We compiled, digitized, and reconstructed from the published literature a comprehensive dataset of 2,621 synaptic signals recorded from the dentate gyrus, CA3, CA2, CA1, subiculum, and entorhinal cortex. For each recording, we annotated the detailed experimental conditions with 75 covariates (Methods; Table 1) and mapped the potential pair of presynaptic and postsynaptic neuron types among 3,120 potential connections identified by Hippocampome.org (Moradi and Ascoli, 2020). While this synaptic database constitutes a uniquely information-rich resource, its quantitative analysis requires solving distinct challenges (Fig. 1). First, researchers record synaptic signals in different modalities (current- or voltage-clamp) and widely diverse experimental conditions, which cannot be directly compared. Second, synaptic measurements can rarely be ascribed to single identified presynaptic and postsynaptic neuron types: in most cases, the mapping is ‘fuzzy’ and matches several potential connections (green arrows in Fig. 1). Third, synaptic data are unavailable for a sizeable minority of potential connections. Additionally, certain experiments only include one synaptic event (e.g., upper right signal in Fig. 1), thus providing no information on short-term plasticity. To solve part of the first challenge (normalizing recording modality and a subset of covariates), we fit all synaptic recordings to the same model via signal simulation. To solve the remaining challenges (normalizing the rest of the covariates, disambiguating potential connections, and inferring missing data), we bring to bear an original strategy based on machine learning.
Modeling comparable synaptic parameters from diverse measures and modalities
Data integration starts with the digitization of published synaptic recordings (Fig. 2a). These signals are diverse in terms of measurement modalities (current vs voltage) and the composition of intracellular and extracellular solutions affecting reversal potentials (Erev). To transform these data into a comparable form, we fitted all digitized signals to a simplified Tsodyks, Pawelzik, and Markram (TPM) model, which represents synaptic properties with 5 parameters (Supplementary Methods)9, 16. These synapse-specific parameters (g, τd, τr, τf, and U) depend on the combination of presynaptic and postsynaptic neuronal types involved and are estimated by fitting the TPM model output to the digitized signals (Fig. 2b). The model also requires a small set of measurements that depend on experimental settings and the properties of the postsynaptic neuron: Erev, the initial value of the membrane voltage (Vm), membrane time constant (τm), and capacitance (Cm). To eliminate the impact of processes causing slow signal fluctuations, we corrected the signals before parametric fitting (Fig. S1 and Methods). The TPM model produced comparable synaptic parameters and normalized the data with respect to synaptic driving force (Vm - Erev) by converting synaptic amplitudes to conductance. Overall, the process reduces data dimensionality by describing every signal with only 5 values.
Construction and validation of a predictive model of all synapses
The fitted parameters for matching potential connections in different experimental conditions reveal a large degree of variation that could be associated with covariates such as animal sex, species, recording and stimulation methods, and temperature (Fig. S2a-d). To normalize the effect of covariates, we trained a predictive deep learning model of the synaptic parameters using a five-layer autoencoder perceptron architecture (Figs. 2c and S3; Methods). Given a potential connection and experimental covariates (i.e., features: Table 1), the models learned to infer the 5 synaptic parameters (i.e., targets). Training converged to stable performance with learned values deviating on average less than 30% from the experimental measurements (Fig. S4a). The model displayed no overfitting and the predicted values (for targets not included in the training set) deviated only marginally more (~ 32%) from the original measurements (Fig. S4a). To assess this performance relative to the reliability of experimental measurements, we consider different experimental values (“targets”) recorded from the exact same nominal conditions (“features”). Those differences can be ascribed to unknown experimental factors, intrinsic biological variability, and random noise. We take such empirical ground-truth range as the “gold standard” to benchmark our model against. In these cases, we calculated the distance of each target value from their average, a measure of experimental fluctuation we call target variability. We compared the target variability with the training accuracy and prediction accuracy, i.e., the distance of model output from seen and unseen targets, respectively. The training and prediction accuracies of our predictive model were remarkably close to the target variability. Testing the predictive power of the model with the jackknife (leave-one-out) method, we found that the vast majority of unitary predictions fell within the 95% confidence interval of the targets, i.e., they were “reliable” (Figs. 3a-b). Specifically, this prediction reliability (PR) ranged from 90% for τr to 96% for U, with intermediate values for g (91%), τd (94%), and τf (94%). By including all synaptic measurements (not just the unitary values, PR was reduced slightly to 88%-94% (Figs. S4b-d). Additionally, comparing the relevant values to sparse estimates available for matching potential connections from a recent CA1 study17 revealed no statistically significant difference for any of the 5 parameters (Fig. S5). Thus, the deep learning model quantitatively predicts the properties of synaptic signals for which experimental recordings are available within the margin of measurement accuracy.
Connectivity matrix completion and synaptic data normalization
Given its demonstrated performance on available data, the predictive model can confidently estimate the synaptic parameters of yet uncharacterized potential connections based on the learned properties of neuronal types. The model can complete the synaptic electrophysiology matrix for all 3,120 potential connections in the hippocampus and entorhinal cortex. Additionally, since the learned neuronal properties are all unique, the model also effectively disambiguates each potential connection: in other words, the predicted synaptic parameters for each pair of neuron types are also all unique. Importantly, the deep learning model can infer synaptic parameters for every potential connection in any desired condition. Applying homogeneous conditions for all potential connections practically normalizes the inferences with respect to the covariates. This study primarily focuses on fast unitary synaptic properties in near-physiological (henceforth “standard”) condition, namely AMPA and GABAA synapses of adult male rats in voltage-clamp at body temperature and with a pipette solution that does not disturb intracellular ionic concentrations (Methods). These so-derived synaptic signals showed a wide range of amplitudes, kinetics, and ST-P across potential connections (Fig. 3c and Suppl. Video). To explore regional differences within the hippocampal formation, we inspected the probability density distributions of all parameter values normalized using the min-max method (Fig. S6a). Interestingly, the range of values in the entorhinal cortex is smaller than in the hippocampus. Moreover, the GABAergic and the glutamatergic synapses had overlapping distributions for g and U but not for the time constants (Fig. S6b), suggesting that these synapse types have similar amplitudes but differ in kinetics and ST-P.
Open access to data and source codes
The normalized and completed synaptic data are broadly applicable to designing experiments in optimal conditions, testing hypotheses, constraining biologically plausible simulations of the entire entorhinal-hippocampal circuit18, and benchmarking machine learning algorithms. We provide 5 synaptic constants for each of 3,120 connections in 32 different settings that include all binary combinations of species (rat or mouse), sex (male or female), age (young or adult), recording method (voltage- or current-clamp), and temperature (room or body). For each parameter we make available the mean, standard deviation, and range over 100 training runs of the deep learning model (Fig. 4a). We also share all implemented tools for unhindered reuse with other datasets. The Synapse Modeling Utility, the preprocessing and analysis code in R, the machine learning library in Python, and the preprocessed machine learning-ready experimental data (2,621 features-targets sets) are all freely available on Hippocampome.org/synapse (Fig. 4b).
Presynaptic and postsynaptic determinants of synaptic physiology
Full data normalization allowed us to compare for the first time the synaptic properties of all potential connections without the influence of confounding variables. To begin the investigation of how the presynaptic and postsynaptic identities combine to define synaptic dynamics, we asked two questions: (1) when a pair of neuron types forms a synapse, which synaptic properties (e.g., amplitude, duration, ST-P) does either side dominantly determine? (2) Does the answer differ for glutamatergic and GABAergic synapses? To answer these questions, we separated the glutamatergic and GABAergic synapses. In each pool, we created two groupings: one based on the presynaptic neuron types, and the other based on the postsynaptic ones. For example, the glutamatergic presynaptic grouping consisted of 38 groups, one for every glutamatergic presynaptic type; each of these groups contains all postsynaptic neuron types that presynaptic type forms a connection with. We then calculated for each group the coefficient of variation (CV) of all 5 synaptic parameters in the standard condition (Fig. 5a). A lower CV indicates less intragroup variation and thus a tighter control of the corresponding grouping on that synaptic property. For GABAergic synapses, the ST-P parameters (but not conductance and kinetics) had significantly smaller CVs if synapses were grouped based on postsynaptic type. For glutamatergic synapses, in contrast, all parameters except U had significantly smaller CVs if synapses were grouped based on presynaptic type. In other words, presynaptic glutamatergic neurons and postsynaptic GABAA receptors are more important determinants of synaptic signals.
Principal covariate effects on synaptic properties
Next, we systematically investigated the influence of experimental covariates on synaptic parameters. Earlier research mainly checked the impact of experimental conditions on synaptic amplitude and kinetics of a limited number of neuron types. Our study also allowed the inclusion of ST-P parameters and systematically covered all potential connections of the hippocampal formation by changing one covariate at a time. All tested covariates had a statistically significant impact on synaptic parameters, but we only report here (Fig. 5b-c) those with a meaningful effect size (> 10%) and emphasize the most substantive ones (> 20%). Our results indicate that g increases more than two-fold and τd decreases 30% when switching from voltage- to current-clamp, from male to female animals, and from gluconate-free to gluconate-containing intracellular solutions. While the change with recording modality agrees with previous studies for example, 19 and we expected a difference by sex, the pronounced impact of gluconate in the pipette solution was surprising. Moreover, current clamp (relative to voltage clamp) and female animals (relative to male) also entailed notably higher τr and lower τf, implying greater propensity towards synaptic depression. In contrast, the opposite trend, conducive to facilitation, was observed with gluconate. Shifting from rats to mice or from room to body temperature affected synaptic properties in the same direction, but to a more modest extent (10–20% effect size), as the male-to-female switch or intracellular gluconate addition, respectively. Reducing [Cl]i substantially increased short-term facilitation at GABAergic synapses, while more modestly slowing down synaptic kinetics which was unexpected based on 20. Other covariates, including to our surprise age, did not affect the parameters substantially. Altogether, remarkably, only two types of variation, differing just in the change direction of τr and τf, could explain the impact of all analyzed covariates irrespective of neurotransmitter type. This observation suggests an interdependence among synaptic parameters.
Synaptic amplitude predicts signal kinetics and the direction of short-term plasticity
Among both glutamatergic and GABAergic types, we noticed that synapses with high amplitude had fast kinetics and demonstrated depressing ST-P. Conversely, synapses with low amplitude had slower kinetics and were facilitating. To visualize these observations, we averaged the model parameters from the 30 synapses with the largest conductance and from the 30 with the smallest one among both glutamatergic and GABAergic groups. We then compared the responses of the four consensus models in standard condition (Fig. 6a and Suppl. Video). The high-amplitude models exhibited short-term depression and short signal duration (half-height width: 2.4 ms for glutamatergic and 3.8 ms for GABAergic), while the low-amplitude models demonstrated short-term facilitation and long signal duration (half-height width: 5.1 ms for glutamatergic and 6.2 ms for GABAergic). Considering all 3,120 connections revealed a significant negative correlation between g and τd and between g and the paired-pulse ratio from baseline of the third synaptic event (AB3:A1), but a positive correlation between g and U, suggesting that high-amplitude synapses have higher resource utilization (Fig. 6b). Facilitation and depression partly depend on interstimulus intervals (ISI) and the measure of ST-P. Testing ST-P at 20 ms ISI and considering AB3:A1, the majority (> 90%) of synapses with amplitude below 0.5 nS facilitated, irrespective of neurotransmitter, while most synapses above 2 nS (glutamatergic) or 3 nS (GABAergic) depressed (Fig. 6c, left). Although the second synaptic events (AB2:A1) tended towards facilitation relative to subsequent signals (e.g., AB5:A1), all ST-P measures consistently transitioned from facilitation to depression as a function of conductance (Fig. 6c, right). Moreover, τf and τr were negatively correlated (Rglu=-0.4, RGABA=-0.1, p < 0.05), indicating that synapses needing a long time to recover their resources tend to reduce their synaptic utilization rate rapidly. Altogether, these analyses suggest that higher synaptic amplitudes predict faster kinetics and a tendency towards depression over facilitation, reflecting coordinated differences in τd and U as well as interdependence of τf and τr.
Presynaptic and postsynaptic molecular expression as a biomarker of short-term plasticity
It is a widespread practice to study synapses based on molecular expression. Chemical biomarkers were not directly among the training features of our predictive synapse models, but were used for mapping mined signals to potential connections8. We employed Hippocampome.org to query neuron types expressing different markers21, 22 and analyzed differences in synaptic properties among neuron types containing (+) or lacking (-) each molecule. Since certain markers are expressed in the presynaptic terminals and others in the postsynaptic dendrites and soma23, we studied the presynaptic and the postsynaptic groups separately (Fig. 7). Considering AB3:A1 as a measure of ST-P and using a 20 ms ISI, we identified two classes of presynaptic markers that respectively predicted synaptic facilitation and depression. Specifically, presynaptic calbindin (CB), cholecystokinin (CCK), and neuropeptide-Y (NPY) expression correlated with facilitation (larger AB3:A1 values). In contrast, calretinin (CR), parvalbumin (PV), and somatostatin (SOM) correlated with depression (smaller AB3:A1 values). The relations of these markers with changes in synaptic amplitude and kinetics were not always statistically significant but generally followed the trends revealed in the previous section: namely, presynaptic expressions predicting short-term facilitation typically demonstrated lower signal amplitudes and slower kinetics and vice versa for those predicting short-term depression. Cannabinoid receptor 1 (CB1) is expressed both on presynaptic and postsynaptic sides24. Since the presynaptic effects were similar to CCK, we only illustrated the postsynaptic effects. Among the postsynaptic markers, both CB1 and serotonin receptor 3 (5HT-3) predicted lower amplitudes and a tendency towards facilitation. Interestingly, CB1 exerted greater impact when partnering with GABAergic than with glutamatergic synapses.
Correlations between neuronal morphology and synaptic parameters
In GABAergic neurons of both hippocampal area CA1 and visual cortex, the kinetics of spontaneous synaptic inputs vary depending on the specific axonal targeting of that same postsynaptic neuron25, 26. We tested similar interactions between input synaptic properties and output axonal patterns throughout the hippocampal formation, not only considering unitary synaptic kinetics, but also conductance and ST-P (Fig. S7). Among GABAergic synapses in CA1, we found significant differences in g, τd, τf, and U, indicating that not only input synaptic duration, but also amplitude and facilitation, vary by output axonal targeting (Fig. S7a). Extending the study to other hippocampal regions revealed significant differences in τd and τf among GABAergic synapses in CA3, and in τr in DG and CA2. Glutamatergic synapses generally demonstrated fewer significant differences. Visualizing consensus traces (Fig. S7b) and synaptometrics differences (Fig. S7c) confirmed these patterns.
In the visual cortex, connection probability correlates with synaptic strength27. Hippocampome.org calculates the probabilities of connections and the average synaptic distance from the presynaptic and postsynaptic soma based on the layer-specific linear densities of the corresponding axons and dendrites 28. Synaptic conductance had a weak but significant positive correlation with the connection probability (RGABA=0.27, RGlu=0.19, p < 0.05). Consistent with dendritic filtering, we also found a significantly negative correlation between g and the synaptic distance from the postsynaptic soma (RGABA=-0.13, RGlu=-0.06, p < 0.05).