Introducing Opal v1
Opal's code and models are now public. We explain how it predicts paired perturbation responses, what the growth benchmarks show, and where it failed.

We've released Opal v1.0.0, including the code, model weights and research preprint. Opal predicts how two perturbations behave together using measurements of their individual effects and a set of other measured pairs from the same assay.
The results are mixed. In the six-task growth benchmark, the largest reduction in prediction error was 28.75%. Jurkat improved by 4.11%, below the 5% required by our promotion rule. That failure remains part of the release.
What Opal learns from
If you have 100 interventions, there are 4,950 distinct pairs to measure. Single-perturbation measurements give you a useful starting point, but simply adding their effects can miss the response of a pair.
Opal learns a correction to a reference prediction. It uses measured pair outcomes during training as well as the measured singles, so those pair-label requirements matter when interpreting the results. The pair model is symmetric: swapping the two interventions preserves its prediction.
We kept growth and expression as separate prediction problems. Growth models return a scalar response for a supported gene pair in a registered CRISPRi assay context. Expression references return a population-mean expression-effect vector in a fixed gene order. Their inputs and evaluations differ, and a good growth score tells us little about expression performance.
Six growth tasks, one failed promotion rule
The growth benchmark compares completion with a validation-selected simple reference for each task. These are relative reductions in mean squared error, with larger values indicating lower error on the registered evaluation population.
| Growth task | Relative MSE reduction |
|---|---|
| K562 | 5.80% |
| Jurkat | 4.11% |
| Environmental 1 | 25.14% |
| Environmental 2 | 28.75% |
| PARPi 2022 | 17.56% |
| PARPi 2023 | 13.95% |
Each fitted procedure uses 16,000 training and 4,000 validation pair labels, measured singles from its own source, and genes supported by training pairs. The scores pool five disjoint held-out blocks per task. We're distributing the original validation-selected fold-0 checkpoint for each task; the pooled scores describe the training procedure, rather than a separate evaluation of that individual checkpoint.
Our registered rule required at least a 5% improvement in every task. Jurkat missed it, so the all-task promotion criterion failed. The larger gains elsewhere don't change that decision. You can inspect every comparator and the evaluation populations in the benchmark ledger.
Transfer was a more direct disappointment. We froze an earlier K562 predictor and evaluated it on four external cohorts. Completion MSE was 1.08% to 2.15% worse than frozen main effects. Later fitting within those exposed sources answered a different question and did not rescue the frozen-transfer result.
That leaves us with evidence for predicting missing pairs within an assay among training-supported genes. It doesn't establish reliable prediction in a new biological study. The records also retain failed expression-composition comparisons, pair-label-budget experiments and numerical-compatibility checks. We haven't established biological discovery, causal understanding or frontier-level superiority.
What you can download
There are three model repositories, each with its own input contract and example:
- Opal Growth contains the six growth checkpoints and their required measured-single and support schemas. It also includes an Opal runtime wheel.
- Opal Norman is a historical expression reference for the Norman K562 CRISPRa assay.
- Opal EmeraldBay is a historical contextual expression reference for registered chemical interventions and cellular contexts.
For growth queries, you need supported genes in the correct assay context. Unknown or unsupported genes, self-pairs and incorrect contexts produce abstentions. Calibrated per-query prediction intervals are unavailable, and the models have no clinical-use validation.
The GitHub release includes the paper, implementation and scientific evidence. We preserved historical protocols and negative results in checksummed archives with restoration tools. Raw biological datasets remain separate from that source distribution.
The licenses also depend on what you're downloading. Original Opal code is MIT licensed, while Norman and EmeraldBay retain recorded CC BY 4.0 source notices. The growth release separates authorization for original Opal contributions from upstream reference terms and discloses remaining source-license scope uncertainty. Read its distribution terms and source notices before reusing those assets.
Checking the files yourself
During preparation, the six growth checkpoints reproduced 17,176 saved predictions exactly, without new fitting or biological scoring. We also downloaded the published repositories, verified their hashes and ran their prediction examples using the downloaded runtime and recorded scientific dependencies.
That checks the behavior of the distributed files. It doesn't add an independent biological replication. The reproducibility guide explains which checks you can run from a public checkout and which historical experiments require separately obtained assets.
If you want to try Opal, start with the example in the relevant model repository. For the methods, full comparisons and failed experiments, the preprint is available alongside the code.
