TL;DR
Paralogous transcription factors (TFs), which are closely related but perform distinct cellular functions, often appear to bind DNA identically based on current models. This study reveals that these paralogs possess subtle, intrinsic differences in DNA binding preferences that are not captured by standard motif models. These intrinsic differences, quantified by a novel computational framework (iMADS), are shown to be a significant factor in explaining why these TFs bind to different genomic locations and ultimately achieve their distinct cellular roles.
What Problem Does This Paper Address?
Transcription factors (TFs) are crucial regulators of gene expression, binding to specific DNA sequences to control which genes are turned on or off. Many TFs belong to families, meaning they share similar DNA-binding domains and often recognize very similar DNA sequences. These "paralogous" TFs, despite their similar DNA-binding motifs, frequently have distinct biological functions within the same cell. Traditionally, it was assumed that these functional differences arose from interactions with other proteins (cofactors) or the surrounding cellular environment (chromatin). This paper tackles the question of whether the TFs themselves, through subtle differences in their inherent DNA-binding preferences, can explain their differential genomic targeting and functional specificity, even when their canonical DNA-binding motifs appear identical.
Methods
The researchers employed a two-pronged approach:
- Genomic-context Protein-Binding Microarray (gcPBM) assays: This high-throughput in vitro method was used to quantitatively measure the binding of purified TF proteins to a custom library of DNA sequences. Crucially, these sequences were derived from actual genomic regions, including flanking sequences, providing a more realistic context than purely artificial sequences. This allowed for the measurement of binding across a wide range of affinities, including medium and low affinity sites that are often missed by other methods.
- iMADS (integrative Modeling and Analysis of Differential Specificity) framework: This novel computational framework integrates the gcPBM data with sophisticated modeling techniques. It includes:
- Core-stratified Support Vector Regression (SVR) models: These models are trained to predict TF binding specificity, with separate models built for different "core motifs" within a TF's binding site. This allows for capturing the nuanced contributions of both the core binding region and flanking sequences.
- Weighted Least Squares Regression (WLSR): This statistical method is used to directly compare the binding specificities of paralogous TFs and quantify the differences, establishing a "prediction band" for sites bound similarly and identifying sites that are differentially preferred by one TF over the other.
Datasets
The study utilized several datasets:
- Custom DNA libraries for gcPBM assays: These libraries were designed to contain putative TF binding sites from the native genomic context of 11 paralogous TFs from 4 distinct human TF families (bHLH, ETS, E2F, and RUNX).
- ChIP-seq data: Publicly available ChIP-seq data for Mad, Myc, Elk1, Ets1, E2f1, E2f4, Runx1, and Runx2 from various cell lines (e.g., H1-SC, K562) were used for training and validating the iMADS models, as well as for comparing in vitro predictions with in vivo binding. These datasets were chosen because they were not used in the gcPBM design, ensuring independent evaluation.
- Somatic mutation data: Data from cancer patients (melanoma, prostate, etc.) were used to analyze the impact of genetic variants on TF binding preferences.
Key Insights and Ideas
- Beyond Canonical Motifs: The most striking insight is that paralogous TFs, even those with identical canonical DNA-binding motifs, possess distinct intrinsic DNA-binding preferences. This challenges the long-held assumption that motif similarity equates to binding similarity.
- The Importance of Low-Affinity Sites: Differences in binding specificity are most pronounced at medium and low-affinity DNA sites. These sites, often overlooked by methods focusing on high-affinity interactions, play a crucial role in differential genomic targeting.
- Flanking Regions Matter: The sequences flanking the core binding site significantly contribute to differential binding specificity, highlighting the need to consider genomic context beyond just the core motif.
- Shape Matters Too: DNA shape features, such as minor groove width and roll, are also differentially preferred by paralogous TFs, suggesting an indirect readout mechanism contributing to specificity.
- iMADS: A Powerful New Tool: The iMADS framework provides a robust computational approach to quantify and model these subtle differences, enabling predictions of differential TF binding that correlate well with in vivo data.
- Genetic Variants and Paralog Competition: Non-coding genetic variants can have differential effects on paralogous TFs, not just by altering a single TF's binding but by shifting the balance of preference between competing family members.
Key Results
- Differential Binding Observed: For 10 out of 11 paralogous TF pairs tested, significant differences in in vitro binding specificity were observed using gcPBM assays, with differences most prominent at medium and low-affinity sites.
- iMADS Performance: The iMADS models demonstrated high prediction accuracy for TF-DNA binding specificity (R²=0.82-0.96) and showed significant correlations between predicted in vitro preferences and in vivo ChIP-seq signals (Spearman correlation up to 0.44 for Mad vs. Myc, 0.44 for E2f1 vs. E2f4).
- ROC Curve Superiority: iMADS models outperformed traditional PWMs (trained on in vitro or in vivo data) in distinguishing between TF-preferred and TF2-preferred ChIP-seq peaks, achieving AUCs as high as 0.9 for E2F proteins.
- Functional Relevance: Differentially preferred genomic sites were associated with distinct gene ontology (GO) terms, suggesting a link between differential binding and functional specificity.
- Somatic Mutation Impact: Cancer-associated non-coding variants showed significantly larger changes in Elk1-Ets1 preference scores compared to common variants, underscoring the phenotypic relevance of these differential preferences.
Limitations and Future Work
- ChIP-seq Noise: The study acknowledges that ChIP-seq data is inherently noisy and subject to numerous technical biases, which can limit the accuracy of in vivo binding comparisons. While iMADS models show good correlation, the inherent limitations of ChIP-seq data prevent a definitive quantification of in vivo binding differences solely attributable to intrinsic specificity.
- In Vitro vs. In Vivo Complexity: The in vitro gcPBM assays, while controlled, do not fully capture the complex in vivo environment, which includes TF concentrations, co-factors, and chromatin structure. The study emphasizes that in vivo binding is a result of many factors, and their work represents a crucial step in deciphering the molecular mechanisms of differential specificity.
- Future Directions: The authors envision the development of more quantitative high-throughput technologies and computational models to further unravel the complexities of differential genomic binding and function of paralogous TFs. They also highlight the potential of their iMADS framework for analyzing the impact of genetic variations on TF binding, providing a new avenue for understanding disease-causing non-coding mutations.
Summary updated