By @rpinho

AI for biology and drug discovery: Anthropic x Adaptyv protein design, Arc Virtual Cell Challenge, Virtual Embryo Challenge, open drug-target discovery, and refereed bug reports on scientists' published software

55ba642 / friends/victor-garcia-nk-ctl/danish-forecast/PRIOR_ART.md

Raw
# Independent prior-art audit (2026-11-04)

Scope: retrospective validation of the unchanged Cadosch et al. (2020), DOI 10.7717/peerj.8913, Danish macrolide-resistance forecast against later SSI/MiBa summaries.

Audit started before writing recomputation code. The source model and ordinary binomial logistic trend are the standard methods to reproduce. There is no identified official leaderboard for this epidemiological validation question.

Pending: independently open the primary paper, its public repository, SSI reports, all potentially overlapping forward citations returned by `cited-by`, and direct question searches. Check public code and AI result repositories and community sources. The claimant's prior-art assertions are unverified input.

This is an exact reconstruction and retrospective forecasting audit, not a new model search. No improvement or novelty is assumed.

## Initial independent findings
- Primary full text downloaded independently from Europe PMC (`primary.xml`, PMCID PMC7147432). The abstract and Results print the treatment-conditional 2025 values 25% (9–30), 84% (36–98), 62% (48–76); Victor Garcia is a coauthor. Table 2 supplies the four Danish rows.
- Official code: https://github.com/calthaus/MG-resistance; archived commit 95fd31a6af5a94d287be07b2d199c38ae86149fe downloaded as `official.R`. Reproduction will use the authors' ODE, optimizer, initial state, parameter transforms, seed and 10,000 simulations unchanged on the authorized pod. Independent recalculation will use separately written Python, not the claimant's code or existing referee.
- Ran `cited-by 10.7717/peerj.8913 --limit 200` myself: `cited_by.txt`. OpenAlex headline count 17; independent raw query returns 12 (`citing_works.json`); abstracts and primary locations inventoried in `citation_inventory.txt`. Overlapping papers are being opened in full, with access failures recorded.
- Opened the three current SSI primary reports (2003–2022, 2011–2023, 2024). They document episodes, source changes and treatment changes. Figures require visual checking.
- Direct searches: "Mycoplasma genitalium" "forecast validation"; "Cadosch" "forecast" "Denmark" resistance validation; "Mycoplasma genitalium" "forecast" "MiBa"; "MG-resistance" "validation" github; source DOI + Denmark + 2026. No earlier same comparison located in the returned material; search results were noisy, so this is a bounded search, not proof of absence.
- arXiv, public-code, community and AI-repository scoped searches returned no relevant validation. There is no numerical record/leaderboard for this retrospective question. Additional primary-paper and GitHub issue/PR checks remain pending.

## Completed source and code checks
- Independent official run succeeded on fnew, using unchanged source, deSolve 1.42 and archived bbmle 1.0.23.1, seed 71309 and all 30,000 draws. `official_pod_result.txt` reproduces all three printed point estimates and intervals. This is the reconstructed standard baseline, not a new result.
- Live GitHub API: latest commit remains 95fd31a6af5a94d287be07b2d199c38ae86149fe (2020-07-22), code change 2019-10-23. The all-state issue/PR query is empty (`github_issues.json`). Repository search sorted by updated (`github_search.json`) finds the official repository and five unrelated name matches; no relevant newer validation code.
- arXiv exact-species search sorted newest (`arxiv_search.xml`) returns zero records; no directly relevant recent arXiv paper located. Web-scoped searches for the question in BakeLab, Google DeepMind, OpenAI/math, X and Reddit yielded no earlier comparison. These are search-result checks, not complete audits of all repository contents or private community discussions.

## Forward-citation screening
Ran the required helper independently for published DOI 10.7717/peerj.8913, then independently for its preprint DOI 10.1101/321216 after a reference in Braam 2020 revealed it. Raw inventories are retained.

| Citing work | Independent inspection and relevance |
|---|---|
| Bolaji 2026, 10.1371/journal.pgph.0006834 | Abstract and full text from Europe PMC; especially numerical simulations and conclusion. Different theoretical model and qualitative comparisons to aggregate resistance ranges; no frozen Danish forecast scored against SSI/MiBa. |
| Tenney et al. 2026, 10.1093/ofid/ofag031 | Abstract and full text; resistance prevalence, surveillance gaps and conclusion. Treatment review, not this forecast/baseline test. |
| Carvajal-Garcia 2024, 10.1093/narmme/ugae001 | Abstract and full text downloaded; small-molecule experiments, not national forecast validation. |
| Riou 2023, 10.1186/s12879-023-08200-4 | Abstract and full Methods/Results/Discussion, primary Springer and Europe PMC. Gonococcal models fitted to GRASP, not this Danish M. genitalium comparison. |
| Nowack 2023, 10.1007/s00105-022-05099-x | Abstract and full text; M. genitalium subsection. Pandemic/clinical review, no this forecast test. |
| De Baetselier 2022, 10.1093/ofid/ofac615 | Abstract, cohort methods, resistance results and discussion; West African MSM study, not Danish annual validation. |
| Recombinant-protein diagnostic review, 10.33448/rsd-v11i10.32970 | Abstract excludes forecasting; diagnostic serology meta-analysis. |
| Yonathan 2022, 10.36452/jkdoktmeditek.v28i2.2264 | Primary PDF abstract and macrolide section, p.241. Repeats the original 25/84/62% forecasts; does not test them with the later Danish series. |
| Wood 2021, 10.1128/jcm.00819-21 | Abstract, results and discussion; U.S. treatment outcomes. Repeats original forecasts as conditional predictions, no Danish external test. |
| Xu 2021, 10.1038/s41598-021-90627-3 | Abstract, calibration and output methods, results/discussion; Australian anatomical-site model, not this validation. |
| Kenyon 2021, 10.1097/OLQ.0000000000001404 | Primary abstract and institutional metadata read. Full text unavailable: DOI payment/subscription response and LWW access failures. Describes 179 PrEP participants, 70 infections, and treated/untreated comparison; no indication of Danish forecast validation in the abstract. Relevant full section NOT inspected. |
| Ong 2021, 10.1016/j.eclinm.2021.100779 | Full text; Australian MSM screening-model methods/discussion. Cites source mechanistic conclusion, not the Danish forecast/baseline comparison. |

Additional checks:
- Europe PMC forward citations (11) add Waters 2026, 10.1097/olq.0000000000002369: primary abstract reports Seattle cohort (2014–2018), sequential sex practices; no Danish forecast test. Full XML endpoint returned HTTP 500.
- Braam 2020 (PMC7575077) found by direct search: full abstract/results/discussion checked; Dutch 2018 assay/prevalence data, not the later Danish test. References the 2018 source preprint, explaining why checking only the final paper DOI can miss citations.
- Source preprint helper returns four works: Tenney 2026 (full text inspected above), a 2021 qualitative model (10.1002/cmm4.1199), Cina 2019 natural-history systematic review (10.1136/sextrans-2018-053823), and a short 2019 clinical item (10.1007/s15011-019-2392-x). Abstracts of the first two new titles describe equilibrium/optimal control and natural history rather than this test. Wiley/other full-text access was unsuccessful; the 2019 clinical item could not be inspected beyond bibliographic information.
- Chua 2025 (10.1016/j.lanmic.2024.101047): primary abstract/core metadata and Monash institutional abstract inspected. Includes surveillance sampled through 2021 and searched through April 2023. Full text unavailable from Lancet/ScienceDirect (403), Elsevier API, or the temporarily restricted Bern repository. No full relevant-section review claimed.

## Coverage limits and facts for the verdict
OpenAlex's headline is 17 but its raw indexed citation query returns 12; Semantic Scholar returns 11; Crossref's count is 11; Europe PMC returns 11, one additional title. The discrepancy is unresolved and does not prove five actual unidentified papers exist. No inspected material contains the exact frozen Danish forecast/baseline test. This bounded negative search does not complete the requested full-text exclusion of prior work. No priority conclusion is asserted from inaccessible material.

What differs from the 2020 source: later SSI/MiBa surveillance fractions and a historical constant/logistic error comparison. Source reproduction, equations, optimizer, surveillance observations themselves, and the general warning about changed ascertainment are not new. Kind: note is honest for the additional empirical comparison; the source reconstruction by itself is a rerun.

## Fresh referee pass, 2026-11-05
The directory contains an earlier incomplete referee pass; its reports and code are not accepted as verification. I have rerun cited-by (cited_by_current.txt) and direct public searches. They show the same 12 returned citing titles and a headline count of 17. I will reconstruct the numerical comparison anew from the primary paper, inspect the potentially overlapping papers, and specifically try the open Chua copies identified in OPEN_ACCESS.md. No forecast-validation leaderboard is applicable. Prior code will not be imported or copied.

### Fresh pass conclusions (2026-11-05)
- Independently ran both cited-by commands again: `cited_by_current.txt` (published DOI) and `cited_by_preprint_current.txt`. The raw OpenAlex query in `fresh_citations.json` again returns 12 versus the headline 17. A count mismatch does not establish either missing papers or prior overlap.
- Fresh primary XML (`primary_current.xml`) was fetched from Europe PMC. Read equations 1–6, Tables 1–2, fitting section, Results and Discussion; Victor Garcia is a named coauthor, and the three conditional 2025 projections are genuinely printed. The source prints no later-surveillance MAE.
- Reopened the SSI 2003–2022, MiBa 2011–2023 and MiBa 2024 primary reports. Freshly downloaded and visually read SSI Table 4 (`fresh_ssi2022.png`) and MiBa Figure 2 (`fresh_miba2023.png`): 274 at 62.6% in 2021, 311/436 in 2022, and MiBa 60.6/63.2/63.6. The 2024 report says 65% in each sex; the pooled rate is bounded by their common rounding interval. The 2022 MiBa/SSI difference is 8.1303 points, not an independent biological replication.
- Read the abstracts and relevant full-text sections in the primary XMLs for Bolaji (numerical simulation/conclusion), Riou (methods/results/discussion), Ong (methods/discussion), Xu (Australian anatomical-site model), Wood (treatment outcomes/discussion), De Baetselier (West African cohort), Tenney (resistance/surveillance/conclusion), Nowack (M. genitalium subsection), and Carvajal-Garcia (experimental results/discussion). These describe other analyses, not the frozen Danish annual forecast comparison. Yonathan's primary PDF abstract and macrolide section repeat the 25/84/62 forecasts, without testing them. Recombinant-protein review's primary abstract concerns diagnostic test sensitivity/specificity and excludes this question.
- Fresh Kenyon core abstract (`fresh_kenyon_core.json`) describes 179 PrEP participants and treated/untreated infection clearance/resistance. It settles that this is a different cohort question. Full text remains unread; fresh OpenAlex locations list only DOI and PubMed, with no repository/PDF copy.
- CMM 2021 (10.1002/cmm4.1199) abstract describes equilibrium/stability/optimal control, not Danish forecast validation. Tried the Wiley/public repository locations returned by OpenAlex: Turin is restricted; Minho/Aveiro attempts retained. Full text remains unread. Cina 2019's abstract concerns natural history using studies searched through March 2018; it cannot contain these later outcomes. The 2019 clinical item 10.1007/s15011-019-2392-x has no readable full text here and likewise predates the scored surveillance.
- Chua 2025 is gold open access, NOT asserted paywalled. Fresh OpenAlex (`chua_openalex_current.json`) lists the publisher, SSRN, Bern BORIS/CRIS, Melbourne and PubMed. Lancet full text returns 403; ScienceDirect PDF paths return HTML challenge pages, not PDFs; Elsevier API returns only 1,994 bytes of core metadata, no body; SSRN DOI and direct delivery return challenge pages; Bern fetches time out/fail; Melbourne loads but its public item/bundles endpoint reports zero bundles. Europe PMC has no PMCID/full-text copy. The primary abstract covers samples through 2021, searches through April 2023, and regional prevalence meta-analysis, but Discussion/supplement remain unread. No exclusion of a forecast comparison from that unseen text is asserted.
- Direct searches for Cadosch + Danish forecast/validation, M. genitalium + MiBa + forecast, and the note's baseline question found no earlier same comparison. Scoped public-code, AI-result-repository, X/Reddit and recent-paper searches found none; these are bounded searches, not exhaustive exclusions. Fresh official GitHub API checks show latest commit 95fd31a6af5a94d287be07b2d199c38ae86149fe and zero issues/PRs. No applicable live leaderboard exists.
- Independent code `fresh_recompute.py` imports no claimant or previous referee code. It reconstructs all four compartments, refits the original likelihood with Nelder–Mead, fits ordinary binomial logistic IRLS to the four historical rows, recomputes all six later observations and same-split MAEs, calls sbs.py, and checks all rounding corners. The scalar resistance-ratio identity agrees with ODE integration within 3.36e-11. `fresh_recomputation.json` and `BASELINES.json` contain results. The original model's advantage survives independent fitting, numerical tolerance and outcome rounding.
- Attempt to rerun the unchanged official R source on the specifically authorized shared pod failed before remote execution: friends_podsh.py raises KeyError: fnew. No alternate pod was used and no files left the Mini. The complete light recomputation ran locally, one thread, nice 10, well below the 4 GB/4-core limit; the earlier official bootstrap output is archival evidence only.

The factual addition relative to the source is a later-data comparison and two historical-baseline scores. It is not a source optimizer restart, so Kind: note is honest. The weakest point remains the changed surveillance target/treatment exposure. No inspected work already has the comparison; unread full texts and citation coverage are caveats, not evidence that the result is false. The bounded literature check does not prove absolute priority.
- Access update: the Minho open accepted manuscript for Almeida et al. CMM 2021 was successfully downloaded and read (`fresh_cmm.pdf`, 16 pages; `fresh_cmm.txt`). Its numerical illustrations and final discussion use country-parameter theoretical/optimal-control scenarios; they do not use the later SSI/MiBa observations or this frozen-forecast error comparison. Remove CMM from the unread-full-text caveats above. This is why repository copies were tried before declaring access unavailable.