Opal Growth
Observed-single prediction of assay-local double-CRISPRi growth
Quick start 路 Evaluation 路 Limitations 路 Provenance
Research preprint 路 Source code 路 The Tiny Intelligence Lab
Maintainer-authorized research release with mixed source terms. Original Opal contributions are MIT licensed. Bundled measured references retain upstream terms and disclosed source-license uncertainty; see distribution terms and source notices. The original manifest is preserved, and release authorization supersedes its historical internal local-only policy without claiming new third-party permissions.
Opal Growth contains six original validation-selected fold-0 checkpoints. Each predicts the mean of two infection-specific gamma measurements for a pair of supported genes within its registered assay context. A calibrated single/main-effect reference combines with a symmetric learned pair residual. Historical "rank 8" is factor width, with algebraic rank at most 16.
Models and inputs
| Task | Required context | Supported genes |
|---|---|---|
horlbeck_K562 |
K562 |
427 |
horlbeck_Jurkat |
Jurkat |
427 |
environment1 |
K562_environment1 |
313 |
environment2 |
K562_environment2 |
313 |
parp2022 |
K562_parp2022 |
539 |
parp2023 |
K562_parp2023 |
537 |
Stored measured singles and support arrays accompany each checkpoint. Queries cannot replace those singles. Support requires at least two distinct TRAIN partners. Unknown or unsupported genes, self-pairs, malformed names and context mismatches produce explicit abstentions.
Quick start
Use Python 3.11-3.14; the local verification used Python 3.14.0. Install CPU Torch first if required, followed by the pinned source and dependencies:
python -m pip install "torch==2.9.1" --index-url https://download.pytorch.org/whl/cpu
python -m pip install -r requirements.txt
python verify_files.py
python example.py
Run these commands from this model repository folder. The requirements install the bundled Opal runtime wheel, built from the pinned source revision, so no private GitHub access is needed. The wheel contains the Python runtime source, but not the full historical experiment archive. No model training runs during installation or prediction. These checkpoints use Opal's custom loaders, not transformers.AutoModel.
example.py checks the original manifest hash, loads the files in this repository, runs a supported pair query and verifies symmetry and finite output. A successful example is an inference smoke check, not a new biological evaluation.
Evaluation
| Growth task | Held pairs across five blocks | Relative MSE reduction against the selected simple reference |
|---|---|---|
| K562 | 16,210 | 5.80% |
| Jurkat | 16,210 | 4.11% |
| Environmental 1 | 9,833 | 25.14% |
| Environmental 2 | 9,833 | 28.75% |
| PARPi 2022 | 29,016 | 17.56% |
| PARPi 2023 | 28,651 | 13.95% |
These are pooled five-fold training-procedure results, not scores of the single exported fold-0 checkpoints. Each procedure uses 16,000 training and 4,000 validation pair labels. Coverage is 100% on the registered populations among training-supported genes. The unchanged criterion requires at least 5% gain in every task; Jurkat fails at 4.11%, so the all-task promotion criterion remains failed.
Earlier frozen external transfer also failed, with completion MSE 1.08-2.15% worse than frozen main effects across four external cohorts. Local retraining on those exposed sources does not rescue that result. benchmark.json retains the complete aggregate comparison; it is not a new score. Local delivery verification previously replayed 17,176 stored predictions exactly without new fits or scores.
Limitations
Assay-local prediction does not establish biological discovery, causal understanding or reliable transfer to independent studies. These checkpoints cannot predict arbitrary cold genes or swap assay contexts. The two infection measurements are dependent assay readings, not independent replications. Calibrated per-query prediction intervals are unavailable. Conditional gene sensitivity is aggregate evidence only. Predictions have no clinical-use validation.
Provenance and license
Sources are Horlbeck 2018, GSE116198; environmental cohorts, GSE312636; and PARPi cohorts, GSE284018. The original manifest binds assay names, contexts, source hashes, checkpoint hashes and the fixed fold-0 deployment choice. Original checkpoint and schema bytes are retained. The historical manifest hash remains unchanged. A separate release authorization records the later project publication decision.
LICENSE.md distinguishes MIT licensing of original Opal contributions from upstream terms for source-derived references. SOURCE_NOTICES.md records attribution and remaining source-license scope uncertainty. This is not a blanket MIT or CC BY grant over every bundled reference value. RELEASE_AUTHORIZATION.json documents the maintainer publication instruction; it does not rewrite the frozen original manifest or supply missing upstream permissions.
Reproducibility and citation
- Original manifest SHA-256:
fa8490abe83e5a5d0d0f8456639d11df16a2a6ca3ed440fb40bfd74459488e72. - Opal source release:
v1.0.0. - Original model files and their hashes: provenance.json.
- Full folder checksums: SHA256SUMS.txt.
- Scientific methods and negative findings: research findings and preprint.
- Citation metadata: CITATION.cff. Cite the source study/dataset as well as Opal.
Prepared by The Tiny Intelligence Lab. Original checkpoint and manifest bytes are preserved; no new model training or biological evaluation was performed during packaging.