Read-across, refused — what it usually takes to get accepted on the second submission.

A read-across rejected by ECHA is not the end of the matter. Most often it is feedback on a justification that was technically present but scientifically incomplete. The reasoning was usually there. It just was not anchored in a way the evaluation could accept.

This piece is about what changes between an original read-across justification that is refused and a revised version that is accepted. The regulatory architecture around read-across has continued to develop. The OECD published its third edition of the Guidance on Grouping of Chemicals (GD 418) in December 2025, explicitly written to integrate NAMs, adverse outcome pathways, omics data, high-throughput screening and (Q)SAR. EFSA published its Guidance on the use of read-across for chemical safety assessment in food and feed in July 2025 (EFSA Journal 23(7):e9586). A systematic analysis of 1,538 ECHA testing proposal evaluations published in the same period (Maertens et al., ALTEX, 2024) provides the most comprehensive empirical view to date of how registrant-submitted read-across is actually evaluated. ECHA's RAAF, originally published in 2015 and extended in 2017 to multi-constituent substances and UVCBs, remains the operational framework for REACH dossier evaluation.

The picture across these documents is consistent. Read-across is no longer accepted as a structural argument supported by predictive tools. It is expected to be a hypothesis-driven scientific argument supported by structured evidence and explicit uncertainty analysis. The frameworks differ in scope. They converge on the standard.

Why original justifications are most often refused

In our experience, the most frequent reason an original read-across is refused is that the justification was built on in silico evidence alone. Structural similarity profilers, (Q)SAR predictions, and analogue identification tools are useful for hypothesis generation. They are not, on their own, sufficient as the empirical basis for a read-across decision under REACH. When the file rests entirely on these outputs, with no in vitro or in vivo data on the source substance, no toxicokinetic comparison, and no mechanistic argument for biological relevance, there is no empirical anchor against which to evaluate the prediction.

The OECD Third Edition is direct on this point. The grouping hypothesis must be supported by lines of evidence that include empirical data on at least one source substance and a documented mechanistic argument for transferability to the target. The (Q)SAR Assessment Framework (OECD, 2024) sets out the four principles against which any individual (Q)SAR prediction is assessed, and it does not present (Q)SAR output as substituting for the read-across justification. The same logic applies in EFSA's seven-step workflow, which separates target substance characterisation, source substance identification, source substance evaluation, and data gap filling as distinct steps. Each requires its own evidence.

The structural argument is the entry point. It is not the conclusion.

What the revised submission usually contains

When a refused read-across is rebuilt successfully, the structural argument typically does not change. What changes is the evidence base around it.

1 · Empirical data on the source substance

The first addition is empirical data on the source substance, evaluated for reliability and relevance. The OECD Third Edition is explicit that grouping hypotheses must be supported by experimental data, and the RAAF assessment elements assume the existence of such data on at least one analogue. Where the source has only legacy or limited-quality information, the revised file has to either upgrade the data set, expand the analogue group to include better-characterised members, or provide a transparent uncertainty analysis that justifies the conclusion despite the limitation.

2 · An explicit toxicokinetic argument

Read-across under REACH assumes that the source substance generates internal exposure profiles in the organism comparable to those expected for the target. In practice this means absorption, distribution, metabolism and excretion characteristics that are sufficiently aligned to make biological transferability coherent. In vitro metabolism data, where available, belong in the file. Where they are absent, the argument has to be constructed from physicochemical reasoning, structural metabolism predictions, and read-across at the metabolic level itself. The toxicokinetic question has to be addressed in some form. The OECD Third Edition treats this as part of the grouping hypothesis.

3 · A mode-of-action argument anchored to an AOP

The OECD AOP framework and AOP-Wiki provide a structure for articulating how a molecular initiating event leads to the apical effect being read across. Where an established AOP exists for the endpoint in question, the read-across argument benefits from being mapped onto it. Where it does not, a structured mode-of-action narrative supported by mechanistic evidence is the substitute. The biological link between chemistry and apical effect has to be visible in the file.

4 · NAM-derived evidence

The fourth addition, increasingly common, is the integration of NAM-derived evidence. In vitro bioactivity data from sources such as ToxCast and Tox21, where curated and contextualised properly, can confirm or challenge the biological similarity hypothesis at the level of key events. The OECD Third Edition was rewritten specifically to incorporate NAMs, AOPs, omics and high-throughput screening into the grouping argument. EFSA's 2025 guidance is explicit about the role of NAMs as tools for strengthening the read-across argument and for analysing uncertainty. ECHA's RAAF predates the NAM acronym, but the framework allows the same evidence to be brought in through the assessment elements that deal with biological plausibility.

5 · Transparent uncertainty analysis

The OECD Third Edition introduces methods to quantify read-across performance and uncertainties as a defined element of the grouping argument. EFSA's seven-step workflow includes uncertainty assessment as a discrete step. The RAAF expects uncertainty to be embedded across the assessment elements rather than addressed in a single closing paragraph. The revised submission usually contains a structured uncertainty discussion that the original did not.

The structural mistake to avoid

The temptation, when responding to a refused read-across, is to load the revised submission with new predictive output: additional in silico tools, expanded structural similarity matrices, more analogues. This is rarely the central problem.

The central problem is usually that the original argument was assembled around the structural hypothesis and asked the reader to infer the biology from it. The revised version inverts that structure. It begins with the regulatory question, which is the endpoint conclusion the read-across is required to support and the level of confidence the regulatory framework demands. It then identifies the biological reasoning that would justify that conclusion if the data were available for the target. It evaluates whether the source data, supported by toxicokinetic and mechanistic evidence, provide that reasoning. The structural similarity argument is then deployed in support of the biological case, not as the foundation of it.

What this means in practice

A read-across refusal is recoverable. It does not necessarily require new vertebrate testing, and it does not necessarily require a new analogue group. What it requires is an honest re-examination of the gap between the predictive argument and the empirical and mechanistic argument, and a rebuild of the file that closes that gap on the page.

Three operational points follow.

1. Plan the empirical scaffold from the outset. When commissioning a read-across, the empirical and mechanistic scaffold should be planned from the outset, not retrofitted after refusal. In silico tools have a defined role in hypothesis generation and applicability domain analysis, but the file should not rest on them. Where empirical data on source substances are limited, this needs to be identified and addressed before the dossier is submitted.
2. Read the refusal carefully. If a read-across is refused, the decision should be read carefully before responding. The assessment elements ECHA cites in the decision identify the specific gaps in the original reasoning. The response should address those gaps, not provide a generic strengthening of the file. Where the decision identifies an evidentiary gap, predictive tools alone will not close it.
3. Build to one standard. The OECD GD 418, EFSA 2025, and the RAAF differ in scope but converge on the same standard. A read-across is a hypothesis-driven scientific argument supported by empirical evidence on the source substance, mechanistic reasoning for transferability, NAM-derived support for biological plausibility where available, and explicit uncertainty analysis. Files built to that standard from the outset rarely need to be rebuilt.

The discipline has matured considerably since the original RAAF in 2015. It now sits at the intersection of structural chemistry, mechanistic toxicology, NAM-derived evidence and uncertainty quantification, with regulatory frameworks across REACH, food and feed, and increasingly cosmetics, all converging on similar expectations. A refused read-across, properly understood, is feedback on the file. The accepted version that follows is usually a more honest piece of science.