DeepDAM inference framework for monosynaptic connectivity inference in CA1 networks.

(a) Challenges associated with inferring connectivity, including incomplete observations and strong recurrent connections. (b) DeepDAM framework. The gray boxes outlined the five major steps of the framework. Note that both the feature extractor and the classifier are shared between synthetic and experimental data. (c) Distributional discrepancy of experimental and synthetic data caused by model mismatch in the model construction step. (d) Synthetic and experimental feature distributions after training the traitional model-based inference framework. (e) Synthetic and experimental feature distributions after training our DeepDAM framework.

DeepDAM framework architecture.

(a) Workflow of the DeepDAM framework. The temporal feature extractor and regressor/classifier are shared between synthetic data and experimental data. Firstly, temporal features of synthetic data and experimental data are extracted by the shared temporal feature extractor. The output space from the temporal feature extractor is the feature space. Secondly, the shared regressor or classifier processes the temporal features and outputs the inferred neural properties. Sup. learning stands for supervised learning. (b) The training workflow of our inference framework. Each dot or cross represented a sample feature. (c) Pseudo code. The equation IDs correspond to the IDs in the main text and d. (d) The function of each unified backbone equation.

Training workflow of our inference framework.

(a) The experimental data and computational model for inferring the existence of monosynaptic connectivity between pairs of neurons. (b) Evolution of feature representations across the training process for both our inference method and the vanilla deep learning method without self-training and domain adaptation. The same fifty randomly selected synthetic connection and non-connection samples are shown in each subplot. (c) The evolution of distance matrix of synthetic and experimental features across the training process for both our inference method and the vanilla deep learning method without self-training and domain adaptation. Distance between two groups was calculated as the average Euclidean distance between each pair of features.

Performance comparison on experimental dataset with ground-truth labels.

(a) Example experimental CCGs and predictions by different methods. Ours: our aggregated version. V-DNN: the vanilla deep learning method. V-DNN-Agg: the aggregated vanilla deep learning method. CCGs were plotted from [-20, 20] ms based on the neuron spike timing with the time bin of 0.2 ms. (b) Performance comparison between 6 different methods. Error bar: standard deviation across 10 random seeds.

Performance stability across different experimental conditions and computational model structures.

(a) Experimental CCG of the same pair of neurons at different recording lengths. (b) MCC, false positives, false negatives of our aggregated method at different recording lengths. Error bar: standard deviation across 10 random seeds. (c) Synthetic CCGs generated from MAT neural networks of different sizes. (d) MCC, false positives, false negatives of our aggregated method at different sizes of MAT neural networks. Error bar: standard deviation across 10 random seeds.

Consistency between the prediction of our aggregated method and two label methods.

(a) All CCGs with evoked spikes in the large dataset. For labeling, the CCGs were calculated based on ‘presynaptic’ spikes that were evoked via juxtacellular stimulation. Spontaneous spikes of the ‘presynaptic’ neuron were discarded. The label is given by label method 1. Color bar denotes the spike count, normalized by the maximal value of all CCGs. (b) Illustration of two label methods. The green line represents the slowly co-modulated baseline (SI Appendix). Briefly, label method 1 tests the lowest p-value of time bins in both directions (pcausal and pfast) while label method 2 tests the p-value of three consecutive time bins in one direction (SI Appendix). (c) Experimental CCGs and predictions of our aggregated method and two label methods. For inference, CCGs were calculated using spontaneous spikes of the ‘presynaptic’ neuron, while evoked spikes and the corresponding stimulus window was discarded from the CCGs. The numbers in the parentheses indicate the count of random seeds out of 10 that resulted in the same inference outcome. (d) MCC between our aggregated method and two label methods on the amibigous dataset. (e) Consistency matrix between our aggregated method and two ground-truth tagging methods on the amibigous dataset. For example, the upper-left grid in the left panel represented the percentage of experimental CCGs in the dataset that both our aggregated method and label method 1 inferred as non-connections.