EVERY STEP OF A TOXICOLOGICAL RISK ASSESSMENT, IN ONE PLACE
The science took an hour. The stitching took three weeks.
IRIS for the reference dose. PubChem for identifiers. ECHA for the REACH dossier. Then ATSDR, IARC, PubMed. Then a spreadsheet named v7_FINAL that is not final.
Two sources disagree on the same endpoint, and nothing on either page tells you which one outranks the other. Three endpoints have no data at all, and the assessment still has to say something about them.
None of that is science.
WHY THIS IS BUILT THE WAY IT IS
A risk assessment is not a report. It is a position someone defends.
When a value is questioned, the question goes to the person who signed the assessment, not to the database it came from. Sometimes years later.
Which is why toxicologists are right to distrust automation. Speed usually arrives by removing the ability to check, and a professional who cannot check cannot sign.
TOXRA.AI is built inside that constraint. Make the stitching disappear. Leave the reasoning visible.
Not a better search engine.
Lookup is the cheap half of the problem. The expensive half starts once the data is in front of you: reconciling it, filling what is missing, testing it against a framework, and turning it into something that survives review.
So this is an assessment instrument, not a search product. One chemical context that follows you from first identifier to signed report, with the seams left visible on purpose.
CHEMICAL SEARCH AND PROFILES
Search once. Every value arrives with its origin attached.
Name, CAS number, SMILES string or DTXSID. What comes back is one assembled profile rather than a list of links: identifiers, physical chemistry, hazard classifications, occupational limits, reference values and literature, from 51 sources in a single view.
Provenance at the moment of retrieval. Every number traces back to its source and the version of that source. That is the difference between a value you found and a value you can defend.
Disagreements surfaced, not quietly resolved. When two sources conflict on the same endpoint, both are shown, with source precedence applied so an authoritative classification is never overridden by a secondary aggregation. The conflict is a finding.
Gaps named with their remedy. Where an endpoint has no data, the profile says so and names the specific OECD, EPA or ICH study that would close it. A blank becomes a decision.
For most chemicals, the study you need was never done.
No amount of searching fixes that. And regulation is moving away from generating it, toward new approach methodologies, which squeezes the assessor from both sides at once: less data, and less licence to go and make more.
What is left is inference from structure. That is legitimate science with one condition attached. A prediction that cannot explain itself is worthless in a submission, however accurate it happens to be.
IN-SILICO NAMS (QSAR)
ICH M7 does not ask for a prediction. It asks for two.
One expert rule-based, one statistical. Two methods that reason differently, run against the same molecule, with their agreement or disagreement reported. That is not bureaucracy. It is epistemology written into regulation.
Law reached the same split about fourteen hundred years earlier.
Pandecta
The rule-based engine
In 533, Justinian's jurists compressed a thousand years of legal opinion into fifty books and named the result from the Greek for all-receiving. Then he forbade commentary on it, because a code anyone may reinterpret is no longer a code.
A structural alert is a statute. PANDECTA reads it that way: this substructure, in this context, indicates this concern, for these documented reasons.
78 active alerts, authored in a clean room from seven families of primary literature. Not ported from a commercial library, and never fitted to the set used to test them.
0.844 sensitivity at the alert layer against Hansen 2009, out-of-sample for the rule layer.
Silence is reported as silence. A structure that fires nothing returns nothing to report inside the domain. It is never returned as safe.
Precedenta
The statistical engine
Common law worked in reverse. From about 1166 no code came first. Judges decided individual cases, those decisions were reported, and the accumulated record became binding on the next case. It grew stronger as the record grew larger.
PRECEDENTA holds no rule about mutagenicity at all. It finds the tested chemicals structurally nearest the query and reasons from them.
Trained on curated experimental data. Assays reconciled to a consistent protocol, structures standardised and de-duplicated, contradictory records resolved rather than averaged, and the result held to a defined chemical space. Validation is external, because a model scored against the data it was fitted to will flatter itself.
ROC-AUC 0.897 on a locked internal hold-out (n=1,228). On a genuinely external set the model never saw in training — the EPA CompTox-genetox Ames panel (n=1,048) — sensitivity is 0.849 at 0.981 negative predictive value. The Hansen 2009 set scores higher, but it is 99.6 percent folded into training, so that figure is resubstitution, not external.
The threshold is sensitivity-biased on purpose, holding false negatives to 7.9 percent on the internal hold-out. A false alarm is resolved by the next assay in the sequence. A missed mutagen travels forward into development. Those are not comparable errors, and the trade is deliberate.
It abstains. Conformal prediction defines a gray zone, and structures falling inside it are returned as inconclusive and routed to expert review rather than classified badly. Within confident calls, specificity is 0.89 to 0.91.
It shows its neighbours. The analogues that drove the result arrive with their measured outcomes and their structural similarity, so the reasoning can be read the way a lawyer reads the cases a judgment rests on.
When they agree, and when they do not.
Pandect and precedent. Rome in 533, England from 1166. Deduction from a code, induction from decided cases. The two traditions were rivals rather than a designed pair, and we are borrowing the contrast rather than claiming otherwise. What neither could do is run both methods against the same molecule and hand you the reconciliation.
Concordance is a strong position and ICH M7 treats it as one. Divergence is the more valuable output, because it identifies the exact compound where judgment is required and shows why. Here is the rule that fired. Here are the analogues that disagree with it.
2 alerts fired
Four of the five nearest tested compounds came back negative, which is why the statistical leg said negative. The enone is why the rule leg said positive. Both are visible, and the assessor decides.
Read-across. Precedent applied deliberately rather than statistically. Analogues named, justification built, the grouping yours to accept or reject. Regulators accept read-across when the rationale is documented, so the rationale is the deliverable.
Carcinogenicity weight of evidence. Never one number. Classifications from multiple authorities, genotoxicity findings, mechanistic evidence and structural concerns, assembled as strands you can weigh rather than collapsed into a verdict.
ASSESSMENTS AND FRAMEWORKS
Knowing the framework is not the same as proving you met it.
The work is walking the clause list and showing, for each clause, what evidence satisfies it and where the record is thin.
TOXRA.AI evaluates against six versioned frameworks at clause level and reports data sufficiency per clause. A framework is not satisfied because the fields are populated. It is satisfied because the right evidence sits in the right place.
Exposure is modelled with EPA standard scenarios feeding straight into chronic daily intake. The calculators are the ones already in use: ELCR, MOE, MOS, RfD and TTC, with dose unit conversion handled rather than done twice by hand. Every figure keeps its audit trail, and every assessment records which version of the framework judged it.
AUTO MODE AND WORKSPACE MODE
Save the time. Keep the pen.
The objection is never whether it is fast. It is what happens to your judgment. A tool that produces something plausible by means you cannot inspect costs you time, because you redo the work by hand before you put your name on it.
So there are two modes, and the difference between them is authorship, not power.
Auto mode. One chemical in, a complete draft assessment out: profile, in-silico predictions, exposure, calculations, framework evaluation and report structure. For triage. Twelve candidate impurities where eleven will be dropped and the question is which one. Labelled a draft, because that is what it is.
Workspace mode. Coming soon The same pipeline opened up, with you in command at every stage. Choose which sources to accept and which to set aside. Adjudicate a flagged disagreement and record why. Override an alert. Set the exposure scenario instead of inheriting a default. Every intervention is captured, so the finished assessment shows where judgment was exercised.
These are not tiers, and not a beginner and expert split. It is the same instrument at two magnifications. Auto to get the shape of a problem in ten minutes. Workspace for the compound that is going to be defended.
REPORTING AND AUDIT TRAIL
The assessment ends when the document survives review.
Built from the workspace, not retyped into it. Everything assembled and every decision made carries through into a source-traced document. The report is not a summary of the work. It is the work, formatted.
Pluck it out, plug it in. Not every assessment is going to live inside TOXRA.AI, and pretending otherwise would be a fiction. Sponsors have templates. Consultants have house formats. Submissions have prescribed structures. So any element lifts out into the document you are actually writing, with its traceability travelling along: a value with its provenance, a QSAR result with its analogues, a clause finding with its justification.
The audit trail. Every figure traceable to its source and version. Every override recorded with its reasoning. Every calculation reproducible, years later, when someone asks.
GLP document review. Documentation checked against GLP requirements, catching omissions during drafting rather than at audit.
TOXRA.AI does not sign assessments.*
It does not replace a qualified toxicologist and it is not built to. Every prediction carries an applicability domain, and the platform says when a query falls outside it. Every automated evaluation is a draft position for a professional to accept, amend or reject. Features in active development are labelled as such rather than presented as finished.
Different frameworks. The same three weeks of stitching.
Pharmaceutical impurity assessment. ICH M7 across a candidate list, not one compound at a time.
Medical devices. ISO 10993 extractables and leachables, at the scale an extractables study actually returns.
Cosmetics and personal care. Safety assessment without new animal data.
Food contact and ingredients. Migration and dietary exposure, with the sources reconciled.
Tobacco and PMTA. FDA CTP tiering across long constituent lists.
CROs and consultancies. Someone else's deadline on top of your own.
The science was never the slow part. Now the rest of it is not either.
One workspace, one traced chain of evidence, one document that holds up. Bring a compound you already know the answer to, which is the only demo worth sitting through.