Computational analysis of the immune infiltrate in tumor biopsies

VILARIÑO M; FERNÁNDEZ E; MAHMOUD Y,; HILL M; GIROTTI MR.; VEIGAS F,; RABINOVICH GA

Conferencia; IUIS ImmunoInformatics Course and Conference.; 2019

Melanoma is the deadliest form of skin cancer. Targeted therapies and immunotherapies have revolutionized melanoma treatment showing unprecedented survival benefits. However, responses are not universal and patients who initially respond to therapy develop resistance. Therefore, establishing which patients would derive clinical benefit from immunotherapy is a compelling clinical question. We aimed to determine predictive molecular biomarkers of response to immunotherapy and to quantify the cellular composition of the immune infiltrate of tumor biopsies to investigate their relationship with response to treatment. We analyzed the gene expression profiles of 109 biopsies and a single-sell RNA-Seq dataset from melanoma patients treated with anti-PD1 therapy. We found that patients who respond to anti-PD1 present an activation of the inflammasome NLRP3 pathway during therapy. We implemented an established computational approach (CIBERSORT) to infer the proportion of 22 immune cells from gene expression profiles and found that NL3P3 expression correlates with the proportion of CD8 T cells, CD4 activated memory T cells and total immune infiltrate. However, this method tend to over-estimate the amount of cell-types present in the sample thus under-estimating true proportions, biasing the results. We developed MIXTURE, a noise-constrained recursive feature selection for support vector regression that overcomes such limitations. MIXTURE was evaluated over simulated and benchmark data sets overcoming limitations of competitive methods in terms of accuracy and estimation bias for both: the true number of present cell types and their proportions in the sample as well as showing superior robustness to collinearity problems. Applying this method in the paired biopsies of melanoma patients we could identify associations between immune infiltrate and outcome that were not revealed by previous methods such as a decreased proportion of M2-macrophages in responders to anti-PD1 therapy. Unlike CIBERSORT estimates, MIXTURE did not reveal significative changes in certain cell populations such as CD4 resting memory T cells or plasma cells and these differences seem to be because our method handles better with collinearity and overestimation of lower proportions. Finally, we present the MIXTURE web application publicly available at: https://cidie-conicet-ucc.shinyapps.io/mixture