Spaces:
Running
Running
Commit History
-
Deploy: promoter library 10 entries across E. coli, plant and mammalian
Tengo Gzirishvili +1: ClaudeSquashed deploy of GitHub main. Sequences machine-extracted from GenBank records' own feature tables, each agreed by 2+ independent accessions (scripts/import_promoters.py). CaMV 35S, EF-1alpha, U6, lac, tac and araBAD added; CMV deliberately excluded pending a canonical-length consensus. Co-Authored-By: Claude <noreply@anthropic.com>9c2665c
-
Admin endpoints to run the real validation + seed against bundled DMS
Tengo Gzirishvili +1: Claude Sonnet 5Closes the loop on the "no fabricated numbers" promise: these are the two routes that actually produce real data, running IN the Space where ESM-2 and the Supabase credentials already live (this dev machine is an Intel Mac — PyTorch dropped x86_64 macOS wheels after 2.2.2, so real ESM-2 can only run where the Space already runs it, not here). - dee/data/dms_fixtures/ — one real, rigorously validated public DMS dataset: TEM-1 beta-lactamase (Stiffler/Ranganathan 2015 Cell), 4,996 single mutants, the SAME reference example bundled with facebookresearch/esm for the exact zero-shot scoring scheme our scorer.py implements. Every single mutation was cross-validated against the reconstructed reference sequence (100% wild-type-residue match) before being committed — the numbering in the original paper is absolute (offset 24), renumbered to local 1-indexed positions matching the bundled sequence, verified twice. - dee.core.benchmark.predict_additive — sum-of-single-site ΔLL predi…598a072
-
Receipts + seed: validation harness, DMS→commons seeding, evidence UI
Tengo Gzirishvili +1: Claude Opus 4.8Turns the glass-box thesis from claims into proof — and gives the empty commons a path to real data. Honest by construction: no benchmark number is ever hand-entered; the app serves only what a real run produced, and ships an explicit "pending" state until then (this sandbox has no torch/ESM, so the numbers get made on a box with the weights). - dee/core/benchmark.py — validation harness: average-rank Spearman ρ + top-decile precision ("of your top picks, how many are genuinely high- fitness?") + cross-dataset medians. Pure numpy, predictions injected → unit-testable without ESM. - dee/core/dms_seed.py — turns published DMS assays into de-identified substitution-prior rows via the SAME k-anonymized aggregate machinery user data uses (each assay = one contributor; a substitution needs ≥ MIN_USERS independent studies to survive). ESM-free, so it can seed the commons with real, citable data on day one; user results refine it after. - scripts/run_benchmarks.py + scripts/seed…8d1e644