gms | German Medical Science

Introduction: The growing demand for molecular analyses in routine clinical settings requires bioinformatic innovations to handle exponentially increasing data complexity. Flow cytometry is widely used to analyze cell distributions on a single-cell-level, but standardized and unbiased analysis on a large scale remains a great challenge. Here, we developed a fast, scalable and user friendly system which uses machine learning to enable reproducible end-to-end cytometry analysis.

Methods: The pipeline starts with automated data annotation using user-defined metadata and panel information. Data is transformed using basic algebra and curve sketching of the corresponding histograms to automatically define intensity cutoffs between positive and negative populations. Gating can be imported from BD FACSDiva or FlowJo or calculated using supervised or semi-supervised machine learning. Then, downstream analysis of fluorescence intensity values and marker frequencies per sample and per group is performed. On a single cell level, clustering and dimensionality reduction including UMAP, T-SNE, PCA and trajectory analysis via diffusion maps are performed. Finally, a browser-based interface allows interactive data analysis without requiring knowledge in bioinformatic programming.

Results: In standardized reference data, T-cell populations of varying sizes were quantified using a total of 110 manually drawn gates. The random forest classifier outperformed other classification methods and resulted in an overall accuracy greater than 96.5%. We then prospectively phenotyped neutrophils from healthy donor blood, paired blood and synovial fluid from patients with active inflammatory arthritis and paired blood and colon biopsies from patients with inflammatory bowel diseases (total: 56 individual flow cytometry files). Machine learning identified a distinct population of neutrophils in colon tissue, characterized by reduced expression of CD16, CD62L, ICAM-1, IL17-RA and CXCR2 and an increased expression of CXCR4 in comparison to peripheral blood. This phenotype reflects a tissue adaptation of neutrophils, as it was identified irrespective of disease activity. In addition, we observed a separate, expanded population of CCR5+ neutrophils in colon tissue. Neutrophils in synovial fluid showed a similar upregulation of HLA-DR and CXCR4 compared to colon. However, upregulation of CCR5 and CD64, and downregulation of CD16 were less pronounced. In contrast, synovial fluid neutrophils showed an upregulation of ICAM-1, which was downregulated in colon. Additionally, synovial fluid neutrophils showed an upregulation of TNF-RII, which was absent in blood and colon tissue.

Conclusion: Here, we presented a software pipeline that leverages machine learning to enable automated analysis of cytometry data. When applied, this technology robustly quantified T cell populations in the blood and allowed us to identify novel, tissue-specific and inflammation associated neutrophil phenotypes.

Disclosure: There are no conflicts of interest relevant to these contents.