This is a preprint, a preliminary version of a manuscript that has not completed peer review at a journal. Research Square does not conduct peer review prior to posting preprints. The posting of a preprint on this server should not be interpreted as an endorsement of its validity or suitability for dissemination as established information or for guiding clinical practice.
A new preprint design is coming!
We have improved readability, clarity, accessibility, and opportunities for engagement.
Research Article
Giorgio A. Ascoli
George Mason University
Corresponding Author
License:
This work is licensed under a CC BY 4.0 License. Read Full License
Declarations:
Declarations
- The author has confirmed that a statement listing potential conflicts of interest or lack thereof is included in the text.
Limited experimental yield, heterogeneous recordings conditions, and ambiguous neuronal identification have so far prevented the systematic characterization of synaptic signals for all connections of any neural system. Introducing a novel strategy to overcome these challenges, we report the first comprehensive synaptic quantification among all known neuron types of the hippocampal-entorhinal network. First, we reconstructed > 2,600 synaptic traces from ~ 1,200 publications into a unified model of synaptic dynamics. We then trained a deep learning architecture with the resulting parameters, each annotated with detailed metadata. The model learned to predict the synaptic properties of all 3,120 circuit connections in arbitrary conditions with accuracy approaching the intrinsic experimental variability. Analysis of normalized data revealed that synaptic signals are controlled by few latent variables associated with specific molecular markers and interrelating conductance, kinetics, and short-term plasticity. We freely release the tools and full dataset of unitary synaptic values in 32 covariate settings via Hippocampome.org.
Keywords
Synaptic physiology, Neuronal transmission, Rodentia, Machine Learning, Computer simulation, Software, Database, Quantitative model, Short-term plasticity
Figure 1
Figure 2
Figure 3
Figure 4
Figure 5
Figure 6
Figure 7
This is a list of supplementary files associated with this preprint. Click to download.
- Table1.docx
List of features used for machine learning.
- SupplementaryVideos.mp4
Synapses with smaller and larger amplitudes tend to undergo short-term facilitation and depression, respectively. We sorted all entorhinal-hippocampal synapses by conductance (high-amplitude to low-amplitude) and simulated each synapse separately in voltage-clamp and standard condition with ISI = 20 ms. We animated the evolution of synaptic currents as a function of conductance, which revealed the transition from depression to facilitation.
- Fig.S1.tif
Signal correction methods. We removed any membrane fluctuations with slow kinetics superposed on the recorded synaptic signals using two methods. (a) We extracted the amplitude (A) and deactivation time constant (τd) of the first synaptic event from a digitized trace to reconstruct the signal by estimating the initiation points. (b) We used a simulated signal to approximate the correction amount at the initiation points of synaptic events (e.g., Δai and Δai+1). Triangulation was then used to linearly approximate the correction amounts of intermediate points (Δax).
- Fig.S2.tif
Impact of covariates on synaptic properties data availability. (a) Synaptic traces from two studies16, 17 recording GABAergic signals from CA1 Axo-axonic to CA1 Pyramidal cells in different species and with different intracellular solutions. Higher chloride concentration ([Cl]i) increased the recorded signal via enhanced driving force (i.e., Vm - Erev). Note large differences in g, τd, and U. (b) GABAergic signals from DG HICAP to DG Granule cells recorded at two different temperatures18. All other covariates were identical. (c) Voltage- and current-clamp recording of glutamatergic synaptic signals between CA3 Pyramidal cells at two different membrane potentials19. Even though the estimated synaptic conductances are almost equal, the rest of the parameter estimates differ substantially. (d) Glutamatergic synaptic currents recorded from CA1 Basket CCK+ neurons using three different stimulation paradigms20. Evoked currents were smaller than miniature and spontaneous ones, because of the minimal stimulation protocol. Note that researchers only reported one synaptic event, preventing the estimation of τr, τf, and U. (e-f) Heatmap representations of the number of data points available for each of 3,120 potential connections among 122 neuron types (rows: presynaptic, columns: postsynaptic). Light pink entries are entries with missing synaptic data (19.7% for τd and 38.5% for ST-P parameters). Black entries mark absence of potential connection. In all panels (a-f), blue and yellow colors represent GABAergic and glutamatergic synapses, respectively.
- Fig.S3.tif
Deep learning model architecture. The deep learning model in this study is a five-layer encoder-decoder perceptron regularized with different techniques. We attained the best results with 128 nodes in the encoder (third) layer. For model regularization, we set a 50% dropout rate for all layers except for the encoder layer, which was set to 5%. Prediction accuracy improved by coupling dropout with unitary max-norm weight constraint and batch normalization.
- Fig.S4.tif
Performance of the deep learning model for all types of stimulation methods. (a) Prediction error of the model monitored after each training epoch using k-fold cross-validation. The prediction error was close to the training error and the difference between the two decreased after each training epoch. (b-c) Comparison of training and prediction accuracies with target variability for all types of synaptic stimulations. These results, including prediction reliability (PR), are comparable to those obtained when only considering unitary stimulation (Fig. 3). (d) Trimmed-mean and interquartile range of target variability and training and prediction accuracies (in SMAPE) for different parameters.
- Fig.S5.tif
Comparison of synaptic parameters with existing sparse estimates in CA1. The human brain project (HBP) has recently estimated synaptic parameters of 16 potential connections in CA1 (Ecker et al., 2020). A pairwise comparison by potential connections and synaptic parameter with our estimates to be included in Hippocampome.org (HCO) detects no statistically significant difference.
- Fig.S6.tif
Range and distribution of synaptic parameters in different anatomical regions. (a) Probability density functions of the synaptic parameter in standard conditions globally normalized with the min-max method across different regions and synapse types. Filled circles denote median parameter values. (b) Distributions of min-max normalized GABAergic and glutamatergic parameter inferences. All synaptic parameters except U are right-tailed. The three time constants, but not g and U, differ by neurotransmitter: relative to GABAergic, glutamatergic synapses have smaller τd and τr, but larger τf.
- Fig.S7.tif
Influence of axonal targeting patterns on synaptic input. We compared synaptic properties grouped by the axonal morphology of the postsynaptic neuron: the OR group had axons in strata oriens and radiatum; OPR group was similar to OR but also had axons in stratum pyramidale; SLM group had axons in stratum lacunosum moleculare; SP group had axons only in stratum pyramidale. We also extended the study to CA2, CA3, and DG with similar grouping, if the regions had equivalent neuron types. SG and SMo in DG are homologous to SP and SLM in CA1. (a) We measured the average difference of synaptic parameters among groups using symmetric percentage distance (SPD). Bold values indicate statistical significance. A positive (negative) value indicates the synaptic parameter is larger (smaller) for G1 than for G2. (b) Simulated signals using averaged synaptic parameters in each group (Vh = -60 mV, Erev = 0 mV, and ISI = 10 ms). (c) To visualize the distance of the groups (Left), we normalized the trimmed-mean measures of g, τd, and AB3:A1 by dividing values to the maximum among all synapses. To compare the similarities within each group (Right), we calculated the CV of every group and across all synapses (gray) in each region.
- Fig.S8.tif
Inter-run variability of the deep learning model predictions. Since deep learning models depend on the (stochastic) order in which the training dataset is presented, we expected a certain degree of variation in inferences among the 100 trained models. In all analyses we reported the mean over the 100 values, but here we report the coefficient of variation (CV) of the model predictions for each synaptic parameter and potential connection. The potential connections and the synaptic parameters varied with respect to the CV. On average, τf had a higher CV than the other parameters.
- Fig.S9.tif
Equivalence of numerical and analytical solutions. We used the NEURON simulation environment to simulate the synaptic current of a typical signal both by solving the differential equations numerically and by using the analytical equations. Moreover, we compared the results to the original ModelDB implementation of the four-state TPM model. All three simulations produced identical results, confirming the accuracy of our solution.
- S.Materials51921gaaNN.docx
Normalized unitary synaptic signaling of the hippocampus and entorhinal cortex
- Methods.docx
Loading...
Badges
Prescreen
This manuscript passed our Prescreen assessment
Complete author information
Appropriate declaration statements
Potential risks to human health
Prescreen
History
CURRENT STATUS: Posted
Version 1
Posted 03 Jun, 2021
No community comments so far
Metrics
Comments: 0
PDF Downloads: ...
HTML Views: ...
Subject Areas
Computational Neuroscience
Learn more about our company.
A new preprint design is coming!
We have improved readability, clarity, accessibility, and opportunities for engagement.
This is a preprint, a preliminary version of a manuscript that has not completed peer review at a journal. Research Square does not conduct peer review prior to posting preprints. The posting of a preprint on this server should not be interpreted as an endorsement of its validity or suitability for dissemination as established information or for guiding clinical practice.
Research Article
Giorgio A. Ascoli
George Mason University
Corresponding Author
Badges
Prescreen
This manuscript passed our Prescreen assessment
Complete author information
Appropriate declaration statements
Potential risks to human health
Prescreen
History
CURRENT STATUS: Posted
Version 1
Posted 03 Jun, 2021
No community comments so far
Metrics
Comments: 0
PDF Downloads: ...
HTML Views: ...
Subject Areas
Computational Neuroscience
License:
This work is licensed under a CC BY 4.0 License. Read Full License
Declarations:
Declarations
- The author has confirmed that a statement listing potential conflicts of interest or lack thereof is included in the text.
Limited experimental yield, heterogeneous recordings conditions, and ambiguous neuronal identification have so far prevented the systematic characterization of synaptic signals for all connections of any neural system. Introducing a novel strategy to overcome these challenges, we report the first comprehensive synaptic quantification among all known neuron types of the hippocampal-entorhinal network. First, we reconstructed > 2,600 synaptic traces from ~ 1,200 publications into a unified model of synaptic dynamics. We then trained a deep learning architecture with the resulting parameters, each annotated with detailed metadata. The model learned to predict the synaptic properties of all 3,120 circuit connections in arbitrary conditions with accuracy approaching the intrinsic experimental variability. Analysis of normalized data revealed that synaptic signals are controlled by few latent variables associated with specific molecular markers and interrelating conductance, kinetics, and short-term plasticity. We freely release the tools and full dataset of unitary synaptic values in 32 covariate settings via Hippocampome.org.
Figure 1
Figure 2
Figure 3
Figure 4
Figure 5
Figure 6
Figure 7
This is a list of supplementary files associated with this preprint. Click to download.
- Table1.docx
List of features used for machine learning.
- SupplementaryVideos.mp4
Synapses with smaller and larger amplitudes tend to undergo short-term facilitation and depression, respectively. We sorted all entorhinal-hippocampal synapses by conductance (high-amplitude to low-amplitude) and simulated each synapse separately in voltage-clamp and standard condition with ISI = 20 ms. We animated the evolution of synaptic currents as a function of conductance, which revealed the transition from depression to facilitation.
- Fig.S1.tif
Signal correction methods. We removed any membrane fluctuations with slow kinetics superposed on the recorded synaptic signals using two methods. (a) We extracted the amplitude (A) and deactivation time constant (τd) of the first synaptic event from a digitized trace to reconstruct the signal by estimating the initiation points. (b) We used a simulated signal to approximate the correction amount at the initiation points of synaptic events (e.g., Δai and Δai+1). Triangulation was then used to linearly approximate the correction amounts of intermediate points (Δax).
- Fig.S2.tif
Impact of covariates on synaptic properties data availability. (a) Synaptic traces from two studies16, 17 recording GABAergic signals from CA1 Axo-axonic to CA1 Pyramidal cells in different species and with different intracellular solutions. Higher chloride concentration ([Cl]i) increased the recorded signal via enhanced driving force (i.e., Vm - Erev). Note large differences in g, τd, and U. (b) GABAergic signals from DG HICAP to DG Granule cells recorded at two different temperatures18. All other covariates were identical. (c) Voltage- and current-clamp recording of glutamatergic synaptic signals between CA3 Pyramidal cells at two different membrane potentials19. Even though the estimated synaptic conductances are almost equal, the rest of the parameter estimates differ substantially. (d) Glutamatergic synaptic currents recorded from CA1 Basket CCK+ neurons using three different stimulation paradigms20. Evoked currents were smaller than miniature and spontaneous ones, because of the minimal stimulation protocol. Note that researchers only reported one synaptic event, preventing the estimation of τr, τf, and U. (e-f) Heatmap representations of the number of data points available for each of 3,120 potential connections among 122 neuron types (rows: presynaptic, columns: postsynaptic). Light pink entries are entries with missing synaptic data (19.7% for τd and 38.5% for ST-P parameters). Black entries mark absence of potential connection. In all panels (a-f), blue and yellow colors represent GABAergic and glutamatergic synapses, respectively.
- Fig.S3.tif
Deep learning model architecture. The deep learning model in this study is a five-layer encoder-decoder perceptron regularized with different techniques. We attained the best results with 128 nodes in the encoder (third) layer. For model regularization, we set a 50% dropout rate for all layers except for the encoder layer, which was set to 5%. Prediction accuracy improved by coupling dropout with unitary max-norm weight constraint and batch normalization.
- Fig.S4.tif
Performance of the deep learning model for all types of stimulation methods. (a) Prediction error of the model monitored after each training epoch using k-fold cross-validation. The prediction error was close to the training error and the difference between the two decreased after each training epoch. (b-c) Comparison of training and prediction accuracies with target variability for all types of synaptic stimulations. These results, including prediction reliability (PR), are comparable to those obtained when only considering unitary stimulation (Fig. 3). (d) Trimmed-mean and interquartile range of target variability and training and prediction accuracies (in SMAPE) for different parameters.
- Fig.S5.tif
Comparison of synaptic parameters with existing sparse estimates in CA1. The human brain project (HBP) has recently estimated synaptic parameters of 16 potential connections in CA1 (Ecker et al., 2020). A pairwise comparison by potential connections and synaptic parameter with our estimates to be included in Hippocampome.org (HCO) detects no statistically significant difference.
- Fig.S6.tif
Range and distribution of synaptic parameters in different anatomical regions. (a) Probability density functions of the synaptic parameter in standard conditions globally normalized with the min-max method across different regions and synapse types. Filled circles denote median parameter values. (b) Distributions of min-max normalized GABAergic and glutamatergic parameter inferences. All synaptic parameters except U are right-tailed. The three time constants, but not g and U, differ by neurotransmitter: relative to GABAergic, glutamatergic synapses have smaller τd and τr, but larger τf.
- Fig.S7.tif
Influence of axonal targeting patterns on synaptic input. We compared synaptic properties grouped by the axonal morphology of the postsynaptic neuron: the OR group had axons in strata oriens and radiatum; OPR group was similar to OR but also had axons in stratum pyramidale; SLM group had axons in stratum lacunosum moleculare; SP group had axons only in stratum pyramidale. We also extended the study to CA2, CA3, and DG with similar grouping, if the regions had equivalent neuron types. SG and SMo in DG are homologous to SP and SLM in CA1. (a) We measured the average difference of synaptic parameters among groups using symmetric percentage distance (SPD). Bold values indicate statistical significance. A positive (negative) value indicates the synaptic parameter is larger (smaller) for G1 than for G2. (b) Simulated signals using averaged synaptic parameters in each group (Vh = -60 mV, Erev = 0 mV, and ISI = 10 ms). (c) To visualize the distance of the groups (Left), we normalized the trimmed-mean measures of g, τd, and AB3:A1 by dividing values to the maximum among all synapses. To compare the similarities within each group (Right), we calculated the CV of every group and across all synapses (gray) in each region.
- Fig.S8.tif
Inter-run variability of the deep learning model predictions. Since deep learning models depend on the (stochastic) order in which the training dataset is presented, we expected a certain degree of variation in inferences among the 100 trained models. In all analyses we reported the mean over the 100 values, but here we report the coefficient of variation (CV) of the model predictions for each synaptic parameter and potential connection. The potential connections and the synaptic parameters varied with respect to the CV. On average, τf had a higher CV than the other parameters.
- Fig.S9.tif
Equivalence of numerical and analytical solutions. We used the NEURON simulation environment to simulate the synaptic current of a typical signal both by solving the differential equations numerically and by using the analytical equations. Moreover, we compared the results to the original ModelDB implementation of the four-state TPM model. All three simulations produced identical results, confirming the accuracy of our solution.
- S.Materials51921gaaNN.docx
Normalized unitary synaptic signaling of the hippocampus and entorhinal cortex
- Methods.docx
Loading...