Chapter IV: Neurons in the human amygdala selective for perceived emotion
4.5.6 A full inventory of neurons in the amygdala that encode perceptual judgment
How representative were the subsets of cells described so far of the entire population of amygdala neurons recorded? We next conducted a permutation analysis on the entire population of cells to assess the likely effect size across the population. This analysis used independent subsets of trials for cell selection and response quantification during each repetition of the permutation. We ran 1000 iterations in total. In each, we randomly selected half of the correct trials to select emotion-selective units and to classify them as either fear or happy selective. Subsequently, we calculated the response indices for the remaining half of the correct trials and all incorrect trials. Again, we calculated the population summary metric (as shown in Figure 4.10) but only using this independent
subset of trials not previously used for selecting the cells. For the null distribution, we did the same permutation test (1000 runs) with randomly shuffled trial labels. But here, we still use half of the trials to select cells and the other half to predict response indices. The complete independence between selection and prediction insured our results against biases and false positives during selection since only out-of-sample errors were calculated.
Out of our total 210 neurons recorded, we considered 185 cells with >0.2Hz firing rate for this analysis. Many cells were reliably selected over the 1000 repetitions (Figure 4.13A, upper panels; 40 and 34 cells were selected in at least 10% of runs for fear and happy conditions, respectively). In contrast, selection was random in the control condition with permuted labels (Figure 4.13A, lower panels; no cells were selected in at least 10% of the runs). Not surprisingly, there was considerable overlap between the cells consistently selected by the present split analysis and the cells selected with all trials (n = 41) as analyzed previously (Figure 4.14, upper panels). In contrast, for the permutation test which randomly shuffled labels, each cell was equally likely to be selected with a probability of 0.05 (lower panels of Figure 4.14); the selected cells were evenly distributed across all 185 cells and across permutation runs (lower panels of Figure 4.13A) and did not show a bias towards those that could be selected with all trials (Figure 4.14). On average, 16.3 ± 3.1 (mean ± SD) units (8.8% of 185) were categorized as fear-selective and 13.5 ± 2.8 (7.3% of 185) as happy-selective, above the chance estimate of 9.25 cells for each category (p < 0.01 for fear-selective and p = 0.077 for happy-selective; Figure 4.13B). In contrast, the control permutation test resulted in 9.2 ± 3.0 units that were fear-selective and 9.4 ± 2.8 units that were happy-selective (Figure 4.13B, middle panels), with no difference between the two categories (p = 0.14) and the chance value 9.25 (p > 0.05 for both). Furthermore, the symmetric shape of the null distribution (see Figure 4.15) showed that the permutation test was not biased.
!
Figure 4.13. Illustration of the split analysis method to compute the population response.
(A) Cells selected across runs. Each black dot means a particular cell was selected. There was substantial consistency of cells selected in the split analysis (upper panels) but cell selection was evenly distributed across cells and runs in the permutation test (lower panels). (B) Summary of the number of cells selected across all runs. Gray and red vertical line indicates the mean of the chance and actual distribution, respectively. The number of cells selected in the split analysis was well above chance while the number of cells selected in the permutation test was near chance. See Figure 4.16 for results.
!
! Fear-selective Repetition Tria l 200 400 600 800 1000 Happy-selective Fear-selective permuted Cell ID Repetition Tria l 50 100 150 200 200 400 600 800 1000 Happy-selective permuted Cell ID 50 100 150 200A
B
0 100 200 300 Fear-selective Nr Runs 0 100 200 300 Happy-selective Nr Runs 0 100 200 300 Fear-selective permuted Nr Runs 0 5 10 15 20 25 0 100 200 300 Happy-selective permuted Nr Cells Selected Nr Runs Cell ID Happy-selective 20 40 60 80 100 120 140 160 180 200 0 0.5 1 Cell ID Happy-selective permuted 20 40 60 80 100 120 140 160 180 200 0 0.05 0.1 Cell ID Fear-selective 20 40 60 80 100 120 140 160 180 200 0 0.5 1 Cell ID Fear-selective permuted 20 40 60 80 100 120 140 160 180 200 0 0.05 0.1 ProbFigure 4.14. Summary of the likelihood of each cell being selected.
In the split analysis (upper panels), cells were consistently selected and there was substantial overlap with cells selected by all trials (shown as red bars at the bottom of each color-map with probability equal to 1). But in the permutation test (lower panels), each cell was equally likely to be selected with the predetermined false discovery rate of 0.05. Also, the selection was not biased towards the cells selected by all trials.
!
We next quantified the responses of the groups of cells selected in each run using the population summary metric as described above (Figure 4.16). The population summary metric is calculated as the difference between the average of response indices from all fearful trials (either correct or incorrect) collapsed across all selected cells and the average of response indices from all happy trials (either correct or incorrect) collapsed across all selected cells (see Eq. 4.5). The population metric here combined both fear and happy selective cells. The population response was significantly different from the null distribution, for both correct trials and incorrect trials (unpaired two-tailed t-test, p < 0.0001). The distribution of the incorrect trials was shifted in the opposite direction relative to the distribution of the correct trials. This also held separately for fear and happy selective neurons (see Figure 4.15 for population metric distributions separately for fear and happy selective neurons). Thus, the neural signals always followed the behavioral response instead of stimulus ground truth, regardless of whether the behavioral response was correct or incorrect. We thus conclude that emotion-selective neurons in the amygdala encode perceptual judgment robustly.
!
Figure 4.15. Population summary metric for amygdala neurons separately for fear- and happy- selective neurons.
(A,C) Fear-selective neurons. (B,D) Happy-selective neurons. (A,B) Histogram representation. (C,D) Cumulative distribution function (CDF) representation. The gray distribution is the null distribution from permutation tests. The red distribution is from correct trials and the blue distribution is from incorrect trials. Both the red distribution and blue distribution shifted away from the null distribution. Importantly, the blue distribution was on the opposite side of the red distribution.
!
A B
Correct Incorrect Permuted
Nr Nr
Probability Probability