High-throughput genomic profiling of tumor-infiltrating leukocytes Presented by: Bryan Shaw

Background TIL are a part of tumors and are well know to have major roles in cancer biology TIL are vital in tumor growth, cancer progression and response to therapy.

Figure 1. Current and emerging techniques for evaluating TIL composition in solid tumors. Aaron M Newman, Ash A Alizadeh High-throughput genomic profiling of tumor-infiltrating leukocytes Current Opinion in Immunology, Volume 41, 2016, 77–84 http://dx.doi.org/10.1016/j.coi.2016.06.006

In silico approaches for TIL profiling Enrichment measures for genes associated with individual cell types Leukocyte have established markers for FACS(Fluorescence-activated cell sorting) or IHC (immunohistochemistry) Algorithmic deconvolution of admixed transcriptomes to resolve composition Generating GEP (gene expression profiles) Derive additional marker genes using in silico nanodissection, a novel machine learning technique for predicting cell type-specific genes from GEP mixture data 

Class Method Input: Marker genes Input: Signature matrix Output: Gene enrichment Output: Cell type proportions Robust to >50% unknown content? Robust to random noise or multi-cell type perturbation? Marker gene expression obtained from mixture samples? Resolution of closely related cell types? Significance analysis for gene enrichment or cell type proportions? Enrichment methods Cluster analysis NA ND N TIL meta-genes Y SPEC ssGSEA Deconvolution methods DSA MMAD CIBERSORT LLSR PERT QP

Marker gene enrichment Using gene clusters is not useful for assigning marker genes as they are sensitive to noise Gene enrichment is not good for discriminating cell subsets but is good for inferring there are some Single Sample Gene Set Enrichment Analysis uses preranked GSEA to analyze gene sets

Gene expression deconvolution Gene expression deconvolution is an emerging technique analogous to in silico flow cytometry The deconvolution methods aim to computationally resolve a GEP into its component cell types (virtual tissue dissection) Expression Deconvolution Requires a signature matrix consisting of marker genes and their expression values Also requires a biological mixture There is also a vector which contains the cell subset of the mixture in the signature matrix m = G x f

Gene expression deconvolution These methods designed to analyze peripheral blood Abbas et al created an iterative least squares approach solving for f with a new signature matrix creation strategy Gong expanded on this work using quadratic programming Using non-negative constraints (cell fractions >0) able to demonstrate advantages of this method

Cell deconvolution for solid tumors Tumors are heterogeneous mixtures of cell types Very challenging requires deconvo methods to be extremely robust due to unknown cell types Digital Sorting Algorithm helped to reduce biological noise infer the expression levels of user-specified marker genes directly from RNA mixture samples Similar cell types are able to be given coefficients and grouped together to more accurately reflect their proportions in sample

Cell deconvolution for solid tumors Using CIBERSORT to enumarate cell proportions based on vector regression Uses linear loss function and L2-norm regulation making it robust to micture content , noise, and highly correlated cell subsets CIBERSORT was able to reveal complex associations between 22 leukocytes subsets from a GEP of 25 tumor types and thousands of tumor samples

Calibration of in silico TIL profiling methods Establish baseline performance, use marker genes (and signature matrices) be validated by analyzing independently generated GEPs the same leukocyte phenotypes different leukocyte phenotypes samples devoid of immune content Deconvolution methods should be robust by resolving all cell types within a given signature matrix; if external datasets are not available, this can be accomplished by leave-one-out cross validation Comparisons against flow cytometry, expression profiling performed on RNA to avoid alterations in cellular representation

Outstanding issues Maximum phenotype distinction is open Results depend on Features of the selected algorithm Robustness of the marker genes Level of unknown content and/or noise Expression profiling platform all methods are expected to benefit from the lower noise levels and increased dynamic range of RNA-seq

Outlook “In silico tissue dissection is an emerging class of techniques for large- scale characterization of tumor cellular heterogeneity…When integrated with complementary data, including somatic and epigenetic alterations, B-cell and T-cell receptor profiles, and imaging, in silico tissue dissection will lead to more comprehensive portraits of human tumors. In the near future, we expect such analyses to transform cancer therapy and management.”