SNPsim: A Beginner’s Guide to Simulating Genetic Variation

Optimizing SNPsim Parameters for Realistic SNP Datasets

Accurate simulation of single-nucleotide polymorphism (SNP) data is essential for testing analysis pipelines, benchmarking tools, and designing experiments. SNPsim is a flexible simulator that lets you model population structure, linkage disequilibrium (LD), allele frequency spectra, and sequencing or genotyping error profiles. This guide walks through key SNPsim parameters and practical strategies to produce realistic SNP datasets that match empirical properties.

1. Define your biological scenario

• Purpose: Choose whether you’re modeling a single panmictic population, multiple subpopulations, admixture, or pedigrees.
• Timescale: Demography (population size changes, bottlenecks, expansions) strongly shapes the site frequency spectrum (SFS). Set effective population sizes (Ne) and generation times to match your study species or human population.
• Recombination landscape: Determine if constant recombination rate is sufficient or if you need hotspots/coldspots.

Assumption: a reasonable default is a constant-rate recombination with Ne tuned to match target heterozygosity.

2. Match allele frequency distribution

• Mutation rate (μ): Set μ to obtain realistic overall diversity (π). For humans, use ~1.2e-8 per site per generation as a starting point. Adjust for non-human species.
• Target SFS: Compare simulated SFS to your empirical SFS. If rare variants are overrepresented, simulate recent population expansion or raise μ; if too few rare variants, simulate bottleneck or lower μ.
• Ascertainment bias: If simulating genotyping-array-like data, apply SNP discovery filters (e.g., minor allele frequency (MAF) cutoffs and discovery sample sizes) to reproduce biased frequency spectra.

3. Reproduce linkage disequilibrium (LD)

• Recombination rate ®: Set r per base per generation. To mimic empirical LD decay, adjust r or include recombination hotspots based on maps (e.g., human genetic maps).
• Window size & marker density: Higher marker density increases observed LD. Simulate marker selection (prune or downsample) to match real dataset density and genotyping patterns.
• Background selection / selective sweeps: If LD is elevated around particular loci in your empirical data, include selection models or localized reductions in Ne.

Practical check: compute r^2 decay vs. distance for simulated data and compare to empirical curve; iterate r and demographic parameters until curves align.

4. Model population structure and admixture

• Number of populations and migration rates: Use island, stepping-stone, or explicit admixture events to reproduce FST and PCA patterns.
• Admixture proportions and timing: Recent admixture increases long-range LD; older admixture gives subtler allele frequency shifts. Tune event times and proportions to match empirical PCA clusters and admixture proportions.
• Sampling scheme: Simulate the same sample sizes per subpopulation as your real dataset to avoid sampling biases.

Validation: compare pairwise FST, PCA clustering, and ancestry proportions to empirical values.

5. Simulate genotyping and sequencing error realistically

• Genotyping arrays: Simulate SNP ascertainment (discovery panels), probe failure rates, and per-SNP missingness correlated with MAF or GC content.
• Sequencing reads: Simulate read depth distribution, base quality profiles, allele balance, and genotype calling thresholds. Include platform-specific error models (e.g., Illumina).
• Missing data: Introduce missingness patterns matching empirical data (random vs. correlated with sample or site).

Tip: run simulated reads through the same alignment and variant-calling pipeline used for real data to capture pipeline artifacts.

6. Introduce realistic selection and functional annotation

• Neutral vs. selected sites: Mix neutral SNPs with those under purifying or positive selection to reflect coding/noncoding proportions.
• Selection coefficients distribution: Use empirically derived distributions for deleterious effects if available.
• Annotation-linked mutation rates: Increase mutation rates or selection intensity in functional regions if needed.

7. Calibration and iterative validation

• Summary statistics: Compare simulated and empirical datasets using multiple summaries: π (nucleotide diversity), Tajima’s D, SFS, LD decay, FST, runs of homozygosity, and site-wise missingness.
• Visual diagnostics: Use PCA, ADMIXTURE/STRUCTURE-like plots, and Manhattan-style LD heatmaps to visually inspect realism.
• Parameter sweeps: Run grid searches over uncertain parameters (Ne, r, μ, admixture timing) and use automated fitting (e.g., ABC or likelihood methods) when possible.

8. Performance and reproducibility

• Scaling: For whole-genome simulations, consider using coalescent approximations or hybrid approaches to reduce compute.
• Random seeds and provenance: Record seeds, software versions, parameter files, and input maps for reproducibility.
• Downsampling: Simulate at higher resolution then downsample markers to match empirical panel sizes.

9. Example recommended defaults (human-focused starting point)

• μ = 1.2e-8 per site per generation
• r = 1e-8–1.2e-8 per base per generation (adjust using recombination map)
• Ne = 10,000 (adjust for demographic history)
• Genotyping array discovery: apply MAF ≥ 0.05 filter in discovery panel of ~100 samples
• Sequencing: mean depth 15–30×, base error rate 0.1–1%

10. Checklist before finalizing simulations

1. SFS roughly matches empirical spectrum.
2. LD decay curve aligns with observed data.
3. PCA/FST reproduce population structure.
4. Missingness/error profiles match platform-specific patterns.
5. Functional/selection signals are included if needed.

Conclusion Iteratively tune mutation, recombination, demographic, and genotyping parameters and validate using multiple summary statistics. Start with realistic defaults, compare simulated outputs to empirical summaries, and adjust parameters until you achieve close agreement.