Divergence in DNA Specificity among Paralogous Transcription Factors Contributes to Their Differential In Vivo Binding

Abstract

Paralogous transcription factors (TFs) are oftentimes reported to have identical DNA-binding motifs, despite the fact that they perform distinct regulatory functions. Differential genomic targeting by paralogous TFs is generally assumed to be due to interactions with protein co-factors or the chromatin environment. Using a computational-experimental framework called iMADS (integrative modeling and analysis of differential specificity), we show that, contrary to previous assumptions, paralogous TFs bind differently to genomic target sites even in vitro. We used iMADS to quantify, model, and analyze specificity differences between 11 TFs from 4 protein families. We found that paralogous TFs have diverged mainly at medium- and low-affinity sites, which are poorly captured by current motif models. We identify sequence and shape features differentially preferred by paralogous TFs, and we show that the intrinsic differences in specificity among paralogous TFs contribute to their differential in vivo binding. Thus, our study represents a step forward in deciphering the molecular mechanisms of differential specificity in TF families.

Keywords: E2F; ETS; RUNX; bHLH; differential protein-DNA binding; divergence in DNA specificity; genetic variation; iMADS; paralogous transcription factors.

Figure 1. Paralogous TFs with Indistinguishable PWMs Show Distinct In Vitro Specificities

(A) Examples of paralogous TF pairs with indistinguishable PWMs: Myc versus Mad (bHLH family), Elk1 versus Ets1 (ETS family), E2f1 versus E2f4 (E2F family), and Runx1 versus Runx2 (RUNX family). Similar PWMs derived from in vivo data are shown in Figure S1. (B) Design of a gcPBM library containing putative binding sites for paralogous TFs, selected from unique (red and blue) and overlapping (purple) in vivo genomic targets, as identified by ChIP-seq. We used the gcPBM assay to quantitatively measure in vitro binding of TF1 and TF2 to all selected genomic sites. Figure S2 provides a more detailed description of gcPBM assays. (C) Direct comparisons of the binding specificities of paralogous TFs for genomic sites (top panels), in contrast to comparisons between replicate gcPBM datasets (bottom panels). Each point in the scatterplot corresponds to a 36-bp genomic region tested by gcPBM. Black points correspond to genomic sequences centered on putative TF binding sites. Gray points correspond to negative control regions, i.e., sequences without binding sites. (D) Squared Pearson correlation coefficient (R²) for all pairs of paralogous TFs in our study (colored bars) and for representative replicate experiments (gray bars). The lower panels show example 3D structures for one TF in each family, as reported in the PDB (Rose et al., 2015). From left to right: Mad:Max (PDB: 1NLW), Ets1 (PDB: 2STT), E2F4:DP2 (PDB: 1CF7), and Runx1 (PDB: 1HJC). See Figure S3 for a comparison of the similarity in DNA binding specificity versus the amino acid identity in the DNA binding domains of paralogous TFs.

Figure 2. Core-Stratified SVR Approach to Model TF-DNA Binding Specificity

(A) To build core-stratified SVR models of TF specificity, we start with normalized gcPBM data, apply an inverse logistic transformation, separate the gcPBM probes by their core motifs, derive features (Figure S4), and train one SVR model for each core. The predictions made by the SVR models are mapped back to a 0–1 range by applying the logistic transformation (see STAR Methods). (B) The core motif sequences for each group/family of TFs are shown. Heatmaps show the DNA sequences for all gcPBM probes with binding intensity above the negative control range. (C) DNA bases that have direct interactions with TF residues, according to available X-ray crystal structures (PDB: 1NKP for Myc/Max; PDB: 2STT for Ets1; PDB: 1CF7 for E2f4/Dp2; and PDB: 1HJC for Runx1). DNA sequences tested in the crystal structure are shown. Core motifs are underlined, and bases that have direct interactions with protein residues are highlighted in bold. (D) Comparison of measured versus predicted DNA-binding specificity for Tad (from nested 5-fold cross-validation test; STAR Methods). The cores used in the core-stratified SVR model are shown. Figure S5 shows similar plots for all other TFs. (E) Prediction accuracy of core-stratified SVR models, assessed as the squared Pearson correlation coefficient (R²) between measured and predicted DNA-binding specificity (from nested 5-fold cross-validation test; STAR Methods). Gray bars show the correlation (R²) between replicate experiments. Figure S6 shows comparison with nearest-neighbor models. (F) Core-stratified SVR models (motivated by the results shown in Figure S7) can be used to make binding specificity predictions for any genomic region.

Figure 3. Modeling Differential DNA-Binding Specificity

(A) Weighted least-square regression (WLSR) is used to fit the gcPBM data of two paralogous TFs (here, Elk1 versus Ets1), and learn a linear or quadratic function $\widehat{f}$, as well as the variance ${\sigma }_{pi}^2$ at every data point (i.e., genomic site) i. (B) WLSR is used to learn the variance structure for replicate gcPBM datasets. (C) By combining the variance learned from replicate data with the WLSR model for paralogous TFs, we compute a “99% prediction band for replicate TFs” (gray), which contains genomic sites bound similarly by the two TFs. Genomic sites outside the prediction band are preferred by one of the two paralogous TFs (red, Elk1; blue, Ets1). The color intensity reflects the quantitative preference score computed according to the WLSR model. Similar plots for all paralogous TFs pairs are shown in Figure S8. (D) Fraction of genomic sites, among the sites tested by gcPBM, which are differentially preferred by the paralogous TFs in our study.

Figure 4. DNA Sequence and Shape Preferences Contribute to the Differential Specificity of Paralogous TFs

(A) Core motif GGAT shows significant specificity preference for Ets1 versus Elk1. The p value shows the enrichment of the core in Ets1-preferred sites (Mann-Whitney U test). (B) 1-mer and 2-mer sequences most differentially preferred by Elk1 or Ets1, among sites with the GGAA core. The p values were computed using the Mann-Whitney U test. (C) Left: schematic of the minor groove width (MGW) and roll structural features. Right: MGW and roll profiles for genomic sites preferred by Elk1 versus Ets1. Asterisks (*) mark the positions within the binding sites (core or flanking region) that are significantly different between the two profiles (p < 10⁻⁵; Mann-Whitney U test). Shaded regions show the 25th–75th percentile ranges at each position. See Figure S9 for comparisons of additional paralogous TF pairs.

Figure 5. In Vitro Binding Preferences of Paralogous TFs Partly Explain Their Differential In Vivo Binding

(A) Genomic region bound in vivo by Elk1, but not Ets1 (according to ChIP-seq data [Ayer et al., 1993]) contains binding sites with high preference for Elk1. (B–D) TF1 versus TF2 in vitro binding preferences, as predicted using our iMADS preference models, have a significant correlation with in vivo binding preferences, as reflected by the log ratios of TF1 versus TF2 ChIP-seq pileup signal (STAR Methods). The Spearman correlation coefficient (ρ) and its statistical significance (p value computed using the asymptotic t approximation [Best and Roberts, 1975]) is shown for each pair of TFs. Due to outlier data points, the scatterplot for E2F factors is limited to peaks with log ratios in the [–3,2.5] interval, and iMADS preference scores in the [–4,4] interval. The full set of peaks, with log ratios in the [–7.79,5.41] interval and iMADS scores in the [–5.89,6.97] interval, are available in Table S5. The full datasets (3,726 peaks for bHLH proteins, 13,004 peaks for E2F proteins, and 2,208 peaks for ETS proteins) were used to assess the correlations and to compute the best fit lines (shown in black). (E–G) Pearson correlation coefficients between the ChIP-seq pileup signals computed from replicate ChIP-seq datasets. All datasets used in this analysis show good correlation, except for the Ets1 ChIP-seq data. Additional analyses of ChIP-seq data quality are shown in Figure S10. (H) ChIP-seq data for Mad and Myc, with peaks sorted in decreasing order of the log ratio of Mad versus Myc signal. Regions of 1,000 bp centered at the peak summits are shown. The data can be used to identify “Mad-preferred” and “Myc-preferred” peaks, selected as the top and bottom N% of peaks, respectively. For different values of N, we tested how well iMADS models can distinguish between the peaks preferred by each TF. (I and J) Receiver operating characteristic (ROC) curves showing the performance of iMADS models of differential specificity, as well as PWM models trained on in vitro or in vivo data, in distinguishing Madfrom Myc-preferred peaks. In vitro PWMs were derived from the same gcPBM data used to train iMADS preference models. In vivo PWMs were trained on the ChIP-seq datasets used for testing (STAR Methods). The area under the ROC curve (AUC) is shown for each model. AUC values vary between 0 and 1, with 0.5 corresponding to a random model. Results are shown for N = 5, 10, and 30. (L and K) Similar to (I and J), but for E2F proteins E2f1 versus E2f4. (M) Similar to (I–K), but for ETS factors Elk1 and Ets1, and showing the results for N = 5 and 10. Additional results are available in Table S6. Additional analyses are shown in Figures S11 and S12.

Figure 6. Analyzing Non-coding Somatic Mutations Using iMADS Models of Specificity and Differential Specificity

(A) Example of iMADS predictions for variant rs786205688. Plots show predicted ETS sites in a 300-bp genomic region centered on the variant (chr5:74893909). Binding sites (1), (2), and (3) are present both in the wild-type (reference) and the mutant (tumor) sequences. Binding site (4) is created in the tumor by the somatic mutation. The color of the binding sites reflects the Elk1 versus Ets1 preferences, as computed by iMADS. Scatterplots show the binding specificities and preferences of sites (4) and (2) (marked by stars) compared with the genomic sites tested by gcPBM in our study. See Figure S14 for the precise steps users can follow to reproduce the predictions shown here for rs786205688. (B) Scatterplot of the absolute change in Elk1 binding score compared with the absolute change in preference score of Elk1 versus Ets1, for noncoding somatic mutations identified in melanoma cancer patients (ICGC dataset SKCA-BR). Ovals highlight sets of mutations discussed in the main text. (C) Boxplot showing the changes in preference score for non-coding somatic mutations identified in different types of tumors, compared with a control set of non-coding variants from the 1000 Genomes Project. The somatic variants were identified from either whole-genome (light orange bars) or whole-exome (dark orange bars) sequencing data from ICGC (STAR Methods). The control variants (gray bar) were randomly selected among common variants with minor allele frequency >0.01 (Zhou and Troyanskaya, 2015). For each dataset, the box shows the median change in preference score and the 25th and 75th percentiles. Whiskers extend to the most extreme data points that are no more than 1.5 times the interquartile range from the box. For all tumor types, preferences changes are significantly larger for somatic mutations than for common variants: one-sided Mann-Whitney U test p value < 2.2 × 10⁻¹⁶.

Figure 7. The iMADS Framework for Integrative Modeling and Analysis of Differential DNA-Binding Specificity between Paralogous TFs

(A) iMADS uses as input gcPBM data for the paralogous TFs, based on a DNA library that contains genomic sites bound by either TF in the cell. (B) iMADS models are trained using support vector regression (to describe the DNA-binding specificity of individual TFs) and weighted least-square regression (to describe the differential specificity between two TFs). (C) gcPBM data and iMADS models can be used to analyze sequence and shape preferences of paralogous TFs (left), to make genomic predictions of binding specificity and differential specificity (middle), and to analyze the contribution of differential in vitro specificity to differential in vivo binding (right).