Introduction

Without the CFTP method, the initial methylation states are sampled independently for each CpG site based on the equilibrium frequencies of its genomic structure (island or non-island). As a result, the initial distribution lacks neighbor correlations of the correlated Single-Site Event (SSE) process. These correlations develop over time during evolution along a branch.

To ensure the that root sequence includes neighbor correlations, we adapted the Coupling From The Past (CFTP) algorithm (see JG Propp and Wilson (1996), J. Propp and Wilson (1998) for details). This method samples methylation states from the model’s equilibrium accounting for the correlated SSE process. Note that in the current implementation CFTP neglects the Island-Wide Event Process and thus in particular the reduction of neighbor-specific correlations within CpG islands after and island-wide event.

Use examples

The CFTP method can be applied when simulating initial methylation data with simulate_initialData or when simulating data evolution along a tree with simulate_evolData with the argument CFTP. The default value for this argument is FALSE but can be set to TRUE to apply the CFTP algorithm initial data or at the root of the tree.

Example: Sample initial data with the sequence of methylation states according to the equilibrium.

# Example with sampled initial methylation frequencies and default parameter values
custom_infoStr <- data.frame(n = c(100, 100, 100),
                             globalState = c("M", "U", "M"))
initialD <- simulate_initialData(infoStr = custom_infoStr, CFTP = TRUE)
#> Simulating initial data with default parameter values
#> Calling CFTP algorithm.

Example: Simulate evolution of data along a tree calling the CFTP method to sample a sequence of methylation states at the root of the tree from the equilibrium.

# Example with customized initial methylation frequencies, customized parameter values 
# and default dt (0.01)
tree <- "((a:1, b:1):2, c:2, (d:3.7, (e:4, f:1):3):5);"
custom_infoStr <- data.frame(n = c(10, 10, 10),
                             globalState = c("M", "U", "M"),
                             u_eqFreq = c(0.1, 0.8, 0.1),
                             p_eqFreq = c(0.1, 0.1, 0.1),
                             m_eqFreq = c(0.8, 0.1, 0.8))
custom_params <- get_parameterValues()
custom_params$mu <- 0.005

evolD <- simulate_evolData(infoStr = custom_infoStr, tree = tree, params = custom_params, CFTP = TRUE)
#> Simulating data at root and letting it evolve along given tree:  ((a:1, b:1):2, c:2, (d:3.7, (e:4, f:1):3):5);
#> Calling CFTP algorithm for data at root before letting it evolve along given tree.

Convergence and Convergence Approximations

Applying the CFTP method increases run time and memmory requirements. This increase depends on the parameterization used for the given simulation. For example, a combination of high \(\iota\) values (leading to the SSE process being mostly independent) and low \(\alpha_{Ri}\) values (leading to a high variation in the 3 rates of independent SSE) leads to some sites having a rate of change through the SSE process of nearly 0, which results in a long runtime of the exact CFTP algorithm. To limit the run time and memmory consumption for such cases, the maximum number of steps used is limited with the argument CFTP_step_limit. If after the limiting number of steps the algorithm has not converged, an approximation is applied and a message will show:

“Steps limit reached. Applying approximation for CFTP.”

The number of steps used by the algorithm is increased by subsequent iterations, so that the number of steps at each iteration \(n\) given by:

\[\text{steps}_n = 10000 \cdot 2^{n-1}\]

The default value for CFTP_step_limit is set to 327680000, allowing the algorithm to try 16 full iterations before applying the approximation.

Visual Examples

To help understanding how CFTP samples an initial sequence of methylation states from equilibrium, we provide plots from our tests. The test used 10 parameter combinations since the model equilibrium depends on them. The summary statistic used is the mean correlation of a central segment of methylation states per genomic structure type (see vignette “Summary Statistics with MethEvolSIM”).

Without CFTP, the initial states of the sequence are sampled independently based on the equilibrium frequencies of each genomic structure. Over time, local correlations between methylation states develop during evolution. With CFTP, the initial states of the sequence are sampled from the equilibrium accounting for the correlated SSE process. Therefore, the local correlations are already present in the initial states of the sequence.

Our test consists in comparing the change in correlations observed in a sequence initiated without CFTP along evolutionary time to the correlations observed in a sequence initiated with CFTP. We expect the mean correlation of the sequence without CFTP to change with evolutionary time until it stabilizes at the equilibrium for the corresponding parameter combination. To track this, we discretize evolution along a tree branch into small steps of length 0.01. We also apply CFTP to the initial data, sampling a sequence of methylation states from equilibrium and computing the corresponding summary statistic. The distribution of the summary statistics computed for the sequence obtained with the CFTP method should match the summary statistics equilibrium reached without the CFTP method after evolutionary time.

In the following figures, the y-axis represents the mean correlation, and the x-axis shows the number of discretized time steps. Each figure corresponds to one of the 10 parameter combinations. For each combination, 10 replicates were simulated, shown in different colors. Dots (with unique shapes for each replicate) indicate the mean correlation at each time step for the sequences initiated without CFTP. For each sequence, the line of the same color represents the value after applying CFTP with the same CpG-island-specific equilibrium probabilities of the methylation states as in the initialization of the non-CFTP simulations. The plots show that CFTP leads to a correlation sampled from a distribution to which the simulations without CFTP converge.

Sampled parameter combination 1

Sampled parameter combination 2

Sampled parameter combination 3

Sampled parameter combination 4

Sampled parameter combination 5

Sampled parameter combination 6

Sampled parameter combination 7

Sampled parameter combination 8

Sampled parameter combination 9

Sampled parameter combination 10

Reproducibility

The test used to generate the visual example is provided as a workflow in MethEvolSIM GitHub Repository, under /extensive_tests/CFTP_convergence. All scripts are provided and can be run independently, or as in our example, using Snakemake. Here, we describe the structure and content of the workflow.

Workflow Requirements

Dependencies are listed on the file environment.yaml. They can be installed previous to running the workflow or, alternatively, a conda environment can be created with:

conda env create -n cftp_testConvergence -f environment.yaml

Then, the Snakemake workflow can be run by:

conda activate cftp_testConvergence
snakemake --use-conda

If using nohup one can activate the environment as in the following command.

nohup conda run -n cftp_testConvergence snakemake &

Provided: environment.yaml, Snakefile, config.yaml and associated scripts.

Steps in the Workflow

All the given parameters to the workflow steps are taken from the config.yaml file.

Rule get_spatial_str run by script get_spatial_str.R. It sets the spatial distribution of the sequence to simulate. The test simulates data on a simplified genomic region alternating island and non-island structures each a given number of times (Str_n) and each with a constant number of CpGs (CpG_n). The output is a file named as methSite_genomicDist under the directory dir.
Rule set_design run by script set_sim_design.R. Takes as input the output of the previous rule. For a number of model parameter combinations (n_sim) and setting a seed (seed) for reproducibility, samples the parameter combinations from the prior distributions as defined in prior_distributions. The output is a file named “design.RData” under the directory dir containing both the spatial structure defined in the first rule and the sampled parameters. Note that this script corrects the values of \(\iota\) and \(\alpha_{R_i}\) in the sampled parameters to be a minimum of 1e-2, as that is the smallest value used in the implementation due to numerical limitations.
Rule run_sim run by script run_sim.R. Takes as input the output of the previous rule. For each parameter combination it simulates a number of replicates (replicate_n) and each replicate is simulated with a seed obtained as the index of the corresponding parameter combination * 1000 + the replicate number for reproducibility. For each replicate it first generates initial data without the CFTP method and simulates evolution along a branch of length branch_length a number of times from start to end and, then calls the CFTP method on the previously generated initial data. The output files are written under the directory dir and named:

"CFTP_testConvergence_paramsID_{parameter_combination_number}_rep_{replicate_number}_{step_n}.RData"
"CFTP_testConvergence_paramsID_{parameter_combination_number}_rep_{replicate_number}_cftp.RData"

Rule compute_meanCor run by script compute_meanCor.R. Takes as input the output of the previous rule. For each branch step output and cftp output it computes the mean correlation of a segment of methylation states per genomic structure type (see vignette “Summary Statistics with MethEvolSIM”). The output files are written under the directory dir and are named:

"summaryStats_CFTP_testConvergence_paramsID_{parameter_combination_number}_cftp.RData"
"summaryStats_CFTP_testConvergence_paramsID_{parameter_combination_number}_rep_{replicate_number}.RData"

Rule plot_results run by script plot.R. Takes as input the output of the previous rule.
The output files are written under a subdirectory Figures generated in the snakemake directory (or the directory from where the R script is run if not using Snakemake), and are named “CFTP_testConvergence_paramsID_{parameter_combination_number}” both in .pdf and .png format.

References

Propp, JG, and DB Wilson. 1996. “Exact Sampling with Coupled Markov Chains and Applications to Statistical Mechanics.” Article; Proceedings Paper. RANDOM STRUCTURES & ALGORITHMS 9 (1-2): 223–52.

Propp, J, and D Wilson. 1998. “Coupling from the Past: A User’s Guide.” Proceedings Paper. In MICROSURVEYS IN DISCRETE PROBABILITY, edited by D Aldous and J Propp, 41:181–92. DIMACS-Series in Discrete Mathematics and Theoretical Computer Science. P.O. BOX 6248, PROVIDENCE, RI 02940 USA: NSF Sci & Technol Ctr; Discrete Math & Theoret Comp Sci; Rutgers Univ; Princeton Univ; AT&T Labs; Bell Labs; Bellcore; AMER MATHEMATICAL SOC.

CFTP Method