Seven anomalies in the first published Wuhan sequences — three weeks of joint analysis, six years of silence from the original authors.
In January 2020, a team led by virologist Yong-Zhen Zhang published the first complete genome of SARS-CoV-2 in Nature. The raw sequencing data — the actual digital readout of genetic material scraped from the lungs of the five earliest known COVID-19 patients — was deposited in a public database under the identifier PRJNA605983.
These samples were collected in December 2019. The patients were admitted to Wuhan Jinyintan Hospital with unexplained pneumonia. The BALF samples — fluid drawn directly from the lungs — were sequenced on high-throughput machines and the output uploaded to NCBI, the US-run international genomic database, on 16 February 2020.
For four years, this dataset sat in the public domain. Researchers downloaded it, extracted the SARS-CoV-2 sequence, and moved on. Very few asked the more basic question: what else is in here?
On 18 June 2026, @BillyBostickson publicly asked whether anyone had run the junction read test on this dataset. I ran it. On 23 June I published the first results on X: zero RBD-Fc junction-spanning reads in either sample, full methodology, raw numbers, reproducible pipeline. Dr. Rogier Louwen — a molecular biologist at CCassured in Breda, who had independently been analysing the companion dataset PRJCA002163 — saw the post and reached out. What followed was three weeks of joint work, spanning two countries, two datasets, and seven findings that the original authors have never addressed.
On 24 February 2020 — forty-four days after the first public release of the SARS-CoV-2 genome — a PLA military researcher named Brigadier-General Yusen Zhou of the Academy of Military Medical Sciences filed a patent for a SARS-CoV-2 vaccine.
The patent number is CN111333704B. The vaccine design: an RBD-Fc fusion protein — the receptor binding domain of the SARS-CoV-2 spike protein, fused directly to a human antibody fragment (IgG1-Fc) without a linker. This is not an obvious design choice. It is a specific technical signature.
The RBD-Fc fusion architecture was not new to Zhou. He had used it before — for his 2017 MERS-CoV vaccine. Based on that prior work, Robert Kadlec (Texas A&M) calculated that the minimum time from concept to patent filing for an RBD-Fc construct is approximately four months. Working backwards from 24 February 2020: Zhou's team would have had to begin work no later than October or November 2019 — before the outbreak was declared, before the genome was published, before the world knew the virus existed.
The same conclusion — independent of any assumptions about lab leaks or intent — is what motivated Louwen to look at the sequence data. If Zhou was working on a SARS-CoV-2 vaccine before December 2019, would traces of that construct appear in the clinical samples processed at the same institute?
When you sequence a patient's lung fluid, you are sequencing everything in that fluid. Mostly human RNA from immune cells, some viral RNA from whatever pathogen caused the infection, and — ideally — nothing else. Any non-human, non-viral sequences in the output represent contamination of some kind. The question is always: contamination from where?
We used standard bioinformatics tools — BWA-MEM for alignment, samtools for analysis — to compare every read in the five datasets against a panel of reference sequences: SARS-CoV-2, the Zhou patent construct, Nipah virus, influenza H7N9, a CRISPR gene-editing vector, and an HPV-16 expression vector. Then we looked at what was there, and — critically — what was there in only some of the samples, but not others.
SARS-CoV-2 is an RNA virus. RNA viruses mutate constantly — at a rate of roughly one mutation per thousand bases per year. Over four months of circulation among patients, you expect to see variation. It is the fundamental signature of an evolving pathogen.
We measured nucleotide identity in 111 assembled SARS-CoV-2 genomes from PRJCA002163 across two independent genomic regions:
RBD (667 nt): 109/111 assemblies at 100.000% identity
nsp12/RdRp (3,212 nt): 96/111 assemblies at 100.000% identity
The RBD is the region encoded by SEQ ID NO. 4 of Zhou Yusen's patent (CN111333704B) — the receptor binding domain fused directly to an IgG1-Fc fragment. At the observed mutation rate, four months of natural circulation would produce ~0.17 substitutions per RBD copy and ~1.07 per nsp12 copy. Across 111 genomes, you would expect roughly 19 substitutions to have accumulated in the RBD alone. The observed count approaches zero.
The phylogenetic tree makes the pattern visible: the genome tree collapses almost entirely into a star, with 83 out of 117 internal branches at effectively zero length — the topology of a clonal population, not a naturally evolving virus spreading through a human population for four months.
In the raw sequencing data from patient WIV05, we found 32 reads containing a specific 43-letter DNA sequence: the SpyCas9 gRNA scaffold — the structural backbone that every CRISPR guide RNA uses to dock with the Cas9 cutting enzyme.
What makes it remarkable is what comes immediately before the scaffold in every single read: a 20-letter spacer sequence — the "address" that tells CRISPR which gene to cut. We identified six distinct spacer sequences across the 32 reads and aligned them against the human genome, the SARS-CoV-2 genome, and the mouse genome.
0 spacers → human genome 0 spacers → SARS-CoV-2 6/6 spacers → Mus musculus
All six spacers target mouse genes: Cd68 (macrophage marker), Sat1 (polyamine metabolism), Mecp2 (epigenetic regulator), and Procr (endothelial Protein C receptor). Each read shows the same architecture: a 20-nucleotide spacer sequence, a PAM motif matching the mouse genome target, and the downstream 42-nucleotide SpCas9 scaffold — the defining signature of a CRISPR guide RNA cassette. These sequences were absent from all four negative control samples (CRA002390, Wuhan University, same patients).
Why there are no SpyCas9 reads — and why that matters. RNP delivery packages SpyCas9 as a pre-formed protein complexed with the guide RNA in lipid nanoparticles. No Cas9 DNA or RNA ever enters the cell — only protein, which is degraded after cutting and is invisible to sequencing. Only the guide RNA cassette leaves a detectable trace. RNP is chosen specifically when the goal is genome editing without leaving forensic traces. The absence of Cas9 reads is the expected signature of this method, not a gap in the evidence. (Dr. Rogier Louwen has performed this exact protocol in the laboratory and confirmed this interpretation independently.)
The dataset exists in two parts. The raw sequencing files were deposited in NCBI in February 2020. A companion set of 114 assembled SARS-CoV-2 genomes was deposited in the Chinese NGDC database (PRJCA002163) in two separate batches.
The second batch — 101 samples (SAMC311431–SAMC311531), submitted January 2021 — carries structured GPS metadata that reads:
38.98°N, 77.11°W — free-text: "China: Wuhan"
Those coordinates are not in Wuhan. They map to Walter Reed National Military Medical Center, Bethesda, Maryland — the United States Army's flagship hospital. The earlier batch (January–March 2020) carries correct Wuhan coordinates. The discrepancy affects 101 of 114 entries, all submitted in a single batch a full year after the original deposit.
We built a phylogenetic tree of 117 SARS-CoV-2 genomes from the first Wuhan patients, calibrated against the two closest known bat coronavirus relatives: ZC45 and ZXC21.
ZC45 and ZXC21 show branch lengths 100 to 1,000 times longer than all WIV assemblies combined. That gap represents decades of natural bat coronavirus evolution — the empirical baseline for what genuine evolutionary distance looks like in this virus family.
83 of 117 internal branches in the WIV tree have length ≈ 0. A natural outbreak spreading through a human population for four months accumulates mutations across every transmission chain. The phylogenetic signature is a branching tree. What we see is a star — nearly all sequences collapsing to a single origin point with zero divergence between them.
RecCA phylogenetic rooting — a molecular clock method that corrects for root-placement bias — assigns 90.2% posterior probability to Lineage A as the ancestral root of the pandemic (vs. 8.0% in unconstrained models). Lineage A is not associated with the Huanan Seafood Market. Lineage B is. The data places the ancestral sequence outside the market cluster.
SARS-CoV-2 has four spike protein insertions that are absent from RaTG13, its closest known bat coronavirus relative. Two of those insertions — the ones unique to SARS-CoV-2 — carry restriction enzyme recognition sequences at their boundaries.
Insert 1 (90 nucleotides, positions 21,740–21,829, absent from RaTG13): flanked by NcoI-compatible sequences. NcoI is a standard restriction enzyme used for directional cloning into expression vectors, generating a 4-nucleotide CATG overhang.
Insert 4 — the furin cleavage site (positions 23,453–23,752, absent from RaTG13): flanked by two NdeI recognition sequences, one 65 nt upstream and one 560 nt upstream of the furin site. NdeI is the standard restriction site for directional insertion into pET-series expression vectors. The 495-nucleotide distance between the two NdeI sites falls within the standard size range for restriction-fragment inserts in coronavirus reverse genetics systems.
The two insertions shared with RaTG13 (Inserts 2 and 3) carry no equivalent restriction signatures. The pattern — cloning enzyme sites at the two SARS-CoV-2-specific insertions, absent at the shared insertions — is consistent with the methodology described in the DEFUSE proposal (Daszak et al., DARPA 2018). We note that restriction enzyme recognition sequences arise naturally in viral genomes, and multiple alternative explanations remain formally possible.
Two pathogens appear in the sequencing data from the earliest COVID-19 patients that have no clinical explanation — and neither is mentioned anywhere in the paper's methods.
H7N9 hemagglutinin (KC853766.1) in WIV07-2:
SARS-CoV-2: 87.8× average depth
H7N9 HA: 1,472× average depth / 100% breadth
H7N9 is 16.8 times deeper per base than the actual disease-causing virus. Coverage is uniform across all 1,683 base pairs of the hemagglutinin gene. The other seven influenza genome segments are effectively absent. A natural H7N9 infection produces reads across all eight segments — finding only segment 4 at extraordinary depth is the signature of a cloned HA insert in a high-copy expression vector (pVAX1-H7N9-HA). The Wuhan Institute of Virology published H7N9 HA DNA vaccine work in mice during 2019–2020. Dr. Louwen independently identified GFP, origin of replication, and antibiotic resistance cassette elements in the companion dataset — the backbone hallmarks of an expression vector.
Nipah virus (AY988601, Bangladesh strain, BSL-4, ~70% case fatality rate): 634 reads in WIV05, 683 reads in WIV07-2. Zhang (2021, Zenodo:4486195) showed these reads carry three construction markers absent from natural Nipah RNA: a 3'-HDV ribozyme, a T7 terminator, and a tetracycline resistance gene — the structural scars of an assembled infectious clone. Both signals are absent from the negative control dataset (CRA002390, Wuhan University, same patients, four MiSeq runs).
Every read produced by a sequencing instrument carries the instrument's identifier in its file header. The five deep-sequenced BALF libraries — the ones containing H7N9, Nipah, and the CRISPR scaffold — all carry flowcell ID v300043428 in every read. That is a BGI identifier. The instrument is a MGISEQ-2000RS.
In the NCBI database, those same five libraries are recorded as Illumina HiSeq 3000 — a different machine, different chemistry, different company. Six independent lines of evidence confirm the mislabeling: the read headers, the SRA alias field containing the original BGI filename, the depositors' own design description ("performed on the MGISEQ-2000RS platform"), the parallel NGDC deposit where the same authors correctly recorded MGISEQ-2000RS, the BGI-proprietary library prep kit listed in the submission, and the selective pattern — the three MiSeq libraries in the same submission are labeled correctly. Only the five MGISEQ libraries are not.
The instrument identity determines which sequencing centre processed these samples and is the anchor for tracing construct signals to their source. Mislabeling it in every international database simultaneously — in a deposit made 15 days after publication, not cited in the paper's data availability statement — systematically removes that traceability.
When you find unexpected sequences in patient data, the first question is always: could these patients simply have been co-infected with multiple pathogens at once? This is a legitimate alternative explanation — and we tested it directly.
Alongside the five deeply-sequenced MGISEQ samples, the same dataset contains a sixth sample: WIV06 (SRR11092056), sequenced on a separate MiSeq instrument. This is critical. If H7N9, Nipah, and the HPV-16 vector were genuinely present in the patients, they should appear in a sample from the same patients regardless of which machine sequenced it.
| Signal | MGISEQ samples | MiSeq control (WIV06) | Conclusion |
|---|---|---|---|
| SARS-CoV-2 | 97–100% breadth | 45% breadth | Present both platforms (different patient load — expected) |
| H7N9 HA | 100% breadth / 1,472× | 4.5% breadth / 0.09× | MGISEQ only |
| Nipah | 32–43% breadth | 8.3% breadth / 0.27× | MGISEQ only |
| HPV-16 vector | 91.5% breadth / 877.9× | 10.5% breadth / 0.3× | MGISEQ only |
SARS-CoV-2 — the real infection — shows up on both platforms. All three of the anomalous signals show up overwhelmingly on the MGISEQ and barely at all on the MiSeq. The "barely at all" levels (4–10% coverage breadth, sub-1× depth) are what bioinformaticians call LAHH noise — Low Abundance High Homology background scatter that every alignment produces by chance. Patient co-infection would produce signal on both platforms. The data does not show that.
This result independently confirms a 2021 finding by @BillyBostickson and Y. Ghannam, who documented the same machine-specific clustering pattern using a different method (Section 13 of their ResearchGate paper). We reproduced it with our own alignment data.
One of the most important contextual facts about this investigation is that we were not the first to find these things. Independent researchers had identified versions of these signals years before our analysis — and none of it was ever formally addressed.
Sucharit Chakraborty (2020) — First documented the Nipah Bangladesh signal in PRJNA605983. No institutional response.
Abouelkhair MA (2020, PeerJ) — Peer-reviewed detection of H7N9 and other non-SARS-CoV-2 sequences in PRJNA605983. Interpreted as co-infection; subsequent structural analysis (pVAX1 architecture, single segment only) supports contamination over co-infection.
Quay, Rahalkar, Jones & Bahulikar (2021, Zenodo:5067706) — Assembled a complete 4,765 bp pVAX1-H7N9-HA plasmid de novo from WIV07-2 reads using SPAdes. Also documented Nipah and Sf9 insect cell markers (Spodoptera frugiperda rhabdovirus). Machine-specific clustering independently noted.
Zhang D. (2021, Zenodo:4486195) — Independent analysis of SRR11092059–62. Identified Nipah with HDV ribozyme + T7 terminator + tetracycline resistance (infectious clone markers) and H7N9 HA under CMV promoter + bgH poly(A) — the canonical pVAX1 architecture.
Bostickson W, Ghannam Y (2021, ResearchGate) — "Proposed Forensic Investigation of Wuhan Laboratories." Section 13: machine-specific clustering of construct signals on MGISEQ vs. MiSeq.
Five years elapsed. The dataset remained publicly available. The original authors published no response, no correction, no clarification. Louwen and I are submitting the first formal scientific comment that attempts to compile all of this into a structured, reproducible, peer-reviewed challenge.
We want to be precise about what this investigation establishes and what it does not.
What the data shows: The earliest published COVID-19 patient samples contain sequences from multiple laboratory constructs — a CRISPR toolkit targeting mouse genes, an H7N9 influenza vaccine gene at extraordinary depth, a Nipah infectious clone, an HPV-16 research vector — that are instrument-specific and absent from the negative control dataset. The assembled genomes from the same cohort show near-complete clonality in the region of a military patent. GPS metadata points to a US military hospital. The sequencing platform was mislabeled in every international database simultaneously.
What the data does not prove: We cannot determine from sequence data alone whether these signals entered the samples during collection, during library preparation, during sequencing, or during database submission. We cannot prove intent. We are not asserting that any specific person did anything deliberately. These are anomalies in a public dataset, submitted by the original authors under their names, that require a scientific explanation.
The original paper — Wu et al. (2020), Nature — remains one of the foundational documents of the COVID-19 pandemic. It announced the pathogen that killed millions of people. The raw data deposited alongside it contains questions that have not been answered in six years. Answering them is not an accusation. It is science.
The joint analysis ran from late June to mid-July 2026. The tools we used — BWA-MEM, samtools, MAFFT, IQ-TREE2 — are freely available. The datasets are publicly accessible through NCBI SRA and NGDC. Every step is documented in a reproducible pipeline on Zenodo. Another team with a modern computer and an internet connection can reproduce every result.
The manuscript was submitted to Nature in July 2026 as a formal comment on the original Wu et al. (2020) paper. A preprint is available on bioRxiv. The full technical analysis — with exact alignment statistics, depth profiles, CIGAR-correct junction detection, and code — is documented separately at sovereignhealthbotanicals.com/wuhan-analysis.html.
Authors: Jasper Vermeer¹ · Dr. Rogier Louwen²
¹Sovereign Health Botanicals, Ghent, Belgium
²CCassured, Breda, The Netherlands
This analysis was conducted independently and is not affiliated with any academic institution or government body. No financial interest in any outcome of the COVID-19 origins debate.
Correspondence: jasper@sovereignhealthbotanicals.com