Reproducibility in ToxIndex

When a regulator asks you to explain how you reached a conclusion, you need to show your work. When they ask you to repeat the analysis, you need to get the same answer. ToxIndex is built to make both of those straightforward.

What we’re building toward

The goal: run a pipeline with the same chemical input, the same data, and the same model versions, and get identical output.

Every piece of evidence in an assessment traces back to a specific source: a study in Yard, a document in Crawler, a model prediction from ToxJobs. The reasoning chain from input to conclusion is recorded and inspectable.

How we approach it

Data pinning. Before a pipeline runs, ToxIndex snapshots the relevant Yard data. The pipeline operates against this snapshot, not the live graph.

Model tracking. Predictions record which model weights were active when the pipeline started.

Parameter recording. Pipeline steps record their input parameters, configuration, and execution context.

Isolation. Each assessment runs independently. No shared state between runs.

Audit trails

Every pipeline run produces an audit trail. It records:

• Pipeline template version and input parameters
• Yard data snapshot version
• ToxJobs model versions
• Execution timestamps
• Step-by-step inputs and outputs
• Warnings and errors

Audit trails are immutable. Once an assessment finishes, its trail can’t be modified. You can export it for regulatory inspection, QA review, or internal documentation.

Evidence provenance

Every data point in an assessment result includes a provenance chain to its original source.

Experimental data traces back to a specific study, publication, or regulatory dossier in Yard. Predictions trace to the specific model, model version, training data, and applicability domain assessment in ToxJobs. Read-across evidence traces to the analogs, similarity metrics, and supporting data. Regulatory context traces to the specific document and section in Crawler.

No black boxes. You can always follow a finding back to where it came from.

Rerunning assessments

You’ve got two options for reruns:

Original conditions. Select the assessment and click rerun. ToxIndex uses the versioned data snapshot and model weights from the original run. You get identical results, confirming the original assessment.

Current conditions. Select the assessment and choose “rerun with current data.” ToxIndex uses the latest Yard data and ToxJobs models. Comparing the two results shows you exactly what changed: new data points, updated predictions, different regulatory context.

This is useful for responding to regulator queries, updating assessments after guidance changes, and validating that platform updates don’t shift existing conclusions.

Why regulators care

ECHA compliance checks can ask registrants to explain their assessment methodology months or years after submission. FDA and EPA reviews may request re-analysis of specific endpoints. Internal QA teams need to verify assessments were conducted correctly.

Documented, reproducible methodology gives you legal defensibility. The assessment can be independently verified by running the same pipeline with the same inputs. As regulatory agencies increase their use of computational methods, this kind of documentation is quickly becoming the expectation.