Optimizing FastSimCoal: Tips for Faster, Accurate Simulations

Optimizing FastSimCoal: Tips for Faster, Accurate Simulations

Overview

FastSimCoal (fastsimcoal2) is a flexible coalescent simulator widely used for modeling complex demographic scenarios and generating expected site frequency spectra (SFS). Efficient, accurate simulations require careful balance between computational speed and statistical precision. This guide gives practical tips to speed up runs, reduce errors, and get reliable parameter estimates.

1. Choose the right SFS and data representation

• Folded vs. unfolded SFS: Use the unfolded SFS if ancestral states are confidently known; it contains more information but requires accurate ancestral allele inference. Use the folded SFS when ancestral state uncertainty could bias results.
• SNP-only vs. full sequences: Simulating only SNPs (SFS input) is faster than simulating full linked sequences. Use SNP-style SFS when linkage is not essential for inference.

2. Reduce model complexity where justified

• Simplify demographic models: Remove parameters that are unsupported by data (e.g., extremely short-duration events, overly granular migration schedules). Fewer parameters reduce run time and improve identifiability.
• Aggregate populations if appropriate: Combine closely related or low-sample populations to reduce dimensions of the SFS.

3. Optimize the number of replicates and simulation length

• Control -n and -N: Use the minimum number of coalescent simulations (-n) and maximum coalescent size (-N) that yield stable likelihood estimates. Start with modest values for exploratory runs and increase for final inference.
• Pilot runs: Run short pilot analyses to gauge variance in likelihoods; scale up replicate counts only when needed.

4. Fine-tune parameter search and optimization

• Smart initial guesses: Supply reasonable starting parameter values to reduce optimization time and avoid local minima.
• Use multiple independent runs: Run optimization multiple times with different seeds to ensure convergence; this is faster than a single extremely long run that may be stuck in a local optimum.
• Adjust ECM settings: Tune expectation-conditional maximization (ECM) iterations to the problem size—fewer iterations per run can speed up exploration, then increase for the final fitting.

5. Parallelize strategically

• Multithreading where available: Use fastsimcoal’s parallel options (e.g., MPI builds) on multi-core systems to distribute simulations.
• Divide-and-conquer: Split bootstrap or replicate sets across nodes/jobs in an HPC environment. Combine results after runs finish.

6. Manage input/output to avoid bottlenecks

• Use local SSDs: Write temporary files and simulation outputs to fast local storage rather than network filesystems.
• Minimize logging verbosity: Disable excessive logging unless debugging; smaller outputs reduce I/O overhead.

7. Improve numerical stability and precision

• Scale parameters appropriately: Rescale time and population size parameters to avoid extremely small or large values that can cause numerical issues.
• Check SFS projection: Properly project sample sizes to reduce empty allele-count cells that inflate variance.
• Monitor likelihood variability: High variance across replicates may indicate insufficient simulations or mis-specified models.

8. Use efficient file formats and preprocessing

• Preprocess genotype data: Filter low-quality SNPs and missing data before constructing the SFS to reduce noise and simulation complexity.
• Compress intermediate files: Use compressed archives for long-term storage but keep active simulation files uncompressed for speed.

9. Validate models and results

• Posterior predictive checks: Simulate data under fitted parameters and compare summary statistics to empirical data.
• Compare nested models: Use likelihood ratio tests or information criteria to justify added complexity.
• Bootstrap parameter uncertainty: Use nonparametric or parametric bootstraps distributed across cores to estimate confidence intervals.

10. Practical workflow example (recommended)

1. Preprocess data: filter SNPs, project SFS, decide folded/unfolded.
2. Run quick exploratory fits with simplified models and low -n.
3. Identify promising models and refine parameter bounds and starting values.
4. Scale up -n and bootstrap replicates; run multiple seeds in parallel.
5. Final fits on high-precision settings; perform posterior predictive checks and report uncertainties.

Quick checklist

• Folded/unfolded SFS chosen
• Model complexity justified