Report

Efficient cortical coding of 3D posture in freely behaving rats

See allHide authors and affiliations

Science  02 Nov 2018:
Vol. 362, Issue 6414, pp. 584-589
DOI: 10.1126/science.aau2013

Posture in the brain

Our understanding of the neural basis of motor control originates in studies of eye, hand, and arm movements in primates. Mimica et al. investigated neuronal representations of body postures in the posterior parietal and frontal motor cortices with three-dimensional tracking of freely moving rodents (see the Perspective by Chen). Both brain regions represented posture rather than movements and self-motion. Decoding the activity of neurons in the two regions accurately predicted an animal's posture.

Science, this issue p. 584; see also p. 520

Abstract

Animals constantly update their body posture to meet behavioral demands, but little is known about the neural signals on which this depends. We therefore tracked freely foraging rats in three dimensions while recording from the posterior parietal cortex (PPC) and the frontal motor cortex (M2), areas critical for movement planning and navigation. Both regions showed strong tuning to posture of the head, neck, and back, but signals for movement were much less dominant. Head and back representations were organized topographically across the PPC and M2, and more neurons represented postures that occurred less often. Simultaneous recordings across areas were sufficiently robust to decode ongoing behavior and showed that spiking in the PPC tended to precede that in M2. Both the PPC and M2 strongly represent posture by using a spatially organized, energetically efficient population code.

More than a century of clinical observations have implicated the posterior parietal cortex (PPC) and related networks as essential for maintaining awareness of the spatial configuration of the body, or “body schema” (1, 2). Consistent with this notion, neurophysiological investigations in head-fixed subjects have identified key roles for the PPC and frontal motor cortices in controlling the positioning of individual effectors, such as the eye, arm, or hand (310). Parallel studies in rodents have demonstrated ostensibly similar functions for the PPC and frontal motor cortex region M2 in spatial orienting (11), movement planning (1214), and navigation (1517), but the field still lacks a quantitative understanding of how the cortex represents posture in freely behaving individuals.

We therefore tracked the heads and backs of 11 rats in three dimensions while recording neural ensembles with dual microdrives targeting deep (>500 μm) layers of the PPC and M2, which exhibit thalamic, cortical, and subcortical connections similar to those of the PPC and premotor areas across mammals (1618). We recorded 729 well-isolated single units in the PPC and 808 units in M2 during 20-min foraging sessions in a 2-m octagonal arena (fig. S1 and movie S1).

By measuring Euler angles (pitch, azimuth, and roll) of the head, pitch and azimuthal flexion of the back, and neck elevation in an egocentric reference frame (Fig. 1A and methods), we found robust tuning curves for all postural features in the PPC and M2, with peak rates often >5 standard deviations (SD) from the shuffled distribution (Fig. 1, B to G). The majority of cells with tuning peaks exceeding the 99th percentile of the shuffled data (fig. S3, A and B) were stable across recording sessions (mean of 56.4% in the PPC and 57.8% in M2) (Fig. 1, B to G, and table S1). Postural tuning was also stable across light and dark sessions (fig. S4), indicating its independence from allocentric landmarks and visually oriented attention (18).

Fig. 1 The PPC and M2 show stable 1D tuning curves for postural features of the head, back, and neck.

(A) (Left) Schematic for the head and three markers along the back (methods and fig. S1). (Middle) Back pitch (blue arrow), neck elevation (orange arrow), and head pitch (red arrow) were calculated relative to the arena floor. Red spheres represent the markers along the back. (Right) Azimuths of the head (dark pink arrow) and back (blue arrow) were measured relative to the body axis vector, from the tail to the base of the neck. Head roll (light pink arrow) was calculated relative to the arena floor. (B) (Left) 1D tuning curves for PPC cells for head posture, measured in two open-field sessions, with the 95% CI for shuffled data shown in gray. (Right) Cumulative frequency curves for tuning stability for each feature (arrowheads mark the 95th percentile of the null distribution; detailed results are in table S1). (C) (Left) Tuning curves for back pitch (top) and azimuth (bottom). (Right) Across-session stability. (D) Same as (C) but for neck elevation. (E to G) Same as (B) to (D) but for M2.

Cells in the PPC and M2 frequently responded to conjunctive postures involving the head, back, or whole body (Fig. 2, A and C, and movies S2 to S6), prompting us to build a generalized linear model (GLM) (methods) to identify features best explaining neural activity. We utilized a forward-search procedure in which egocentric posture variables and their derivatives, as well as allocentric features including head direction, running direction, and spatial location, were added until the cross-validated model performance no longer improved significantly (19) (methods).

Fig. 2 The PPC and M2 are tuned to combinations of head, back, and neck positions.

(A) Example PPC cells tuned to combinations of head, back, and neck positions. Conjunctive representations produce single-firing fields in the 2D rate maps; maximal firing rates (in hertz) are indicated above each map (top). 3D animal models (bottom) depict postures to which cells were tuned. Cell 1 preferred whole-body flexion and head roll to the right; cell 2 fired during rearing, with firing driven by the interaction of head pitch with neck elevation. (B) Distribution of behavioral tuning in the PPC as determined by the GLM (see the color-coded legend and table S2 for detailed results). (C) Examples of postural tuning in M2 cells. Cell 3 (top right) fired when the head, back, and neck were raised vertically; cell 4 was tuned to leftward head roll and back flexion during sharp turns. (D) Distribution of coding properties for 808 M2 cells.

This approach indicated that the largest fractions of cells in the PPC (n = 237, 32.5% of 729 cells) and M2 (n = 316, 39.1% of 808 cells) were driven by postural features of the head, including interactions (e.g., between pitch and azimuth), conjunctions of head posture and neck height, and movement (Fig. 2, B and D). Substantial fractions of cells were also tuned to back posture or movement (n = 69, 9.5% in the PPC; n = 75, 9.3% in M2), as well as elevation or movement of the neck (n = 43, 5.9% in the PPC; n = 84, 10.4% in M2) (Fig. 2, B and D, and table S2). Smaller percentages of cells exhibited whole-body tuning, being driven by combinations of head, neck, and back angles [n = 29, 4.0%, Z = 7.9, P < 0.001 in the PPC (large-sample binomial test with expected null probability P0 of 0.01); n = 27, 3.3%, Z = 6.5, P < 0.001 in M2].

Running speed (n = 38 cells, 5.2% in the PPC; n = 26 cells, 3.2% in M2) and self-motion (n = 15 cells, 2.1%, Z = 2.7, P < 0.01 in the PPC; n = 4 cells, 0.5%, Z = −1.3, P > 0.95 in M2) (Fig. 2, B and D; fig. S5; and table S2) accounted for considerably less of the population than posture (fig. S5). Weaker still were allocentric signals, including head and running direction (Z = −0.67, P > 0.85 in the PPC; Z = −0.56, P > 0.81 in M2) and spatial location (Z = −2.16, P > 0.99 in the PPC; Z = 0.86, P > 0.19 in M2), which did not reach significance in either area (Fig. 2, B and D, and table S2).

The statistical model indicated that the main features driving cells in the PPC and M2 related to posture (46.2% in the PPC; 58.7% in M2) as opposed to movement (5.6% in the PPC; 3.6% in M2). We tested this further by splitting recording sessions on the basis of movement velocity or posture and found that tuning curves for posture remained virtually identical regardless of movement status, whereas tuning to movement varied unreliably when split by posture (fig. S6). Postural tuning was thus expressed independently of movement, but not vice versa.

Previous studies showed anatomical organization for body and facial movement in parietal and motor areas in various mammalian species (2023), so we assessed whether postural tuning was also topographical. Head representation in M2 was concentrated at anterior [χ2(4) = 57.1, P < 0.001; Yates corrected χ2 test] (Fig. 3A) and medial [χ2(4) = 110.6, P < 0.001] locations, whereas back posture predominated at the posterior [χ2(4) = 98.1, P < 0.001] and lateral [χ2(4) = 105, P < 0.001] poles (Fig. 3, A and B). In the PPC, anteromedial sites adjacent to M2 showed the strongest back tuning [χ2(3) = 29.9, P < 0.001, anterior-posterior gradient; χ2(4) = 12.5, P < 0.05, medial-lateral], whereas posterior-lateral regions responded primarily to head posture [χ2(4) = 47.5, P < 0.001, anterior-posterior; χ2(4) = 52.4, P < 0.001, medial-lateral], producing a coarse mirroring of head and back representation across the PPC and M2 (Fig. 3A).

Fig. 3 Head and back posture were organized topographically across the PPC and M2, and the PPC led cross correlations between areas.

(A) Dorsal view of the cortex with boundaries delineating primary and secondary motor cortices (M) and somatosensory (S), retrosplenial (R), posterior parietal (P), and visual (V) cortices. The magnified view (right) shows recording locations (gray dots; 41 sites in M2, 40 sites in the PPC), and shading indicates tuning for the head (red) and back (blue). The black dot represents the bregma, with distance marked in millimeters. (B) Percentages of cells in M2 (top) and the PPC (bottom) driven by head and back positions. For all comparisons, the actual distribution of tuning differed significantly from theoretical distributions that assumed a constant proportion of tuned cells across bins. (C) Four cell pairs in the PPC and M2 showing stable z-scored cross correlations, with the PPC preceding M2. Dashed and solid lines represent the temporal offset of the cross-correlation peaks and time zero, respectively. Gray-shaded areas indicate ±6 SD of the shuffled data. (D) The normalized cross correlation for all cell pairs shows a negative peak for the PPC relative to M2 for positive and negative correlations. Shading indicates the 99% CI.

Because the PPC and frontal motor cortices form an extended network supporting spatial movement planning and decision-making (2427), we asked whether structured correlations existed between spikes recorded simultaneously across areas (n = 5 rats). We screened for cells with significant interregional signal correlations (methods) and identified 1017 positively and 182 negatively correlated pairs in one recording session (n = 15 sessions) and 758 positively and 141 negatively correlated pairs in a second session (n = 14 sessions) the same day. The average normalized positive cross correlations indicated a consistent peak, with the PPC preceding M2 by 50 ms across sessions [−72 ms, −32 ms, bootstrapped 99% confidence interval (CI) for session one; −71 ms, −31 ms for session two] and negative correlations peaking at −85 ms [−190 ms, +16 ms] in session one and −25 ms [−140 ms, +82 ms] in session two (Fig. 3, C and D).

To next address whether population activity was sufficient to reconstruct behavior, we reduced the behavioral dataset from six dimensions (three axes for the head, two for the back, and one for neck elevation) to two by using Isomap (28). This rendered posture for the head, back, and neck on a two-dimensional (2D) surface, or “posture map,” with each pixel corresponding to a particular bodily configuration (fig. S7). We chose a session with 37 PPC and 22 M2 neurons recorded simultaneously to train a uniform prior decoder to predict the animal’s dynamic position on the posture map on withheld segments (Fig. 4A and movies S7 and S8). Decoder performance, on average, exceeded the shuffled distribution by >45 SD (Fig. 4B).

Fig. 4 Ensemble decoding of posture in the PPC and M2 reveals a nonuniform distribution of tuning.

(A) (Top left) Four snapshots taken within 1 s as the animal came down from rearing and bent rightward. (Top right) Corresponding posture maps illustrate the log posterior distribution of the animal’s posture estimated by using PPC and M2 cell ensembles. Actual posture is marked with a green “O,” and the maximum likelihood is color coded from yellow to black. (Middle) Timeline indicating error, in Isomap pixels (pix.), over a 20-min recording. (Bottom) Five examples of distinct poses (time points are listed below) and attendant Isomaps illustrating real and decoded postures. (B) Decoder accuracy as a function of cell sample size (red dots), with the null distribution above (black dots). The shaded area indicates ±3 SD. (C) (Left) Cumulative occupancy on the Isomap showed that the longest dwell times were in the low center of the map, corresponding to foraging. The dashed oval delineates the high-occupancy area where the animal spent >50% of the session. (Right) The average decoding error on the Isomap was smaller in the low center of the map than elsewhere. The dashed oval is the same as in the left panel. (D) (Left) The percentage of cells representing the six postural features (black dots) was significantly higher for low- versus high-occupancy bins in the decoding session. (Right) The same analysis for all animals in the study. (E) Decoder error was below chance for low- and high-occupancy regions of the posture map and was significantly smaller for the high- than for the low-occupancy area. The line near the top of the graph indicates the mean ± 3 SD for shuffled data. Bar graphs in (D) and (E) indicate mean ± SEM. **P < 0.01; ***P < 0.001.

Cumulative occupancy on the posture map was dominated by epochs when the animal was on all fours with its head lowered (i.e., foraging) (Fig. 4C, left). We found significantly fewer cells tuned to these high-occupancy, or “default,” postures (Fig. 4C, dashed oval), whereas less-visited postures were represented more densely by the ensemble (t10 = 4.82, P < 0.01, Welch’s two-sided t test) (Fig. 4D, left). The same pattern was observed across animals (t10 = 7.74, P < 0.001) (Fig. 4D, right, and fig. S8), suggesting that receptive fields were distributed on the basis of occupancy. Despite this anisotropy in representation, decoder performance was significantly better than chance for all postures, with smaller error for high- than for low-occupancy postures (t74 = 6.21, P < 0.001) (Fig. 4E).

The finding that cell populations in parietal and frontal motor cortices represent 3D posture robustly and in larger proportions than other behavioral features complements and extends decades of study on the positional coding of single effectors in stationary animals [e.g., (5, 2931)]. The predominance of postural tuning in our data may reflect the myriad kinematic computations that must be solved to coordinate whole-body movement during free behavior. It is also consistent with a functional division of labor in which higher cortical areas specify body position and goals (3234) whereas descending motor pathways and subcortical nuclei control movement dynamics more directly (3539).

The topographical distribution of postural tuning for the head and back appeared to follow a functional organization identified in earlier microstimulation studies in anaesthetized animals (22, 23). We found mainly head and back representation in the PPC and M2, but it is possible that posture for the entire body overlays the cortical surface, including primary somatosensory and motor cortices. More broadly, it remains to be established whether postural signals are generated in the cortex specifically or whether they are inherited from other regions. Our cross-correlation analyses also suggest that a temporal structure exists for postural representations across areas, with the PPC operating upstream from M2, though such an ordering could shift in the context of different tasks (26).

Our use of 3D tracking additionally revealed that speed and self-motion tuning in the PPC (13, 14, 40) were likely overestimated in previous studies using 2D tracking of rodents, owing to insufficient resolution to disambiguate posture from movement. Tracking the back allowed us to detect neural tuning to flexion of the trunk, indicating that vestibular signaling (41) alone could not explain the postural coding in our recordings. For both the back and the head, the arrangement of postural tuning peaks was notably nonuniform and appeared to be optimized for the duration for which postures were occupied (Fig. 4D and figs. S3, A and B, and S8). Previous theoretical works considering optimal coding strategies in sensory systems (42, 43) suggested that the range of the stimulus spectrum visited most should be encoded by more cells with narrower tuning widths, but this was not the case in our data. Rather, we found that proportionately less of the network was dedicated to default states in which the animals spent more time. This arrangement allowed high-fidelity decoding of the entire range of postures while minimizing metabolic demand on the cells, making it both precise and efficient. Together, our results strongly support the notion that the PPC-M2 network plays a key role in representing the dynamic organization of the body in space, or body schema, postulated more than 100 years ago (1, 2).

Supplementary Materials

www.sciencemag.org/content/362/6414/584/suppl/DC1

Materials and Methods

Figs. S1 to S8

Tables S1 and S2

References (4551)

Movies S1 to S8

References and Notes

Acknowledgments: We thank B. McNaughton and E. Moser for helpful comments on the manuscript; K. Haugen, K. Jenssen, E. Kråkvik, H. Obenhaus, R. Gardner, T. Feyissa, K. Hovde, H. Kleven, M. Gianatti, and H. Waade for technical and IT assistance; J. Jeon, J. Adams, and NeuroNexus for assistance in drive design; G. Olsen and M. Witter for assistance with anatomical delineations; and S. Eggen for veterinary oversight. Funding: This study was supported by research grants from the European Research Council (“RAT MIRROR CELL,” starting grant agreement 335328), the Research Council of Norway (FRIPRO Young Research Talents, grant agreement 239963), the Kavli Foundation, and the Center of Excellence scheme of the Research Council of Norway (Center for Neural Computation). Author contributions: J.R.W. and B.A.D. designed the experiments. B.M. and T.T. conducted the experiments. B.A.D. designed the analyses. B.A.D., B.M., V.P.T.N.C.S.B., and J.R.W. performed the analyses. J.R.W. and B.M. wrote the paper with assistance from T.T. and B.A.D. Competing interests: No competing interests declared. Data and materials availability: Datasets validating the main findings and conclusions of the paper, as well as the 3D tracking graphical user interface and support folders, are available in the Norwegian national research data archive (44).
View Abstract

Stay Connected to Science

Navigate This Article