softwareDevelopment/GWAS

GWAS Explorer — User Manual: Settings, What They Mean, What to Pick

This is a practical guide, not a code walkthrough. For every setting in the app, it explains what it controls, what happens if you set it too high/too low, and what range to actually use for a typical rice breeding panel.

1. Before you touch any settings — data checks

• —Genotype coding must be consistent across the whole file: either all 0/1/2 or all -1/0/1. Don't hand-mix them.
• —Missing data: only -9, -99, -999 are read as missing. If your file uses a different missing code, recode it before uploading, or it'll be silently treated as a real genotype call.
• —MAF filtering happens before any model runs — a marker dropped here never appears in any plot or the Excel output.

2. Tab 1 — Core run settings

MAF Threshold — default 0.05, typical range 0.03–0.10

Drops markers whose minor allele frequency is below this value.

• —Too low (e.g. 0.01): keeps very rare-allele markers, which are statistically unstable — a handful of individuals can swing the p-value a lot, inflating false positives.
• —Too high (e.g. 0.15+): throws away real, usable markers, especially in panels with several minor subpopulations.
• —Recommendation: 0.05 is a safe default for most diversity/breeding panels (100+ lines). Drop to 0.03 if your panel is small (<80 lines) and you don't want to lose markers; raise to 0.08–0.10 for very large, diverse panels (300+ lines) where rare-allele noise is more of a problem.

Fallback Top-N — default 15, range 10–20

Only used if the permutation threshold fails to compute. It's a safety net, not a real significance rule — don't rely on it to decide anything. Leave at default; if it's actually being used, something upstream (usually too few markers or too few permutations) needs fixing instead.

Permutations for LOD threshold — default 100, range 100–200

Number of phenotype shuffles used to build the empirical significance cutoff (Churchill & Doerge method). This is the real basis for calling a marker significant.

• —Too low (<50): threshold estimate is noisy — rerunning the pipeline can give you a visibly different threshold each time.
• —Too high (>200, capped internally anyway): diminishing returns, just slower.
• —Recommendation: 100 is fine for exploratory work. Use 200 (the max) for a threshold you're going to report/publish — it's more stable and the app caps compute anyway, so there's no real reason to use fewer.

Manual LOD threshold — default 0, i.e. off (automatic)

Overrides the permutation threshold with a fixed LOD value you set yourself.

• —When to use: only when you have a specific external reason to fix the threshold — e.g. matching a threshold used in a companion paper/QTL study, or comparing runs on unequal marker sets where the permutation threshold isn't comparable across them.
• —Typical values if you do use it: LOD 3–4 is the conventional QTL mapping range; for GWAS with thousands of markers you'd generally want it higher (LOD ~4–6) to stay close to a Bonferroni-equivalent stringency.
• —Default advice: leave at 0 and let permutation set it — that's the statistically defensible choice for a given dataset.

Minimum PVE % — default 0, i.e. off

Extra filter: a marker must explain at least this % of phenotypic variance in addition to passing the LOD threshold.

• —When to use: when you specifically want to report only large-effect MTAs (e.g. for marker development / MAS candidates), not every marker that's merely statistically significant.
• —Typical range if enabling: 5–10% for a moderately polygenic trait; 10–15%+ if you specifically want major-effect QTNs only.
• —Caution: setting this too high (>20%) on a polygenic trait can wipe out every hit — most real QTNs for complex traits explain a few percent each, not 20%.

3. Model selection — which of the 8 to run

You don't need all 8 every time.

• —Always include one baseline + one kinship-corrected model so you can see the difference structure correction makes: OLS + EMMAX (or MLM) is the minimum sensible pair.
• —MLM vs EMMAX: MLM adds fixed kinship-PC covariates on top of the kinship random effect; EMMAX uses kinship alone. If your panel has clear subpopulation structure (e.g. indica/japonica mix), MLM's extra Q covariates usually help. If it's a fairly uniform breeding population, EMMAX alone is often enough and slightly less conservative.
• —GEMMA: the most rigorous single-locus model (per-marker REML), but the slowest. Use it as a confirmatory check on your top hits from MLM/EMMAX rather than as a first-pass scan on a huge marker set.
• —FarmCPU / BLINK: multi-locus methods — better power when several QTNs are linked or when polygenic background masks individual effects under single-locus models. Slower, and iterative, so results can shift slightly run-to-run depending on cofactor settings (see below). Good second-pass once you know roughly what OLS/MLM/EMMAX are finding.
• —mrMLM / FASTmrMLM: also multi-locus, but with an explicit two-stage screen → joint-fit → backward-elimination design. Useful if you want an independent multi-locus method to cross-check FarmCPU/BLINK hits against (different QTN-selection logic → different false-positive profile). FASTmrMLM trades some rigor (drops kinship in the joint-fit stage) for speed — use it for quick multi-locus scans, mrMLM for the more careful version.

Practical recommendation for a standard run: OLS + MLM + EMMAX + FarmCPU + GEMMA (the app's default selection) gives you baseline, single-locus-corrected, and multi-locus coverage without running everything. Add BLINK/mrMLM/FASTmrMLM only when you want to cross-validate a specific set of candidate MTAs.

4. Cofactors / Advanced tab

FarmCPU & BLINK — pseudo-QTN iteration

Max FEM/REM iterations — default 8, range 5–12 How many rounds of "scan genome → pick pseudo-QTNs → re-check them" the method runs before stopping (it also stops early if the QTN set stops changing).

• —Too few (<4): may stop before the pseudo-QTN set has converged, under-correcting for background genetic effects.
• —Too many (>12): rarely helps — if it hasn't converged by ~10 iterations it usually won't, and you're just adding runtime.
• —Recommendation: 8 is a good default; drop to 5–6 for a quick look on a large marker set, raise to 10–12 only if you notice the results changing meaningfully when you bump it up from 8 (a sign it hadn't converged).

FarmCPU bin size (bp) — default 1,000,000 (1 Mb), range 500 kb – 2 Mb Genome is chopped into bins this wide; only the single best candidate per bin becomes a pseudo-QTN each round.

• —Too small (e.g. 100 kb): can select several near-duplicate pseudo-QTNs from the same LD block as if they were independent — inflates the covariate set with redundant markers.
• —Too large (e.g. 5 Mb): can merge two genuinely distinct, linked QTLs into one bin and only keep one of them.
• —Recommendation: 1 Mb is a reasonable default for rice (LD generally decays within a few hundred kb in most cultivated panels). If your LD decay plot (Tab 4) shows LD extending past ~1 Mb, consider widening the bin to 1.5–2 Mb; if LD decays very fast (<200 kb), you can narrow it to 500 kb.

BLINK LD r² pruning cutoff — default 0.7, range 0.5–0.9 Two candidates in the same local LD block are treated as redundant (and the weaker dropped) once their r² exceeds this.

• —Too low (e.g. 0.3): over-prunes — treats markers that aren't really redundant as duplicates, potentially discarding a real second QTN that happens to sit near a stronger one.
• —Too high (e.g. 0.95): under-prunes — keeps near-identical markers as separate pseudo-QTNs, which can waste degrees of freedom and slow convergence.
• —Recommendation: 0.7 is the standard LD-pruning cutoff used in most GWAS pipelines (similar to PLINK's common defaults). Only move it if you have a specific reason — e.g. very high average LD in your panel, where 0.8–0.9 avoids over-pruning.

mrMLM & FASTmrMLM — two-stage screening

Stage-1 screening p-value threshold — default 0.01, range 0.001–0.05 How lenient the initial scan is about calling something a "candidate" worth testing jointly in Stage 2.

• —Too strict (0.001): may exclude real but modest-effect QTNs before they even get a chance in the joint-fit stage.
• —Too lenient (0.05): floods Stage 2 with weak candidates, most of which get backward-eliminated anyway — just adds runtime and noise.
• —Recommendation: 0.01 is a sensible middle ground; loosen to 0.02–0.05 if you suspect real signals are being screened out (e.g. a highly polygenic trait with many small effects), tighten to 0.001–0.005 for a cleaner, more conservative candidate list.

Max candidate QTNs carried into Stage 2 — default 20, range 10–30 Caps how many Stage-1 candidates get fit jointly.

• —Too low (<10): may cut off real candidates before backward elimination gets a chance to sort them out, especially on a trait with many markers passing Stage 1.
• —Too high (>30): joint model becomes large and can get numerically unstable, especially with a small sample size relative to candidate count.
• —Recommendation: 20 works for most panels. Scale down toward 10–15 if your sample size is small (<100 lines) — you don't have the degrees of freedom to support a 20-marker joint model reliably.

Stage-2 backward-elimination drop threshold — default 0.05, range 0.01–0.1 In the joint fit, the weakest surviving candidate is dropped (and the model refit) if its p-value exceeds this, repeated until everyone left clears the bar.

• —Too strict (0.01): can eliminate real co-segregating QTNs that only look weak because they're sharing variance with a correlated candidate.
• —Too lenient (0.1): leaves weaker/noisier candidates in the final set.
• —Recommendation: 0.05 (standard significance level) is the sensible default; there's rarely a good reason to move it unless you're deliberately being more/less conservative about the final candidate list.

Kinship-PC covariates (per model) — default 3 for MLM/mrMLM/FASTmrMLM, 0 for GEMMA; range 0–6

Number of top kinship eigenvectors used as fixed "Q" structure covariates.

• —0: disables it — the model relies on kinship (K) alone, or on the method's own structure handling (e.g. GEMMA's per-marker REML).
• —Too many PCs (8–10): starts absorbing real trait-associated genetic variance into the covariates, reducing power to detect true associations (over-correction).
• —Recommendation: 3 PCs is the standard starting point and works for most panels with 2–4 broad subpopulations. If a PCA plot (Tab 4) shows more than ~3–4 distinct clusters, consider 4–5 PCs. For GEMMA, leaving it at 0 is usually right — GEMMA's per-marker REML already handles structure, and adding PCs on top is somewhat redundant. Never go above 6 unless you have a very large, strongly structured panel and have checked that it doesn't kill all your signal.

5. Quick-reference table

If in doubt: run once with all defaults, look at the QQ plots (λ close to 1 = well-controlled) and the PCA plot (how many clusters?), then adjust kinship-PC count and MAF threshold based on what you see — those two have the biggest practical effect on your results.