Toxicity screening and toxicogenomics

Classical toxicology observes apical endpoints: an animal develops a lesion, loses weight, or does not. These outcomes are what regulation is written around, but they are slow, expensive, insensitive to low-frequency effects and demanding of animals. Toxicogenomics substitutes an earlier, richer measurement — the molecular response of cells or tissue to exposure — and asks whether that response predicts the endpoint.

The signature idea and its arithmetic

A compound perturbs transcription in a pattern reflecting the pathways it engages: oxidative stress response, DNA damage response, nuclear-receptor activation, mitochondrial dysfunction. Reference collections pairing compounds with expression profiles allow a new compound’s profile to be compared against those of substances with known effects, so mechanism can be inferred by similarity rather than by hypothesis. The same data support a dose-response reading: transcriptional changes are measured across concentrations and the point of departure is estimated from the dose at which coordinated gene-set responses begin. Empirically, transcriptomic points of departure often fall near those derived from apical endpoints in the same study design, which is the quantitative basis for treating them as a screening-level surrogate.

The concordance problem

The difficulty is that a molecular response is evidence of perturbation, not of harm. Cells mount stress responses to exposures they fully recover from, so a signature can flag a compound that would never cause injury in an animal. In the other direction, toxicities that depend on properties the assay does not have — metabolic activation by liver enzymes absent from the cell line, immune involvement, accumulation over months, effects on a tissue not represented — are missed entirely. Concordance between in-vitro signatures and in-vivo outcomes is therefore substance- and endpoint-dependent, good for some mechanisms and poor for others, and reporting a single overall accuracy figure for the approach obscures that structure.

Extrapolation is the second half of the problem. An assay reports a nominal concentration in a well, while a regulatory decision needs an external dose. The concentration actually reaching the cell differs from the nominal one because compounds bind plastic, serum protein and lipid, and translating a free concentration into an administered dose requires reverse dosimetry through a pharmacokinetic model. This step, rather than the assay, is frequently where the uncertainty concentrates.

What the programmes established

Tox21, the collaboration between US federal agencies, screened a library of thousands of chemicals across a large panel of cell-based assays in quantitative high-throughput format. Its lasting contributions are a public dose-response dataset of unusual size and a demonstration of what the format can and cannot do: it prioritises chemicals efficiently and identifies pathway activity, and it does not reproduce whole-organism toxicity on its own. The broader new-approach-methodology framework now under active regulatory development combines such assays with organotypic culture systems, exposure modelling and read-across from structurally related substances, with validation against existing animal data as the acceptance criterion.

The realistic present position is that these methods are established for prioritisation, mechanistic interpretation and reducing the number of animal studies needed, and are still being qualified — endpoint by endpoint — for replacing them in regulatory submissions.