The limitations of conventional preclinical species are well recognized. A compound can appear reassuring in animal studies and yet expose an unexpected liability in the clinic. That does not make animal studies irrelevant; rather, it tells us that no single model can adequately represent the complexity of human toxicity. New approach methodologies (NAMs) give us the opportunity to complement conventional approaches with human-relevant, organ-specific systems that interrogate the mechanisms we ultimately care about.

The real opportunity, however, is not in treating NAMs as a collection of individual assays. It is in building an integrated safety cascade in which complementary biological signals are considered together. Drug-induced liver injury (DILI) provides a compelling example. BSEP inhibition, mitochondrial dysfunction, cellular phenotypes and responses in three-dimensional human liver microtissues each capture different dimensions of hepatic liability. When these signals are considered together, our ability to distinguish potential human liabilities improves substantially. The broader lesson is equally important: when toxicity is mechanistically heterogeneous, our predictive strategy must be capable of capturing that heterogeneity.

This is where AI and machine learning become genuinely useful. The objective is not to replace biological understanding with an algorithm, but to extract information from biological complexity that conventional decision rules cannot readily capture. By integrating data from multiple NAMs, machine-learning models can identify patterns that may not be apparent when each endpoint is considered independently. More importantly, they offer a route towards connecting biological responses with exposure, allowing us to ask not simply whether a compound produces a response, but whether that response occurs within a clinically meaningful range. That is a fundamentally more useful question when making decisions about compound progression and risk.

But we can go further. Toxicity is rarely the consequence of a single molecular event and our assays should not be constrained to measuring only a handful of predefined endpoints. Whole-genome transcriptomics gives us a much wider window into what a human cell is actually doing in response to a compound. When these data are generated across human-relevant liver, cardiac and kidney models and combined with well-annotated reference compounds, they begin to form something much more powerful: a mechanistic map against which new compounds can be interpreted. AI/ML can then help translate these complex molecular signatures into predictions of exposure, points of departure and biological similarity. The question becomes less about detecting toxicity after the fact and more about recognizing the molecular trajectory towards toxicity early enough to act.

This, in our view, is the more interesting direction for safety science. We are moving from isolated toxicity flags towards integrated, mechanism-informed predictions that bring together human biology, quantitative exposure and computational intelligence. But what does this look like when applied in practice? How can transcriptomic signatures be converted into meaningful points of departure? How can AI distinguish a biological response from a response that represents genuine clinical risk? And how much more can we learn when we look across organs rather than studying each liability in isolation? These are the questions we explore in the webinar below. Watch the video to learn how in vitro safety NAMs, transcriptomics and AI/ML are being brought together to push toxicity prediction towards a more human-relevant and mechanistic future.