Self-motion processing in visual and entorhinal cortices: inputs, integration, and implications for position coding

Abstract

The sensory signals generated by self-motion are complex and multimodal, but the ability to integrate these signals into a unified self-motion percept to guide navigation is essential for animal survival. Here, we summarize classic and recent work on self-motion coding in the visual and entorhinal cortices of the rodent brain. We compare motion processing in rodent and primate visual cortices, highlighting the strengths of classic primate work in establishing causal links between neural activity and perception, and discuss the integration of motor and visual signals in rodent visual cortex. We then turn to the medial entorhinal cortex (MEC), where calculations using self-motion to update position estimates are thought to occur. We focus on several key sources of self-motion information to MEC: the medial septum, which provides locomotor speed information; visual cortex, whose input has been increasingly recognized as essential to both position and speed-tuned MEC cells; and the head direction system, which is a major source of directional information for self-motion estimates. These inputs create a large and diverse group of self-motion codes in MEC, and great interest remains in how these self-motion codes might be integrated by MEC grid cells to estimate position. However, which signals are used in these calculations and the mechanisms by which they are integrated remain controversial. We end by proposing future experiments that could further our understanding of the interactions between MEC cells that code for self-motion and position and clarify the relationship between the activity of these cells and spatial perception.

INTRODUCTION

The brain’s ability to estimate self-motion, or the speed and direction of one’s movement, is essential to survival. Across the animal kingdom, striking examples of self-motion processing occur in both invertebrate and vertebrate species. Desert ants can continuously integrate locomotion cues (i.e., “step counting” or pedometry) to keep track of their current distance from a home nest site (Wittlinger et al. 2006; Wohlgemuth et al. 2001). When flying, honeybees and budgerigars balance optical flow between their two eyes to control speed and calibrate the distance to barriers (Bhagavatula et al. 2011; Kirchner and Srinivasan 1989). Gerbils and funnel-web spiders can return home in a straight line after venturing from a nest site in complete darkness (Görner and Claas 1985; Mittelstaedt and Mittelstaedt 1980). These behaviors all depend on accurately estimating one’s own velocity from some subset of the available external sensory and internal idiothetic cues.

The study of the neural substrates for self-motion remains challenging. This is due in part to the highly multimodal nature of self-motion. When an animal moves, it generates feedback from multiple modalities, most notably from visual, motor, and vestibular systems, that must be integrated to form a unified self-motion percept (Fig. 1). The problem is further complicated by the fact that computing self-motion often requires the extraction of higher-order features, such as the heading of an optic flow field, from primary sensory input. Sensory representations of these high-order features can then support speed and direction estimates, allowing animals to compute a unified self-motion percept based on multiple sensory modalities.

Fig. 1.

Components of the self-motion signal. Locomotor, optic, and vestibular motion cues carrying information about linear and angular velocity (v) must be integrated to form a unified self-motion percept.

However, the principles for integrating multiple sensory representations to generate a unified self-motion percept remain incompletely understood. Here, we focus on recent work addressing how neurons in the visual and medial entorhinal cortices represent self-motion signals derived from different sensory modalities. Because these two brain regions contain multimodal neural representations, they may serve as exemplary circuits to study the algorithms underlying the integration of sensory and idiothetic cues for computing self-motion estimates. We begin by discussing visual motion processing in the primate and rodent visual cortex, emphasizing that the larger body of primate work can be used as a model for advancing our understanding of the rodent system. We then review the properties of and some of the primary sources for self-motion coding in the medial entorhinal cortex (MEC). We consider the implications of these data for cellular mechanisms by which self-motion inputs might be integrated to estimate an animal’s current spatial position, a process known as path integration. We end by proposing how future work can identify causal links between specific cell types and spatial perception and address the ways in which animals use self-motion cues to estimate position.

VISUAL MOTION PERCEPTION IN PRIMATES

Visual input is a primary sensory source for extracting information regarding self-motion. Initial studies of motion processing focused on the representation of visual motion in the primate cortex. Progress in addressing this topic set a type of gold standard for identifying how a given brain region encodes multimodal information relevant to computing self-motion and demonstrating causality between encoding sensory features of self-motion and behavioral motion perception. Although we do not have the space to fully cover this topic here, we will provide a brief overview of its history, as it provides a critical foundation for work on self-motion processing in rodents.

PRIMATE VISUAL AREA MT IS SPECIALIZED FOR MOTION PROCESSING

Throughout the primate visual system, neurons encode various types of visual motion. Extensive anatomic and physiological data point to a multistage hierarchical processing of visual motion (Fig. 2) (Felleman and Van Essen 1991). Specialized cortical circuits for the analysis of motion begin in V1, where neurons show orientation and direction selectivity, responding only when an oriented stimulus such as a bar or grating moves in a particular direction. However, the functional specialization of motion processing is often ascribed to neurons in the middle temporal (MT) area of extrastriate cortex, a recipient of V1 input (Dubner and Zeki 1971; Zeki 1974). Functionally, MT has a columnar structure, with each column responding to visual motion of a particular direction in a form- and cue-invariant manner (Albright 1992; Albright et al. 1984; Maunsell and Van Essen 1983; Movshon et al. 1985; Van Essen et al. 1981). Nearly all MT neurons respond to first-order motion (∼99%), defined as spatiotemporal variations in luminance, and many respond to second order motion (∼87%), defined as spatiotemporal variations that do not consist of luminance (e.g., contrast, depth) (Albright 1992). In addition, the majority of MT neurons are also tuned to the speed of a drifting bar (∼80%) (Maunsell and Van Essen 1983). This speed tuning of MT neurons becomes more invariant to spatial frequency as the spatial frequency content increases (Priebe et al. 2003), suggesting that MT is specialized to combine multiple spatial frequency components of an object to compute its speed, rather than responding to the individual components in isolation.

Fig. 2.

Overview of primate and mouse visual systems. A: anatomic locations of primate visual areas (adapted from Rokszin et al. 2010 with permission). B: hierarchical structure of the primate visual system. Areas specialized for processing of motion and heading direction are highlighted (adapted from Rokszin et al. 2010 with permission). C: mouse visual areas. Anterograde tracers of 3 different colors were injected into 3 locations in primary visual cortex (V1). Preserved retinotopy was apparent in 9 extrastriate visual areas. Subsequent imaging studies found differences in preferred temporal and spatial frequencies across these extrastriate areas. Most prominently,anterolateral visual area (AL) prefers fast-moving stimuli, whereas posteromedial visual area (PM) prefers slow-moving stimuli (adapted from Wang and Burkhalter 2007). D: connectivity between the mouse visual areas, which was quantified based on tracer experiments. Network analysis revealed 2 modules of connected areas, which bear some similarities to the dorsal and ventral streams of primate visual cortex (reprinted from Wang et al. 2012 with permission). 36p, posterior area 36; ; A, anterior visual area; AM, anteromedial visual area; FEF, frontal eye field; FST, fundus of the superior temporal area; LIP, lateral intraparietal area; LIPd, lateral intraparietal area (dorsal); LIPv, lateral intraparietal area (ventral) LGN, lateral geniculate nucleus; LI, laterointermediate visual area; LM, lateromedial visual area; MM, mediomedial area; MST, medial superior temporal area; MT, middle temporal area; PIP, posterior intraparietal area; P, posterior visual area; PM, posteromedial visual area; PO, parieto-occipital sulcus; POR, postrhinal cortex; RL, rostrolateral visual area; RSA, retrosplenial agranular area; S1, primary somatosensory cortex; STP, superior temporal polysensory area; TE, anterior inferior temporal complex; TEa, anterior inferior temporal cortex (anterior); TEO, posterior inferior temporal cortex; TEp, anterior inferior temporal cortex (posterior); VIP, ventral intraparietal area; V2–4, 2nd, 3rd, and 4th visual cortices, respectively; 7a, visual area 7a in parietal cortex.

MICROSTIMULATION EXPERIMENTS ESTABLISHED CAUSALITY BETWEEN MT ACTIVITY AND MOTION PERCEPTION

Evidence for the causal role for MT in processing visual motion came from experimental work in which microstimulation of MT altered a monkey’s perception of motion direction. To demonstrate this principle, a classic set of studies used a task in which a field of randomly moving dots contained a certain percentage of dots that moved coherently in one direction. Although this stimulus contained no moving object, it elicited a motion percept whose strength depends on the coherence of the random dot movement. Indicating that activity in MT was associated with performance on this task, lesions of MT impaired the ability of a monkey to report the direction of motion of the coherent dots (Newsome and Paré 1988), MT neurons showed selectivity for the speed of the moving dots (Britten et al. 1992), and the activity of MT neurons correlated with perceptual thresholds for motion detection and speed discrimination (Liu and Newsome 2005). Critically, experiments then established causality between MT encoding and perceptual performance on this task by demonstrating that microstimulation of an MT column biased the monkey’s perception of motion toward that encoded by the stimulated neurons (Salzman et al. 1990).

PRIMATE VISUAL AREAS MST AND VIP COMPUTE HEADING FROM OPTIC FLOW PATTERNS

Unlike object motion, self-motion generates complex full-field optic flow stimuli from which heading direction and movement speed can be computed. The small- to medium-sized receptive fields of V1 and MT leave these areas less suited to encode full-field optic flow. However, the hierarchical transformation of visual information results in larger receptive fields in neurons located in upper visual areas, marking these regions as potential candidates for encoding full-field visual stimuli (Desimone and Ungerleider 1986; Komatsu and Wurtz 1988; Tanaka et al. 1986; Van Essen et al. 1981). Consistent with this idea, neurons in two such upper visual areas in the primate, the medial superior temporal area (MST) and ventral intraparietal area (VIP), are tuned to specific heading directions of optic flow and receive input from area MT (Fig. 2) (Britten 2008; Duffy and Wurtz 1995, 1991). Causally linking these responses to perception, microstimulation of MST neurons biased the monkey’s reports of heading perception based on optic flow (Britten and van Wezel 2002, 1998). Interestingly, a subset of neurons in MST and VIP also encoded vestibular signals, another primary sensory modality for deriving self-motion estimates (Ben Hamed et al. 2003; Page and Duffy 2003). Psychophysical and physiological work has shown that both visual information and vestibular signals can mediate the neural responses in MST and VIP and the subsequent perception of motion, depending on the type of stimulus used and the statistics of eye and head movements (Britten 2008; Gu et al. 2008; Gu et al. 2007; Gu et al. 2006). This indicates that MST could serve as a neural substrate for the integration of visual and vestibular cues for calculating self-motion percepts in the primate. However, other regions are also likely involved. Although MST neural signals correlate with heading direction, as a neural population, they often perform worse at discriminating heading than the monkey itself. Even so, the MT, MST, and VIP areas are part of the “dorsal stream” of visual processing that guides actions, and thus, motion processing in these regions is likely critical to visually guided navigation (Goodale and Milner 1992; Mishkin et al. 1983; Nassi and Callaway 2009).

VISUAL MOTION PROCESSING IN RODENTS

Developing technologies, in particular the advent of newer genetic tools for circuit analyses, have highlighted the rodent as a useful model system for studying cortical processing. Although the visual systems of primates and rodents differ, for example, in visual acuity, the underlying algorithms for calculating self-motion based on visual input may follow similar principles across species (Huberman and Niell 2011). Here, we review the basic anatomy of rodent visual cortex and describe how self-motion modulates response properties of visual cortex neurons.

PRIMARY AND SECONDARY VISUAL AREAS OF RODENT CORTEX

As in the primate visual system, primary visual cortex sends projections to a set of extrastriate visual cortical areas in the rodent (Fig. 2). Although there is no rodent analog of MT, MST, or VIP, there are at least nine secondary visual areas that receive direct V1 input (Garrett et al. 2014; Wang and Burkhalter 2007), and each of these areas has its own retinotopic map (Fig. 2). Anatomic connectivity suggests parallel organization of rodent and primate visual pathways, with the medial/anterior and lateral/posterior areas analogous to the dorsal and ventral primate visual pathways, respectively (Fig. 2) (Wang et al. 2012). Furthermore, some of these extrastriate areas are tuned to specific parameters of visual motion (Andermann et al. 2011; Marshel et al. 2011). Although it remains challenging to assign specific functions to specific rodent extrastriate areas, like motion processing to MT in primates, multiple studies have demonstrated that area AL (anterior and lateral to V1) prefers fast-moving stimuli (high temporal frequency, low spatial frequency), whereas area PM (posterior and medial) prefers slow-moving stimuli with finer detail (low temporal frequency, high spatial frequency) (Andermann et al. 2011; Glickfeld et al. 2013; Marshel et al. 2011). In particular, neurons in AL could serve as a neural substrate for the optic flow component of self-motion calculations, as locomotion often results in fast-moving visual stimuli (Andermann et al. 2011). If so, then rodent area AL could serve a computationally similar role to primate MST or VIP. However, the possibility also exists that some of the visual motion processing that occurs in extrastriate areas in primates could occur earlier in the visual pathway in rodents, as mouse V1 has a higher proportion of neurons selective to global motion patterns over individual motion components compared with primate or cat (∼10% in mouse vs. ∼1% in primate/cat) (Palagina et al. 2017).

THE INFLUENCE OF LOCOMOTION ON RODENT VISUAL CORTEX: SOURCES OF LOCOMOTION SIGNALS TO V1

Like primate regions MST and VIP, multiple sensory signals related to self-motion influence the activity of visual neurons in rodents. However, whereas the work in primates has demonstrated the presence of multimodal representations for optic flow and vestibular signals in higher-order visual areas, rodent work has primarily focused on how locomotion impacts visual processing in V1. Locomotion-related signals arrive in V1 from motor cortex (Leinweber et al. 2017; Miller and Vogt 1984) and the mesencephalic locomotor region of brain stem via the basal forebrain (Lee et al. 2014). In addition, thalamic projections to V1 from LP and dLGN have been shown to encode combinations of optic flow and the animal’s running speed (Roth et al. 2016). Thus, there are multiple sources of locomotion-related information to rodent visual cortex from both cortical and subcortical areas.

THE INFLUENCE OF LOCOMOTION ON RODENT VISUAL CORTEX: AROUSAL EFFECTS

Mice are more active when they are aroused. Therefore, the extent to which neurons directly encode self-motion is confounded by the effects of general arousal. Multiple studies have demonstrated that the aroused brain state associated with locomotion has profound effects on visual processing. When compared with quiescence, periods of locomotion desynchronize the cortex (Bennett et al. 2013; Polack et al. 2013), increase the gain of visual responses (Bennett et al. 2013; Fu et al. 2014; Niell and Stryker 2010; Polack et al. 2013), and reduce surround suppression (Ayaz et al. 2013), increasing spatial integration. Arousal also modulates visual cortex activity during spontaneous behavior. Unsupervised methods showed that the largest component of spontaneous behavior correlates with the traditional arousal measures (locomotion, whisking, and pupil size) (Stringer et al. 2018). Large-scale calcium imaging revealed that ∼60% of visual cortex neurons were significantly correlated or anti-correlated with this arousal-related component, alternating their activity on the scale of seconds to tens of seconds in a push/pull manner (Stringer et al. 2018). Neuromodulatory input likely plays an important role in generating these activity patterns, notably via long-range cholinergic projections from the basal forebrain (Arroyo et al. 2012; Fu et al. 2014; Pinto et al. 2013; Polack et al. 2013). In addition, several studies have implicated local inhibitory circuitry in brain state modulation (Fu et al. 2014; Jackson et al. 2016; Pakan et al. 2016; Polack et al. 2013).

Recent studies have attempted to dissociate the effects of locomotion and arousal on visual cortex activity by recording both locomotion and pupil size and examining periods of disagreement between the two or using air-puffs to induce arousal without causing locomotion (Reimer et al. 2014; Vinck et al. 2015). Here, pupil size is taken as the measure of “true” arousal, whereas locomotion is viewed as a separate variable that is correlated with arousal. Periods when the pupil was dilating were found to be associated with desynchronization of visual cortex (depolarization and reduction in variance of subthreshold membrane potential, reduction of low-frequency oscillations, and reduction of correlations between neurons) and enhancement of visual responses (Reimer et al. 2014; Vinck et al. 2015). In contrast, locomotion led to increases in firing rate in anticipation of and during movement (Vinck et al. 2015). Thus, arousal leads to a desynchronized network state in which sensory responses are more reliable, whereas locomotion can cause firing rate changes that do not occur during pure fluctuations in arousal. These locomotion-specific changes may be more directly related to self-motion processing. Pupillometry may be an effective way of teasing apart these effects.

THE INFLUENCE OF LOCOMOTION ON RODENT VISUAL CORTEX: RUNNING SPEED

In addition to brain state effects, recent evidence has raised the possibility that locomotor speed itself is encoded in V1. One study that provided evidence for this principle recorded V1 units while mice ran along a virtual corridor (Saleem et al. 2013). First, the authors found many V1 cells that were strongly tuned to running speed in the dark. Next, the mice ran in a “closed loop” session, where the mouse’s movement along the virtual corridor was directly tied to the locomotion of the mouse on the treadmill. This match between the mouse’s locomotion and the visual feedback creates the impression of moving through the virtual environment. Immediately after the closed loop session, the mice ran an “open loop” session where the visual cues were instead driven by the previous session’s locomotion trace. In these sessions, the link between locomotion and visual feedback was broken, allowing the researchers to separately assess the contributions of visual and locomotor speed to V1 firing rates. V1 neurons responded to linear combinations of visual and running speed with weights that varied across the population but were typically positive. Importantly, neurons varied their firing rate smoothly with running speed rather than in a binary fashion. This close tracking of running speed suggests true speed encoding rather than discrete changes between “active” and “quiescent” brain states. In future studies, it will be important to further separate out arousal effects by recording pupil size and other facial features. However, despite providing compelling evidence that locomotor speed is encoded in rodent V1, these results do come with the caveat that all vestibular input was removed by head fixation. To our knowledge, no studies have looked for speed-tuned cells in V1 in freely moving animals, which could serve as an interesting direction for future work.

THE INFLUENCE OF LOCOMOTION ON RODENT VISUAL CORTEX: VISUOMOTOR MISMATCH

Another way in which locomotion influences V1 activity is via the “visuomotor mismatch” response, in which V1 neurons respond robustly when the visual scene is briefly halted while the animal is running (Keller et al. 2012). These mismatch responses could potentially be used to detect moving objects while the animal is running (Keller et al. 2012; Zmarz and Keller 2016). Consistent with this interpretation, mismatch neurons have localized receptive fields (Zmarz and Keller 2016). Further supporting the existence of a dedicated population for visuomotor mismatch, axons from the lateral posterior nucleus of the thalamus in V1 responded to conditions in which locomotor speed exceeded visual speed (Roth et al. 2016). However, fewer responded to conditions in which visual speed exceeded locomotor speed. This asymmetry in the thalamic input could lead to nonlinearities in the way speed cells combine these cues in V1 as well as higher brain regions such as hippocampus and parahippocampal cortices.

The mismatch responses described above could be well fit by a model in which visual speed was encoded in inhibitory input (negative weight) and locomotor speed was encoded in excitatory input (positive weight) (Attinger et al. 2017). This contrasts with the positive weights reported by Saleem et al. 2013. The apparent disagreement between these studies could be resolved if V1 neurons heterogeneously encode locomotion depending on the exact experimental conditions. For example, brief mismatches between locomotion and vision (Keller et al. 2012) might be probing a different network state than the long periods of visuomotor decoupling used by in Saleem et al. 2013. Importantly, visuomotor mismatch cannot fully explain the influence of locomotion on V1 responses, as some V1 neurons are tuned to running speed in the dark (Saleem et al. 2013). In summary, locomotion exerts a profound influence on the activity of V1 neurons, including a combination of brain state, visuomotor mismatch, and speed-tuning effects.

SELF-MOTION CODES IN THE MEDIAL ENTORHINAL CORTEX

The studies reviewed above revealed multimodal representations of self-motion in visual cortex. Along with other inputs, these codes could contribute to the formation of spatial maps in the hippocampus or the entorhinal cortex. These brain regions are critical for spatial navigation (Maaswinkel et al. 1999; McNaughton et al. 2006; Parron and Save 2004), a complex process that requires synthesizing information regarding external landmarks with information about the animal’s movement relative to these landmarks. The process of integrating self-motion to estimate position is called path integration, and it is this component of navigation that we will focus on in the remainder of this review, as it is the component that directly relies on self-motion processing.

Path integration involves calculating a position estimate by continuously integrating self-motion cues (such as motor efference, optical flow, or vestibular information) to estimate one’s position and calculate vectors connecting one’s current and previous locations (Etienne and Jeffery 2004; Etienne et al. 1996; Mittelstaedt and Mittelstaedt 1980). Although path integration involves multiple cortical and subcortical structures, the hippocampus and entorhinal cortex appear to play critical roles, with lesions of these structures causing significant path integration deficits (Jacob et al. 2017b; Parron and Save 2004; Van Cauter et al. 2013). Lending support to the idea that MEC is critically involved in path integration was the electrophysiological discovery of MEC grid cells, which fire in multiple periodic spatial locations despite constant changes in an animal’s running speed and direction (Fyhn et al. 2004; Hafting et al. 2005; Maaswinkel et al. 1999; McNaughton et al. 2006). Grid cells appear immediately upon exploration of new environments, pairs of grid cells maintain their spatial phase relationships across environments, and, like path integration calculations, grid cells accumulate error that can be corrected by input from sensory landmarks (Fyhn et al. 2004; Hafting et al. 2005; Hardcastle et al. 2015; McNaughton et al. 1996; Yoon et al. 2013). Causally linking grid cell activity with path integration-based navigation, however, has remained challenging due to the lack of genetic markers corresponding to specific MEC cell classes (Sun et al. 2015; Tang et al. 2014). Even so, compelling evidence has emerged that strongly supports this idea. Manipulations that alter the structure or presence of the grid code but leave the majority of other MEC signals intact do significantly impact the accuracy of path integration (Allen et al. 2014; Gil et al. 2018). Moreover, pointing to a link between path integration and grid cell representations, recent work demonstrated that the firing patterns of grid cells and behavioral path integration-based estimates of position weigh locomotor versus visual cues to a similar degree when these two cues conflict (Campbell et al. 2018). Taken together, this body of work strongly implicates the MEC, as well as the spatial maps within it, in the ability to path integrate.

Although path integration is certainly not the only function of MEC, there is a considerable amount of ongoing research into how MEC processes self-motion information and how this self-motion information combines with landmark inputs and network dynamics to create grid cell firing patterns. Self-motion itself is encoded within the MEC by speed cells whose firing rate is correlated with the animal’s running speed (Hardcastle et al. 2017; Hinman et al. 2016; Kropff et al. 2015) and head direction cells that increase their firing rate when an animal faces a particular direction (Sargolini et al. 2006; Taube et al. 1990a, 1990b). Whether these cells provide the self-motion information that is integrated by grid cells to estimate position is an important open question. In the following sections, we focus on self-motion codes in the MEC, describe the known sources of this information, and consider future directions for answering how the MEC might support navigation through the integration of self-motion information.

SPEED CODING IN THE HIPPOCAMPUS

The earliest descriptions of speed coding in the parahippocampal formation came from electrophysiological recordings in the hippocampus (McNaughton et al. 1983). Hippocampal neurons were classified as either “theta” or “complex spike” cells, with both groups shown to increase their firing rate with running speed until reaching a saturation point at the highest running speeds. Later work quantified speed tuning more thoroughly across the hippocampal population, with the majority of recorded principal cells tuned to running speed (∼69%) (Wiener et al. 1989). Interestingly, this tuning was highly heterogeneous. Whereas some cells fired maximally at high running speeds, subsets of cells fired maximally at low running speeds, and many were optimally tuned to a particular intermediate running speed (Wiener et al. 1989). Thus, it became clear even at this initial stage that speed or its behavioral correlates played a key role in determining the firing rates of hippocampal principal neurons, and in these neurons a variety of tuning curve shapes existed for the relationship between firing rate and running speed.

SPEED CODING IN THE MEC

The encoding of speed and the heterogeneous nature of this coding also occur upstream of the hippocampus in MEC. Multiple works have identified populations of MEC neurons that strongly encode running speed. Initial reports suggested the presence of a small population (∼15% of MEC neurons) of speed cells, which were proposed to form a dedicated cell class that did not significantly overlap with other functionally defined MEC cell types (e.g., grid or head direction neurons) (Kropff et al. 2015). This population of speed cells maintained their speed tuning across different environments and contained an overrepresentation of fast-spiking interneurons (Kropff et al. 2015; Pérez-Escobar et al. 2016). The initial report of speed cells only considered neurons that showed a significant linear increase in their firing rate with running speed. Using more unbiased, model-based approaches, later work discovered that many cells in MEC (estimates ranged from 30 to 80%) show speed tuning (Hardcastle et al. 2017; Hinman et al. 2016). This speed tuning is highly heterogeneous, as the relationship between running speed and firing rate can be positive or negative, and the shape can take linear, saturating, and nonmonotonic forms (Hardcastle et al. 2017; Hinman et al. 2016; Kropff et al. 2015). It is also important to note that most of the speed-tuned cells in MEC have only weak correlations between running speed and firing rate (<0.3; higher correlations exist but are rare and often putative fast-spiking cells; see Kropff et al. 2015) and are conjunctive for position or head direction coding (Hardcastle et al. 2017). This mixed selectivity in speed coding offers advantages when downstream neurons must decode multiple discrete states from population-level activity (Barak et al. 2013; Fusi et al. 2016; Hardcastle et al. 2017; Rigotti et al. 2013).

AROUSAL COULD INFLUENCE MEASURES OF SPEED CODING

As previously mentioned, an important caveat to keep in mind when considering speed coding in the brain is that arousal and other motivational and emotional states are highly correlated with locomotion. A recent study using large-scale silicon probe recordings revealed that arousal-related signals were present in every brain region that was recorded (Stringer et al. 2018). Despite these important findings, the relative influence of arousal and running speed on MEC and hippocampus firing rates remains incompletely understood. Initial work on speed cells controlled for arousal by training rats to run at speeds set by the experimenter using a moving boxcar (Kropff et al. 2015). However, later studies primarily investigated speed coding during random foraging (Hardcastle et al. 2017; Hinman et al. 2016; Pérez-Escobar et al. 2016), raising the possibility that some speed tuning identified by these studies reflected changes in arousal rather than pure speed tuning. Moving forward, it will be important to distinguish the effects of arousal versus running speed in entorhinal self-motion codes. As has been elucidated in visual cortex (Fu et al. 2014; Reimer et al. 2016), neuromodulatory input and local inhibitory circuitry are strong candidates for driving behavioral state effects in MEC.

INPUT FROM THE MEDIAL SEPTUM IS AN IMPORTANT SOURCE OF LOCOMOTOR SPEED INFORMATION FOR THE MEC

Given that MEC lies several synapses away from primary sensory receptors, what sources might provide the inputs necessary for MEC neurons to encode running speed? One major source of speed-modulated input to the MEC comes from the medial septum diagonal band (MS) (Fig. 3), which sends strong projections to MEC, hippocampus, and other parahippocampal structures (Amaral and Kurz 1985; Unal et al. 2015). Stimulation of glutamatergic MS neurons elicits running in mice, whose speed increases with stimulation frequency (Fuhrmann et al. 2015). In addition, glutamatergic cells in MS are strongly tuned to running speed in the absence of optic flow (Justus et al. 2017). This suggests that the MS could serve as a prominent source of locomotion-based speed information to the MEC. Notably, inactivation of the MS selectively disrupts grid cells (Brandon et al. 2011; Koenig et al. 2011), although it remains to be determined which aspects of the MS input are essential for grid cell periodicity (Carpenter et al. 2017).

Fig. 3.

Locomotor speed input to medial entorhinal cortex (MEC) comes from the medial septum diagonal band (MS). A: images from the Allen Brain Explorer showing the anatomic locations of the MS and the MEC. B: the MS provides running speed information that is encoded in the firing rate of glutamatergic projections and in the frequency of theta rhythms. Speed-tuned glutamatergic input is integrated most effectively by MEC pyramidal (P) cells (thick green line) (Justus et al. 2017). MS fast-spiking (FS) cells strongly inhibit MEC FS cells (thick magenta line) (Fuchs et al. 2016; Justus et al. 2017). Cholinergic projections target a small subset of MEC pyramidal cells and interneurons (Justus et al. 2017). C: connectivity among excitatory cell types and FS interneurons in MEC. Two intermediate excitatory cell types, intermediate pyramidal (IM-P) and intermediate stellate (IM-S), were described based on electrophysiological and morphological parameters (Fuchs et al. 2016). Other inhibitory cell types, including SOM and 5HT3, are omitted for simplicity; for more complete data, see Fuchs et al. 2016. Note the lack of connectivity among stellate cells (Couey et al. 2013; Fuchs et al. 2016; Pastoll et al. 2013), suggesting that communication between stellate (S) cells could be mediated by speed-tuned interneurons, and the lack of connectivity between FS and pyramidal cells despite the prevalence of speed tuning in FS cells and the strength of MS glutamatergic projections to pyramidal cells. Connectivity rates between connected excitatory cell types were typically 5–10%. Dashed lines indicate pairings that were not tested in (Fuchs et al. 2016).

CELL TYPE-SPECIFIC CONNECTIVITY BETWEEN THE MS AND MEC

Speed information from the MS arrives in MEC in multiple formats via different types of projection neurons (Fig. 3). Axons from VGluT2-positive MS neurons monosynaptically connect to superficial MEC pyramidal, stellate, and fast-spiking interneurons, with this input integrated most effectively by pyramidal cells (Justus et al. 2017). This can then result in sustained depolarization of pyramidal cells when input from VGluT2 neurons is strong, which occurs during high running speeds (Justus et al. 2017). Counterintuitively, pyramidal cells have very low connectivity rates with MEC fast-spiking cells (Fuchs et al. 2016), which are frequently speed tuned (Kropff et al. 2015; Pérez-Escobar et al. 2016). This highlights the possibility that multiple inputs might carry speed information to MEC, or that multiple subpopulations of MEC neurons might independently encode speed. In addition to glutamatergic, speed-tuned projections, the MS sends GABAergic and cholinergic projections to MEC, both of which play an important role in generating theta rhythms (Chapman and Lacaille 1999; Hangya et al. 2009; Hasselmo 2006; Yoder and Pang 2005). GABAergic projections primarily target MEC interneurons (Fig. 3) (Fuchs et al. 2016; Gonzalez-Sulser et al. 2014). However, it is currently unknown whether cholinergic or GABAergic projection neurons are tuned to running speed. Moreover, whether any of the MS projections to MEC are used by grid cells or by animals to path integrate remains a matter of debate (see Fig. 6). What inputs provide the locomotor signals to MS also remains incompletely understood. Possible sources include subcortical structures such as the hypothalamus, in which regions that initiate locomotion when electrically stimulated do project to the MS (Fuhrmann et al. 2015; Sinnamon 1993), and the mesencephalic locomotor region of the midbrain (Lee et al. 2014) (see additional subcortical structures involved in processing self-motion).

Fig. 6.

Cellular sources of self-motion information to medial entorhinal cortex (MEC) grid cells. Unsolved problems are in italics. Landmark information is another critical input that was not discussed in this review.

MS-GENERATED THETA RHYTHMS ARE ANOTHER POTENTIAL SOURCE OF SPEED INFORMATION

The MS drives hippocampal and entorhinal theta rhythms, ∼4- to 12-Hz oscillations measured in the extracellular field potential that have been the subject of intense research over the past six decades (Buzsáki 2002; Buzsáki and Moser 2013; Petsche et al. 1962; Vanderwolf 1988). Theta rhythms have been proposed to organize parahippocampal activity into temporal chunks, in which cell assemblies carry information about the mnemonic order of events, places, or items and facilitate the transfer of information between brain regions (Buzsáki 2002). They are also relevant to self-motion calculations, as theta oscillations accompany spatial navigation (Buzsáki and Moser 2013; Maurer and McNaughton 2007), and their frequency and amplitude positively correlate with running speed (Giocomo et al. 2011; Jeewajee et al. 2008; Maurer et al. 2005; McFarland et al. 1975). This speed modulation of theta oscillations can impact the spiking dynamics of the subset of MEC neurons that fire at theta frequency, with their interspike interval decreasing with running speed (Hinman et al. 2016).

CHALLENGES TO THE ROLE OF θ-RHYTHMS IN SELF-MOTION PROCESSING IN THE MEC

The prominence of oscillations in the theta frequency band seen in rodents does not apply to all mammalian species, however. In primates and bats for example, MEC neurons show grid firing patterns during navigation despite the absence of continuous theta frequency oscillatory activity (Jacobs et al. 2013; Killian et al. 2012; Yartsev et al. 2011). Although not continuous, the bouts of theta observed in humans are speed modulated, showing significantly higher power during movement compared with immobility (Aghajan et al. 2017). This suggests that theta could still provide speed information in species without continuous theta, with the signal complemented in MEC by speed input from other sources. Supporting this idea, inactivation of the medial septum can enhance the firing rate speed signals of some MEC neurons (Hinman et al. 2016). Together, this points to the importance of contributions from other inputs in driving the response properties of MEC speed cells, such as proprioceptive feedback, motor efference copies regarding movement, or sensory features such as optic flow or whisking information (Chorev et al. 2016; Kerr et al. 2007; McNaughton et al. 2006).

ADDITIONAL SUBCORTICAL STRUCTURES INVOLVED IN PROCESSING SELF-MOTION

Although we have focused on the MS because it is one of the primary inputs to the entorhinal cortex and hippocampus, many other subcortical areas are involved in processing self-motion. This is perhaps unsurprising, since, from an evolutionary perspective, self-motion processing is an important basic survival skill that likely preceded the development of the cortex. Speed tuning has been found in the habenula and interpeduncular nucleus of the rat (Sharp et al. 2006). However, other than a small population of cells in the lateral habenula (∼10%), much of the speed tuning reported in this study was temporally coarse on the time scale of minutes. This type of activity is less useful for path integration and could instead reflect slowly fluctuating brain states that co-vary with running speed. An additional subcortical brain region that likely plays a significant role in self-motion processing is the mesencephalic locomotor region (MLR) in the brainstem. Located in the midbrain, it was first discovered in cats, when Shik et al. (1969) found that electrically stimulating it evoked locomotion. More recent studies have leveraged modern genetic tools to dissect the roles of distinct MLR cell types and nuclei in mice (Caggiano et al. 2018; Josset et al. 2018; Roseberry et al. 2016). Cell type-specific optogenetic and chemogenetic studies showed that activating MLR glutamatergic cells initiates locomotion. An important distinction was made between the cuneiform nucleus (CfN) and the pedunculopontine nucleus (PPN), both of which are part of the MLR. CfN but not PPN cells were critical for activating high-speed gaits (gallop and bound), whereas glutamatergic cells in both brain regions were involved in activating low-speed gaits (walk and trot) (Caggiano et al. 2018; Josset et al. 2018). The basal ganglia project to MLR nuclei, and this connection seems to be critical for the initiation and termination of locomotion (Roseberry et al. 2016). The MLR strongly projects to the basal forebrain (Lee et al. 2014), which encompasses the MS. Therefore, this pathway could be a critical component of the locomotor speed input to hippocampus and parahippocampal structures. Thus, it is important keep in mind that there are likely other important subcortical loci of self-motion processing that have yet to be fully described.

VISUAL INPUT TO MEC: CONNECTIVITY WITH VISUAL CORTICES

What other sensory features might contribute self-motion cues to MEC? In addition to locomotion, recent data highlight a role for the visual system in providing relevant velocity information to MEC. Although there are no direct visual subcortical projections to MEC, there are several routes by which visual self-motion information could arrive via direct or feedfoward projections from visual cortex (Fig. 4A). Topographically organized projections directly from visual cortex target MEC, with dorsal MEC receiving ∼20–30% of its cortical input from visual areas and ventral MEC receiving only ∼2% (Fig. 4B). These projections originate mostly in lateral and medial secondary visual areas (layers II and VI), but there are also some projections from layer VI of primary visual cortex (Burwell and Amaral 1998). Of the secondary visual areas, projections to MEC are strongest from areas LM, LI, P, and POR (Wang et al. 2011). Interestingly, areas LM and LI prefer high temporal frequency stimuli, and so they could be used to process fast-moving visual input during running (Fig. 4C). Area P provides the strongest input to MEC of any secondary visual area, but its coding properties have not been studied, suggesting an interesting direction for future work. Areas AL and AM are orientation and direction selective and preferentially respond to fast-moving stimuli. Although these coding features suggest that AL or AM may be important for self-motion processing, neither region provides much direct input to MEC. This information could instead reach MEC indirectly. Despite these recent insights, however, the degree to which rodent primary and secondary visual areas represent full-field optic flow stimuli and how this information could reach the MEC remains unknown, topics that future work could address.

Fig. 4.

Routes by which visual information reaches medial entorhinal cortex (MEC). A: strength of projections from 10 visual areas to MEC and lateral entorhinal complex (LEC; data adapted from Wang et al. 2012 with permission). MEC receives the majority of its direct visual input from lateral/posterior secondary visual areas. B, top: schematic of projections from visual areas to MEC. Thickness of lines is proportional to connection strength reported in A. B, bottom: axons from primary and secondary visual cortex innervate dorsal MEC (dMEC) much more densely than ventral MEC (vMEC) (Burwell and Amaral 1998). C: mean orientation selectivity index (OSI), direction selectivity index (DSI), preferred spatial frequency (SF), and preferred temporal frequency (TF) of 7 visual areas studied in (reproduced from Marshel et al. 2011 with permission). Note that the area with strongest projections to MEC (area P) was not included in this study. In general, more work is needed to identify specific streams of self-motion information from visual cortex to MEC. D: the influence of darkness on MEC self-motion codes. Top: firing rates and slopes of MEC speed cells both decreased in darkness (Pérez-Escobar et al. 2016). Middle: firing rates and slopes of speed-modulated grid cells decreased in darkness (Chen et al. 2016). Bottom: frequency and slope of local field potential theta oscillations in the MEC decreased in darkness (Chen et al. 2016). 36p, posterior area 36; A, anterior visual area; AM, anteromedial visual area; LI, laterointermediate visual area; LM, lateromedial visual area; MM, mediomedial area; MST, medial superior temporal area; MT, middle temporal area; PIP, posterior intraparietal area; P, posterior visual area; PM, posteromedial visual area; PO, parieto-occipital sulcus; POR, postrhinal cortex; RL, rostrolateral visual area; RSA, retrosplenial agranular area; S1, primary somatosensory cortex; STP, superior temporal polysensory area; TE, anterior inferior temporal complex; TEa, anterior inferior temporal cortex (anterior); TEO, posterior inferior temporal cortex; TEp, anterior inferior temporal cortex (posterior); VIP, ventral intraparietal area; V2–4, 2nd, 3rd, and 4th visual cortices, respectively; 7a, visual area 7a in parietal cortex;

EVIDENCE FOR THE IMPORTANCE OF VISUAL INPUT FOR MEC SPEED AND POSITION CODES

Taken together, the work reviewed above suggests that MEC receives visual information from a variety of sources, which could then potentially support the computation of self-motion estimates by MEC neurons. Consistent with this idea, theoretical work has demonstrated that it is possible to extract sufficient linear and angular velocity information from behaviorally realistic optic flow fields to support grid cell path integration (Raudies et al. 2012). This work also demonstrated that the distance and direction of environmental boundaries could be accurately estimated from these flow fields (Raudies and Hasselmo 2012). In support of these theoretical proposals, experimental work examining MEC coding in rats navigating in virtual reality demonstrated that border cells, which fire at a high rate near environmental boundaries (Savelli et al. 2008; Solstad et al. 2008), can be driven by visual cues alone (Aronov and Tank 2014), and multiple studies have shown that rodents can integrate optic flow to estimate distance traveled (Campbell et al. 2018; Kautzky and Thurley 2016).

Recent experimental work in mice also strongly suggests that visual inputs are necessary for the full expression of MEC coding properties. In complete darkness, with olfactory, auditory, and tactile cues carefully eliminated, the firing patterns of grid cells rapidly disorganized (Chen et al. 2016; Pérez-Escobar et al. 2016). These findings conflict with earlier reports of grid cells persisting in darkness (Hafting et al. 2005). The disagreement could reflect species differences between mice and rats in the relative influence of landmark versus self-motion cues on grid cells or in the ability to use nonvisual cues as landmarks. More thorough attempts at quantifying the influence of visual cues on rat grid cells are needed to fully resolve this discrepancy. Moreover, although the above data support a key role for visual input in generating MEC grid cell firing patterns, visual information alone appears insufficient to drive grid cell representations, as passive transport through the environment disrupts grid cells in rats, even when visual information remains intact (Winter et al. 2015b).

Darkness also strongly influenced MEC speed signals (Chen et al. 2016; Pérez-Escobar et al. 2016). Although some speed tuning remained intact, speed cell firing rates significantly decreased, as did the slope of firing rate with respect to running speed. Somewhat surprisingly, the frequency and slope of theta oscillations also decreased (Chen et al. 2016). This unequivocally demonstrates that visual input contributes to MEC speed coding. At the same time, the residual speed coding in darkness shows that locomotor input is also important and that visual cues combine with idiothetic cues to generate the speed signal in MEC.

HOW VISUAL INPUT COMBINES WITH IDIOTHETIC CUES TO GENERATE SPATIAL MAPS

The above data strongly support the multimodal nature of the inputs that generate path integration-based position estimates in MEC. Recent studies have examined the way in which these inputs combine in place and grid cells using manipulations of visual virtual reality environments (Campbell et al. 2018; Chen et al. 2013). Chen et al. (2013) found that 80% of place fields were altered by the removal of visual cues, and 25% of fields could be driven by visual cues alone. In addition, when the experimenters reduced the gain of the transformation between the animal’s running speed and movement of the visual cues, place fields shifted toward the beginning of the virtual track. The authors emphasized the nonlinearity of this transformation, as place field shifts did not grow with distance since the start of the track. Campbell et al. (2018) tested a number of gain change values and found that, depending on the magnitude of the gain change, grid fields either shifted or remapped. These nonlinear dynamics were well captured by a model in which landmark inputs correct path integration errors up until a critical threshold. Furthermore, Campbell et al. (2018) found that visual and locomotor speed inputs influenced both MEC speed signals and behavioral path integration estimates in an asymmetrical manner, where the response depended on which of the two inputs was larger. That is, when the gain was increased (visual cues move faster), both speed cells and behavioral distance estimates responded more to visual cues and vice versa. Together, these studies support the view that spatial maps in the hippocampus and entorhinal cortex are driven by a combination of landmark and self-motion inputs, each of which can derive from multiple sensory modalities, and elucidate the dynamics of this integration process. However, they suffer from the drawbacks of head-fixed virtual reality, which eliminates vestibular input and limits the movement of the animal. Further work is needed to extend these findings to situations in which the animal is freely moving and the full complement of sensory and idiothetic input is intact.

DIRECTIONAL CODING IN THE PARAHIPPOCAMPUS

To calculate self-motion when freely moving, an animal needs access not only to speed information but also information regarding its direction of movement. In MEC, directional input comes from the head direction system. Although this system has been extensively reviewed previously, we provide a brief overview of this input here (Clark and Taube 2012; Taube 2007).

THE HEAD DIRECTION SYSTEM OUTSIDE MEC

The head direction signal follows a hierarchical organization that is hypothesized to originate via attractor network circuitry in the dorsal tegmental nucleus and lateral mammillary nuclei (Clark and Taube 2012). These deep brain structures then project to the postsubiculum via the anterior thalamic nucleus (ATN), and the postsubiculum sends input to MEC (Fig. 5) (Taube 2007). Head direction cells require vestibular input (Muir et al. 2009; Stackman et al. 2002; Stackman and Taube 1997) but align to visual landmark cues when these cues are rotated (Taube et al. 1990a; Yoder et al. 2011), with the integration of visual landmarks into the head direction system requiring feedback to the lateral mammillary nucleus from the postsubiculum (Fig. 5) (Calton et al. 2003; Clark and Taube 2012; Yoder et al. 2015). Despite the strong control visual landmarks can exert upon head direction cell responses, visual and motor cues are insufficient to re-establish head direction responses after vestibular lesions, even after many weeks (Stackman and Taube 1997). This demonstrates that visual and motor cues cannot provide the angular velocity information needed to maintain directionally selective firing, although they could do so in theory. Importantly, this does not rule out the possibility that heading-selective responses to optic flow fields, which exist in primates (Duffy and Wurtz 1995; 1991), might be found in the mouse visual system, a potentially interesting line for future research.

Fig. 5.

Overview of the head direction input to medial entorhinal cortex (MEC). A. images from the Allen Brain Explorer showing the anatomic locations of the anterior-dorsal nucleus of thalamus (ADN), the postsubiculum (POS), and the MEC. B: a schematic of the head direction system, adapted from Taube (2007) with permission. The ADN provides head direction input to POS. This information is fed forward to MEC. Feedback from POS to the lateral mammillary nuclei is critical for the influence of visual landmarks on head direction cells (Calton et al. 2003; Yoder et al. 2015). The attractor circuitry that generates the head direction signal is thought to reside in the lateral mammillary nuclei and dorsal tegmental nucleus, which receive vestibular input from the brainstem (Clark and Taube 2012; Taube 2007).

HEAD DIRECTION CELLS IN MEC

In MEC, head direction cells are thought to provide an allocentric direction signal to grid cell position estimates. Supporting this idea, lesions or inactivation of the ATN profoundly disrupt grid cell periodicity (Winter et al. 2015a). However, the degree to which cells in MEC with head direction selectivity serve as a functionally dedicated and discrete cell class remains somewhat unclear. Although some pure head direction cells exist in the MEC (Sargolini et al. 2006), several studies instead raise the possibility that the majority of head direction signals in MEC show mixed selectivity for both directional and position coding. Head direction signals can be conjunctive with grid cell firing patterns, particularly in layers III and V of MEC (Sargolini et al. 2006), and work using a model-based approach to examine directional coding reported a high degree of mixed head direction-position selectivity in the superficial layers of MEC (Hardcastle et al. 2017). In addition, hippocampal manipulations that severely disrupt grid cell firing patterns reveal clear head direction tuning in the same neurons, possibly indicating that grid cells receive head direction input that is usually masked by stronger spatial firing correlates (Bonnevie et al. 2013). Taken together, these data support the notion that head direction signals provide the directional information used in path integration calculations. This does not rule out a role for other very recently discovered directional signals in providing a complementary signal for self-motion and position estimates. These cells include neural populations in the subiculum tuned to axis-of-motion and, during navigation through linear routes, bidirectionally tuned neurons in the dysgranular retrosplenial cortex (Jacob et al. 2017a; Olson et al. 2017).

HEAD DIRECTION VERSUS MOVEMENT DIRECTION

As reviewed above, many signs point to head direction cells as providing the angular component of path integration to grid cells. However, there is an important limitation to this view. As was recently pointed out by Raudies et al. (2015), it is movement direction and not head direction that is relevant for path integration calculations. Nevertheless, it is head direction and not movement direction that is encoded most strongly by entorhinal neurons (Raudies et al. 2015). In the authors’ hands, grid cell models using real rat trajectories could not produce grid cells when head direction was used instead of movement direction in the velocity input. It remains to be seen whether adding landmark input to these models could correct the errors introduced by using head direction instead of movement direction. Nevertheless, this is an important caveat that still needs to be resolved (Fig. 6).

CONCLUSIONS AND FUTURE DIRECTIONS

Self-motion consists of both linear and angular velocity, with the three primary cue types coming from the motor, visual, and vestibular systems (Fig. 1). Pioneering work on the perception of motion in the primate visual system, using a combination of anatomy, electrophysiology, stimulation, inactivation, and sophisticated behavioral experiments, provided a gold standard for current and future work on self-motion processing in rodents. Work in the rodent visual system and its role in representing cues related to self-motion has also gained momentum in recent years. Multiple extrastriate visual areas in the mouse are specialized for specific parameters of visual motion (Fig. 2), and recently discovered coding properties in secondary rodent visual areas mark these regions as potentially critical to self-motion calculations. One attractive candidate is area AL, which contains neurons that are effectively driven by fast-moving stimuli that occur during running. These visual signals influence self-motion coding in MEC (Fig. 4), where path integration calculations using self-motion signals are thought to occur. Self-motion input to the MEC also comes from the medial septum and head direction system (Figs. 3 and 5) as well as other sensory, proprioceptive and motor efference signals not discussed in this review (Chorev et al. 2016; Lackner and DiZio 2005; Mao et al. 2011). Together, these signals combine in a manner not yet fully understood, generating a unified self-motion percept and position estimate encoded in the spatially periodic firing patterns of grid cells (Fig. 6).

Despite these advances, many questions remain unresolved. The possibility remains that there are other, currently unknown, information sources for MEC self-motion and position estimates, and further work is needed on the routes by which visual self-motion information reaches MEC. Another major unsolved problem is the source of speed input to the grid cell population (Fig. 6). Arguably, the ideal speed input would linearly increase its firing rate with running speed. However, MEC speed representations are highly heterogeneous (Hardcastle et al. 2017; Hinman et al. 2016; Kropff et al. 2015). This poses a potential challenge for attractor-based network models of grid cell formation, which typically utilize a linear speed input to coherently drive the translation of the grid pattern (Burak and Fiete 2009; Bush and Burgess 2014; Pastoll et al. 2013; Si et al. 2014). It remains unknown whether this class of models could effectively leverage nonlinear speed signals to translate the network level activity patterns proposed to generate grid cell firing patterns. Because many of the linearly tuned MEC speed cells are fast-spiking interneurons (Kropff et al. 2015; Pérez-Escobar et al. 2016), and MEC stellate cells do not connect directly to each other but rather via inhibitory interneurons (Fig. 3) (Couey et al. 2013; Fuchs et al. 2016; Pastoll et al. 2013), models based on excitatory-inhibitory interactions in which inhibitory interneurons receive speed-tuned input could partially resolve this issue (Pastoll et al. 2013; Shipston-Sharman et al. 2016).

However, in addition to the issues associated with nonlinearity, the heterogeneity and weakness of the speed signals in MEC pose issues for grid cell models. To derive a robust speed signal from these heterogeneous and weak inputs, grid cells would have to sum input from many cells, many of which are conjunctive for variables other than speed (Hardcastle et al. 2017). Alternatively, grid cells could selectively integrate input from only the most highly tuned speed cells. This places other demands on the network, namely for connections to selectively form between highly tuned speed cells and grid cells. As a final possibility, grid cells could receive speed input from outside the MEC. For example, the total activity of place cell inputs could provide speed information, as the number of place fields that are entered per unit time is roughly proportional to the animal’s movement speed. Direct evidence for this possibility is lacking, but reanalysis of data in which MEC cells were recorded while the hippocampus was inactivated could reveal whether hippocampal inputs influence speed tuning in MEC (Bonnevie et al. 2013). Recording many grid and speed cells simultaneously, which is now possible using high-density silicon probes, and examining correlations between speed signals and the position of grid cell firing fields could provide clues as to the relationship between these two cell classes (Jun et al. 2017). For example, if MEC speed signals drive grid cell path integration calculations, one would expect variations in the population speed signal to predict variations in grid field location from trial to trial.

A major open question is whether MEC grid or speed cells are critical for self-motion guided navigation or whether path integration is subserved by other mechanisms. The MEC as a whole, along with brain regions such as medial septum and parietal cortex, is necessary for accurate path integration involving a translational component (Jacob et al. 2017b; Parron and Save 2004). The angular component of path integration does not appear to require the MEC, however, since head direction cells in thalamus were unaffected by MEC lesions (Clark and Taube 2011). Moreover, it remains challenging to provide causal links between specific MEC cell types and path integration due to the lack of genetic markers for functionally defined MEC cell types and the intermingled nature of these functionally defined cell types (Sun et al. 2015; Tang et al. 2014). These features thus present major roadblocks to establishing a clear link between grid, speed, head direction, or border cells and behavior using recent neuroscience tools, such as cell type specific optogenetics, chemogenetics, or regionally selective electrical manipulations. However, alternative approaches have made significant headway in addressing the role of grid cells in path integration. For example, the neural spiking dynamics of grid cells during navigation are consistent with path integration, as grid patterns accumulate error that can be corrected by interactions with environmental landmarks (Hardcastle et al. 2015). In addition, mice lacking GluA1-containing AMPA receptors or NMDA glutamate receptors show impairments in path integration and have disrupted grid but intact MEC head direction and speed coding (Allen et al. 2014; Gil et al. 2018).

An alternative approach that could be used to assess the importance of these cells to self-motion guided behavior would be to correlate entorhinal cellular activity with perceived location in the absence of landmark cues. Although this approach provides only correlational evidence, it has been fruitful in studying the neural basis of visual perception (Britten et al. 1992; Liu and Newsome 2005). Virtual reality tasks could be useful here, as they allow many trials to be run in quick succession while precisely measuring the perceived location of the animal either via trained responses or spontaneous behavior. Although head fixation can impact vestibular signals, introducing a caveat to the use of head fixation often associated with virtual reality, recent technological designs allow the preservation of head-fixed conditions with the integration of two-dimensional (2D) head rotation (Chen et al. 2018). Both this approach and virtual reality setups in which the animal is not head fixed and can freely rotate its body result in intact 2D grid, head direction, and border and speed cell firing patterns and could serve as useful approaches for examining the connection between entorhinal activity patterns and behavior (Aronov and Tank 2014; Chen et al. 2018). More generally, as was shown in the primate visual system, well-designed behavioral tasks coupled with causal manipulations of brain activity will be crucial in revealing the links between neural activity and spatial perception.

Emerging technologies for recording large populations of neurons, such as high-density silicon probes, should also aid in untangling interactions between cell types (Fig. 6) (Jun et al. 2017). These types of approaches could be combined with virtual reality or conducted in freely moving animals. One advantage of large-scale recording is that the number of neuron pairs grows exponentially with the number of neurons recorded, greatly facilitating the analysis of interactions between neurons that code for different variables. Furthermore, neural codes (such as for position and speed) can be more accurately measured from neural populations than single neurons, allowing for better analysis of how these codes interact both within and across brain regions. For example, high-density probes could illuminate how speed and position coding relate along the dorsal-ventral axis of the entorhinal cortex by facilitating simultaneous recordings from grid cells across a large portion of this axis (Jun et al. 2017). A progressive decrease in the strength of velocity inputs to grid cells along the dorsal-ventral MEC axis would be consistent with the mechanism used by computational models to generate the increase in grid spatial scale observed along the dorsal-ventral MEC axis (Burak and Fiete 2009; Hafting et al. 2005; Maurer et al. 2005; McNaughton et al. 2006; O’Keefe and Burgess 2005). Large-scale recordings from multiple brain regions during behavioral tasks that require the integration of self-motion cues could thereby elucidate the mechanism by which these cues are integrated, the relationships between the critical cell types involved, and the extent to which these various signals support behavior.

GRANTS

This work was supported by funding from The New York Stem Cell Foundation, Whitehall Foundation, NIMH MH106475, a Young Investigator Award from the Office of Naval Research, the Simons Foundation, a Klingenstein-Simons Fellowship, and a James S. McDonnell Foundation Scholar award to L. M. Giocomo and an National Science Foundation Graduate Research Fellowship and Baxter Award to M. G. Campbell.

DISCLOSURES

No conflicts of interest, financial or otherwise, are declared by the authors.

AUTHOR CONTRIBUTIONS

M.G.C. and L.M.G. prepared figures; M.G.C. and L.M.G. drafted manuscript; M.G.C. and L.M.G. edited and revised manuscript; M.G.C. and L.M.G. approved final version of manuscript.