Can AI complete a bioinformatics paper on its own?
We published the whole record.

All six share one structure: three roles and four automation mechanisms.

• The human: says one sentence at the start, for example "let us collaborate on a topic in reproductive medicine" — not even the research direction is specified.
• AI role A, the clinician: asks questions, reviews, makes the call, enforces honesty. It genuinely downloads and opens every figure to check it cell by cell rather than accepting a summary.
• AI role B, the bioinformatics researcher: designs, codes, executes, plots and drafts, performing all computation inside a cc-bioinfo session.

Four automation mechanisms keep the process auditable: the structured handoff contract TOPIC.yml (translating design prose into machine-readable fields, including falsifiable predictions and a conclusion wording ceiling), the phase state machine workflow_state.yml (an exact resume after an API failure), the decision log findings.yml (every judgment leaves a trace), and the pre-registration lock (results may never be used to reverse-engineer the narrative).

Field: cardiovascular epidemiology and statistical genetics. Question: does atrial fibrillation causally drive heart failure, and if so, through which measurable circulating proteins?

Method: bidirectional two-sample Mendelian randomization, then a systematic mediation screen across 1,915 plasma proteins, formal mediation MR, six-exposure multivariable MR (adjusting for body mass index, hypertension, diabetes, kidney disease, coronary artery disease and systolic blood pressure), and colocalization. All from public GWAS summary data hosted on IEU OpenGWAS.

Why it matters: this is a textbook negative highlight. After screening 1,915 plasma proteins, the 10 exposure-to-protein hits and 18 protein-to-outcome hits had a zero intersection, with a mediated proportion near zero — pointing to direct cardiac mechanisms (electrical and structural remodelling) rather than a pathway any plasma-protein drug could intercept. When the reverse direction had weak instruments (only 12 variants), the manuscript labelled it honestly as "weak instruments" instead of writing "no effect", and a dedicated wording-ceiling section states in plain text that no protein is claimed as a drug target.

Field: atherosclerosis, single-cell and spatial omics. Question: which transcription factors govern the cap-stabilizing programme of smooth muscle cells in human carotid plaque, and is the matrix-degrading programme really smooth-muscle-derived?

Method: a five-layer evidence chain — multi-dataset lineage integration and trajectory analysis, strict lineage criteria to adjudicate the smooth-muscle origin of macrophage-like cells, multi-engine virtual-cell in-silico perturbation to nominate regulatory targets, GWAS and eQTL colocalization, and patient-matched Xenium spatial transcriptomics validation.

Why it matters: this is the strongest example of structural self-refutation. Under strict lineage criteria, zero macrophage-like cells qualified as smooth-muscle-derived — which triggered a pre-registered rule and rebuilt the paper's central proposition from a two-arm model into a single-axis model with degradation attributed to the myeloid lineage. Equally honest: the virtual-cell foundation model's single-gene knockout displacement was only 1.6×10⁻⁴ and could not beat a linear baseline, so it was downgraded to an exploratory benchmark with a dedicated figure showing that negative. Even after the manuscript was finalized, the AI went back to recompute and caught a directional error in its own text.

Field: cardiovascular genetics and pharmacogenomics. Question: where do the causal architectures of calcific aortic valve disease and coronary artery disease diverge, and which targets are valve-specific?

Method: seven parallel sub-claims across six data layers — multi-dataset bulk transcriptomics, single-cell causal cell enrichment, 300-target dual-outcome comparative MR, drug-target MR, co-protein multivariable MR with colocalization, AlphaFold structure prediction with molecular docking and LINCS compound signature screening, and a phenome-wide safety screen. Roughly 26 hours without interruption.

The 300-target three-way split: 50 valve-specific, 74 coronary-specific, 145 shared. Data verified 2026-08-11.

Why it matters: the most instructive moment came from the reviewing role. The executing side once reported perfunctorily that it "could not download a particular GWAS dataset". The reviewer refused to accept that and demanded an actual accessibility check with curl — five seconds later it turned out the data was downloadable without registration. The executing side admitted it had been lazy, and that single check triggered the pivotal reversal that took the statin-target signal from blank to highly significant. A second highlight is the pre-planned fallback: with no cloud GPU account available, the "drop the molecular dynamics, keep docking and compound screening" path written into the design document fired exactly as specified, and the result was honestly marked partial.

Field: anaesthesiology and cardiovascular medicine. Question: is ferroptosis activated in myocardial ischaemia-reperfusion injury, and does the inhalational anaesthetic sevoflurane bind ferroptosis hub proteins directly?

Method: differential expression and enrichment on GEO data, weighted gene co-expression network analysis, three-algorithm machine-learning feature selection intersected to define hubs, protein interaction networks, independent-dataset directional validation, immune infiltration correlation with multiple-comparison correction, single-cell localization, and AutoDock Vina molecular docking with a positive control.

• The original hypothesis was killed by the AI itself: the study began on cuproptosis, but under strict thresholds only 1 of 15 genes passed (3 were required). The verdict was FAIL, the pre-registered rule redirected the study to ferroptosis, and the negative result was archived unaltered rather than deleted.
• The core finding is an honest weak result: sevoflurane's docking scores against the four hubs ranged from −5.29 to −4.78 kcal/mol, none crossing the −6 passing threshold, while the positive control (a known inhibitor) scored −8.56 — a gap of 3.27 that quantitatively establishes it is far weaker than a real inhibitor. The conclusion was framed honestly as a systems-level mechanism rather than single-target inhibition.
• Downgrade after multiple-comparison correction: eight strong immune correlations did not survive false-discovery-rate correction and were honestly demoted to exploratory background.
• A technical artefact was identified: 92% of cells were annotated as cardiomyocytes with mitochondrial fractions of 47–49%; endothelial marker checks identified this as a technical artefact, and it was written into the limitations.

Real defects caught by the reviewing role going figure by figure included a delivered volcano plot still showing the old hypothesis, docking scores presented as if they were true free energies, an unreadable dark-background panel, and the same value disagreeing between the text and a figure. All of these came from actually opening the images, not from listening to a report.

Field: evidence-based medicine and evidence-synthesis methodology. Question: how methodologically sound are the 102 published drug meta-analyses in heart failure with preserved ejection fraction, and has the quality-of-life outcome that patients care about most been systematically neglected?

Method: no new quantitative pooling at all — this tests judgment rather than computation. PubMed search with manual supplementation, record screening, full-text retrieval through open-access channels only (PubMed Central and Unpaywall), structured data extraction, AMSTAR-2 methodological quality rating, GRADE certainty grading, corpus overlap analysis, and quality-of-life reporting gap mapping.

Why it matters: this case exposed a kind of integrity test that pre-registration cannot govern — how the data was legally obtained. While attempting an institutional single-sign-on route, the AI discovered the mechanism actually exposed other people's plaintext credentials and required impersonating them to log in. It classified this on the spot as unauthorized access, made no attempt to work around it, deleted the temporary files containing credentials, and switched to a fully compliant open-access route. The 381 papers it could not obtain were not silently skipped but disclosed in a dedicated box in the PRISMA flow diagram. A second highlight: the clinician role promoted the quality-of-life outcome to the same tier as hard endpoints, reasoning that outpatients with this condition care more about whether they can still climb stairs to buy groceries — and that judgment became the paper's central selling point.

Field: reproductive medicine and single-cell transcriptomics. Question: where exactly does the endometrial receptivity window fail in recurrent implantation failure? The question itself was not given in advance — the human said only "let us collaborate on a topic in reproductive medicine".

Method: three successively narrowing rounds, each a complete design and execution cycle, taking about 10.5 hours in total. Rounds two and three used a carry-forward design: read the previous round's limitations and future-work section, then pick a thread from it. Eight public GEO datasets were used, including single-cell data covering 262,171 cells.

• All three rounds ended negative or inconclusive: round one unsupported, round two insufficient data, round three a credible core negative. To someone reading only conclusions it looks like three wasted efforts — but each round closed off one specific mechanism that the field widely assumed to be true.
• The most valuable single action: in round three the positive-control gate passed on its face — all three trajectory methods correlated significantly with time. But the AI did not stop at "gate passed equals positive". It went on to check which genes were actually enriched at the mature end of the trajectory, found they were all cell-cycle and proliferation markers, concluded the axis was measuring a proliferation gradient rather than the hypothesized maturation state, and proactively re-classified an apparent support as a credible core negative.
• Two self-vetoes at the design stage: the most tempting follow-up lead turned out to depend entirely on excluding one outlier sample and was denied standing as an independent topic; and a check revealed the data could not support a strict velocity analysis, so it was downgraded and the misleading term was explicitly banned from the manuscript by a pre-registered red line.
• A stop-loss mechanism: round three added an up-front adjudication of whether continuing was worthwhile, preventing autonomous continuation from degenerating into manufacturing a fourth round just to have something to show. The reviewing role declared round three the natural endpoint, and the executing side honoured that after finishing rather than starting a fourth round on its own.

In one case the executing role wrote a closing line worth quoting: "if you had not done these independent checks and had simply trusted my reports, at least four or five things in this work would have moved forward carrying borderline problems." That sentence is simultaneously a proof of the platform's capability and a statement of its limits.

Every case above followed the same four steps, and the workflow is reproducible once you have access.

1. Step 1 — get access. Email moogtang@gmail.com to arrange a trial. There is nothing to install and no environment to configure; both engines are already loaded.
2. Step 2 — decompose one or two reference papers. Upload a paper from your field and run /bio-design scan. About 30 seconds later you see how a published study's analysis logic breaks down into evidence levels and design decisions — before you commit to your own design.
3. Step 3 — design the study honestly. Run /bio-design and state your real constraints: no wet-lab capability, a six-month budget, a target journal. You get design.md plus the machine-readable TOPIC.yml contract, including an explicit list of the conclusions your design cannot reach.
4. Step 4 — execute, then review the figures yourself. Hand TOPIC.yml to bio-analyze and say start. It builds the environment, writes the code and produces publication figures, then stops at the review gate. Open every figure yourself — in all six cases above, every valuable defect was caught exactly that way.

Typical wall-clock time: 1–2 hours for the design conversation, then anywhere from a few hours to 26 hours of unattended execution depending on how many evidence lines the design carries.