Machine learning predicts nucleosome binding modes of transcription factors

Kishan Kc, Sridevi K. Subramanya, Rui Li, Feng Cui

Introduction

In the eukaryotic nucleus, DNA is organized into nucleosomes—complexes where DNA is wrapped around histone octamers. This structural organization regulates gene expression by controlling the accessibility of DNA to transcription factors (TFs). While some TFs can bind to DNA even when it is nucleosome-bound, others are blocked. Understanding these "binding modes"—whether a TF prefers the accessible ends of the nucleosomal DNA or can target the more occluded dyad region—is critical for deciphering the regulatory code of the genome. However, experimental determination of these modes via techniques like cryo-EM is resource-intensive. To address this, researchers developed ProtGauss, a machine learning framework designed to predict TF-nucleosome binding preferences directly from protein sequences.

Methodology: The ProtGauss Framework

The core innovation of ProtGauss lies in its representation of protein sequences. Traditional methods often average feature vectors across a sequence, which can obscure localized biochemical signals. ProtGauss instead treats each protein as a multivariate Gaussian distribution.

The ProtGauss pipeline: from protein sequences to multi-label binding preference classification.

The workflow begins by decomposing protein sequences into overlapping 3-mers, which are then mapped to 100-dimensional vectors using the ProtVec embedding model.

Workflow for generating feature matrices from amino acid sequences using ProtVec embeddings.

Instead of a simple mean, the model calculates both the mean (μ) and the covariance matrix (Σ) for each protein. This "Gaussian twist" captures the variance and texture of biochemical properties along the sequence. These distributions are compared using a specialized similarity kernel within a Support Vector Machine (SVM) to predict multiple binding preferences simultaneously, including End, Periodic, Groove, Dyad, Gyre, and Orientational preferences.

Results and Biological Significance

ProtGauss was trained on a high-quality dataset of 167 TFs and validated against in vivo human data. The model achieved a micro-AUPR of 0.61, significantly outperforming baseline models that rely on simple averaging. The binary classifier for the "End Binding" (EB) mode was particularly robust, achieving 75% AUPR.

Evolutionary Specialization

Large-scale application of the model across five species (human, mouse, fruit fly, nematode, and yeast) revealed a striking evolutionary trend. In yeast, nearly 99% of TFs are predicted to be "end-binders," preferring the most accessible DNA. In contrast, only 88% of human TFs follow this pattern. This suggests that higher eukaryotes have evolved specialized "middle-binders" to navigate the complex chromatin landscapes required for multicellular development.

The Role of SOX and HOX Families

The model identified that TFs involved in body plan development, such as the SOX and HOX families, are enriched for non-end-binding modes. This finding is consistent with structural data indicating these proteins' ability to interact with more internal regions of the nucleosome, facilitating their role as pioneer factors.

Discussion and Future Directions

The success of ProtGauss demonstrates that the variation of biochemical properties within a protein sequence is a powerful predictor of its functional interactions. By preserving this information through Gaussian representations, the model provides a scalable way to map the TF-nucleosome interactome.

Current limitations include the relatively small size of experimental training sets. Future work may enhance predictive power by incorporating 3D structural data from tools like AlphaFold or considering the specific shape and mechanical properties of the target DNA sequences. Nonetheless, ProtGauss represents a significant step toward a predictive understanding of how the "gatekeepers" of the genome interact with its structural packaging.

Summary updated

Abstract

Background Most transcription factors (TFs) compete with nucleosomes to gain access to their cognate binding sites. Recent studies have identified several TF-nucleosome interaction modes including end binding (EB), oriented binding, periodic binding, dyad binding, groove binding, and gyre spanning. However, there are substantial experimental challenges in measuring nucleosome binding modes for thousands of TFs in different species. Results We present a computational prediction of the binding modes based on TF protein sequences. With a nested cross-validation procedure, our model outperforms several fine-tuned off-the-shelf machine learning (ML) methods in the multi-label classification task. Our binary classifier for the EB mode performs better than these ML methods with the area under precision-recall curve achieving 75%. The end preference of most TFs is consistent with low nucleosome occupancy around their binding site in GM12878 cells. The nucleosome occupancy data is used as an alternative dataset to confirm the superiority of our EB classifier. Conclusions We develop the first ML-based approach for efficient and comprehensive analysis of nucleosome binding modes of TFs.
BibTeX
@article{kc2021machine,
  author = {Kishan KC and Sridevi K. Subramanya and Rui Li and Feng Cui},
  title = {{Machine learning predicts nucleosome binding modes of transcription factors}},
  journal = {BMC Bioinformatics},
  year = {2021},
  doi = {10.1186/s12859-021-04093-9},
}
Figure