Fig 2: Identification of a reduced set of TRs. (a) PCA (FactoMineR R package) on TR log2-expression across cell lines in each dataset. In the upper panels the scree plots show the selection of the 5 first principal components (37 to 44% of the total variance). The lower panels show the projection of the 50 best TRs taken as variables on the 2 first principal axes in each dataset. For each dataset, absolute value of correlation coefficient > 0.5 with each of the axes was taken as meaningful. (b) Venn diagram showing the selected TRs common to the 2 datasets. (c) Distributions of mean expression and of expression variation coefficient of the 318 selected TRs across cells in each dataset associated with boxplots. They show moderate expression of the majority of TRs with variation coefficient across cells mostly between 20 and 50% (inter-quartile range). (d) Boxplots showing ESRRA expression and variability in the BC cells of CCLE and GSE58135 datasets. (c, d) Expression is log2 of upper-quartile normalized TPM (for CCLE) or FPKM (for GSE58135) values.

Fig 3: Model features across all ERRa-activated genes and all replicated analyses using two expression datasets. (a) Flowchart of the modeling procedure used to select TRs. (b) Boxplots showing individual model R-squared values and the number of TRs selected by the model, i.e. with non-0 coefficient, among 318 ones in each model computed 10 times for 68 and 63 genes with the CCLE and GSE58135 dataset respectively. Red lines indicate the thresholds used for further selection (R2 > 0.6 and number of selected TRs > 159). The lower panel shows histograms of ESRRA coefficient values in models. Most of the values are in the interval ]-0.05, 0] leading to 206/680 (30%) and 177/630 (28%) models including ESRRA with a positive coefficient for CCLE and GSE58135 data respectively. The green rectangles indicate positive ESRRA coefficients. (c) Example of model results for one gene with the two datasets: scatterplots of predicted vs true gene expression and values of TR coefficients given by the model. Data are log2-transformed and the identity line is shown. The TRs selected in one model have non-0 coefficient values and include ESRRA with a positive coefficient. (d) ERRa-activated genes with suitable expression model in CCLE dataset: number of model replicates giving suitable results and expression of these 17 genes given as boxplot across the 51 breast cancer cells. (e) Model characteristics across the 10 replicates for the 17 ERRa-activated genes with at least one suitable expression model in CCLE dataset.

Fig 4: Transcriptional regulators associated with ESRRA in sPLS expression models of ERRa-activated genes. (a) Venn diagram of TRs with model coefficient > 0 and p-value < 0.05 identified from the ERRa-activated genes with reliable sPLS models in CCLE and GSE58135 datasets. The 24 selected TRs commonly identified in the two datasets are ordered according to the mean proportion of genes (mean gene prop, > 0.2) over the CCLE and the GSE58135 datasets for which the TR has positive coefficient in the 10 model replicates computed with one dataset. HAT: histone acetyltransferase, FACT: facilitates chromatin transcription, HMT: histone methyltransferase, chrom: chromatin, ACF: ATP-dependent chromatin assembly factor, TF: transcription factor. (b) Expression boxplots for the 24 selected TRs (same order as table above). Expression is upper-quartile normalized by sample and log2-transformed. (c) Genes including at least one of MYBBP1A, DDX21 or NFKB1 as selected TR in addition to ESRRA in their sPLS models.

Fig 5: Comparison of the expression level of ESRRA mRNA after co-culture in IMR-90 cells in the empty insert (control) (A,B) and the insert with lung cancer cells [NCI-H1703 (C,D) and NCI-H522 (E,F)], * p = 0.05, ** p = 0.005.