Statistical power of phylo-HMM for evolutionarily conserved element detection

Xiaodan Fan, Jun Zhu, Eric E. Schadt, Jun S. LiuBMC Bioinformatics 2007 · DOI

TL;DR

This paper investigates the statistical power of the phylo-HMM model for identifying evolutionarily conserved elements in genomic sequences. The authors found that the number of species, evolutionary distances, conservation ratio, and expected length of conserved elements are crucial factors, while tree topology and substitution models have a lesser impact. This work provides practical guidelines for selecting species and parameters in comparative genomics studies.

What Problem Does This Paper Address?

Identifying functional elements within genomes is a fundamental challenge in molecular biology. Evolutionarily conserved elements, regions that evolve more slowly than the rest of the genome, are strong indicators of functional importance. Comparative genomics, by aligning and comparing sequences from related species, is a powerful approach for this task. The phylo-HMM (phylogenetic Hidden Markov Model) is a sophisticated method that models molecular evolution along both the genome and the phylogenetic tree, accounting for substitution rate variation. However, before this study, its statistical power – its ability to correctly identify conserved elements – had not been rigorously evaluated. This paper aims to fill that gap by systematically assessing the factors that influence phylo-HMM's power, providing guidance for its effective application.

Methods

The core of this study is a simulation-based evaluation of the phylo-HMM model's statistical power. The authors first estimated a "baseline" set of phylo-HMM parameters from real genomic data (promoter regions of dog, human, mouse, and rat). They then generated simulated DNA sequences based on this baseline, systematically varying individual parameters or groups of parameters to explore their impact on detection power.

For each simulation, they generated a true state sequence (conserved or nonconserved) and then simulated the alignment of modern-day genomes based on a given phylogenetic tree and substitution model. The phylo-HMM was then used to infer the hidden state sequence from these simulated alignments. The power was assessed by comparing the inferred states to the true states, using metrics like sensitivity and specificity, and visualized through Receiver Operating Characteristic (ROC) curves. They also compared phylo-HMM's performance to a simpler method, PID (Percent Identity), to benchmark its effectiveness.

Datasets

The study utilized promoter regions from four species: dog, human, mouse, and rat. These regions were defined as 2000 bp upstream to 500 bp downstream of annotated transcription start sites. Orthologous gene clusters were identified using MLAGAN for alignment. The authors did not introduce new benchmark datasets but rather used existing genomic data to derive parameters for their simulations.

Key Insights and Ideas

  • Species selection is paramount: The number of species and their evolutionary distances are the most critical factors for phylo-HMM's power. This suggests that carefully choosing related species with appropriate divergence is key.
  • Conservation matters: The ratio of substitution rates between conserved and nonconserved regions (conservation ratio) and the expected length of conserved elements significantly impact detection power. Regions that are truly conserved and of a reasonable length are easier to find.
  • Topology and substitution models are secondary: The specific branching order of the phylogenetic tree and the complexity of the nucleotide substitution model have a relatively minor impact on the model's power. This is a practical insight, suggesting that simpler topologies and models can often suffice.
  • Balancing branch lengths is beneficial: For topologies with balanced branch lengths, power is generally higher. Unbalanced topologies, where some branches are very short, can reduce the effective number of genomes and thus decrease power.
  • Alignment quality is robustly handled: The phylo-HMM model's power is surprisingly robust to moderate levels of alignment error, suggesting that it can still perform well even with imperfect alignments.

Key Results

  • Phylo-HMM outperforms PID: Under baseline conditions, phylo-HMM achieved a median sensitivity of 0.91 at a median specificity of 0.94, significantly outperforming the PID method (median sensitivity 0.73 at the same specificity).
  • Optimal branch lengths exist: For the baseline tree, there's an optimal range for branch lengths (around 0.6-1 substitution per site). Branch lengths that are too short or too long can decrease power.
  • More genomes generally mean more power: Increasing the number of species generally increases sensitivity, especially for shorter branch lengths.
  • Conservation ratio is a major determinant: Power decreases dramatically as the conservation ratio approaches 0.6, highlighting the importance of this parameter.
  • Expected element length is crucial: Longer conserved elements lead to better ROC curves with smaller variation.

Limitations and Future Work

The authors acknowledge several limitations:

  • Simulation assumptions: The simulations assumed sequences were generated by phylo-HMM and that true, ungapped alignments were available. Real-world scenarios involve gaps and potentially more complex evolutionary processes.
  • Fixed conservation state: The current model assumes a fixed conservation state for a site across all species, which might be too restrictive. More flexible models allowing for species-specific differences in conservation states are needed.
  • Simplified functional element modeling: The model focuses on conserved elements but could be extended to directly model specific functional elements like transcription factor binding sites.
  • Goodness-of-fit test: A validated goodness-of-fit test for phylo-HMM on real data is still needed.

Future work could involve developing more flexible models, integrating alignment and conservation analysis, and directly modeling functional elements.

Summary updated

BibTeX
@article{fan2007statistical,
  author = {Xiaodan Fan and Jun Zhu and Eric E. Schadt and Jun S. Liu},
  title = {{Statistical power of phylo-HMM for evolutionarily conserved element detection}},
  journal = {BMC Bioinformatics},
  year = {2007},
  doi = {10.1186/1471-2105-8-374},
}
Figure