Limb-Enhancer Genie: An accessible resource of accurate enhancer predictions in the developing limb

Remo Monti, Iros Barozzi, Marco Osterwalder, Elizabeth Lee, Momoe Kato, Tyler H. Garvin, Ingrid Plajzer-Frick, Catherine S. Pickle, Jennifer A. Akiyama, Veena Afzal, Niko Beerenwinkel, Diane E. Dickel, Axel Visel, Len A. PennacchioPLOS Computational Biology 2017 · DOI

Introduction

Enhancers are short segments of non-coding DNA that act as primary "switches" controlling gene expression during development. Identifying these regions is a significant challenge in genomics. Traditionally, researchers have relied on single epigenomic marks, such as the acetylation of histone H3 at lysine 27 (H3K27ac), to locate potential enhancers. However, this approach is prone to high false-positive rates, as many regions marked by H3K27ac do not exhibit enhancer activity when tested in vivo. While integrative computational methods have been developed to address this, they often lack transparency or accessibility for experimental biologists.

Methods

To overcome these limitations, the authors developed Limb-Enhancer Genie (LEG), a machine learning framework designed to predict limb-specific enhancers with high accuracy. LEG integrates two primary categories of data:

  1. Chromatin Features: Over 50 datasets, including DNase-seq (accessibility), ChIP-seq for various histone modifications (e.g., H3K27ac, H3K4me1), and binding data for co-activators (p300) and architectural proteins (CTCF, Cohesin).
  2. Sequence Features: Predicted clusters of evolutionarily conserved transcription factor (TF) binding sites for TFs known to be active in the developing limb.

The framework utilizes an ensemble machine learning approach, training four types of classifiers: LASSO logistic regression, Random Forests (RF), and Support Vector Machines (SVM) with both linear and radial kernels. These models are combined using ridge regression and a "Sum of Ranks" (SOR) approach to generate a final prediction score. A critical innovation in LEG is the DNase I enrichment score, which compares nuclease accessibility in the limb against a "headless embryo" baseline to distinguish tissue-specific regulatory elements from general ones.

Key Insights and Innovations

The study highlights several key findings regarding enhancer prediction:

  • The Power of Contrast: The DNase I enrichment score was the most impactful feature. By contrasting limb accessibility with a non-limb baseline, the model effectively filtered out "ubiquitous" regulatory elements that often act as decoys.
  • Biological Interpretability: Using LASSO and Random Forests allowed for feature ranking. The results confirmed that while p300 and H3K27ac are strong indicators, their predictive power is maximized when combined with tissue-specific accessibility and sequence motifs.
  • Synergy of Data Types: While chromatin data is generally more predictive of in vivo activity, the addition of conserved TF motifs provided a significant performance boost, identifying enhancers that might be missed by chromatin marks alone.

Results and Validation

LEG demonstrates superior performance compared to existing state-of-the-art tools like EnhancerFinder and EMERGE, achieving a median AUROC of 0.875 and an AUPRC of 0.582. The model also showed a significant reduction in false positives; for instance, models trained on the full feature set had less than 5% overlap with CTCF-only sites (insulators), which simpler models often misidentify as enhancers.

To validate the model's utility, the authors tested four high-scoring new predictions near the Hand2 gene using mouse transgenic reporter assays. Three out of four predictions were confirmed to drive reproducible reporter expression in the developing limb. Furthermore, the authors have established a user-friendly web interface (leg.lbl.gov) and a UCSC Genome Browser track hub to make these predictions accessible to the research community.

Conclusion

Limb-Enhancer Genie represents a significant advancement in the functional interpretation of genetic studies. By integrating diverse epigenomic and sequence data with a novel tissue-specific enrichment strategy, LEG provides a robust and accessible resource for identifying limb enhancers. While currently limited by the size of available "gold standard" training sets and the granularity of sub-tissue data, the underlying strategy is likely applicable to other tissues and developmental stages.

Summary updated

BibTeX
@article{monti2017limb,
  author = {Remo Monti and Iros Barozzi and Marco Osterwalder and Elizabeth Lee and Momoe Kato and Tyler H. Garvin and Ingrid Plajzer-Frick and Catherine S. Pickle and Jennifer A. Akiyama and Veena Afzal and Niko Beerenwinkel and Diane E. Dickel and Axel Visel and Len A. Pennacchio},
  title = {{Limb-Enhancer Genie: An accessible resource of accurate enhancer predictions in the developing limb}},
  journal = {PLOS Computational Biology},
  year = {2017},
  doi = {10.1371/journal.pcbi.1005720},
}
Figure