AstraZeneca has announced a research collaboration with Tremont AI focused on applying artificial intelligence to preclinical toxicology screening — the process of identifying whether a drug candidate is likely to cause harmful side effects before it ever reaches human clinical trials. The partnership adds to a wave of pharmaceutical industry deals in 2026 pairing established drugmakers with AI startups specializing in narrow, high-value parts of the drug development pipeline.
Why toxicology screening is such a costly bottleneck
Drug development remains an extraordinarily expensive and failure-prone process, and toxicity issues discovered late — sometimes only after a drug has already entered human clinical trials — represent some of the most costly failures a pharmaceutical company can experience, both financially and in terms of lost development time that could have gone toward more viable candidates. Traditional toxicology testing relies heavily on animal studies and laboratory assays that are time-consuming and do not always accurately predict how a compound will behave in humans.
What AI changes about the process
AI models trained on large datasets of known chemical structures and their associated toxicity outcomes can, in principle, predict which structural features of a new drug candidate are likely to cause problems — flagging risky candidates earlier, before a company invests heavily in further development and testing. This approach does not eliminate the need for traditional safety testing, but it can help companies deprioritize the riskiest candidates earlier in the pipeline, focusing limited resources on compounds more likely to clear safety hurdles later.
Part of a broader AI drug discovery wave
The AstraZeneca-Tremont partnership joins a crowded field of similar deals announced through 2026, including Sanofi’s licensing of an AI-platform mRNA program from ScionxBio, Astellas’s roughly $400 million computational drug discovery partnership with AI Proteins, and Eli Lilly’s $2.75 billion collaboration with Insilico Medicine alongside a separate $1 billion investment in an AI drug discovery lab with Nvidia. Taken together, these deals suggest that nearly every major pharmaceutical company now views AI partnerships as a standard part of its drug discovery strategy rather than an experimental side project.
The evidence is still mixed
Despite the enthusiasm reflected in deal volume, AI-driven drug discovery has produced a mixed track record so far. Insilico’s rentosertib became the first generative-AI-discovered drug to enter Phase 3 trials, a genuine milestone, but Neumora’s lead AI-identified drug candidate failed two separate Phase 3 studies around the same period — a reminder that AI’s ability to identify promising candidates earlier in the pipeline does not guarantee those candidates will ultimately prove safe and effective in large human trials, which remain the gold standard test no algorithm can bypass.
What skeptics want to see
Industry skeptics argue that toxicology-focused AI partnerships like AstraZeneca’s deal with Tremont will only prove their value once a drug developed using this earlier AI screening process successfully completes the full clinical trial pipeline and reaches approval — a process that typically takes many years even under accelerated timelines. Until then, deal announcements represent promising bets on a technology with genuine scientific plausibility, rather than proven, validated outcomes.
What’s next
AstraZeneca and Tremont AI have not disclosed specific drug candidates or therapeutic areas the collaboration will initially target, nor a timeline for when the partnership’s AI toxicology models might influence an actual development decision. As with most AI drug discovery partnerships announced this year, the real test will come years down the line, when candidates screened using this technology either succeed or fail in human trials — the only outcome that will definitively validate whether earlier AI-based toxicity prediction actually improves overall drug development success rates.
Photo: Tima Miroshnichenko / PEXELS via Pexels
