Figure 1. MIEL compares the epigenetic landscape of multiple cell populations and can be used to detect active epigenetic drugs across cell lines and drug concentrations.

(A) Flowchart of MIEL pipeline. Fixed cells were immunostained for the desired epigenetic modifications and imaged. Nuclei were segmented based on DNA staining (Hoechst 33342 or DAPI) and texture features were calculated from the pattern of immunofluorescence. The relative similarity of multiple cell populations was assessed by calculating the multi-parametric Euclidean distance between populations centers, and represented in 2D following MDS (distance map). Discriminant analysis is used to functionally classify whole cell populations based on their multi-parametric centers. SVM classification is used to separate single cells in each population and estimate populations overlap. (B) Table showing the fraction of epigenetic drugs in each functional category identified as active by either MIEL analysis employing texture features derived from images of GBM2 cells stained for H3K9me3, H3K4me1, H3K27ac, H3K27me3, or by intensity-based analysis using the same modifications (see Materials and methods). (C,D) Quadratic discriminant analysis using texture features derived from images of GBM2 cells treated with either DMSO or 85 active compounds (two technical replicates per compound; 38 DMSO replicates) stained for H3K9me3, H3K27me3, H3K4me1, H3K27ac. (C) Scatter plots depicting the first two discriminant factors derived from features of all four histone modification images for each cell population. (D) Confusion matrix showing classification results of discriminant analysis. Left column details number of compounds or DMSO replicates for each category in the test set (one replicate per compound). Numbers represent the percent of compounds classified correctly (diagonal) and incorrectly (off the diagonal).

Figure 1—figure supplement 1. Overview of the epigenetic drug library used in this study.

(A) Pi chart showing functional classes of epigenetic drugs used in the study. (B) Table detailing the molecular targets of epigenetic drugs used in the study.

Figure 1—figure supplement 2. MIEL improves detection rates and effect size compared to intensity based methods.

(A) Table showing the fraction of epigenetic drugs in each functional category identified as active by either MIEL analysis employing texture features derived from images of GBM2 cells stained for either H3K9me3, H3K4me1, H3K27ac, H3K27me3 or by intensity-based analysis using individual modifications (see Materials and methods). (B) Bar graph depicting the average fold change in Euclidean distance from DMSO replicates induced by drugs from several functional categories as calculated using mean intensity or using texture features derived from images of individual histone modification (Mean ± SD; p-values calculated by ANOVA using Tukey’s HSD for multiple comparisons and shown in table; n for each category is shown in ‘A’).

Figure 1—figure supplement 3. Intensity based functional classification shows reduced classification accuracy compared with MIEL.

(A) Quadratic discriminant analysis using average fluorescent intensity derived from images of GBM2 cells treated with either DMSO or 85 active compounds (two technical replicates per compound; 38 DMSO replicates) stained for H3K9me3, H3K27me3, H3K4me1, H3K27ac. Scatter plots depicting the first two discriminant factors derived from features of all four histone modification images for each cell population. Confusion matrix showing classification results of the discriminant analysis. Left column details number of compounds or DMSO replicates for each category in the test set (one replicate per compound). Numbers represent the percent of compounds classified correctly (diagonal) and incorrectly (off the diagonal). (B) Quadratic discriminant analysis using texture features derived from images of GBM2 cells treated with either DMSO or 85 active compounds (two technical replicates per compound; 38 DMSO replicates) stained for H3K9me3, H3K27me3, H3K4me1, H3K27ac or DAPI. Confusion matrix showing classification results from discriminant analysis. Number of compounds or DMSO replicates for each category (one replicate per compound) is as shown in ‘b’. Numbers represent the percent of compounds classified correctly (diagonal) and incorrectly (off the diagonal).

Figure 1—figure supplement 4. MIEL can distinguish between HDAC inhibitors with different specificity.

(A) Bar graph showing expression level (FPKM) of HDAC and SIRT genes in GBM2 cells obtained from RNA sequencing. (B, C) Quadratic discriminant analysis using texture features derived from images of GBM2 cells treated with either DMSO or 45 active compounds (two replicates per compound, 38 DMSO replicates) and stained for H3K9me3, H3K27me3, H3K4me1, H3K27ac. (B) Scatter plot depicting the first two discriminant factors derived from features of all histone modification images for each cell population. (C) Confusion matrix showing classification results for the discriminant analysis (test set: one replicate per compound; 10 DMSO replicates). Numbers represent the percent of correctly (diagonal) and incorrectly (off the diagonal) classified cell populations.

Figure 1—figure supplement 5. Low concentration valproic acid treatment induces epigenetic and transcriptomic changes distinct from that of known HDAC inhibitors.

(A) Bar graph showing average fold change in average intensity resulting from 24 hr treatment of GBM2 cells with DMSO, valproic acid (15 µM), SAHA (3 µM) or TSA (1 µM) (Mean ± SD; n = 6 technical replicates). (B) Heat maps showing log2 of fold change in expression (RNA sequencing) of select differentially expressed genes. (C) Distance map depicting the relative Euclidean distances between the multiparametric centroids GBM2 cells treated for 24 hr with either DMSO, valproic acid (15 µM), SAHA (3 µM) or TSA (1 µM). Left: Distances calculated using texture features derived from images of H3K9me3, H3K27ac and H3K27me3 marks. Right: Distances calculated using FPKM values of all expressed genes (13,119 genes; FPKM > 1 in at least one sample). R denotes Pearson correlation coefficient.