showed that visual and haptic stimulus caused similar time course activation of the lateral occipital cortex and argue that there is mounting evidence that visual and haptic systems share the same neural substrate in encoding the structure of objects. Also, Amedi et al. (2002) report an overlap between visual and haptic processing in regions regarded as purely visual. Furthermore, Saito et al. (2006) showed that tactile stimuli activate V1 in well trained players of the chinese game of Mah-Jong. This is in line with Zangaladze et al. (1999), who demonstrated that optimal tactile performance in normally sighted subjects has an intimate relationship with the visual cortex dynamics. Interestingly, Johnson and Frostig (2016) found long, horizontal axons linking the barrel cortex to the visual cortex in rats. Even though the visual and tactile systems are closely related, we believe that the apparently earlier V1 response in comparison to that of S1 in our results may likely arise from different sized neural samples, but further work should explore possible alternative routing of tactile information to visual areas.

Functional NAs lack a strict definition, especially regarding timescales and the understanding of how interactions on the scale of single neurons relate to behavioral outcomes (Harris, 2005; Kreuz et al., 2013; Russo and Durstewitz, 2017). In this work, we used a method that characterizes NAs based on synchronous modulations in neuronal firing rate that does not support causality or time-delayed inferences. Also, it is not possible to discriminate between NAs formed by internally generated mechanisms from NAs resulting from common external inputs (Lopes-dos Santos et al., 2013). Moreover, linear PCA-based methods are fast and of high temporal resolution, but have limitations in face of the intrinsically non-linear and flexible brain dynamics (Peyrache et al., 2010; Lopes-dos Santos et al., 2013; Russo and Durstewitz, 2017). Despite the limitations due to non-stationary data, establishing the statistical threshold of assembly membership via the Marčenko-Pastur distribution considerably reduces the computational burden in comparison with surrogate methods. While this approach is much more computationally tractable than a surrogate approach, it identifies significant rate correlations which are unlikely to arise by chance, but without directly quantifying their statistical significance. Nevertheless, our findings (e.g. that NA units carry more task-related information than non-NA units) support that this approach can reveal physiologically relevant neural correlations. In this way, methods that incorporate information theory (Lopes-dos Santos et al., 2015) or statistical frameworks (Russo and Durstewitz, 2017) may enhance the assessment of distributed cortical interactions. Nevertheless, similarly to previous works (Peyrache et al., 2010; Lopes-dos Santos et al., 2011, 2013; Almeida-Filho et al., 2014; Gulati et al., 2014; Bower et al., 2015), the PCA+ICA method for characterizing NAs revealed the formation of task-related groups of neurons that are more informative about tactile discrimination than random groups of neurons, which is suggestive of functional neural interactions. Further work shall refine functional NA characterization, also considering the behavioral distinctions with which animals navigate in the scenario (Grün, 2009; Long and Carmena, 2011).

In summary, in this work we analyzed a complex dataset, recorded simultaneously from four different cortical regions in seven rats engaged in an active tactile discrimination task, which is, to the best of our knowledge, original in the literature. In addition to highlighting the neural dynamics and information content from primary sensory (S1 and V1) and associative areas (PFC and PPC), we characterize the dynamic formation of NAs, a hallmark of distributed neural interactions. Taken together, we believe our results contribute to shed light on the neural mechanisms underlying intelligent behavior.

CD designed the study, performed data analysis and wrote the manuscript; AK designed the study, performed experiments and wrote the manuscript; MdS performed experiments; FL provided crucial feedback on data analysis and manuscript; RM designed the study, performed data analysis and wrote the manuscript.