Repository navigation
Releases
Opal v1.0.0
Opal v1.0.0
Opal's first research source release provides expression and growth prediction code, explicit input/support contracts, the research paper, frozen protocols and compact scientific evidence. GitHub contains no trained weights or biological reference banks; model artifacts have a separate Hugging Face release.
The retained six-task growth benchmark reports pooled five-fold procedure results, not scores of the single exported fold-0 checkpoints.
| Task | Selected simple MSE | Completion MSE | MSE reduction |
|---|---|---|---|
| K562 | 0.00190091929 | 0.00179069074 | 5.80% |
| Jurkat | 0.000507630148 | 0.000486775245 | 4.11% |
| Environmental 1 | 0.000551581712 | 0.000412898098 | 25.14% |
| Environmental 2 | 0.000727352832 | 0.000518265563 | 28.75% |
| PARPi 2022 | 0.000308097243 | 0.000254004302 | 17.56% |
| PARPi 2023 | 0.000302940513 | 0.000260679297 | 13.95% |
Jurkat misses the registered 5% threshold, so the all-task promotion gate failed. Earlier frozen-weight K562 transfer also failed: completion had 1.08–2.15% higher MSE than frozen main effects, and minimum-single references performed better. Later fitting on exposed local sources does not reverse that result. Failed composition, context-sharing, low-label and numerical-compatibility experiments remain documented.
Opal supports assay-local prediction among training-supported genes. It establishes no biological discovery, causal understanding, frontier-level superiority, reliable independent-study transfer or calibrated biological prediction intervals. Technical pair effects remain confounded with biological interactions.
The six retained growth checkpoints replayed 17,176 predictions exactly during release preparation, with no new fits, scores or test-target access. Reproducibility separates source-free checks from local biological replay. The public checkout lacks assets needed to regenerate every historical biological score. Validation records release checks.
Original code is MIT licensed; third-party notices preserve external terms. Trained models have a separate Hugging Face distribution through TheTinyIntelLab. No archive DOI or arXiv identifier is claimed.
Release assets
- Python wheel and source distribution for installation.
- Research-source archive with the code, paper and preserved historical records.
- Evidence archive with protocols, provenance, aggregate results and negative findings.
- Manuscript PDF and standalone TeX source bundle.
- File manifest and SHA-256 checksums.
Validation
Release-integrity CI passes on the tagged source revision across Linux and Windows with Python 3.11 and 3.14. The complete preparation suite recorded 1,223 passed, 3 skipped and 1 expected failure; runtime source is unchanged. See the validation asset for check scope and local rerun limitations.
The release distributions pass content and nested-archive audits. The standalone TeX bundle compiles, active documentation links resolve, and the installed wheel passes the source-only release checks. Computational replay preserves original model behavior and does not establish biological generalization.
Model downloads
- Opal Growth: six assay-local growth checkpoints with required reference schemas.
- Opal Norman: historical Norman expression reference.
- Opal EmeraldBay: historical contextual expression reference.
Model cards link to the v1.0.0 source and manuscript. Model-specific terms apply; the growth release distinguishes maintainer authorization for original Opal contributions from upstream reference terms and disclosed source-license uncertainty.
Opal 0.1.2: comparative research release
Opal evaluates population-mean combinatorial response prediction when matched single effects are observed. This experimental research release packages the comparative study, scoped reference inference, replicated-study tooling and retrospective measurement diagnostics.
The nine uploaded assets are the validated clean snapshot of commit 53d6e83fc251b789e90fbd2a95c050cd3e200592. They include the wheel, source distribution, research source, compact evidence, permitted Norman/EmeraldBay reference exports, prepared manuscript and standalone TeX source, manifest and SHA256SUMS. Raw expression matrices, private plans and unlicensed exports are excluded. Code is MIT; the permitted dataset-derived reference content retains CC BY 4.0 terms and attribution.
Validation: 213 full local tests; all four Linux/Windows Python 3.11/3.14 CI jobs pass fresh CPU dependency installs, 106 integrity tests and the standalone synthetic assay cycle. Local installed-wheel checks also verify 65 saved reference predictions, support guards and the synthetic quickstart outside the checkout. Actual LNCaP bulk diagnostic fields replay exactly from the installed wheel.
Scientific limits remain explicit. On the exposed Combi-seq retrospective benchmark, the pair-kernel candidate loses to the selected incidence reference. The LNCaP bulk diagnostic measures reported-repeat agreement at separate times and fits no new model. It does not verify independent culture initiation. No original biological advancement gate passes. Later companion evidence discloses residual-cosine summaries as a post-access descriptive amendment beyond the narrow initial bulk error-metric scope; the original artifacts remain preserved.
The prepared manuscript is supplied as a comparative research document. No arXiv identifier, DOI, public registration, peer review or prospective confirmation is claimed. GitHub publication was explicitly approved by the owner; Zenodo archival and arXiv submission remain separate.
CI: https://github.com/TheTinyIntelLab/opal/actions/runs/37955608011
See THIRD_PARTY_NOTICES.md, docs/next_cycle/execution_status.md, docs/next_cycle/retrospective_combi_20261009.md and docs/next_cycle/eldiaz_processed_diagnostic.md for claim and source boundaries. The release manifest binds the immutable build snapshot; later companion records are retained in the repository.