Chemical Fingerprints of Emotional Body Odor

Chemical communication is common among animals. In humans, the chemical basis of social communication has remained a black box, despite psychological and neural research showing distinctive physiological, behavioral, and neural consequences of body odors emitted during emotional states like fear and happiness. We used a multidisciplinary approach to examine whether molecular cues could be associated with an emotional state in the emitter. Our research revealed that the volatile molecules transmitting different emotions to perceivers also have objectively different chemical properties. Chemical analysis of underarm sweat collected from the same donors in fearful, happy, and emotionally neutral states was conducted using untargeted two-dimensional (GC×GC) coupled with time of flight (ToF) MS-based profiling. Based on the multivariate statistical analyses, we find that the pattern of chemical volatiles (N = 1655 peaks) associated with fearful state is clearly different from that associated with (pleasant) neutral state. Happy sweat is also significantly different from the other states, chemically, but shows a bipolar pattern of overlap with fearful as well as neutral state. Candidate chemical classes associated with emotional and neutral sweat have been identified, specifically, linear aldehydes, ketones, esters, and cyclic molecules (5 rings). This research constitutes a first step toward identifying the chemical fingerprints of emotion.

S2B. Additional testing to explore bipolar pattern in happy chemical cluster: statistical analyses To assess correspondence between chemical clusters and experimental conditions, as a first step, a statistical model was built that can accurately and reliably distinguish between fear and neutral conditions. As a second step, this model was then used to distinguish between the HF and HN sub-clusters. The reasoning behind it was that if a model that can use emotional states to distinguish between emotion conditions can similarly distinguish between chemical clusters, then it follows that both experimental conditions and chemical clusters follow similar emotional patterns. In other words, donors who showed overlap in terms of chemical composition of sweat between happy and fearful conditions, possibly also showed such overlap in profile of pertinent emotional states experienced during happy and fearful conditions. Cross-validated regularized models, optimal for doing an exploratory analysis that uses many variables and intended to generalize to new/unseen data [1], were used. Tuning parameters, as used in regularization, dictate how many variables can be used in a regression model, and how large the model's slopes can be [1]. Whereas cross-validation involves building a model using parts of the data, and testing it on the remaining data [1]. By combining cross-validation and regularization, it is possible to find optimal tuning parameters, which create models that generalize to unseen data. It is recommended to use two values of tuning parameters: one in which the model performs best on unseen data, and one which results in the simplest model whose accuracy is no more than 1 standard error from the most accurate model [1]. Models based on these optimal tuning parameters are then tested on the unused 'test' subset, and if they perform well, they can be said to accurately and reliably predict experimental conditions.
Only single-item emotional variables were used in the analysis, due to the results from PCA being inconclusive. 1.
Step 1: As a first step in the analysis, cross-validated regularized logistic regression was used to distinguish between conditions using the different emotional states. Then, the models suggested by the regression were applied to the chemical clustering data. Data from fear and neutral conditions were divided into two subsets, with each subset holding one observation per participant -from one of the conditions -which was randomly assigned per subset. This was done to better approximate the statistical constraints of the chemical clustering data, where there is only one observation per participant. Then, using only one of the two subsets, the 'training set', a 10-fold cross validation was used to assess which tuning parameter is to be used for regularization. Following this, the most accurate and least complex models that were built using the training set were tested on the testing set. Those that performed well were then used to predict clustering data.

2.
Step 2: Following model selection, the models that best distinguish between fear and neutral emotion conditions at the test subset were used to predict chemical sub-clusters HF and HN. If they did with high accuracy, then it was concluded that chemical clustering and experimental conditions follow similar emotional patterns. If they did not a cross-validated regularized logistic regression would be performed on the chemical clustering data. In that case the emotional patterns that distinguish chemical clusters were assessed using additional data on experimental conditions. If the affective patterns that distinguish between chemical clusters do not similarly distinguish between experimental conditions, and vice versa, then chemical clusters and experimental conditions do not follow the same emotional patterns.
S3: Study to address social factor during happiness induction.
After the completion of the study reported in the main text, we conducted a separate study comparing self-reported emotions of 24 male donors who were individually seated in both the happiness and neutral condition, against the results from 84 male donors collated from 4 previous studies with 2-4 males seated together in the happiness condition. This was done to explore whether emotions experienced from emotion induction of happiness are different depending on whether males in the happy condition are seated alone or together. We also addressed another limitation in that study by counterbalancing condition order (happy first, neutral first). As can be seen in Figure S3 no differences were found between the new study and the combined data from the previous studies on self-reported emotion. Figure S3. Scatter plot of self-reported emotions in happiness condition (top panel) and neutral (bottom panel) on PC1 and PC2 comparing scores from 4 studies (n=84) in which multiple men sat together and happiness condition was always second (pink) vs. a study (n=24) in which men were sitting alone and condition was counterbalanced (with happiness condition being first or second; turquoise).