Data and code availability

All code, data, and figures are distributed under open licenses with persistent citation handles. Code is MIT; the Zenodo data deposit is CC0 1.0 with a DOI that resolves indefinitely; each figure ships a standalone reproduction gist archived to Software Heritage as a content-addressed SWHID.

Paper

The preprint describing the pipeline, benchmark, and catalog. An author-deposited copy — the PDF plus JATS XML and figure files — is archived on Zenodo under the same DOI.

Preprint (bioRxiv): 10.64898/2026.10.18.759859
Author copy (Zenodo): zenodo.org/records/23476946
License: CC BY 4.0

Code

The full project — data pipelines, agent prompts, build scripts, and the source for this site — lives in one repository under the MIT License.

Repository: github.com/Deliverome-Project/accessible-surfaceome
Concept DOI (all versions): 10.5281/zenodo.22116981
License: MIT

Each tagged GitHub release is archived on Zenodo through the GitHub integration. The concept DOI above resolves to the latest release, and each release also has its own version DOI, listed on its Zenodo record. Cite the concept DOI for the software in general, or a version DOI to pin a specific snapshot.

Data

Datasets are deposited on Zenodo under CC0 1.0. The deposit includes:

SurfaceBench — the curated 147-protein benchmark with ground-truth labels, per-model verdicts, and the curation rationale.
Triage results — one row per human protein-coding gene with the agent's verdict, reason, and per-call metadata.
Deep-dive records — a per-gene JSON for every gene that received a full evidence assembly, each carrying its evidence chain and its deep-dive tier.

Concept DOI (all versions): 10.5281/zenodo.20805383
License: CC0 1.0

Cite the concept DOI, which always resolves to the latest release. Each numbered data release also has its own version DOI, listed on the Zenodo record and by the API at /v1/releases; cite a version DOI to pin an exact snapshot.

The Zenodo DOI is wired into the figure generator (scripts/figures/embed_figure_gist_metadata.py) so every figure PDF/PNG carries the dataset citation in its metadata as soon as it's rendered. Release 1.3.0 of the deposit contains:

triage-runs-genome-with-reasoning.tsv — every triage agent call across the protein-coding genome (33,728 rows: the 19,324-gene NCBI-context pass plus 14,404 rows from the PubMed-context rescue pass, each labelled by a triage_pass column), with verdict, reason, confidence, and free-text reasoning.
triage-benchmark-with-reasoning.tsv — the 147-protein SurfaceBench (7,938 rows: 147 genes × 18 model-variant cells × 3 replicates) with truth labels joined to every triage call.
deep-dive-preset-membership.tsv — catalog preset membership for every deep-dived gene (5,332 rows), computed with the predicates documented on the API page.
deep_dives_1.3.0.tar.gz — the 5,332 per-gene deep dives: for each gene the record JSON, its evidence ledger and its Markdown brief, plus a manifest.tsv.
deep-dive-reproducibility-replicates-v1.tar.gz — the run-to-run reproducibility replicates behind Supplementary Figure 16 (not published records).
figure-source-data_1.3.0.tar.gz — the source table behind every data figure in the manuscript, the four canonical figure-input tables (including db_optimized_cutoffs.tsv), and a manifest.tsv giving each file's figure, sha256, reproduction gist and SWHID.

Between releases, the API keeps a write-once history of every gene: each distinct state it served is archived as a numbered revision at /v1/genes/{SYMBOL}/revisions, and the numbered data releases, with their Zenodo version DOIs, are listed at /v1/releases. See the API page for details.

Per-figure reproduction

Each published figure ships with a public GitHub Gist containing a standalone reproduction script and a short README pointing to the canonical data source. Scripts declare their dependencies inline using inline script metadata so a reader can execute the figure with uv run make_<figure>.py — no environment setup, no pip install step.

Each gist is archived to Software Heritage (via Save Code Now) and cited as a content-addressed swh:1:rev:<sha> of the gist's HEAD commit — recorded per figure in the repository's swhid_map.json. The revision pins the exact script + data snapshot a figure was rendered against and resolves from the SWH archive even if the gist is later deleted. The underlying dataset's durable handle is the Zenodo DOI above.

For convenience, every figure file also carries citation handles in its metadata: PNG outputs embed the gist URL in the Source tEXt chunk and the dataset Zenodo DOI in Subject; PDF outputs use the analogous Subject + Keywords fields. The handles travel with the file across downstream contexts — slide decks, blog posts, supplementary uploads. Read with exiftool figure.png.