ttubiana/HEV-ORF1-models

Hepatitis E virus ORF1 (nsp1) — AlphaFold2 model collection

1,178 AlphaFold2 predictions of the HEV ORF1 (nsp1) replicase, packaged so a static web app can render the 3D model, the predicted aligned error (PAE) matrix and the multiple sequence alignment without a server.

• —open the viewer: <https://tubiana.github.io/ORF1viewer> (this dataset is its data root)
• —repository — app + pipeline code, no data: https://github.com/tubiana/tubiana.github.io
• —dataset repo: https://huggingface.co/datasets/ttubiana/HEV-ORF1-models
• —generated 2026-08-29T10:20:00Z by scripts/prepare_data.py (preset pages)

Layout — the repo root is the app's data root

Total payload 984.3 MB. All payloads are gzip/WebP-lossless where the numbers matter (PAE images are index images: no colour-space quantisation of the values themselves).

Use it from the viewer

https://tubiana.github.io/ORF1viewer/?dataBaseUrl=https://huggingface.co/datasets/ttubiana/HEV-ORF1-models/resolve/main

Every path inside manifest.json is relative to that root. Other overrides (see the repo README): VITE_DATA_BASE_URL at build time, window.__ORF1_DATA_BASE_URL__ in index.html, or localStorage['orf1.dataBaseUrl'].

Note — if this repo is private, the browser cannot read it from a static site: make it public, or serve the payload from storage that allows plain HTTP GET (institute storage, Buckets, …).

Reading the numbers without the app

PAE images are lossless: a pixel value is an index into manifest.pae.lut (balanced, 33 levels, max 33.0 Å, documented tolerance ±1.5 Å):

import json, urllib.request, numpy as np
from PIL import Image

root = "https://huggingface.co/datasets/ttubiana/HEV-ORF1-models/resolve/main"
man = json.load(urllib.request.urlopen(f"{root}/manifest.json"))
lut = np.asarray(man["pae"]["lut"], dtype=np.float32)

m = man["models"][0]
img = Image.open(urllib.request.urlopen(f"{root}/{m['paePath']}"))   # mode 'L' (8-bit)
pae = lut[np.asarray(img, dtype=np.int16)]                          # Angstrom, shape (len, len)
print(m["id"], pae.shape, round(float(pae.max()), 2))

Integrity, per model: verify.points = (i, j, Å_original) (0-based matrix indices) and verify.decoded = the same cells decoded back through the LUT. In this snapshot: 1178 models / 28,272 points, max abs error 1.0 Å, 0 over the limit.

The annotation CSV (source of truth for domains)

metadata/dataset_ORF1s_1178_reviewed_111724.csv — ;-separated (it is not comma-separated), 1179 data rows, UTF-8, 22 columns:

genbank, genbank_nucl, generic_hostname, Genogroupe, sequence_size, border_MetY, border_FABD-like, border_HVR, border_domX, border_Hel, border_RdRp, species_y, size_MetY, size_FABD-like, size_HVR, size_domX, size_Hel, size_RdRp, host, organism, strain, isolate

Domain borders arrive as border_<Domain> = "(start-end)" (1-based, e.g. "(1-467)") with matching size_<Domain> columns; genbank is the protein accession that keys every model in manifest.json.

Domains are taken verbatim from this CSV — the pipeline and the app never cluster, segment or merge them. HVR is a hypervariable stretch, so it is excluded from the domain count in the UI while remaining annotated and coloured everywhere else. Accessions absent from the CSV simply have no domains; nothing is invented for them.

Composition

• —lengths: min 1393, median 1704, max 1783 aa
• —mean pLDDT: 77.2 (range 53.1–80.7)
• —models with domain annotation: 1,178 / 1,178
• —CSV coverage: 1179 annotated accessions, 1,178 with an AlphaFold model (1 without: AQN78288.1)

• —hosts (top 8): human (731), swine (113), rat (55), wildboar (54), rabbit (53), synthetic (37), rodent (27), avian (25)

Verification after download

DIR=$(mktemp -d)
hf download ttubiana/HEV-ORF1-models --repo-type dataset --local-dir "$DIR"
( cd "$DIR" && sha256sum -c metadata/SHA256SUMS.txt )   # every file, payload included

Provenance & license

See metadata/provenance.json (source tree, AlphaFold pipeline preset, LUT, integrity summary, artifact counts, app version). Code and data are released under the MIT license; the underlying sequences are public GenBank records — cite the original accessions and the AlphaFold2 publication for any scientific use.