RNA Atlas Explorer

Interactive, configurable web explorer over the Ribonanza-2 A–H prediction-atlas mining results. Replaces the static PowerPoint decks (per_letter_top10_novel.pptx, etc.) — you tune the selection arguments live (length, novelty, SHAPE support, motifs, pseudoknot, source, per-letter top-N) and the candidate list + per-fold deep view update instantly. Built for picking high-value cryo-EM targets.

Status

• v1 (this): curated set of 7,757 folds (already fully mined: motifs + SHAPE + sparse novelty). Filtering/ranking is 100% client-side; structures + reactivity load lazily per fold.
• v2 (planned): scale to the 414,633 high-confidence strict index via the precompute pipeline (motifs + SHAPE cheap; USalign novelty ~1–2 days on LSF). See plan.

Run it

eval "$(mamba shell hook --shell bash)"; mamba activate rna   # needs h5py + pyarrow + gemmi + numpy
cd /groups/das/home/zouinkhim/atlas_explorer
python serve.py --port 8765          # then open http://<host>:8765/

A tiny static server is required (browsers can't fetch the local structure/reactivity files over file://). It only serves bytes — all filtering happens in the browser.

Layout

config.json              # machine-specific absolute base paths — GITIGNORED, copy from config.example.json
build_feature_table.py   # assembles data/*.json from the mined TSVs (run once / on update)
build_static.py          # exports dist/ (data + gz structs + react) for S3 hosting
serve.py                 # local dev server: lazy /structs/<id>.cif and /react/<id>.json
web/                     # index.html, app.js, style.css, viz_style.js, config.js, lib/3Dmol-min.js
data/                    # folds.json (table), motifs.json (deep-view spans) — no absolute paths
DEPLOY.md                # github.io shell + S3/CloudFront data + shared-token runbook

Hosting

• Local / internal: python serve.py (reads structures + reactivity live from /groups; no gate).
• Shareable (off-network): shell + all data on one S3 prefix (s3://rnanix/atlas_explorer/) served via CloudFront (HTTPS), gated by a single passcode. build_static.py stages the whole site into dist/; users open the CloudFront URL, enter the passcode (it becomes the ?t= token a CloudFront Function validates on data paths). Source code lives in a public GitHub repo (version control only — it does not serve the site; JaneliaSciComp/rna_atlas_website, public since ~2026-08-03). Secrets/keys never live here — .infer_api/DEPLOYED.local.md/ config.json are all gitignored and injected at deploy time (see CLAUDE.md "Gotchas"). There is no Anthropic API key anywhere in this deployment, shared or server-side — each user of the in-page assistant sets their own key in their own browser. Full runbook in DEPLOY.md.

config.json (not committed)

All absolute paths (/groups/...) live in config.json so nothing internal is committed or served to the browser. Copy config.example.json → config.json and fill in mined_dir, the two struct_bases (A-E / F-H curated CIF dirs), metadata_parquet, react_override, and the per-letter hdf5 map. serve.py builds each structure path as struct_bases[AE|FGH]/<seq_id>.cif at request time; data/ only ever holds scalar features + motif spans.

Data provenance (curated 7,757)

Built from lsf/20260612_rna_motif_chaitanya/:

• selection.tsv — id, design_sequence, cif path, pLDDT, gpde, sublibrary, letter
• fold_metadata.tsv — length, source, pseudoknot, r_2a3_ispaired, openknot, overlap_ae_tm1, …
• summary/motifs_labeled.tsv — per-motif type + residue spans
• summary/motifs_shape_gated_AH.tsv — per-motif 2A3 protection (all A–H) → shape_ok/mean_prot_2a3
• summary/{per_letter_candidates,top10_novelty_v341,top10b_novelty_v341,pk_candidates}.tsv — continuous best_tm1 vs v341 (only ~105 folds scored; the rest get best_tm1 = null until v2)
• name TSVs — human-readable names

Reactivity is read on demand: A–E from the cmuts HDF5 (r_norm, sliced by sub_start), F–H from the design-aligned react_override_fgh40.parquet (only the len>40 coverage subset exists). F–H sequences are derived from the CIF (they are empty in selection.tsv).

Configurable arguments

Length, pLDDT, clashscore, novelty (best_tm1 ≤, only-scored toggle, overlap-vs-AE), SHAPE support (require shape_ok, r(2A3) ≤), motifs (require tertiary / rare / specific types), pseudoknot, source/letter, ranking key + top-N (overall or per letter).

Verification

build_feature_table.py output reproduces the per-letter deck exactly: with has-best_tm1 + length>40 + rank best_tm1 ascending + top-10 per letter, the explorer yields the same 80 seq_ids as summary/per_letter_candidates.tsv (checked 80/80).

Note: the deck's "SHAPE-supported" label was generous — several per-letter picks have negative 2A3 protection (motif residues more reactive). The explorer exposes the real mean_prot_2a3 and a strict protection-based shape_ok, so a genuine SHAPE gate is now possible (it is not applied by default, to match the deck's pool).

Rebuild the table

python build_feature_table.py            # -> data/folds.json, motifs.json, paths.json