Immunoinformatic methods to study the mechanisms of resistance to immunotherapy in melanoma

MAHMOUD, YAMIL; SEGOVIA, MERCEDES; VEIGAS, FLORENCIA; BALZARINI, MÓNICA; ROCHA, DARIO; PEREZ SAENZ, JUAN MANUEL; MARIÑO, KARINA; RABINOVICH, GABRIEL; HILL, MARCELO; FERNÁNDEZ, ELMER A.; GIROTTI, MARÍA ROMINA

Congreso; nternational Conference on Bioinformatics ? Iberoamerican Bioinformatics Society and the Master in Bioinformatics; 2019

PEDECIBA-UdelaR

Targeted therapies and immunotherapies have revolutionized melanomatreatment. However, responses are not universal and patients who initiallyrespond to therapy develop resistance1. We implemented computational tools toanalyze public datasets with the aim of studying predictive biomarkers ofresponse and resistance to immunotherapy. We analyzed the gene expressionprofiles of a single-cell RNA-Seq dataset of 48 biopsies (16,291 cells) and bulkRNA-Seq from 109 biopsies from melanoma patients before and duringtreatment with anti-PD-1/anti-CTLA-4 therapy2,3. We applied establishedmethods for single-cell analysis such as Seurat4 and Monocle5 to study theexpression profiles of different immune cell populations. Our analysis showedthat macrophages and CD8 T-cells of non-responders upregulate a specificglycosylation-related signature. In addition, the trajectory analysis showed theappearance of a novel sub-group of macrophages associated withimmunosuppression and resistance during treatment. On other hand, our bulkRNA-Seq analysis on melanoma biopsies show that inflammasome activation isrequired for anti-PD1 response. We used CIBERSORT6, a machine learningdeconvolution approach based on SVR, to infer the immune cell compositionfrom bulk gene expression profiles. We found that patients who respond to antiPD1 therapy present increased proportion of effector immune populationsduring treatment. However, this method tends to overestimate the amount ofcell types present in the sample thus underestimating true proportions, biasingthe results. We therefore developed MIXTURE, a new algorithm that overcomessuch limitations implementing a noise-constrained recursive feature selectionfor SVR. We evaluated MIXTURE over simulated and benchmark data setsovercoming limitations of competitive methods in terms of accuracy andestimation bias. By applying MIXTURE on melanoma biopsies, we identified associations between the immune infiltrate and outcome that were not revealedby previous methods such as a decreased proportion of immunosuppressiveM2-macrophages in responders to anti-PD1 therapy. MIXTURE is available forthe wider scientific community as a web application and as an R package7.