Fig. 4: Validation and representative output of the deep learning model for unitary signals.

(a) Probability distribution functions of the mismatch between model output and experimental data for each of the 5 TPM parameters (targets). The mismatch is quantified by the distance, measured by symmetric absolute percentage error (SAPE), of the model output relative to targets used during training (training accuracy) and targets left out of training set (prediction accuracy). These two measures to the intrinsic variability of the available experimental data grouped based on features (target variability). Synapses are plastic and stochastic entities, hence if neuron pairs from the same presynaptic and postsynaptic types are recorded several times in the same conditions, several values will typically be obtained experimentally, while model outputs will be constant. In other words, target variability is the distance of a training data point from the average of targets with identical features. This measure of variability in the experimental dataset defines the ideal limit of model accuracy. Training accuracy quantified the learning capacity of the model. Prediction accuracy measures the inference power of the model by calculating the distance of model output with a ground truth, which is the experimental data not seen by the model during jackknife (leave-one-out) procedure. The overall similarity of distributions indicates that the model achieved a level of accuracy comparable to the reproducibility of corresponding experimental data. The prediction reliability (PR) is the proportion of model outputs falling within the 95% confidence interval of the experimental data with identical features. (b) Prediction and training accuracy are highly correlated for all 5 parameters, suggesting the absence of overfitting. (c) Simulated synaptic dynamics based on original training targets (black dotted traces) with identical features, and deep learning model inferences (solid lines). Model predictions are remarkably close to the training data even though experimental data showed variability. Simulated synaptic traces (V_h = −35 mV, GABA_A E_rev = −70 mV, and AMPA E_rev = 0 mV) showed a wide range of amplitudes and kinetics as well as different forms of short-term plasticity. Glutamatergic and GABAergic examples are provided for every area involved in the tri-synaptic circuit.