Discovering a less-is-more effect to select transcription factor binding sites informative for motif inference

Jinrui Xu, Jiahao Gao, Mark GersteinbioRxiv 2020 · DOI

TL;DR

This paper introduces a new method, MOBI, for selecting transcription factor (TF) binding sites that are more informative for inferring TF binding motifs. By moving beyond simply selecting sites with strong binding signals and instead prioritizing "less crowded" and shorter regions, MOBI significantly improves motif inference accuracy by 30-70%.

What Problem Does This Paper Address?

Inferring transcription factor (TF) binding motifs is crucial for understanding gene regulation. A common approach is to use ChIP-seq data to identify TF binding sites and then infer motifs from these sites. However, the authors argue that simply selecting sites with strong ChIP-seq signals is insufficient. This is because:

  1. Spurious Signals: Strong binding signals don't always mean the TF's target motif is present. They can be due to experimental artifacts or other factors like permissive DNA sequences.
  2. Interfering Motifs: Sites with strong signals often contain binding motifs for other TFs (co-binding TFs). These "interfering motifs" can confuse the inference process, leading to inaccurate motifs for the target TF.

The problem is that the standard method of selecting the strongest binding sites overlooks these issues, leading to suboptimal motif inference.

Methods

The core of the paper is the development of the MOBI (Motif Inference using Binding site selection) approach. MOBI introduces two key innovations:

  • Crowding Score (C-score): Instead of just looking at the strength of a single TF's binding signal, MOBI quantifies how "crowded" a binding site is by counting the number of other TF binding sites that overlap it. A high C-score indicates a crowded site, which is deemed less informative.
  • Binding-Crowding Score (BC-score): This score combines the traditional binding signal rank with the C-score rank, allowing for a more nuanced selection of informative sites. By adjusting a weight parameter (w), researchers can prioritize sites with lower crowding.
  • Optimized Binding Site Length: The authors also found that shorter binding sites (e.g., 40 bp instead of the conventional 200 bp) can further reduce the impact of interfering motifs. MOBI incorporates tuning of this length parameter (l).

Essentially, MOBI selects sites that are both less crowded by other TFs and are shorter, aiming to isolate the true binding motif of the target TF.

Datasets

The study utilized publicly available ChIP-seq data from the following sources:

  • ENCODE portal: For human cell lines (GM12878, K562, and others) and fly (Drosophila melanogaster).
  • modERN and modENCODE projects: For worm (Caenorhabditis elegans).
  • Mock IP controls: For fly and worm samples, used to assess the impact of different control strategies.

In vitro TF binding motifs (from SELEX, PBM, and B1H experiments) were collected from the Cis-bp database for evaluating motif inference accuracy.

Key Insights and Ideas

  • "Less-is-more" effect: The most striking insight is that selecting fewer, more carefully chosen binding sites (less crowded, shorter) leads to better motif inference results, rather than using all strong signals.
  • Crowding is a major confounder: The presence of many other TF binding sites in a region is a strong indicator that the site might not be specific to the target TF, even if the signal is strong.
  • Interfering motifs are a significant problem: Co-binding TFs can mask or distort the signal of the target TF's motif, making it difficult to infer accurately.
  • Optimal parameters are context-dependent: The ideal weight for the crowding score (w) and the optimal binding site length (l) can vary between species and cell types, highlighting the need for a tunable approach.

Key Results

  • Substantial Improvement: The MOBI approach consistently improved motif inference accuracy by 30-70% across different species (fly, worm, human) compared to conventional methods.
  • BC-score Superiority: Using the BC-score for site selection outperformed using binding signal alone or C-score alone in motif enrichment.
  • Shorter Sites Help: Reducing binding site length (e.g., from 200 bp to 40 bp) significantly improved motif inference, especially for human TFs.
  • DREME Benefits Most: The DREME motif inference tool showed the most significant gains from MOBI, particularly for human cell lines, suggesting its algorithm benefits from cleaner input data.
  • MOBI Database: The authors generated a database of inferred motifs for 1,207 TFs using their MOBI approach, which they claim are more accurate than existing databases.

Limitations and Future Work

  • Parameter Tuning: While MOBI offers improved accuracy, it requires tuning two parameters (w and l), which might add complexity. The paper suggests that optimal parameters can be identified through cross-validation.
  • Transplanted C-scores: The study found that C-scores calculated from one sample are not always robust for inferring motifs in another, suggesting that sample-specific C-score calculation is important.
  • Genome Accessibility as Proxy: While genomic accessibility showed some correlation with C-scores, it did not outperform C-scores for motif inference, indicating that the "crowding" aspect is more directly informative.
  • Future Work: The authors suggest that MOBI could be further improved by integrating with more inference tools and by leveraging data from newer, higher-resolution ChIP techniques (like ChIP-exo and ChIP-nexus). They also propose that the MOBI database can facilitate evolutionary analyses of TF binding.

Summary updated

Abstract

ABSTRACT Many statistical methods have been developed to infer the binding motifs of a transcription factor (TF) from a subset of its numerous binding regions in the genome. We refer to such regions, e.g. detected by ChIP-seq, as binding sites. The sites with strong binding signals are selected for motif inference. However, binding signals do not necessarily indicate the existence of target motifs. Moreover, even strong binding signals can be spurious due to experimental artifacts. Here, we observe that such uninformative sites without target motifs tend to be “crowded” -- i.e. have many other TF binding sites present nearby. In addition, we find that even if a crowded site contains recognizable target motifs, it can still be uninformative for motif inference due to the presence of interfering motifs from other TFs. We propose using less crowded and shorter binding sites in motif interference and develop specific recommendations for carrying this out. We find our recommendations substantially improve the resulting motifs in various contexts by 30%-70%, implying a “less-is-more” effect.
BibTeX
@misc{xu2020discovering,
  author = {Jinrui Xu and Jiahao Gao and Mark Gerstein},
  title = {{Discovering a less-is-more effect to select transcription factor binding sites informative for motif inference}},
  year = {2020},
  doi = {10.1101/2020.11.29.402941},
  howpublished = {bioRxiv},
}
Figure