The canonical two-introduction paper. Reconstructs the phylogenetic network of the ~2 early SARS-CoV-2 lineages (A and B, differing at two linked positions) from genomes collected before March 2020, then uses coalescent-based epidemic simulations to ask how often a single introduction would be expected to generate the observed topology plus case-count/diversity pattern versus two independent introductions. Concludes lineage B likely entered humans around 18 Nov 2019 (23 Oct–8 Dec range) with lineage A’s introduction following within weeks, favoring (Bayes factor) at least two zoonotic spillovers over one. A 2024 erratum corrected a code error in the Bayes-factor calculation (it had overestimated support for multiple introductions); the corrected number still nominally favors two introductions but is much smaller, and this correction itself became the anchor for several later critiques (Weissman, McCowan) arguing the true corrected number favors one introduction.
relevance_note: the load-bearing quantitative case that the phylogenetic/genomic record alone supports (at least) two independent zoonotic spillovers rather than one lab-associated introduction.
Results — early genomic structure
O-4 - Early SARS-CoV-2 split into two lineages A and B differing by two linked substitutions
Analysis of the early genome collection (787 near-full-length lineage A/B genomes sampled by 14 Feb 2020) shows two dominant co-circulating haplotypes separated by two linked positions. This two-mutation gap with two large clades is the structural fact the rest of the paper reasons about.
Link to original
O-5 - The 20 apparent A-B intermediate genomes are low-coverage and lab-clustered
Twenty genomes contained either T28144C or C8782T but not both (C/C or T/T). Examination showed depth of coverage at or below ~10x at the key positions and lab-specific shared rare mutations — the empirical pattern the authors use to argue these are not genuine circulating intermediates.
Link to original
H-6 - The A-B intermediate haplotypes did not genuinely circulate in humans
Load-bearing sub-claim: if the intermediates were real, a single introduction that mutated within humans is back on the table; if they are artifacts, the two-mutation gap must be bridged by two introductions.
Link to original
A-4 - Low-coverage lab-clustered intermediates indicate contamination, not real transitional genotypes
General methodological principle: true low-frequency circulating variants are not expected to co-occur preferentially with lab-specific rare mutations nor to sit at positions with borderline read depth; contamination and assembly artifacts do exactly that. Since all 20 intermediate genomes exhibit at least one of these tells (>=10x-or-worse coverage at 8782/28144, same-lab shared rare mutations), they should be discounted as evidence for a circulating C/C or T/T intermediate. Removing them removes the principal genomic support for a single introduction whose descendants mutated into A and B within humans, so the two-mutation gap must instead be bridged across the human-population boundary — i.e. by separate introductions.
Validity verdict
status: approved; reason_if_not_false: checked.
The statement’s conclusion is the negative claim that these 20 genomes provide no evidence of a genuine within-human A-to-B intermediate. Reconstructed step: (premise) true low-frequency circulating variants are not expected to preferentially co-occur with lab-specific rare mutations or sit at borderline read depth, whereas contamination/assembly artifacts produce exactly that pattern; (premise) all 20 intermediates carry at least one such tell; (therefore) each of the 20 is better explained as artifact than as a real intermediate, so they do not support a circulating C/C or T/T genotype. Traced myself: conditional on the premise that these features are the recognized artifact signature and that all 20 exhibit one, an artifact explanation screens off the intermediate signal, so the collection provides no independent evidential support for a real transition. Probed the obvious defeater — a genuine variant that merely happens to be low-coverage — but the premise (co-occurrence with lab-specific shared mutations, not just low depth) is what defeats it, and it is granted here. The statement stops at “no evidence of a genuine within-human transition”; the further body leap to “therefore separate introductions” is not part of the claim being scored (and would have its own unsampled-intermediate defeater). As scoped, the inference holds.
Link to original
Results — molecular dating
O-6 - Molecular-clock dating places tMRCA in mid-December 2019 and lineage B primary infection around 18 Nov 2019
Dating used recCA rooting (a reconstructed recombinant common ancestor across 15 non-recombinant sarbecovirus regions) as well as unconstrained rooting. Median lineage-B tMRCA 15 Dec, lineage-A tMRCA 20 Dec; lineage B primary case predated lineage A in 64.6% of the posterior, by a median 7 days.
Link to original
Results — epidemic simulations
A-3 - Single introduction almost never reproduces the two-lineage two-mutation topology, favoring two or more introductions
The inference: simulate SARS-CoV-2-like epidemics with FAVITES-COVID-Lite on a scale-free network of 5 million individuals using a customized SAPHIRE compartmental extension (doubling time 3.47 days, 95% HDI 1.35-5.44), generate coalescent trees, and simulate mutations along them. Under a single introduction, the most common outcome is one large basal polytomy (>=100 lineages), seen in 47.5% of runs; a topology matching the observed two-clade / two-mutation-gap structure arises in only 0.1% of runs when the ancestral haplotype is the intermediate C/C, and 0.5% when it is an ancestral lineage A or B haplotype with a large polytomy plus the two-mutation gap. Comparing the likelihood of the observed topology under one vs two introductions gives BF = 61.6 (recCA rooting) and 60.0 (unconstrained rooting).
Validity caveat (load-bearing for scoring): a 2024 erratum corrected a code error that had inflated the reported frequency of single-introduction topologies, so the corrected Bayes factor is much smaller (still nominally favoring two introductions but by roughly an order of magnitude less), and independent reanalyses (Weissman, McCowan) argue the intended/corrected model actually favors a single introduction. The argument’s direction is therefore genuinely contested, and its strength is model- and ascertainment-dependent (simulations cover a pre-ascertainment period that could not be empirically validated).
Original
Realistic epidemic simulations reproduce the observed pattern of two distinct early lineages separated by two mutations (each forming a basal polytomy) in only 0.1-0.5% of runs under a single introduction, yielding a Bayes factor of about 60 in favor of two or more introductions over one.
Validity verdict
status: corrected; reason_if_not_false: checked.
The general inferential schema is valid and elementary: if the observed topology is much rarer under a single introduction than under multiple, its likelihood ratio is evidence for multiple introductions. What fails is the specific quantitative claim baked into the statement. This is not merely a “premise is false, step is sound” case that would leave it approved: the load-bearing computational step (the simulation frequencies feeding the Bayes factor) is the very thing a 2024 erratum corrected for a code error, shrinking the reported
Link to original0.1-0.5% / BF60 by roughly an order of magnitude, and independent reanalyses (Weissman, McCowan) argue the corrected model can point the other way. So the “BF ~60 favors two introductions” claim does not hold as stated; only the weaker “the two-lineage/two-mutation basal topology is unusual under a single introduction and thus is evidence for multiple introductions” survives, with magnitude (and possibly sign) unresolved — the classic “proves” → “is evidence for” downgrade. Corrected the statement accordingly; step 8 must price the actual, erratum-corrected and disputed magnitude.
H-7 - No substantial cryptic human circulation of SARS-CoV-2 before December 2019
Simulated epidemics matched to the data show a median of three cumulative infections at the tMRCA (95% HPD 1-18) and median zero hospitalizations by 1 Dec 2019, arguing against a long pre-December cryptic phase.
Link to original
Discussion
H-5 - Early A-B diversity reflects two or more separate zoonotic introductions, not one
Central claim of the paper. Contestable and contested: a 2024 erratum reduced the supporting Bayes factor, and independent reanalyses (Weissman, McCowan) argue the corrected model actually favors a single introduction.
Link to original
A-5 - Both lineages spread without within-human adaptation, implying a pre-adapted animal reservoir
If two independently introduced lineages both achieved sustained spread with no shared adaptive changes, then human-transmissibility did not have to evolve after the jump; it pre-existed in the source population. A reservoir carrying already-human-capable virus makes repeated spillover a series of Bernoulli trials rather than a singular fluke: the authors estimate about eight introductions (95% HPD 2-23), most going extinct (77.8% of simulated introductions), were needed to yield the two that established. This raises the prior on two-or-more successful zoonotic introductions relative to a scenario requiring a single unique jump.
Validity verdict
status: approved; reason_if_not_false: checked.
Chain: (a) both lineages spread with no shared adaptive changes → (b) human-transmissibility pre-existed the jump rather than evolving in humans → (c) with a pre-adapted source and repeated human-reservoir contact, each contact is a Bernoulli trial, so multiple successful introductions become plausible → (d) prior on two-or-more introductions rises relative to a single unique jump. Traced myself: (a)→(b) holds on the premise as granted; (c) is the substantive move, and it checks out probabilistically — a higher per-contact success probability p increases the ratio P(>=2 successes)/P(exactly 1) across many trials, so pre-adaptation does specifically raise multiple-vs-single, not just overall spillover odds. This is scored against H-5 (number of introductions), where the inference holds. Noted defeater for the record, not load-bearing for this scope: “no in-human adaptation needed” is symmetric between an animal reservoir and a lab-passaged/pre-adapted source, so the “animal reservoir” wording does not by itself discriminate zoonosis from lab origin — but that bears on step 8’s cross-hypothesis pricing, not on whether the single-vs-multiple inference is valid. Holds as stated.
Link to original