Distributed subthreshold representation of sharp wave-ripples by hilar mossy cells
- Graduate School of Pharmaceutical Sciences, The University of Tokyo, Japan
- Laboratory for Systems Neurophysiology, RIKEN Center for Brain Science, Japan
- Laboratory for Neural Computation and Adaptation, RIKEN Center for Brain Science, Japan
- Department of Mathematical Informatics, Graduate School of Information Science and Technology, The University of Tokyo, Japan
- Institute of AI and Beyond, The University of Tokyo, Japan
- Center for Information and Neural Networks, National Institute of Information and Communications Technology, Japan
Figures
Figure 1 with 4 supplements
Spatiotemporal diversity of CA3 sharp wave-ripple (SWR)-induced ΔVms in mossy cells (MCs).
(A) (Top) Schematic illustrations of whole-cell Vm recordings from five MCs together with local field potential (LFP) recordings from the CA3 region. (Bottom) Recorded cells were confocally identified by immunostaining for intracellularly injected biocytin (red) in a Nissl-counterstained slice (blue). (B) Representative traces of five MC Vms and CA3 LFPs. The horizontal ticks on the left of the Vm traces represent –60 mV. The traces during the three SWRs are expanded in time in the bottom boxes. (C) The LFP traces of 945 SWRs in a representative dataset were dimensionally reduced to two dimensions using the multidimensional scaling (MDS) algorithm based on the root mean square errors (RMSEs) (i.e. dissimilarity) between each SWR waveform. Each dot indicates a single SWR event. The SWR waveforms connected to the red dots correspond, indicating that the SWR waveforms differ depending on the location of the MDS plot. That is, visually similar SWRs are plotted proximally by the MDS algorithm. (D) A pseudocolor map for the amplitudes of the SWR-induced depolarizations (ΔVm) in five MCs during a total of 102 SWRs, demonstrating the heterogeneity of Vm responses among MCs across SWRs. (E) Weak correlations between the SWR amplitudes and ΔVms of MCs. Each dot indicates a single SWR event. ΔVms and SWR amplitudes were Z-standardized for each recording electrode before the data were pooled. R=0.25, n=28,463 SWR events in 87 mossy cells in 23 slices, from 14 mice.
Figure 1—figure supplement 1
Estimated total number of excitatory neurons in each subregion in the rat hippocampus.
Because the mossy cell (MC) numbers were very low, the information had to be compressed in the MC layer during sharp wave-ripple (SWR) backpropagation from the CA3 region to the dentate gyrus (DG).
Figure 1—figure supplement 2
Sharp wave-ripple (SWR) onset-triggered average of ΔVm of mossy cells (MCs).
The ΔVm of the MCs was averaged from –0.25 to 0.25 s with respect to the SWR onset. n=28,360 ΔVms from 87 MCs. The black line indicates the average trace of ΔVm. The gray area demonstrates SD.
Figure 1—figure supplement 3
Lack of sharp wave-ripples (SWR)-induced depolarization of mossy cells (MCs) isolated from the CA3 region.
(A) The CA3 region and the hilus were surgically dissected before recordings. (B) A representative Vm trace of an MC during SWRs. The bar located on the left of the Vm trace indicates the position of –60 mV.
Figure 1—figure supplement 4
Sharp wave-ripples (SWR) frequency does not vary during recording time.
The SWR frequency was calculated in 60 s bins. n=31 recordings (only data recorded for long periods, such as more than 5 min) from 21 slices, p=0.0630, Jonckheere-Terpstra test.
Figure 2
Diversity in sharp wave-ripple (SWR)-induced mossy cell (MC) ΔVms in vivo.
(A) MCs were current-clamped in a urethane-anesthetized mouse, while the local field potentials (LFPs) were recorded from the CA1 region. (B) The recorded neurons were confocally visualized by the detection of intracellularly loaded biocytin (red) in a section immunostained with anti-GluR2/3 (green), an MC marker, and counterstained with NeuroTrace Nissl (blue). Scale bar: 100 µm. The high-magnification images on the right indicate that the recorded neuron was positive for GluR2/3 and that thorny excrescences were observed on proximal dendrites. Scale bar: 10 µm. (C) (Top) Representative Vm traces of an MC. (Middle) A raw LFP trace recorded simultaneously from the CA1 region and its bandpass-filtered trace (100–250 Hz). The bottom panel indicates the wavelet spectrogram of the raw LFP trace. The traces during two SWRs are expanded in time in the right insets. (D) Weak correlations between the SWR amplitudes and ΔVms of MCs during SWRs. R=0.1053, n=756 SWR events in six MCs, from 6 mice.
Figure 3 with 3 supplements
Machine learning-based prediction of sharp wave-ripple (SWR) traces from Vms in mossy cells (MCs) in vitro/in vivo.
(A) Design of the neural network, which consisted of five layers connected unidirectionally with fully connected (FC) links. The number of Vms (left input) varies between 1 and 5, depending on the number of MCs. In this neural network, the SWR and the corresponding Vm of MCs are linked and trained. Thus, by using the Vm of the MC as input, the SWR waveform is obtained as output. (B) Five examples of original SWR traces (black) and SWR traces predicted from real Vm datasets (green) and shuffled Vm datasets (gray). All traces are merged at the bottom. Note that the vertical axis is scaled for each waveform. (C) The results show that the difference between RMSEreal and RMSEsurrogate is statistically significantly different from 0. RMSEsurrogate was calculated as the average of the waveforms that were shuffled 100 times for each SWR. n=178 SWRs from one recording dataset of MC quintets, *p = 2.11 ×10–9, paired t-test. (D) Root mean square error (RMSE) calculated between the original and predicted traces was decreased as the number of simultaneously recorded cells used for prediction increased. One cell: n=28,463 SWRs from 87 single MC recordings; 2 cells: n=35,914 SWRs from 113 MC pair recordings; 3 cells: n=19,874 SWRs from 67 trios; 4 cells: n=4,403 SWRs from 18 quartets; 5 cells: n=271 SWRs from 2 quintets, F=8.28, p=2.51 × 10–6, one-way ANOVA. *p<0.05, **p<0.005, Fisher’s LSD test. (E) The t values determined by the paired t-test are plotted in black, and the p-values for the corresponding t values are plotted in red, indicating that the Vms of one to five MCs led to a significantly accurate prediction. The t value and p-value were also calculated from in vivo dataset as in C and plotted on the right. (In vitro) One cell: n=28,463 SWRs from 87 single MC recordings; 2 cells: n=35,914 SWRs from 113 MC pair recordings; 3 cells: n=19,874 SWRs from 67 trios; 4 cells: n=4,403 SWRs from 18 quartets; 5 cells: n=273 SWRs from 2 quintets. (In vivo) n=757 SWRs from 6 mice, *p=0.024, t=2.2619, paired t-test. (F) Four examples of original SWR traces (black) and SWR traces predicted from real Vm datasets (green) and shuffled Vm datasets (gray). All traces are merged in the bottom panel. Note that the vertical axis is scaled for each waveform.
Figure 3—figure supplement 1
Reproducibility of 120–250 Hz coherence increases as the number of mossy cells (MCs) increases.
The wavelet coherence was computed between original sharp wave-ripple (SWR) traces and SWR traces predicted from real Vm datasets. The wavelet coherence values were averaged over the number of SWRs and were Z-scored within each dataset. One cell: n=28,463 SWRs from 87 single MC recordings; 2 cells: n=35,914 SWRs from 113 MC pair recordings; 3 cells: n=19,874 SWRs from 67 trios; 4 cells: n=4,403 SWRs from 18 quartets; 5 cells: n=271 SWRs from 2 quintets, F=7.84, p=5.24 × 10–6, one-way ANOVA. *p<0.05, **p<0.005, Fisher’s LSD test.
Figure 3—figure supplement 2
Differences in prediction accuracy for each mossy cell (MC), slice, and animal.
A single MC was input into the neural network model, and the root mean square error (RMSE) of the original sharp wave-ripple (SWR) and the predicted SWR was calculated (RMSEreal). In addition, the RMSE of the original SWR and the RMSE of the averaged SWR after shuffling each SWR 100 times were calculated (RMSEsurrogate), and the difference between them was calculated (RMSEsurrogate – RMSEreal). Each plot corresponds to a single MC. When data were obtained from multiple slices of an animal, the slices were plotted in different colors. Mouse ID corresponds to Supplementary file 1. Although there is some variation in prediction accuracy between individuals, slices, and cells, 66 out of 87 MCs had y>0, i.e., good prediction accuracy, and this accounted for 75.9% of the total. The average ΔRMSE for each mouse was calculated, and these values were significantly different from 0 (n=14, *p=0.0041, Student’s t-test). This result shows that even if each animal is considered an independent sample, a statistically significant difference is obtained, and it can be concluded that the results obtained from this model generally have common characteristics.
Figure 3—figure supplement 3
Design of in vivo neural network model.
Input and output sizes were set to 4001 because the sharp wave-ripple (SWR) of in vivo has a longer SWR than in vitro.
Figure 4 with 1 supplement
Correlation between proportions of predictable sharp wave-ripple (SWR) by single mossy cell (MC) and spatial distribution or fundamental neuronal properties.
(A) Distribution of the % predictable SWR from 87 MCs. The % predictable SWR was computed from root mean square errors (RMSEs) by using a single MC prediction. (B) The relative locations of MCs in the hilus are plotted in a semicircular diagram. Each dot indicates a MC, and its color represents the % predictable SWR. (C) Spatial bias of (B) was evaluated by spatial entropy and compared to its chance distribution. The real entropy value is indicated by an arrow. (D, E) Correlations between % predictable SWR and electrophysiological properties. Each dot indicates a single MC. Cm: p=0.3709, R=–0.0971; Rm: p=0.1754, R=–0.1466, n=87 MCs.
Figure 4—figure supplement 1
Correlations between % predictable sharp wave-ripple (SWR) and mossy cell (MC) EPSPs.
EPSP amplitude and EPSP frequency were calculated from each MC. EPSP amplitude was averaged within each MC. Each dot indicates a single MC. EPSP amplitude: p=0.8979, R=0.014; EPSP frequency: *p=0.0008, R=–0.3526, n=87 MCs.
Figure 5
Biased prediction of sharp wave-ripples (SWRs) by mossy cells (MCs).
(A) (Left) The local field potential (LFP) traces of 945 SWRs from a representative quadruple recording dataset were dimensionally reduced using the multidimensional scaling (MDS) algorithm. Each dot indicates a single SWR event, and its color represents whether the prediction rate is significant (red) or not (gray). (Right) The spatial bias of the prediction rates in the MDS space was evaluated by spatial entropy. Lower entropies indicate higher spatial biases of the prediction rate. In each MC, the entropy was compared to the change distribution obtained by 1,000,000 surrogates in which the prediction rates were shuffled within the MC. The real entropy value is indicated by an arrow with its p-value. (B) The same calculations as those in (A) were repeated for all 87 cells, and their p-values were plotted. The red dots indicate MCs with significantly lower entropy values. (C) Distribution of the correlation coefficient of the root mean square error (RMSE) score sets (non-predictable SWR as 0, predictable SWR as 1) between 113 MC pairs, which were obtained from 2 quintuple, 8 quadruple, and 15 triple recording datasets.
Figure 6 with 2 supplements
Prediction specificity of sharp wave-ripples (SWRs) by mossy cells (MCs).
(A) The representative quintuple recording in a total of 178 SWRs. The top five rows demonstrating the SWRs that each MC can predict. Its color represents whether the SWR was predictable (red) or not predictable (gray). The mean predictivity across five MCs is shown in the bottom rastergram. (B) The Venn diagram demonstrated the image of the range of SWRs that each MC can predict out of the recorded SWR. (C) The % predictable SWR was increased as the number of simultaneously recorded cells used for prediction increased. The linear summation is indicated by a solid line (blue). One cell: n=28,463 SWRs from 87 single MC recordings; 2 cells: n=35,914 SWRs from 113 MC pair recordings; 3 cells: n=19,874 SWRs from 67 trios; 4 cells: n=4,403 SWRs from 18 quartets; 5 cells: n=271 SWRs from 2 quintets, *p=1.00 × 10–38, Jonckheere-Terpstra test. (D) Distribution of the overlap % per total SWR between 113 MC pairs, which were obtained from 2 quintuple, 8 quadruple, and 15 triple recording datasets. (E) Correlations between SWR amplitudes and number of overlapped MCs. The regression line is indicated by solid line (gray). The data were obtained from 2 quintuple, 8 quadruple, and 15 triple recording datasets, F=5.38, *p=0.0007, one-way ANOVA; *p=1.59 × 10–5, Jonckheere-Terpstra test.
Figure 6—figure supplement 1
Visualization of overlapped sharp wave-ripple (SWR).
The number in each area represents the number of SWRs. In total, 178 SWRs were recorded, with 71 predictable SWRs, thus the total number of value is 71.
Figure 6—figure supplement 2
Sharp wave-ripple (SWR) properties are maintained regardless of mossy cell (MC) overlap.
Correlations between SWR properties and number of overlapped MCs. SWR duration: p=0.2249, F=1.45; Ripple frequency: p=0.2928, F=1.26; fast Fourier transform (FFT) power (120–250 Hz): p=0.3781, F=1.07; area under the curve (AUC) of FFT power (120–250 Hz): p=0.4364, F=0.96, one-way ANOVA. These values were obtained from 2 quintuple, 8 quadruple, and 15 triple recording datasets.
Author response image 1
(A) CA3 SWRs propagate backward simultaneously to the MC and GC, with the input being sent to the MC (Diagram A). In this case, inhibiting input from the GC to the MC via the DCG-IV pathway may allow comparison of the magnitudes of the CA3-to-MC and GC-to-MC inputs. However, our previous studies that recorded EPSPs in the MC or GC during CA3 SWR showed that the SWR-to-MC pathway is significantly faster than the SWR-to-GC pathway in terms of time lag (Modified Figure 1,3 from Ouchi et al., 2017). This implies that, when a barrage of MC EPSPs occurs due to a SWR, the fast component likely originates from direct input from CA3 to MC rather than via GC. Regarding the slow component, input from GC may be involved. However, based on the causal diagram A, both components originate from the CA3 SWR. Therefore, even if there were SWR driven input from the GC to the MC EPSP in this study, it would not significantly undermine the argument that the signal can reconstruct the SWR because it originates from the same source. (B) A case in which input from GC to MC occurs without being caused by the CA3 SWR (Diagram B). In Figure 1 —figure supplement 3, severing the connection between CA3 and the hilus eliminates the input from SWR to MC. Therefore, the non SWR driven input from GC to MC is considered to have little effect.
Additional files
-
Supplementary file 1
Recording time and number of sharp wave-ripple (SWR) events in each slice.
Some data were recorded multiple times from a single slice, and the quality of the recordings was confirmed by checking the access resistance and intrinsic electrophysiological properties of the mossy cells (MCs) did not change significantly between recordings. When there were multiple recordings, each value was entered in the same cell. The data thus obtained were concatenated in machine learning and treated as a single dataset. Note that 190801_2 and 190927_1 have overlapping slice IDs and different cell numbers. In this case, the MCs to be recorded are common, but the recording times (i.e. SWR events) used for the actual analysis do not overlap.
- https://cdn.elifesciences.org/articles/97270/elife-97270-supp1-v2.docx
-
MDAR checklist
- https://cdn.elifesciences.org/articles/97270/elife-97270-mdarchecklist1-v2.docx
Download links
A two-part list of links to download the article, or parts of the article, in various formats.
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
Distributed subthreshold representation of sharp wave-ripples by hilar mossy cells
eLife 14:e97270.
https://doi.org/10.7554/eLife.97270
