A computed estimate of an environmental endpoint is a planning instrument. It tells you which substances in a portfolio are likely to need a study, which study to commission first, and where a data gap is going to become a problem later. It is not a substitute for the study where a dossier, a classification or a risk assessment depends on the value, and treating it as one is the mistake that costs the most time, because the gap reappears during review. This article is about the division between those two roles, and the documentation that makes the first role defensible.
What the frameworks actually permit
Regulatory systems do not reject computed estimates. They constrain the conditions under which one can replace a test. Under REACH, Annex XI sets out the adaptation routes, which include validated computational models and read-across within a group of structurally similar substances, each with its own requirements. The OECD principles for the validity of such models name what has to be in place: a defined endpoint, an unambiguous algorithm, a defined domain of applicability, appropriate measures of fit and predictivity, and, where possible, a mechanistic interpretation.
The practical reading of all of this is that a computed estimate is acceptable when the model's validity, the substance's position inside its applicability domain and the reliability of the prediction are documented, and when the endpoint and the regulatory purpose match. Outside those conditions it is screening information, which is still worth having.
Where a computed estimate earns its place
- Portfolio triage. Several hundred substances, a limited testing budget, and a need to know which twenty are likely to drive the registration workload.
- Early development. Choosing between candidate molecules while the choice is still cheap, before a study programme exists at all.
- Study planning. Picking the test species, the concentration range and the design, so the study that is commissioned is the right one.
- Gap mapping. Listing which endpoints will need data for which tonnage band or market, with a first estimate of each, so the sequence and budget can be planned.
- Grouping hypotheses. Identifying which substances might form a defensible category for read-across, which is then developed properly rather than guessed.
Where it does not
- Any value that enters a classification decision, a safety data sheet section as a reported property, a risk assessment ratio or a dossier endpoint, unless the adaptation route is met and documented.
- Substances outside the model's applicability domain, which includes most structurally unusual chemistry, surfactants, organometallics, ionisable substances at environmental pH and multi-constituent or UVCB substances.
- Endpoints that depend on fate as much as on intrinsic properties, where the computed value for the parent substance says little about what is actually in the compartment.
- Degradation and transformation behaviour as a basis for a conclusion of rapid degradability, which has specific test requirements.
Document the prediction, not just the number
A prediction recorded as a number in a spreadsheet is worthless six months later, because nobody can reconstruct what it was. What makes it reusable is the same information a reviewer would ask for:
| Item | Why |
|---|---|
| Substance identity as a structure, with the form modelled | Salt, parent or dissociated species changes everything |
| Endpoint, species and test guideline the estimate refers to | "Toxic to algae" is not an endpoint |
| Model, version and the training set's character | A reviewer cannot assess a model that is only named |
| Applicability domain outcome, with the reason | Inside or outside, and on what basis |
| The prediction with its uncertainty | A point value implies precision that does not exist |
| Analogues used, if read-across | Identity, measured data and the justification for similarity |
| Intended use of the estimate | Screening or adaptation, stated at the time |
The standard formats for exactly this are the model and prediction reporting formats, QMRF and QPRF. Using them even for internal screening work costs little and means an estimate can later be promoted to a dossier argument if the conditions are met, instead of being redone.
Read-across is a scientific argument, not a lookup
Grouping substances and using a measured value from one for another is the adaptation route most often attempted and most often rejected. The rejections are usually not about chemistry but about the argument: structural similarity asserted rather than justified, no account of how the differences would affect the endpoint, no bracketing, a single analogue, or analogue data that turn out to be unreliable. A read-across case that holds up names the structural basis, the mechanistic reasoning for why the endpoint should behave similarly, the data quality of the source substances, and the direction in which the difference would be expected to push the result.
Limits of the statement
- Nothing in this article constitutes an environmental or toxicological assessment. Those assessments, and the decisions about whether an adaptation route applies, belong with a qualified assessor and depend on the specific substance, endpoint and regulatory context.
- Acceptance practice differs between jurisdictions, programmes and endpoints, and it changes. The adaptation routes and current guidance for your specific submission are what govern, not a general article.
- A model that performs well on a published validation set may still be unsuitable for your substance class, which is what the applicability domain assessment is for.
- Screening is only useful if its conclusions are revisited. An estimate made early and never re-examined against later measured data is an assumption carried silently through a programme.
Frequently asked questions
Can a computed estimate replace an ecotoxicity study?
Only where a recognised adaptation route applies and its conditions are documented, for example a valid model with the substance inside its applicability domain, or a properly justified read-across. Otherwise the estimate is screening information, which has real value for planning but not for the dossier endpoint.
What makes a computational prediction defensible?
A defined endpoint, a documented model and version, an explicit applicability domain outcome, a stated uncertainty, and a record of the form of the substance modelled. The QMRF and QPRF reporting formats exist for exactly this purpose.
Why is read-across rejected so often?
Because similarity is asserted instead of argued. A case that holds names the structural basis, explains mechanistically why the endpoint should behave similarly, assesses the quality of the source data, and addresses how the structural differences would be expected to affect the result.
Which substances are poor candidates for a computed estimate?
Those outside the usual applicability domains: surfactants, organometallics, substances ionised at environmental pH, multi-constituent and UVCB substances, and structurally unusual chemistry generally. For these, the estimate's main use is deciding which study to run first.
Is it worth documenting screening predictions formally?
Yes. The documentation costs little at the time and is the difference between an estimate that can later support a regulatory argument and one that has to be regenerated from scratch.