Specific patterns of neural activity in the hippocampus after massed or distributed spatial training

Training with long inter-session intervals, termed distributed training, has long been known to be superior to training with short intervals, termed massed training. In the present study we compared c-Fos expression after massed and distributed training protocols in the Morris water maze to outline possible differences in the learning-induced pattern of neural activation in the dorsal CA1 in the two training conditions. The results demonstrate that training and time lags between learning opportunities had an impact on the pattern of neuronal activity in the dorsal CA1. Mice trained with the distributed protocol showed sustained neuronal activity in the postero-distal component of the dorsal CA1. In parallel, in trained mice we found more active cells that tended to constitute spatially restricted clusters, whose degree increased with the increase in the time lags between learning trials. Moreover, activated cell assemblies demonstrated increased stability in their spatial organization after distributed as compared to massed training or control condition. Finally, using a machine learning algorithm we found that differences in the number of c-Fos positive cells and their location in the dorsal CA1 could be predictive of the training protocol used. These results suggest that the topographic organization and the spatial location of learning activated cell assemblies might be critical to promote the increased stability of the memory trace induced by distributed training.

In this framework we asked whether and how changes in the frequency of training in the spatial version of the Morris water maze (sMWM) might affect cell activity, focusing on the dorsal CA1 (dCA1) of the hippocampus (HPC), with the working hypothesis that different patterns could be observed after massed and distributed learning experiences.To this aim we compared mice trained with a massed protocol with mice trained with a distributed protocol that has been shown to induce a more stable spatial memory 9 .In the massed group, six training sessions were administered consecutively with an intersession interval of 10 to 15 min.The distributed training procedure consisted of the same number of sessions but distributed over three consecutive days, with a 4 h within-day interval.In these training conditions both groups show similar performance when tested 24 h after the last training session but only mice trained in the distributed protocol are able to recall the correct platform location when tested 14 days after training 9 .To assess changes in the pattern of cell activity we focused on the dCA1 of the HPC for its critical role in spatial memory 25 but also because it has been shown to exhibit a high functional heterogeneity [26][27][28][29][30][31] .c-Fos expression in the dCA1 was assessed after learning with the two training protocols, comparing differences in the pattern and stability of the topographical organization of labeled neurons.The analysis demonstrated different regional distribution of c-Fos positive cells in massed and distributed trained mice.Moreover, the findings support the view that distributed training promotes increased clustering and topographical stability in learning activated cells.
A thigmotactic strategy is often deployed by animals in the initial stages of the acquisition as an alternative to a spatial strategy 33 therefore the comparison of the thigmotactic behavior in S1 and S6 in the two experimental groups was used to assess increased proficiency at this spatial task.The statistical analysis revealed a significant difference between S1 and S6 for both groups (mMWM: t (5) = 2.301, p = 0.0349; dMWM t (6) = 2.478, p = 0.0240, paired t-test), demonstrating that with training progression mice reduced the time spent doing thigmotaxis (Fig. 1d).
Learning induced c-Fos activity in the dCA1 shows subregional specificity depending on the training protocol.To investigate changes of c-Fos expression after massed and distributed training in the MWM we focused on the CA1 of the dorsal hippocampus, because it has been shown to be critical for spatial MWM acquisition 25 .To perform a spatially detailed count of c-Fos positive cells from the pyramidal cell layer, we draw up to 12 contiguous regions of interest (ROIs) on each section of c-Fos-stained dCA1, as shown in Fig. 2a,b.The mean number of c-Fos positive nuclei per ROI in the dCA1 was higher in trained mice (mMWM: 26.6 ± 5.6; dMWM: 29.3 ± 3.0) compared to controls (20.4 ± 2.4), although the difference did not reach statistical significance (one-way ANOVA: F (2,24) = 1.988, p = 0.158).Next, we analyzed the pattern of expression of c-Fos positive nuclei along the antero-posterior and proximo-distal axes of dCA1, in light of their significant functional heterogeneity [26][27][28][29][30] (Fig. 2a,b).To this aim, we generated 3D heatmaps reproducing the number of c-Fos positive cells in each ROI within the dCA1.As shown in Fig. 3a control mice exhibited an overall lower number of activated cells, compared to the massed and the distributed groups, and a fairly homogenous distribution throughout the dCA1.On the contrary, the number of c-Fos positive cells was not evenly distributed in mice trained in the MWM.Moreover, the analysis revealed regional differences depending on the intertrial interval used to train the mice.In particular, the massed trained mice showed higher level of expression in the medioproximal portion of dCA1 (Fig. 3b) while mice trained with the distributed protocol revealed higher levels in the postero-distal component of the dCA1 (Fig. 3c).To deepen our analysis and quantitatively compare c-Fos activity in mice trained with the two protocols we grouped the ROIs along the antero-posterior and proximodistal axes (see "Methods").The one-way ANOVA revealed significant differences in the mean c- training protocol in the postero-central and postero-distal compartments when compared to controls (p < 0.01, Tukey HSD) (Fig. 3d).Overall, these findings confirm a functional heterogeneity in the dCA1, demonstrating that the regional pattern of neuronal activity in the dCA1 is sensitive to the time intervals between training sessions.
Topology of neuronal activity in the dCA1 after massed and distributed training.Next, we verified whether the training protocol could also have affected the topological organization of c-Fos positive cells.As shown in Fig. 4a, the classification of the ROIs based on their c-Fos positive cell number revealed a higher frequency of ROIs in the intervals between 41 and 80 cells in the distributed compared to control group (*p ≤ 0.05; **p ≤ 0.01; Mann-Whitney) and in the interval 31-40 in the distributed compared to massed group (#p ≤ 0.05; Mann-Whitney).Additionally, using a custom made MATLAB script (see "Methods"), we identified, among the ROIs with a high cell count (c-Fos + cells ≥ 40), only those that neighbor (within 5 Euclidean distance) at least five other high-count ROIs, showing that in trained mice these ROIs were in greater numbers (18.83 ± 11.12 massed group; 17.2 ± 4.5 for distributed and 4.9 ± 2.7 for control) (Fig. 4b-d), with a significant difference between the distributed and the control groups (t (19) = 2.494, p = 0.022 dMWM vs. controls; t (18) = 1.702, p = 0.106, mMWM vs. controls; t (11) = 0.14; p = 0.89 dMWM vs. mMWM).These results suggest that the distributed training protocol, more efficient in optimizing memory, promotes the clustering of learning-induced active cells within the dCA1.
Training-induced unique c-Fos expression pattern.Using a custom-made MATLAB script (see "Methods") we analyzed the activation pattern in the massed and in the distributed groups compared to controls.Considering the mean of c-Fos positive cells per ROI in controls (c-Fos + = 23) as a threshold for all the three groups, we searched for ROIs activated over the threshold with the same topological position in at least 5 out of 6 animals in the massed, 6 out of 7 in the distributed and 13 out of 14 in the control group.We could not find any above-threshold ROI sharing the same position in the animals of the control group (Fig. 5a), this held true even considering shared ROIs in only 5 out of 14 mice of the control group.Conversely this analysis revealed that 4 above-threshold ROIs and 11 above-threshold ROIs shared the same position in in the massed and the distributed group, respectively (Fig. 5b,c).Coherently with the regional activation induced by the training protocols, these common ROIs were specifically localized in the medial compartment of dCA1 for the massed and in the posterior portion of dCA1 for the distributed group, as shown in the 3D reconstruction in Fig. 5.
These results led us to ask whether these characteristic activation patterns could be robust enough to predict the training procedure.We computed accuracy using a Quadratic Discriminant (QD) classifier, a supervised machine-learning algorithm for data classification.When trained with all the variables (see "Methods"), the algorithm showed a remarkable capability to discriminate between massed and distributed trained animals (accuracy = 78%).Notably, when we repeated the procedure adding the control group the level of accuracy decreased (accuracy = 41.4%),pointing out how the absence of a specific c-Fos expression pattern in the control group reduced the discrimination ability of the algorithm.The analysis of feature importance highlighted that the most informative variable was the antero-posterior coordinate (section number) confirming that subregional differences in the level of c-Fos activity in the trained animals was a critical feature to distinguish the groups (Fig. 6).

Discussion
In the present study we contrasted c-Fos expression after massed and distributed training in the spatial version of the MWM to outline possible differences in the learning induced pattern of neural activation in the dCA1 of the HPC in the two training conditions.Three main results were obtained: 1. Distinct regional distribution of c-Fos positive cells after massed and distributed training; 2. Increased clustering of cell activity in mice trained with the distributed as compared to mice trained with the massed protocol; 3. Increased stability across animals of spatially patterned organization of distributed training induced active cells assemblies.Moreover, using a Quadratic Discriminant algorithm we found that differences in the number of c-Fos positive cells and their location in the dCA1 could be predictive of the training protocol used.It is well known for many types of learning that distributed training, involving longer inter-trial intervals, leads to a more robust memory formation as compared to massed training 2,5,6,9 .In the present study we compared massed and distributed training in the spatial version of the MWM.Two different training protocols were used, a within day protocol in which the six training sessions are massed with a 5-10 min inter-trial interval, and an across-three-days protocol that has been recently demonstrated to increase memory stability 9 , reproducing the increased memory performance observed in humans when inter-repetition lags increase 11 .Independently on the training protocol, mice of the two groups reduced the distance and the time to reach the platform.Consistently with previous evidence 5,8,9,32,34,35 distributed training did not affect learning, as there was no difference in terms of the total distance moved and the latency to find the platform as the sessions progressed.These finding are in agreement with the suggestion that the ability of distributed training to increase memory cannot be attributed to an effect of the training protocol on learning, but rather on the consolidation process [9][10][11] .
Although we could not find any behavioral differences in the acquisition phase, marked differences could be detected in the pattern of c-Fos expression in the dCA1 of the HPC depending on the training protocol.Analysis of c-Fos labeled cells did not highlight significant differences in the overall number of active cells between the two experimental groups, however we found a change in the regional distribution of c-Fos labeled cells.Our analysis revealed that increasing lags between learning opportunities changes the pattern of activity over the proximo-distal axis as well as over the antero-posterior axis of dCA1.In massed trained mice, in fact, we observed a higher number of c-Fos expressing cells in the medio-proximal component of the CA1, while distributed training induced activity was found mainly in the postero-distal domain of the CA1.Although the hippocampal networks sustaining memory within the CA1 are traditionally viewed as constituted by homogeneous populations of cells, heterogeneity has been demonstrated both at a structural and a functional level [36][37][38] .The structural heterogeneity is well established based on the differential projections from the entorhinal cortex, with the lateral part projecting preferentially to the distal component of the CA1 and the medial part projecting mainly to the proximal CA1 38,39 .The cellular compactness and width of the deep and superficial CA1 pyramidal cells sublayers also show a regionalization, decreasing and increasing respectively along the proximo-distal axis 26 .This structural heterogeneity is paralleled by functional differences on the proximo-distal axis.Relevant to us is the finding that cells in the distal component display a higher place field accuracy compared to those in the proximal component 28 .Such a functional gradient is paralleled by an opposite radial gradient in the place field stability 27 and in the recruitment of CA1 pyramidal neurons during HPC sharp wave ripples 40 , suggesting a higher dynamicity and accuracy of cells in the proximal CA1 as compared to those in the distal component that would show a coarse representation of the environment but with a higher steadiness 26 .Interestingly the different pattern of activity in mice trained with the massed or the distributed protocols suggest a progressive shift along the proximo-distal axis with the increased time lag between learning opportunities.The proximal component is indeed activated after massed training, whereases the distal is activated after a distributed training (Fig. 3).This result would be consistent with behavioral evidence in rodents and humans that memory representation after distributed learning is more stable than after massed learning 9,11 .
The second interesting observation raised by the quantification of c-Fos labeling was that the topographic distribution of learning activated cells in the dCA1 varies depending on the training protocols.In particular the analysis performed revealed a group difference in the number of ROIs expressing higher c-Fos labeling.This observation, together with the clustering of ROIs expressing a high number of c-Fos positive cells in the distributed groups, as compared to control mice or mice trained with the massed protocol, strongly suggests a training protocol dependent clustering of learning activated cells.Two different lines of evidence support the possibility of a learning induced clustering of cell activity.A handful of electrophysiological studies for example reported a cluster type organization of place cells, demonstrating that neighboring pyramidal cells in the CA1 www.nature.com/scientificreports/have overlapping place fields [41][42][43] , but see also 44 .More relevant are findings relative to the organization of IEGs labeled cells after training, demonstrating increased IEGs activity in adjacent cells, in the CA1, after exposure to a novel environment and tighter clusters when exposure to the same environment is repeated 45,46 .A similar learning dependent increase in clustering of active cells was observed also in the dorsal striatum as instrumental learning progresses 24 .Although a causal role of such clustering in memory formation has never been established, our findings support this hypothesis, by showing that it is more pronounced increasing the time interval between sessions when the stability of the memory trace is enhanced.Finally, we also found a difference in the stability of the topographic distribution of activated ROIs in the three experimental groups.A few studies addressed the question relative to the topography of learning activated cells in the HPC as well as in other brain regions [47][48][49] .The increased topographic stability of active cells assemblies we found comparing the two training protocols is in agreement with previous evidence demonstrating a spatially patterned organization of amygdala neurons activated by training in a fear conditioning task 47,48 as well as fMRI data reporting a stable topography of activation for specific memories in the human HPC 50,51 .Thus our data support the hypothesis that increased stability of the memory trace might depend on the spatial location of active cell assemblies within dCA1 47,51 .Finally, we were able to discriminate mice trained with the different training protocols based on a quadratic discriminant algorithm for data classification, using information relative to the pattern of expression of learning induced c-Fos positive cells.These findings suggest that optimization of memory might depend upon specific hippocampal neuronal population codes that vary depending on the stability of the memory trace.It should be mentioned, that we analyzed c-Fos expression in the dHPC after training in the spatial version of the MWM, however, increased IEGs expression in the hippocampal complex has been reported also after training in the cue version of the task, although a causal role of this activity in memory has been debated 52,53 .In the future, it would be very interesting to test whether the changes in cell ensemble activation we observed in the present study could be generalized to the training frequency rather than the specific information to be acquired.
It has been hypothesized that the memory stability induced by distributed training depends on a more efficient stabilization of the memory trace 10 .The evidence presented in this study suggests that the topographic organization and the spatial location of learning activated cell assemblies might be relevant to this process.Moreover, it supports the view that activity in the distal component of the dCA1 might be important in determining the stability of the representation.

Subjects.
The experiments were conducted on naïve CD1 male mice (Charles River, Italy).Mice were 8 weeks old, weighting 34-38 g, at the onset of the experiments.Animals were always housed in groups of three mice in standard cages (26.8 × 21.5 × 14.1 cm), with water and food ad libitum, under a 12 h light/dark cycle (7 a.m./7 p.m.) and constant temperature (22 ± 1 °C).Behavioral training and testing were conducted during the light period (from 9:00 am to 5:00 pm).The maximum effort was made to minimize animal suffering.Procedures were conducted under the authorization N. 450/2018-PR from Italian Ministry of Health, according to Italian (DL.116/92) and European laws and regulations on the use of animals in research, and NIH guidelines on animal care.

Morris water maze.
The apparatus consisted of a circular pool (110 cm diameter and 40 cm high) filled up to 5 cm from the edge with water at 22 ± 1 °C.To allow mice tracking, water was black-colored with non-toxic paint (Giotto, Italy).Black curtains surrounded the pool, and several visual cues were attached on them at the distance of round 50 cm from the pool.A white light (100W) was positioned above the apparatus facing the ceiling.Four additional lights (1.9 W) were positioned facing the cues to light them up.The pool was ideally divided into four identical quadrants.Four starting positions (labeled N, S, E, and W) were located equidistantly around the edge of the maze.A Plexiglas platform (10 cm diameter), covered with wire mesh to avoid slipping, was positioned 20 cm away from the pool wall in one of the four quadrants.The quadrant where the platform was kept during the training was randomized and was named the "Target" quadrant, whereas the other three quadrants were named "Right", "Left" and "Opposite", relative to the Target.
Mice of each group were handled for 3-4 min/day, for 5 consecutive days, before the beginning of the experiments.The spatial version of the Morris Water Maze task (MWM) consisted of two different phases: familiarization and training (Fig. 1a).The familiarization phase was the same for both the massed (mMWM) and the distributed (dMWM) protocols and consisted of one session of three consecutive trials (intertrial interval: 20 s).During the familiarization phase, no cues were attached on the curtains and the platform protruded 1.5 cm above the water surface.During the training phase, the platform was submerged 0.5 cm beneath the surface of the water and visual cues were attached on the black curtains surrounding the pool.The mMWM consisted of six consecutive training sessions (intersession interval: 10-15 min) of three trials (intertrial interval: 20 s), whereas dMWM consisted of six training sessions distributed over 3 days, with 2 sessions per day (intersession interval: 4 h) of three trials (intertrial interval: 20 s) 9 .In each trial, animals were randomly released from one of the three non-target quadrants.Mice in mMWM were kept in their holding cage for 1 h before the starting of the training sessions, whereas mice in dMWM were kept in their holding cage for 1 h before each training session.During intersession intervals, mMWM mice returned to their holding cage until the next session.Control animals were kept in their holding cage for the whole length of training sessions.
All trials were recorded by a camera located over the pool and videos were acquired and analyzed by an automated tracking system (AnyMaze 5.0, Stoelting).Escape latency (in seconds) and distance travelled to reach the platform (in meters) were scored.Thigmotaxis was defined as the swimming time within the zone closer to the pool wall (10 cm from pool wall).each animal was deeply anaesthetized with a mixture of Zoletil (500 mg/kg, Virbac Italia) and Xylazine (100 mg/ kg, Bayer, Germany) and transcardially perfused with 40 ml of saline solution (NaCl 0.9%) followed by 40 ml of 4% formaldehyde in PBS (4 °C).Brains were rapidly removed and post-fixed for 24 h in 4% formaldehyde and then transferred to 30% sucrose in PBS until sectioning.40 μm-coronal sections were obtained using a cryostat (Leica CM 1520, Leica Microsystems, Wetzlar, Germany) and were stored at − 20 °C in cryoprotectant solution.For the c-Fos-immunoreactivity (c-Fos-IR), free-floating sections were incubated for 5 min in 3% hydrogen peroxide in PBS and rinsed three times in PBS containing 0.1% Triton X-100 (PBST).After 1 h of incubation in PBST containing 1% BSA and 1% NGS sections were incubated overnight in anti-phospho-c-Fos rabbit monoclonal antibody (5348S; Cell signaling Technology, USA) diluted 1:8000 in PBST-BSA-NGS, at 4 °C and with constant orbital rotation.Sections were washed three times in PBST and incubated in biotinylated secondary antibody diluted 1:500 in PBST-1% BSA (goat anti-rabbit IgG; Vector Laboratories, USA) for 2 h at room temperature.After three washes in PBST, sections were transferred for 1 h in avidin-biotinylated peroxidase complex diluted 1:500 in PBST (ABC Kit; Vector Laboratories, USA) and rinsed three times in PBS.Finally, the reaction was visualized using nickel intensified diaminobenzidine (DAB peroxidase substrate kit, Vector Laboratories, USA).The reaction was stopped after exactly 4 min washing with 0.1 M PBS (pH 7.6).
Sections were mounted on slides, dehydrated through a graded series of alcohols, cleared and coverslipped.
Sections from groups to be directly compared were processed at the same time and using the same conditions and reagents to reduce variability.In all experiments, the number of cells displaying c-Fos immunoreactivity was measured in dCA1, taking twelve alternating brain sections (from − 1.22 mm to − 1.92 mm relative to bregma), encompassing the whole antero-posterior axis.Regions were defined according to the mouse brain atlas 54 .
Digital reconstruction and c-Fos quantification.Digital images were acquired at 10X magnification, using a microscope (Nikon Eclipse 80I) equipped with a CCD camera.After images acquisition, single images of 2560 × 1920 pixel were stitched together in a mosaic view of the entire dCA1 region and counting of the stained nuclei was carried out using the public domain software ImageJ (http:// rsb.info.nih.gov/ ij/).Briefly, for each section, stained nuclei were automatically detected based on their intensity of staining relative to background and their size.Detailed counting of c-Fos positive cells was performed drawing from 8 to 12 contiguous regions of interested (ROIs), covering the whole dCA1 on each section.Each ROI had a 200 µm width and ROIs were placed consecutively along the dCA1 proximo-distal axis (distal CA1: away from the dentate gyrus, close to the subiculum; proximal CA1close to dentate gyrus and CA2) 36,55 .The anchor point for the first ROI was always set at the beginning of the dCA1 along the midline of the brain, using the ventricle as the starting point marker.For each experimental subject, the frequency of ROIs in each interval was then calculated and reported as percentage of the total number of ROIs for that subject.Regional c-Fos expression was evaluated dividing the whole dCA1 in eight different subregions: antero-distal (AD), antero-central (AC), medio-distal (MD), medio-central (MC), medio-proximal (MP), postero-distal (PD), postero-central (PC), postero-proximal (PP).Each subregion includes 16 ROIs except for the medio-proximal compartment (10 ROIs), according to the dCA1 shape.

Analysis of c-Fos expression pattern.
Data obtained from digital image reconstruction and c-Fos quantification were exported in a tab-delimited spreadsheet and imported in MATLAB.MATLAB heatmap representation was used to obtain the 3D heatmap representation of c-Fos positive cells for each ROI and section (Fig. 3).Custom-made scripts (Mata and Matrixa) were used to identify the ROIs with a cell count equal or above 40 cells (threshold chosen based on the differences observed in the ROIs frequency analysis) and neighboring other ROIs (within 5 Euclidean distance) with the same cell count (Fig. 4 b-d).Finally custom-made scripts (Histcut_23; Pixels_m_23; Common_23) were used to highlight the ROIs with a cell count equal or above 23 (mean of c-Fos positive cells per ROI in control animals) and with a shared position within each experimental group (Fig. 5).All scripts are included as supplementary materials.

Machine learning based classification.
A MATLAB based library was used to train the classification model (https:// www.mathw orks.com/ help/ stats/ choose-a-class ifier.html).The model is a supervised machine learning algorithm called Quadratic Discriminant (QD) that is based on discriminant analysis learning.
We applied the QD classifier adopting the function cross_val_score with cv = 10 to perform tenfold crossvalidation (available in MATLAB R2022a).The classifier was trained using the information from each subject about the c-Fos positive cells number along the anterior-posterior axis (section number), c-Fos positive cells number along the proximo-distal axis (ROI number), total number of c-Fos positive cells from each hemisphere (cell count hemisphere 1; cell count hemisphere 2), total number of c-Fos positive cells averaged between the two hemispheres (cell count both hemispheres), the presence and the number of the ROIs above threshold (ROIs expressing c-Fos + cells ≥ 23), number of the ROIs above threshold adjacent to each other from hemisphere 1, 2 and the average between the two hemispheres (neighboring cell hemisphere 1; neighboring cell hemisphere 2; neighboring cell both hemispheres).To calculate feature importance in classifier decision, we used the attribute feature selection that expresses the fraction of relative importance for each feature.

Figure 1 .
Figure 1.Learning curves for mice trained in the spatial version of the MWM with massed or distributed training protocols.(a) Timeline of the two MWM training procedures.(b) The graph shows the escape latency to reach the platform during training in the MWM in the massed (grey; N = 6) and in the distributed (orange; N = 7) groups (**p ≤ 0.01, S1 vs. S6, massed and distributed, Tukey HSD).(c) The graph shows the distance travelled during training in the MWM in the massed and distributed groups (**p ≤ 0.01, S1 vs. S6, massed, Tukey HSD).(d) The graph shows the time spent in thigmotaxis in S1 and S6 for both massed and distributed groups (*p ≤ 0.05, S1 vs. S6, massed and distributed, paired t-test).Graphs represent mean ± SEM.

Figure 2 .
Figure 2. Schematic representation of dCA1 reconstruction.(a) Schematic representation of antero-posterior (dark green = posterior; green = medial; light green = anterior) and proximo-distal axes (dark pink = distal; pink = central; light pink = proximal) along dCA1.Numbers indicate anteroposterior coordinates relative to bregma (in mm, according to 54 ).(b) Representative posterior dCA1 sections from control, massed and distributed mice, stained for c-Fos.The boxes along the proximo-distal axis represent the position of the ROIs used to count c-Fos positive cells.

Figure 5 .
Figure 5. ROIs above threshold with shared position in the different experimental groups.(a) Controls, (b) massed and (c) distributed.Right panels, 3D brains reconstruction showing the localization of the shared abovethreshold ROIs (yellow ROIs) within the dCA1.

Figure 6 .
Figure 6.Machine-learning based classification of the training procedure.Histogram depicts the relative importance of the top 10 features (Section number, Neighboring cell hemisphere 2, Cell count hemisphere 1, ROI, Cell count hemisphere 2, ROIs ≥ 23cells, ROI number, Neighboring cell hemisphere 1, Neighboring cell both hemispheres, cell count both hemispheres) in the Quadratic Discriminant model (Kruskal-Wallis algorithm features importance scores).