# Reproduction results

> **CORRECTION (2026-09-28).** The first R1 pass analyzed HA in the WRONG READING FRAME (frame 0; the
> GISAID HA alignments carry a 2-column frame-2 offset, same as NA). MEME on scrambled codons manufactured
> a spurious "avian HA episodic selection" signal (80/85 sites). After reframing HA to frame 2 and rerunning,
> **that signal disappears.** The corrected results are below; the earlier claim is RETRACTED. The R2
> single-clock run also produced nonsense (tMRCA 1332 CE) and was replaced by `chronaeon autoclock`
> multi-clock deconvolution (R2b). Both errors were caught by the adversarial-verification workflow + the
> autoclock cross-check.

## R1 (corrected) — Site-level episodic selection (`hyphaeon meme`), HA in frame 2

**Run:** SLURM array `1337114` (HA, reframed) + retained NA/PB1 from `1337086`. `--no-tree` (TN93), `--cpu`,
datamonkey partition. HA & NA use frame-2 `.inframe.fasta` (`scripts/01_reframe_na.py`); PB1 is frame-0 clean.

### FDR-significant episodic-selection sites (q≤0.05)

| gene | N.Am avian | Eurasian avian | swine | human | equine | canine |
|------|-----------:|---------------:|------:|------:|-------:|-------:|
| **HA (frame 2, corrected)** | **0** | **0** | 1 | 1 | 0 | 0 |
| HA (frame 0, INVALID — retracted) | 80 | 85 | 112 | 59 | 2 | 0 |
| NA (frame 2) | 5 | 1 | 2 | 1 | 0 | 0 |
| PB1 (frame 0, correct) | 0 | 0 | 0 | 1 | 0 | 0 |

(HA variable codons after reframing: N.Am avian 243, Eurasian 309, swine 403, human 331, equine 150, canine 133.)

### Interpretation — M2 is NOT supported by this test

1. **In the correct frame, avian HA shows ~zero FDR-significant episodic sites** — and so does every other
   host (max 1). The earlier "avian HA is not quiescent" finding was entirely a frame-0 codon-scrambling
   artifact. Corrected, `hyphaeon meme` does **not** contradict the paper's "little directional selection in
   birds"; the near-zero avian-HA result is broadly *consistent* with the paper.
2. **PB1 control unchanged** (~0 everywhere) — correct behavior in both passes; it was frame-0-clean.
3. **NA unchanged** (low in all hosts) — NA was correctly reframed in the first pass.
4. **Caveat worth flagging to ourselves:** even human/swine HA show only ≤1 FDR site across ~330–400 variable
   codons. Human seasonal H3 HA is the textbook positive-selection case, so ≤1 site suggests our
   `hyphaeon meme --no-tree` pass may be **underpowered** here (TN93 distances, no ML tree, harsh FDR, or the
   general model variant not tuned for within-host-lineage HA). This is a limitation of our test, not evidence
   about the biology. R1c (below) resolved this with a tree-based rerun.

### R1c — tree-based rerun RESOLVES the power question (FINAL)
Reran MEME with the authors' per-host tree PRUNED to the alignment tips (`scripts/04_prune_trees.py`,
`run_memetree_array.sbatch`, array 1337127), replacing --no-tree/TN93. FDR-significant sites (q≤0.05),
tree vs TN93, are **essentially identical**:

| host | class | TN93 var/p≤.05/q≤.05 | TREE var/p≤.05/q≤.05 |
|------|-------|---------------------:|---------------------:|
| N.Am avian | avian | 243/22/**0** | 243/21/**0** |
| Eurasian avian | avian | 309/29/**0** | 309/30/**0** |
| swine | mammal | 403/41/**1** | 403/43/**1** |
| human | mammal | 331/43/**1** | 331/43/**1** |
| equine | mammal | 150/6/**0** | 150/6/**0** |
| canine | mammal | 133/12/**0** | 133/12/**0** |

**The tree-vs-TN93 choice was NOT the power limiter** — avian HA is genuinely ~0 FDR episodic sites under the
authors' topology. The near-zero result is real, not a distance-model artifact. **M2 is DEFINITIVELY withdrawn:**
an independent site-level episodic test agrees with the paper that avian HA carries little detectable positive
selection. Caveats retained: (a) trees are a shared per-host backbone (identical across HA/NA/PB1), not
HA-specific segment trees — MEME re-optimized branch lengths, so topology is theirs but HA divergence isn't;
(b) uniformly low counts (≤1 even in human H3 HA) suggest the `general` model variant / FDR is conservative
for within-host-lineage HA — which biases toward MISSING selection, so a ~0 avian result is a safe, not
overstated, conclusion.

## R2b (corrected) — `chronaeon autoclock` multi-clock deconvolution, HA

**Run:** SLURM array `1337115` on frame-2 HA. Replaces R2 single-clock (tMRCA 1332 CE nonsense from pooling
divergent avian sublineages). Per-community rates (subs/site/yr):

| host | k | community rates (μ) | tMRCA range |
|------|--:|---------------------|-------------|
| N.Am avian | 2 | 1.30e-3, 1.57e-3 | 1961–1964 |
| Eurasian avian | 5 | 6.6e-4, 1.62e-3, 7.44e-3, 8.86e-3, 6.63e-3 | 1964–2007 |
| swine | 2 | 2.30e-3, 5.45e-3 | 1958–1996 |
| human | 3 | 3.13e-3, 2.21e-3, 4.34e-3 | 1967–1980 |
| equine | 6 | 3.2e-4 … 7.22e-3 | 1977–2012 |
| canine | 2 | 1.60e-3, 1.84e-3 | 2002–2014 |

### Interpretation — M4/m5
- All rates are now in the plausible influenza HA range (~6e-4 to 9e-3), tMRCAs sane (1958–2014). autoclock
  fixed the single-clock pathology (Steven's call).
- **N.Am avian HA clock (~1.3–1.6e-3) is NOT slower than swine (2.3–5.5e-3) or human** — the single-clock run
  had wrongly deflated avian to 1.4e-4. So the reviewer's M4 worry (that our clocks would rescale and flip the
  avian:swine reassortment-rate ordering) is **not** supported: our avian and mammalian HA clocks are within a
  factor of ~2–4 of each other, not orders apart. The paper's clock-scaling explanation for the
  events-vs-rate discrepancy is not obviously wrong.
- **BUT this is still not a clean test of M4:** we clocked single-gene HA with TN93/--no-tree, not the paper's
  per-segment TreeTime clocks on their trees. To actually challenge the reassortment-rate ordering we'd need
  the paper's reference-segment clocks (TreeSort scales by the *reference* segment rate) on matched trees.
  **R2b status: our clocks are now internally sane, but M4 remains INCONCLUSIVE as a challenge to the paper.**

## Net effect on the review
- **R1 → retracted** (frame bug); M2 critique withdrawn pending a powered rerun.
- **R2 → R2b**; clocks now sane; M4 worry not supported, but not a clean test → inconclusive.
- The paper looks **better**, not worse, after correction: neither M2 nor M4 lands with this data.

## R3 (M3) — skipped (no TreeSort summary trees available). Documented blocker.

## Dataset-curation audit (2026-09-28)
Audited the input GISAID per-host HA alignments for curation problems independent of our pipeline:
- **Host labels:** 0 wrong-class isolates (no cross-host contamination). `A/Turkey/8543/2019` in the human
  file is Turkey-the-country (human seasonal), correctly placed.
- **Duplicates:** 0 duplicate headers. Identical-sequence clusters (77 na_avian, 76 swine) are all from
  DISTINCT strain names = independent isolates sharing a conserved HA (expected under dense sampling), not
  redundant accessions. hyphaeon meme collapses 100%-identical seqs by default regardless.
- **Dates:** 100% dated, all in 1963–2024, none out of range.
- **Ambiguity:** 0 sequences >10% N/gap.
Conclusion: authors' curation (hosts, dates, dedup, sequence quality) is sound. The only data-handling issue
was the 2-column frame-2 offset, which is a consumer-side requirement (we mishandled it initially), not a
curation error. No curation finding added to the review.