Measured labels. Visible limitations.

Evidence behind
every signal.

Predictions compared with published Tox21 measurements. The outcomes, the errors, and the conditions that shape the result.

Read the test report
Distinct molecular parent structures1,000
10,623 observed assay labels12 biological endpoints

Retrospective experimental assay evidence.
Not a clinical or chemical safety validation.

01 / Know the test set

Three sources.
One transparent cohort.

598NIH final evaluationExternal source
285NIH leaderboard testExternal source
117Recorded held-out subsetHistorical benchmark

Exactly 1,000 distinct stereo-free fragment parent structures. The 883 external-source compounds are reported separately in the PDF. Reference labels are measured assay outcomes; missing observations are excluded per endpoint.

02 / Compare the models

Performance,
with the uncertainty.

Each assay contributes equally to the macro average. ROC-AUC measures ranking; average precision emphasizes retrieval of active assay observations.

Macro ROC-AUC

Mixed 1,000-compound cohort
GCN
0.780
GINE
0.777
GATv2
0.755
ECFP-RF
0.737
ECFP-LR
0.686
0.00.51.0

Higher is better for ranking. This is not percentage accuracy.

Full model comparison · mixed 1,000-compound cohort
ModelROC-AUCAverage precision95% bootstrap AUC interval
GCNGraph neural network0.7800.280[0.754, 0.803]
GINEGraph neural network0.7770.258[0.749, 0.802]
GATv2Graph neural network0.7550.274[0.725, 0.783]
ECFP-RFFingerprint baseline0.7370.269[0.709, 0.765]
ECFP-LRFingerprint baseline0.6860.250[0.651, 0.720]
03 / Verify the deployed system

The service matches
the saved models.

All 1,000 molecules also passed through the application API in ten sequential batches.

3,000/ 3,000

Model–molecule parity comparisons

36/ 36

Service checks passed

2.98 × 10⁻⁷

Maximum absolute score difference

Local CPU tests. Sequential timings are not production throughput or a load-test guarantee.
Emerald molecular orbit sculptureValidation study / 2026
19 pages of measured evidence

The full picture.
Ready to inspect.

Assay-level tables, ROC and precision-recall curves, real measured examples, calibration, uncertainty and model disagreement.

04 / Interpretation guide

What the numbers
can tell you.

Research limitations
01

Read activity in a specific assay.

Labels indicate activity in named biological assays. The 0.5 threshold is exploratory; neural scores have not been calibrated as probabilities of harm.

02

Keep uncertainty conditional.

Bootstrap intervals describe uncertainty for the tested weights and compounds. They do not measure variability across independent neural training runs. Chemical dependence and scaffold overlap can affect them.

03

Trace the reference evidence.

The cohort manifest preserves source IDs, structures and observed labels. External-only results remain separate from the historical held-out subset.

NIH Tox21 Challenge data