ISLET: individual-specific reference panel recovery improves cell-type-specific inference

Abstract

We propose a statistical framework ISLET to infer individual-specific and cell-type-specific transcriptome reference panels. ISLET models the repeatedly measured bulk gene expression data, to optimize the usage of shared information within each subject. ISLET is the first available method to achieve individual-specific reference estimation in repeated samples. Using simulation studies, we show outstanding performance of ISLET in the reference estimation and downstream cell-type-specific differentially expressed genes testing. We apply ISLET to longitudinal transcriptomes profiled from blood samples in a large observational study of young children and confirm the cell-type-specific gene signatures for pancreatic islet autoantibody. ISLET is available at https://bioconductor.org/packages/ISLET .

Keywords: Cell-type-specific differential expression; Deconvolution; Individual-specific reference panel; Temporal measures.

Fig. 1

An overview of our proposed method ISLET (Individual Specific celL typE referencing Tool). A ISLET takes repeatedly measured bulk RNA-seq data, cell type proportions (known or estimated), and disease status as the algorithm input. Additional covariates are optional. B By a hierarchical mixed-effect modeling, ISLET can iteratively retrieve individual-specific and cell-type-specific gene expression reference panels. The fixed effect is the group-level average and the random effect is the individual-level deviance from the group mean. C Given the individual-specific reference panel, ISLET can conduct test to identify cell-type-specific differentially expressed genes (csDEG)

Fig. 2

ISLET accurately estimates individual-specific gene expression reference panels. A Scatterplot showing ISLET estimated reference panel versus the true reference panel, in log scale, when using the true cell type proportions as the input. B Scatterplot similar to panel A but using the estimated cell type proportions as the input. C Normalized mean squared error (NMSE) in reference panel estimation, from three methods: TOAST, TCA, and ISLET, at various effect sizes. Log fold change (LFC) used in simulation: 0, 0.5, 0.75, 1, 1.25, and 1.5. D NMSE at various samples sizes per group (25, 50, 75, and 100), comparing three methods. E NMSE stratified by gene expression level (low expression: $\le$160 and high expression: > 160). N = 20 simulations are conducted in each scenario

Fig. 3

ISLET improves testing accuracy in cell-type-specific differentially expressed genes (csDEG) identification. A True discovery rate (TDR), receiver operating characteristic (ROC), sensitivity versus false discovery rate (FDR), and sensitivities at various effect sizes are shown from left to right. Results for one cell type, at sample size of N = 25 subjects per group with log fold change (LFC) = 0.5, are shown for the first three panels. Intended FDR level of 0.1 (vertical line) in the third panel. The fourth panel shows the averaged sensitivity across all cell types, at various LFCs. Benchmarked methods include CARseq, TOAST, TCA, DESeq2, and csSAM. B Displaying the same metrics as in A except sample size per group is N = 75. C Averaged AUC values across all cell types, at various combinations of sample sizes and LFCs. $N=20$ simulations are conducted for each simulation scenario described above

Fig. 4

Simulation studies for type I error control under the null. A Histograms of the observed p-values under the null, when no gene is csDEG. The p-values shown above histograms are from the Kolmogorov-Smirnov (KS) test under the null hypothesis, for which the contrasted distribution is Uniform [0, 1]. The larger the p-value, the more uniform the distribution is. B Quantile-quantile plots of the same p-values as in A but on the $-{log}_{10}$ scale. Methods return well-calibrated small p-values will stay close to the red diagonal line. $N=20$ simulations are conducted each setting

Fig. 5

Application of ISLET in TEDDY bulk RNA-seq data. A Heatmap of p-values for pathway enrichment analysis on csDE results from ISLET, DESeq2, and TOAST. B Dynamics of csDEG identified by ISLET slope test