Review Papers Against Their Own Results

Visual summary (signal → process → inference)

• Signal: temporal autocovariance in neutral allele-frequency changes between generations (Cov(Δp_t, Δp_s)) caused by heritable fitness variation that persists via LD.
• Mechanism: neutral alleles stochastically associate with genetic backgrounds that affect fitness; recombination and segregation decay those associations, producing predictable autocovariance shapes dependent on VA(s), recombination R and initial LD.
• Inference: equate theoretical standardized variance-covariance (S) to observed sample Q and use method-of-moments (least squares) to estimate VA(1) and F(=1/2N), and to estimate the fraction of variance in allele-frequency change attributable to linked selection (G, G').

Primary code & data: GitHub repository and FigShare with simulation sets enable reproduction and testing (see citation below).

Key conceptual diagnostics (figures & intuition)

Below are schematic diagnostic plots you should compute from temporal genomic data before applying the method-of-moments estimators: (A) standardized variances Var(Δp_t)/[p(1−p)] per generation; (B) temporal autocovariance matrices Q (pairwise Cov(Δp_t,Δp_s) normalized); (C) LD-score (mean R^2) vs per-SNP temporal autocovariance; (D) residuals of model fit showing where migration/fluctuating selection violate assumptions.

Critical evaluation — strengths, limitations, and blindspots

• Strengths (theory + practice): formal multilocus derivation linking temporal autocovariance to additive genic/genetic variance and recombination; bridges population and quantitative genetics; validated across broad simulation grid; provides open code and simulation data for reproducibilityBuffalo & Coop 2019 derivations & data
• Important limitations / empirical pitfalls (paper acknowledges many):
  • Assumes a closed, constant-size population — migration/gene flow can create spurious covariance (seasonal migration, admixture) and differentially impact low-recombination regions;
  • Relies on additive, time-invariant effect sizes (a_l constant) — fluctuating or changing selection over short timescales can produce negative autocovariances and bias G estimates unless modeled carefully;
  • Proxying selected-site heterozygosity dynamics with neutral-site SSH (SSH_n) relies on polygenicity and weak selection — may fail with large-effect sweeps or when selected-site frequency dynamics differ from neighboring neutrals;
  • Finite sampling (sequencing depth, sample size) produces shared sampling noise between adjacent timepoints that biases covariance estimates and must be corrected as shown in the Appendix;
  • Ignored cross-association term (E[D_{t,k} D_{s,l}] for k≠l) can become important under Hill–Robertson interference when selected sites are tightly linked or numerous, shifting from V_a to V_A as the determining variance in autocovariance (paper documents this via simulations).
• Practical consequences: the method is best applied to moderate sample sizes with consecutive-generation sampling where possible, good recombination maps or local LD estimates, and after evaluating gene flow and sampling-bias diagnostics (PCA, admixture analysis). The method-of-moments estimator uses the first few generations after selection onset (the paper uses five) when the SSH proxy is most accurate.

Concrete recommendations for empirical application

1. Design: sample consecutive generations for a subset of timepoints (e.g., 1–5 sequential generations) to accurately estimate Var(Δp) without large shared sampling noise; use a mixed-duration sampling scheme for longer experiments.
2. Pre-checks: run PCA and admixture tests to detect migration/admixture; if present, model or partition samples to avoid confounding as the paper warns.
3. LD inputs: where possible compute empirical R^2 between focal neutral SNPs and neighboring polymorphisms and compute A(R;t;s) via Equation 55 in the Appendix rather than assuming the drift–recombination balance formula.
4. Sampling noise correction: apply the bias correction from the Appendix (Equations 98–99) or use split-library strategies to eliminate shared technical noise between adjacent timepoints.
5. Model choice: if evidence of selective interference/HRi or tight linkage exists (many nearby selected sites), prefer using empirical V_A(s) (additive genetic variance including LD) to predict temporal autocovariance rather than assuming genic V_a(s) alone.
6. Fluctuating selection: test for negative temporal autocovariances across seasons/time blocks — paper shows this as a robust signal of selection flips and recommends G_abs to quantify total covariance magnitude.

How to apply the paper’s pipeline to your time-series genomic data (stepwise)

1. Compute allele-frequency changes Δp_t across t generation intervals for M putatively neutral SNPs.
2. Compute sample standardized covariance matrix Q (pairwise-complete, bias-corrected for sampling noise).
3. Calculate per-window empirical A(R;t;s) by summing pairwise R^2×(1−r)^{s−t} over neighboring variants (Equation 55).
4. Formulate q⃗ = VA(1)_b · a⃗ + F · b⃗ + ε⃗ (Equation 20) and estimate VA(1)_b and F via least squares (GMM-like), using bootstrap over windows for CIs.
5. Compute G and G' or G_abs to quantify proportion of variance attributable to linked selection (Equations 24–27).
6. Validate: simulate under inferred VA(1) and N to check whether observed Q is recovered (paper used msprime + forward sims; code available).

Open problems and avenues for improvement

• Integrate cross-association terms (the k≠l LD products) into analytic theory to handle HRi regimes and tightly linked selected sites.
• Extend method-of-moments to allow time-varying effect sizes a_{t,l} (fluctuating selection) and incorporate seasonal sampling explicitly.
• Develop likelihood/composite-likelihood methods using per-SNP empirical LD vectors (akin to LD-score regression) as suggested in Discussion, to improve efficiency and localization.
• Incorporate migration/admixture explicitly (e.g., jointly modeling drift+admixture components) to avoid false positives in natural populations with gene flow.

Conclusion (concise)

Buffalo & Coop (2019) present a clear, testable, and useful theoretical + simulation framework showing that linked selection during polygenic adaptation leaves a distinctive temporal autocovariance signal in neutral allele-frequency changes. Their estimators are practical and reproducible (code/data provided), but empirical users must guard against migration, sampling noise, strong interference, and fluctuating selection — all of which the authors explicitly discuss and partially quantify. When applied carefully, temporal autocovariance opens a powerful window to estimate additive genetic variance for fitness from genomic time series and to partition drift vs selection contributions over ecological timescales.