Article Figures & Data
Figures
- Figure 1.
Example traces of a single-firing and multi-firing hDRG-N. A, Example traces of a single-firing hDRG-N at −100 pA, rheobase, and 400 pA over rheobase. B, Example traces of a multi-firing hDRG-N at −100 pA, rheobase, and at 100 pA over rheobase. Scale bar, 100 ms, 20 mV. C, Zoom in of rheobase trace of cells in A and B showing the short AP duration and short FSL of the single-firing hDRG-N (black) and the long AP duration and long FSL of the multi-firing hDRG-N (red).
- Figure 2.
Intrinsic properties and firing phenotypes of hDRG-N neurons in single versus multi-firing cells. Intrinsic properties and firing phenotypes were compared between single and multi-firing cells. A, Cell diameter, B, capacitance, C, resting membrane potential (RMP), D, input resistance (Rin), E, rheobase, F, normalized rheobase, G, first-spike latency (FSL), H, percent delayed firing, FSL > 100 ms, I, sag percentage at −100 pA current input (Sag_R), J, percentage of cells with visible sag, K, percentage of cells that have rebound firing. L, Percentage of cells with spontaneous activity (SA) at rest, M, percentage of cells with spontaneous activity when held at −45 mV. N, AP peak. O, AP amplitude. P, AP threshold. Q, AP After hyper-polarization (AHP). R, AP rise time, S, AP decay time, T, AP half-width. U, Max AP rise slope. V, Max AP decay slope. *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001 by Mann–Whitney for A–G, I, N–V. By Fisher's exact test for H, J–M.
- Figure 3.
Pearson’s correlation matrix of hDRG-N features. Features compared included size, capacitance (Cap), presence of rebound firing, spontaneous activity at rest (SA_rest), spontaneous activity at −45 mV (SA_45), resting membrane potential (RMP), input resistance (Rin), rheobase (Rheo), normalized rheobase (Norm_rheo), first spike latency (FSL), sag ratio (sag_R), presence of visible sag, AP peak, AP amplitude, AP threshold, AP rise time, AP decay, AP half-width, afterhyperpolarization (fAHP), rise slope (R_slope), decay slope (D_slope), and multi-firing.
- Figure 4.
Machine learning model flow chart. Analysis was performed on 94 hDRG recordings, split using Monte Carlo cross-validation (MCCV), five different models run, and features extracted. Finally, we ran this analysis again using recordings of mouse DRG (mDRG).
- Figure 5.
Performance of chosen models across randomized train–test splits. A, Accuracies of logistic regression, k-nearest neighbors, random forest, linear support vector machine, and eXtreme Gradient Boost on Monte Carlo simulation (1,000 randomly created 80-20 train–test splits) were compared with violin plots of model accuracy when trained and tested on standardized hDRG-N data. Plot represents the range of accuracies, and interior box and whisker plot shows quartiles and median. B, Table showing performance metrics of all models.
- Figure 6.
Performance of chosen models across randomized train–test splits. Performances of logistic regression, k-nearest neighbors, random forest, and linear support vector machine on Monte Carlo simulation (1,000 randomly created 80-20 train–test splits) were compared with violin plots of model accuracy when trained and tested on standardized/unstandardized dataset, hue represents model, plot is cut to represent the range of accuracies, and interior box and whisker plot show quartiles and median. A, Unstandardized dataset. B, Standardized dataset. C, Table showing performance metrics of both groups of models.
- Figure 7.
Feature importance metrics across logistic regression, random forest, and support vector classifier for standardized and unstandardized data. Mean coefficient weights or feature importance, when applicable, were extracted from the iterations in the Monte Carlo simulation. Graphs on the left represent models with unstandardized data while right use standardized. A, B, Logistic regression. C, D, Random forest. E, F, Support vector classifier.
- Figure 8.
Feature importance metrics across logistic regression, random forest, and support vector classifier for standardized data. Mean coefficient weights or feature importance, when applicable, were extracted from the iterations in the Monte Carlo simulation and ordered. Absolute value of coefficients is displayed for logistic regression and support vector classifier. A, Logistic regression. B, Random forest. C, Support vector classifier. D, XGBoost.
- Figure 9.
Monte Carlo Shapely Additive Explanations (SHAP) Beeswarm plots. Beeswarm plots for A logistic regression. B, k-nearest neighbors. C, Random forest. D, Support vector classifier. E, XGBoost.
- Figure 10.
Monte Carlo Shapely Additive Explanations (SHAP) values. A, Logistic regression. B, k-nearest neighbors. C, Random forest. D, Support vector classifier. E, XGBoost.
- Figure 11.
PCA with data normalization. A, PCA plot. B, Explained variance for each PC. C, Feature importance for PC1. D, Feature importance for PC2. E, Feature importance for PC3.
- Figure 12.
Intrinsic properties and firing phenotypes of mDRG-N neurons in single- versus multi-firing cells. Intrinsic properties and firing phenotypes were compared between single- and multi-firing cells. A, Cell diameter, B, capacitance, C, resting membrane potential (RMP), D, input resistance (Rin), E, rheobase, F, normalized rheobase, G, first-spike latency (FSL), H, percent delayed firing, FSL > 100 ms, I, sag percentage at −100 pA current input (Sag_R), J, percentage of cells with visible sag, K, percentage of cells that have rebound firing. L, Percentage of cells with spontaneous activity (SA) at rest, M, percentage of cells with spontaneous activity when held at −45 mV. N, AP peak. O, AP amplitude. P, AP threshold. Q, AP after hyper-polarization (AHP). R, AP rise time, S, AP decay time, T, AP half-width. U, Max AP rise slope. V, Max AP decay slope. *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001 by Mann–Whitney for A–G, I, N–V. By Fisher's exact test for H, J–M.
- Figure 13.
Pearson’s correlation for mDRG properties. Features compared included size, capacitance (Cap), presence of rebound firing, spontaneous activity at rest (SA_rest), spontaneous activity at −45 mV (SA_45), resting membrane potential (RMP), input resistance (Rin), rheobase (Rheo), normalized rheobase (Norm_rheo), first spike latency (FSL), sag ratio (sag_R), presence of visible sag, AP peak, AP amplitude, AP threshold, AP rise time, AP decay, AP half-width, afterhyperpolarization (fAHP), rise slope (R_slope), decay slope (D_slope), and multi-firing.
- Figure 14.
hDRG-N models are predictive of multi-firing in mDRG-N with high accuracy. Twenty mDRG-N recordings were applied to models developed for hDRG-N. We are showing the A confusion matrix for LR model. B, Confusion matrix for KNN model. C, Confusion matrix for the RF model. D, Confusion matrix for SVC model. E, Confusion matrix for XGBoost. F, Accuracy, precision, recall, and F-1 scores for mDRG-N compared with each model.
Tables
hDRG Donors Sex Age (Years) BMI Race Number of recordings 1 F 22 22 White 1 2 F 51 30 White 13 3 M 23 26 Hispanic 14 4 M 67 34 Native 7 5 M 55 19 White 2 6 M 27 24 Hispanic 2 7 M 21 22 White 13 8 F 37 18 Hispanic 18 9 M 51 27 Unknown 3 10 F 23 27 Hispanic 21 94 Total recordings from 10 donors Human DRG tissue was obtained from ethically consented, recently deceased donors. This analysis used tissues from 4 female and 6 male donors aged 21–67 years.
In this issue
