ICTD: A semi-supervised cell type identification and deconvolution method for multi-omics data

Abstract

Traditional deconvolution methods infer the relative proportions of predefined cell types through either regression-or enrichment-based approaches. However, there are several challenges that remain unsolved in the current formulations, including (1) identifying the Immune/Stromal (I/S) cell types that truly exists in a tissue, (2) identifying the marker genes for each cell type that are specifically expressed by one or a few I/S cell types in a TME, (3) co-linearity among I/S proportions due to their co-infiltration. We have developed a novel semisupervised deconvolution method namely ICTD (Inference of Cell Types and Deconvolution), addressing the three challenges via (i) developing a Bi-Cross Validation (BCV) based matrix rank test to assess the significance of the existence of cell types and signature genes, (ii) utilizing a constrained Non-negative Matrix Factorization (NMF) to handle co-linearity caused by co-infiltration of multiple cell types. ICTD is shown to have largely improved prediction accuracy of tissue compositions, and it is capable of identifying novel or sub-cell types with specific functional activity. M k =1,.., ×K [ , ] = 1, if gene is marker of the cell type , and 0 otherwise. The constraint matrix is then enforced upon the regular NMF formulation to guarantee similarity of the signature matrix with the constraint matrix, namely, we solve by min S,P (‖X − S • P‖ 2 + • tr(S(1 − C T ))) . (5) Estimate cell type specific functions. For each cell type detected in (4), ICTD screens the matrix rank of the expression profile of the cell type and functional marker genes through the samples of different level of the cell types. Marker genes of a varied cell type specific function are identified if they form at least one distinct dimension comparing to the cell type markers only in the samples with high level of the cell.

Effects of Molecular Iodine/Chemotherapy in the Immune Component of Breast Cancer Tumoral Microenvironment

“…Deconvolution studies were performed with CIBERSORT [19], which accurately quantifies the relative levels of different types of immune cells within a complex mixture of gene expression, and we used GED-IT to predict the cell type composition of tissue samples. We also used ICTD to deconvolute and identify immune cells [20].…”

Acid-Base Homeostasis and Implications to the Phenotypic Behaviors of Cancer

“…We have utilized our in-house deconvolution method, ICTD (identification of cell types and deconvolution) to estimate the relative proportions among 21 immune and stromal cell types in each TCGA sample (60).…”

SSMD: a semi-supervised approach for a robust cell type identification and deconvolution of mouse transcriptomics data

“…In this study, we define a cell type is "transcriptomically identifiable" if its ground-truth proportion =×% > and estimated as " =×% > have high correlation, i.e.. B =×% > , " =×% > C = 1 − and is substantially small, where " =×% > is the th row of " , ( ×% , and 8 as the number of "identifiable" cell types. A strong condition for a cell type to be identifiable is that it has uniquely expressed genes [24]. Here we provided a comprehensive mathematical derivation of the relationship between cell type unique expression and identifiability of cell proportion in the Supplementary Notes.…”