Human Pol II promoter recognition based on primary sequences and free energy of dinucleotides

Jianyi Yang, Yu Zhou, Zu‐Guo Yu, Vo Anh, Liqian ZhouBMC 2008 · DOI

TL;DR

This paper introduces a novel computational method for identifying human Pol II promoter sequences by analyzing dinucleotide free energy and employing multifractal analysis, Z-curve, and global descriptor techniques. The proposed approach achieves high prediction accuracies, outperforming existing methods and demonstrating its utility in the challenging task of promoter prediction.

What Problem Does This Paper Address?

Identifying promoter regions in DNA is crucial for understanding gene regulation and expression. Eukaryotic RNA polymerase II (Pol II) promoters, in particular, are complex and their computational prediction remains a significant challenge. Existing methods often struggle with low specificity or sensitivity, partly because they rely on short, conserved regulatory sequence elements (RSEs) that can appear elsewhere in the genome, leading to false positives. The authors aim to improve promoter prediction accuracy by exploring more sophisticated sequence analysis techniques beyond traditional motif-based approaches.

Methods

The researchers propose a multi-faceted approach to promoter recognition, integrating four distinct methods:

  • Multifractal Analysis (MFA) and Analogous Multifractal Analysis (AMFA): These techniques are applied to numeric sequences derived from the free energy of dinucleotides. MFA analyzes the measure representation of these sequences, while AMFA examines the integrated time series. These methods extract parameters related to fractal dimensions, mass exponents, and spectrum singularities.
  • Z-curve (ZC): This method represents DNA sequences as curves in a 3D space, allowing for the analysis of nucleotide frequencies and their positional dependencies. The paper utilizes various Z-curve representations, including codon-position-dependent frequencies and phase-specific dinucleotide frequencies.
  • Global Descriptor (GD): Inspired by protein sequence analysis, this method provides a comprehensive description of DNA sequences by considering composition, transition, and distribution of nucleotide attributes. It also incorporates the Entropy Density Profile (EDP) model.

A total of 141 parameters are extracted from these four methods. The authors then employ Fisher's linear discriminant algorithm to classify sequences and evaluate the performance of different combinations of these parameters.

Datasets

The study utilizes two primary datasets:

  • Human Pol II Promoter Sequences: Obtained from Release 90 of the Eukaryotic Promoter Database (EPD). This dataset consists of 1856 sequences, each 499 bp upstream and 100 bp downstream of the transcription start site (TSS).
  • Human Genome Non-Promoter Sequences: Derived from the Exon/Intron Database (EID). This set comprises 1000 intron sequences and 500 exon sequences, with fragments of 600 bp selected.

For evaluation, the authors also use experimentally verified promoter sequences from human chromosome 22.

Key Insights and Ideas

  • Beyond Motifs: The paper strongly advocates for moving beyond simple motif-based approaches, which are prone to false positives due to the short and non-unique nature of RSEs.
  • Multifractal Properties of DNA: The study suggests that multifractal analysis can reveal subtle but important structural information within DNA sequences that is predictive of promoter regions.
  • Parameter Redundancy: Combining various methods doesn't always lead to proportional increases in accuracy. The authors observe that some parameters exhibit redundancy, indicating that fewer parameters can achieve comparable results.
  • Global Descriptor Efficacy: The global descriptor method, despite using a relatively small number of parameters (36), demonstrates significant predictive power, particularly in identifying experimentally verified promoters.
  • Robustness of the Method: The proposed approach shows robustness even with reduced training set sizes, suggesting its reliability.

Key Results

  • High Discriminant Accuracies: The combined methods, particularly with 117 parameters, achieved high discriminant accuracies, reaching 90.43% for the training set and 89.79% for the test set.
  • Improved Performance: When compared to five other existing promoter prediction tools using a test set of 186 promoter and 150 non-promoter sequences, the authors' method (specifically M3, which uses 117 parameters) demonstrated superior sensitivity, specificity, accuracy, and correlation coefficient.
  • Chromosome 22 Validation: The global descriptor method (M1, with 36 parameters) successfully identified 17 out of 18 experimentally verified promoter sequences from human chromosome 22.
  • M1 as a Practical Choice: The global descriptor method (M1) is highlighted as a practical and effective approach due to its good performance with a significantly smaller parameter set.

Limitations and Future Work

  • Parameter Optimization: While the paper explores combinations of methods, it acknowledges that not all subsets of the 141 parameters were tested, and further optimization might be possible.
  • Computational Cost: The paper mentions that combining all 141 parameters yields satisfactory results but notes that the number of parameters is "too high," implying a trade-off between accuracy and computational efficiency.
  • Real Genome Prediction: The authors acknowledge the difficulty of directly applying their method to real human genome sequences due to the lack of clear promoter markings. They suggest using datasets like the ENCODE pilot project data as a next step towards addressing this challenge.
  • Further Contributions: The authors express their intention to contribute to the challenging problem of promoter prediction in the human genome in their future work.

Summary updated

Abstract

The high accuracies achieved suggest that the methods of this paper are useful for understanding the difficult problem of promoter prediction.
BibTeX
@inproceedings{yang2008human,
  author = {Jianyi Yang and Yu Zhou and Zu‐Guo Yu and Vo Anh and Liqian Zhou},
  title = {{Human Pol II promoter recognition based on primary sequences and free energy of dinucleotides}},
  booktitle = {BMC},
  year = {2008},
  doi = {10.1186/1471-2105-9-113},
}
Figure