FLOW CYTOMETRIC AND ARTIFICIAL INTELLIGENCE APPROACH TO DIAGNOSTIC MARKERS FOR B-CELL LYMPHOMAS

B-cell Non-Hodgkin lymphomas (B-NHL) are a wide and highly heterogeneous group of malignancies, characterized by different morphology, phenotype, genotype, aggressiveness and response to therapy. Immunophenotypic characterization is a key element for their classification in order to guide to the correct therapeutic plan. Although the phenotypic study is routinely performed by immunohistochemistry (IHC), we verified that the use of flow cytometry (FC) could bring several advantages. Applied to samples from various sources derived from patients affected by the most common mature B-NHL, FC allows the quantitative expression of multiple markers to be evaluated by analyzing millions of cells simultaneously and easily defining clonal populations. This results in the ability to almost perfectly isolate each neoplastic subpopulation and study it in its uniqueness. This feature distinguishes FC analysis from IHC, which, by presenting the global appearance of the sample being examined, preserves the architecture of the tissue while revealing all mixed clones. In addition, like IHC, FC can provide information about intracellular antigens. In addition, FC can be applied to peripheral blood (PB) samples of leukemia stage lymphoma. In any case, FC provides faster and less biased results than IHC because the data are expressed quantitatively. However, the large amount of data generated by FC is very complex to analyze as a whole. In particular, it is difficult to correlate each marker with the precise diagnosis of a disease. Artificial intelligence (AI) can help by using sophisticated software and computing systems to compare and stratify all this data in a short time. In this way, AI provides more manageable data with a two-dimensional representation that is easier to interpret. In a previous study, we applied machine learning (ML) algorithms to a large dataset of B-NHL immunophenotypes to generate a robust and clinically applicable prediction system. This system would also allow us to overcome the time-consuming, optimize the use of antibodies and standardize a clinically applicable predictive system by establishing a panel of antibodies to be systematically used in a multiparametric immunophenotypic analysis of samples performed with a high-complexity flow cytometer.We then applied additional intracellular markers to a homogeneous case series of 615 tissue samples whose diagnoses, all confirmed by histologic analysis, were grouped into 8 major categories of B-NHL patients. The Predictive Power Score (ppscore) method allowed us to assess the impact of each marker in defining each lymphoma category. Considering that a ppscore greater than 0.22 (the baseline score) is statistically significant for discriminating diagnostic categories, we surprisingly noticed the discriminatory power of intracellular markers not commonly used in a multiparametric immunophenotypic approach to lymphoma diagnosis, such as IRF4 and Bcl6.The role of these markers was validated by combining each one with all the others in a classification tree, resulting in a structural relationship tree that separates the entire database into quasi-homogeneous groups of lymphomas. Finally, the UMAP dimensionality reduction technique, we observed that the 8 lymphoma categories were substantially grouped and separated in clusters. The results obtained demonstrate how the use of surface and intracellular markers allows us to define the major categories of B-NHL with a high degree of accuracy. 10 or less markers seem to be sufficient to achieve an adequate classification capability. Nevertheless, a greater number of markers, combining intracellular with unconventional markers (CD305, CD81), increases the ability of UMAP to separate different entities. In conclusion, it is conceivable that the implementation of AI applied to FC could contribute significantly to an optimal diagnostic process in B-NHL, where histopathological examination remains the gold standard.