In-silico immune cell deconvolution of the airway proteomes of infants with pneumonia reveals a link between reduced airway eosinophils and an increased risk of mortality

Charles J Sande; Jacqueline M Waeni; James M Njunge; Martin N Mutunga; Elijah Gicheru; Nelson K Kibinge; Agnes Gwela

To determine whether the deconvolution proteome data set contained sufficient resolution to distinguish individual immune cell phenotypes, we used protein expression levels from the data set to visualise phenotype segregation by nonmetric multidimensional scaling (NMDS) . The analysis showed that major immune cell phenotypes including B-cells, T cells, natural killer (NK) cells, dendritic cells(DC), monocytes (MO), basophils, eosinophils and neutrophils could be distinctly segregated on the basis of differential protein expression (figure 2a). We then set out to identify individual proteins that could be used to accurately distinguish major and sub immune phenotypes (sub phenotypes defined as lower-level hierarchies of the major phenotypes -e.g. plasmacytoid and myeloid DCs within the major DC phenotype) within the deconvolution data set. Using the recursive feature elimination-paired random forest (RF) algorithm, boruta 13 , we identified classification features for 33 major and sub immune cell phenotypes (supplementary table 1 ). Since only a subset of these feature proteins are likely to be expressed within the specific context of the inflamed airway, we set out to determine which of the features identified by the model were also expressed in the airways of children admitted with pneumonia. For some cell types such as monocytes/macrophages and neutrophils, a substantial proportion (>30%) of the phenotype classification features from the feature selection model were also expressed in the airway, while for others (e.g. NK cells), a lower proportion of the classifiers were Next, we undertook a more detailed analysis of the phenotype classification features that were identified by machine learning to determine whether they could be used to recapitulate the phenotypes in an unsupervised prediction analysis. We reasoned that the classification features of a particular phenotype would be broadly related to its functional properties and that when they are subjected to an independent unsupervised enrichment analysis, they would successfully predict the phenotypes from which they were initially derived. Using the unsupervised enrichment platform, enrichR 14 To enhance the power of the identified features to resolve the constituent cell phenotypes of a complex sample, we generated a phenotype classification profile, in which all the classification features of a particular phenotype were aggregated and their combined expression was plotted relative to all other phenotypes. An example of this analysis is shown in figure 3D , where the combined expression of all monocyte-specific features (plotted in red) was compared to the expression level of the same proteins in all other phenotypes (plotted in gray). The results of this analysis showed that these proteins were expressed at a significantly higher level in monocytes compared to all other phenotypes (p<0.0001). Similar analysis for all other phenotypes is presented in supplementary figure 1. We then used t-SNE dimensional reduction analysis to visualise the segregation of major and sub phenotypes on the basis of the phenotype classification features (figure 4A). The results of this analysis showed that the feature classifier proteins could clearly distinguish most immune cell types; phenotypes like plasma cells (B.plasma), pDC, mDC, eosinophils, basophils, neutrophils and different monocyte sub phenotypes were clearly distinguishable from the rest of the phenotypes. Some subphenotypes such as cen