Response letter to Oka et al.’s letter to the editor

Oka et al. appreciate the integrated design of our analyses¹, which combined case and control experiments using in vitro human-derived fibroblasts and in vivo behavioral assays in a mouse model. To characterize the biochemical complexity of LPV, we performed lipidomic profiling of vulvar fibroblasts and identified 122 unique lipid mediators. Principal component analysis (PCA) was used solely as a descriptive, visual tool to summarize similarities in lipid mediator profiles across treatment conditions by reducing these variables into two or three dimensions. Specifically, no hypothesis tests, correlations, or models were built upon the PCA analysis.

The authors request that the percentage of variation explained by the principal components should be reported and this was printed in Figure 3B, but left off of Figure 3A. A reproduction of Figure 3A with the PCA axes labeled with their percentage of explained variation appears here (Fig. 1). With 2 principal components, 41.1% of the variability in lipids are explained while 3 components bring the total explained to 50.5%. A scree plot indicates that the first principal component contains the most substantial share of variation explained.

Fig. 1.

2D principal component analysis (PCA) graph for the profiles of 122 lipids detected in human vulvar fibroblast with the PCA axes labeled with their percentage of explained variation. Strains were treated for 48 h with IL-1β [500 pg/mL] (red) or a combination of IL-1β [500 pg/mL] and Ibuprofen [10 μg/mL] (green), Celecoxib [1 μM] (blue), or Zileuton [10 μM] (purple).

All inferential analysis presented employ feature selection for biologically meaningful lipids, as these authors suggest. We separately analyzed prostaglandins, HETEs, HODEs, and specialized pro-resolving lipid mediators.

We attempted to use Feature agglomeration networks or optimized LOESS regression with positive ratio as recommended by these authors^2,3 for implementing them in our current context without substantial analytical and programming effort. The first paper is a method for finding spatially cohesive groups of pixels that contain faces. There are no spatial relationships or other known networks in the lipid features which allow this method to be applied to our data. The second paper is a method of choosing 1000 genes from a single cell gene expression experiment. There is no simple corollary to the concept of positive ratio in the gene expression example; in single-cell RNA, an expression of 1 or more means that the single-cell expression is detectable. In contrast, the lipid measures have a range of measured values from 0.0003 to 156. Still, an alternative threshold could be chosen, but the algorithm as described orders the features by their residuals and chooses the top 1000. We have only 122 lipids and any choice of how many to choose for this would need to be based on wanting a specific number of lipids to analyze. We have already chosen specific lipids to analyze in our paper because of their biological importance and have not used any feature selection or dimension reduction in our inference comparing the effects of NSAID treatment on lipid mediators.

In summary, we appreciate the careful and enthusiastic reading of our work and agree that further analysis of our lipid data using non-linear methods is an avenue for future investigation. For the small-sample experiments presented here, however, we believe that focusing on individual lipid mediators of biological interest is the most informative analytical approach.