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

e28bba7 / friends/victor-garcia-nk-ctl/nmds-fig2b/PRIOR_ART.md

Raw
# Prior art

Started 2026-11-04 22:46 UTC. This file precedes research code.

Read first: `../friends_queue/GOAL.md`; also read `_codex/RULES.md`,
`_codex/FRONT.md`, the Victor scoreboard and queue. Only Victor's own papers
are eligible. No outreach, uploads or remote compute. Resource cap: Mini,
four cores, 4 GB, nice 10; total deadline 2026-11-05 04:46 UTC.

## Selection and inherited results

- Scoreboard rank 1, 2025 Figure 2D stress 35.8, was already improved to
  16.8196634149424955 in `friends_board_victor_1031`; do not repeat.
- Rank 2 is a fitted quadratic coefficient and explanatory fit, not a held-out
  prediction record. Its data and fit were already examined; do not relabel it.
- Rank 3 / review T3.12 is killed queue V02; do not repeat.
- Next eligible printed numerical objectives: Figure 2A/B/C, stresses
  16.8 / 34.8 / 23.4. Each requires its exact original profile data and distance
  convention. AnaCoDa's exact fitted costs were previously not located.
- Pathogen targets are excluded by this job. Previously killed and failed
  targets, and non-Victor-authored mathematical branches, are excluded.

## Public source checks, 2026-11-04

Paper: https://doi.org/10.1101/2025.06.03.657573,
*Negative epistasis limits current codon optimization approaches* (2025).
Victor Garcia is a coauthor. The primary PDF and supplement will be inspected
locally from existing public-source archives, with checksums and provenance.

Live web queries for the exact title, DOI + code/Shah/Rudorf, title +
AnaCoDa/MDS, DOI + GitHub/stress returned no same-target improved number.
Direct bioRxiv web-tool opens failed (tool internal access error); public API
and shell HTTP reads will be attempted. This is not evidence of absence.

The inherited scoreboard's prior-art file records public GitHub searches,
agent-repository trees, recent papers and community checks on 2026-11-01.
They are historical evidence, not a substitute for live checks before a
positive claim. Live GitHub searches sorted by update, open issues/PRs if a
source repository is found, recent papers and public community searches,
and relevant agent repositories remain to be checked.

Best located public numbers for Figure 2A/B/C: printed 16.8 / 34.8 / 23.4;
no independently comparable later value established. Standard method:
two-dimensional Kruskal nonmetric multidimensional scaling, with the paper's
stated profile distances. Reproduction has not yet run. Do not substitute
Hamming distances, a new translation-cost fit or a different stress definition.

The archived independent referee `friends_ref_victor_1031/VERDICT.md` was
subsequently read. It supersedes the old non-reproduction statement:
native R/MASS defaults give 35.75613323 on Figure 2D and ordinary tighter
settings/restarts attain the 16.82 result. That target remains excluded.
Its archived runtime recipe permits an isolated native R extraction here.

## Completed source checks and reproduction

Read-only HTTP fetches succeeded; `FETCH_LOG.jsonl` contains dates, byte counts
and SHA256 receipts. The current bioRxiv API returns only v1 (2025-06-06),
`published: NA`. The fresh PDF caption on page 5 prints all three target
numbers; page 15 defines Euclidean raw profile distances and Kruskal NMDS.

GitHub repository queries sorted by updated (`"Negative epistasis"`,
`"657573"`, `codon epistasis 2026`, `AnaCoDa Gees Garcia`) found only unrelated
repositories or no hits. No original analysis repository, open change to its
analysis or same-data successor was located. This does not prove none exists.
The live public trees of BakeLab/Epi-results, both Google DeepMind AlphaEvolve
repositories, openai/math, kisonecat/openai-math, plasma-ai/erdos and
vinid/einstein-arena had no path matching codon, AnaCoDa, isoMDS or epistasis.
**Limitation:** GitHub truncates the openai/math recursive tree; its absence
of hits is only a partial-tree check. Repository source content was not
exhaustively searched. Large trees were not retained; their hashes, hit lists
and truncation flags are in the fetch log.

Recent arXiv query (2025-11-04 through 2026-11-04, codon AND epistasis)
returned one broad protein-evolution review, not a same-data NMDS result.
Exact-title/DOI searches including public X/Reddit indexes and current 2026
queries found the preprint, its institutional talk and bibliographic mirrors,
with no later numerical configuration located. No account was created and
no post, private community or message was accessed.

Best publicly located methods for these finite targets remain ordinary
Kruskal NMDS. Native R 4.6.2, MASS 7.3-66, all defaults, directly reproduce
34.8056749 → printed 34.8 for Shah and 23.396544 → printed 23.4 for Rudorf.
Unlike the earlier Hamming search, the standard method was run first here.
With only tighter documented convergence settings the classical starts reach
17.0483775 and 13.5073137. A fixed 100 random starts reaches 16.9693902 and
13.4830828. Credit belongs to the established R/MASS/Kruskal method; this job
claims improved feasible values for the printed objectives, not a new method.

Primary rate data:
- Shah & Gilchrist 2010, Table S2, Rc column:
  https://doi.org/10.1371/journal.pgen.1001128.s007 .
- Rudorf & Lipowsky 2015, Table S2, 0.7 h^-1 and 2-1-2 pathway:
  https://doi.org/10.1371/journal.pone.0134994.s003 .
- Original sequences, fresh download identical to archived DC3:
  https://www.biorxiv.org/content/biorxiv/early/2025/06/06/2025.06.03.657573/DC3/embed/media-3.csv?download=true .

## Final recheck, 2026-11-04 22:56–22:57 UTC

Repeated bioRxiv API, supplement listing, updated-sorted GitHub repository
queries for title/DOI/codon+MDS and the new value prefixes, and all seven
agent-repository trees (`results/final_recheck.log`, `FETCH_LOG.jsonl`).
No new version or later comparable result was located. The same partial-tree
limitation for openai/math remains. Final web queries again returned the
preprint and its bibliographic/institutional listings. Sciety's activity feed
lists only 6 June 2025:
https://sciety.org/articles/activity/10.1101/2025.06.03.657573 .
The conclusion is bounded by these searches; no claim of global optimality
or exhaustive priority is made.

## Independent t2 referee, 2026-11-04

The material above was supplied by the claimant and is unverified input. This
review will open the primary PDF and rate supplement, reconstruct the released
sequences and Euclidean rate-profile distances independently, run standard
MASS defaults first, then evaluate the supplied witness with independent code.
The claim concerns an unsupervised finite objective, not predictive accuracy.
Live DOI opening in the web tool failed (internal access error); read-only HTTP
and the archived public PDF will be used. The printed 34.8 is the baseline to
check, not a presumed record. No improvement search is planned by this referee.

Independent live checks: exact-title/DOI/stress/GitHub and value-prefix searches
returned the preprint, bibliographic mirrors and an institutional talk; no later
comparable number was established. Read the official CRAN MASS source:
https://raw.githubusercontent.com/cran/MASS/master/R/isoMDS.R and
https://raw.githubusercontent.com/cran/MASS/master/src/MASS.c .
Its defaults and percentage Stress-1 formula support the convention, but do not
establish that the claimant computed it correctly. Fresh bioRxiv API metadata
and fresh PLOS S2 downloaded; evidence/fetch_receipts.json records provenance.
This review makes no priority claim; it cannot finish the numerical baseline
reproduction because the prescribed pod is unavailable. No substitute compute
or restricted mathematical variant was used.

## Resumed independent review, 2026-11-04
The current instruction explicitly authorizes light computation on the Mini
(4 cores, 4 GB, nice 10). Prior pod unavailability is therefore not a blocker.
I opened the archived primary PDF directly: page 1 names Victor Garcia,
page 5 Figure 2B prints 34.8, pages 14–15 specify Euclidean raw rate profiles
and Kruskal NMDS. Live title/DOI searches again locate the same preprint.
The existing prepared scripts have not been accepted as numerical evidence.
I will run native MASS defaults before evaluating the proposed coordinates,
and independently implement and certify the objective from the primary data.
No predictor is being evaluated: SBS and held-out splits are inapplicable.

## Final independent result and live recheck
Native R 4.6.2 / MASS 7.3-66 reproduced 34.8056748605161; independent
PAVA and Arb certify the submitted map at 16.969390168877389… percent.
The PDF's actual vector circles also reproduce that native default map.
Public metadata, updated-sorted GitHub title/DOI searches, issue search and
seven agent-result trees were rechecked; receipts are in
`evidence/final_public_recheck.json`. GitHub title/DOI results were unrelated;
the broad numeric issue query also returned unrelated results. No original
analysis repository or comparable later objective value was located.
No relevant paths were found in the seven result trees; openai/math's tree
is truncated. Final exact-title/DOI/value web searches again found the
preprint and listings, not a comparable later value. This bounded check is
not an exhaustive priority claim, and the verdict judges the printed-value
comparison only. Established Kruskal/MASS methods receive full credit.