Why AI’s real value in drug development lies in the unglamorous work
Steven Bulera, Corporate Vice President, Chief Scientific Officer: Safety Assessment, Charles River Laboratories, explores various use-cases of AI in improving drug discovery.
It’s difficult, if not impossible, to escape the headlines and reports detailing the extent to which the biomedical research industry is investing in artificial intelligence (AI). A November 2025 update by McKinsey & Company (1) suggests that, in the pharma industry alone, the AI market is projected to grow from more than $4 billion this year to $ 25.7 billion by 2030.
Most of the industry’s largest companies are well beyond the days of referring to AI as an aspirational goal and actively deploying it across their companies. Eli Lilly, for instance, teamed up this year with Chai Discovery to design novel biologics and last year launched TuneLab, a collaborative project with smaller biotechs that it believes will better predict drug outcomes. Meanwhile, AstraZeneca has developed its own open-source generative AI framework, dubbed Reinvent, that it says halves the time needed to identify molecular structures that could become new drug candidates.
Still, while the rapid development and wide application of AI and big data technologies are profoundly changing how biomedical research companies operate with the tantalising possibility that it might be possible one day to dramatically shrink the long timetables of 10+ years it now takes to get a drug to market, the reality is that the industry is still years away from making the AI promise a reality.
In their eagerness to jump on the fast-moving AI train and capitalise on its popularity, companies are in danger of overstating AI’s current capabilities, at least where their programmes are concerned. This is particularly true in early drug discovery, where some of the biggest AI headlines are unfolding. From digital twins to quantum computing to the various iterations of Google’s AlphaFold, the industry is awash with amazing ideas and approaches that could bring about a paradigm shift in how quickly and effectively we identify tomorrow’s drugs. The unfortunate truth is that rushing immature technology into clinical settings risks patient safety and could set the entire industry back for years. This race to the finish line could end up damaging credibility and adoption when failed implementations undermine trust in AI across the sector. Similar to teenager Jesse Gelsinger (2), whose death from a massive immune reaction to a recombinant adenoviral vector set the gene therapy field back a decade and ushered in a “dark age” for gene therapy research, racing a drug to market on the weight of an AI-generated candidate could be detrimental to the entire industry if the technology isn’t fully validated. In other words, overhyping AI produces unrealistic expectations that can be hard to walk back in this competitive environment.
The real value of AI today lies in the more downstream work: optimising study design, reducing animal testing, and automating documentation and evaluations for human review. These are the immediate, tangible efficiencies where AI delivers measurable value. Instead of chasing flashy breakthroughs, companies should focus on securing these foundational wins; the slow-and-steady approach that’s both safer and smarter.
How AI could make toxicology studies more efficient, right now
Thousands of safety studies are conducted each year to determine whether a drug meets the criteria for human testing. The assessment of risk, side effects, and adverse events are specifically defined by the length of exposure to a treatment or the time elapsed during a study. The time frames can’t be modified. You can’t wave a magic AI wand and compress the duration of the study, because laboratories need a defined period of time to learn how animals react to drugs. Whether you are conducting a six-month study or 12-month study, it generally needs to last as long as the human clinical trial.
But AI can be leveraged to help laboratories get studies started more quickly and report data faster at the completion of the study. Though still in its infancy, companies, including ours, are devising ways to use AI and other machine learning technologies to generate study data plans faster and to charge AI models with helping them use resources more efficiently. AI can help reduce much of the duplication inherent in a system that has relied on humans for decades. At the conclusion of the study, AI is also a potentially useful tool for taking a first pass at analysing study data. A study director would have to review the findings for accuracy, of course, but the work done by an AI model could ultimately save a lot of time.
It is also possible to use AI to assess study endpoints, the specific, pre-defined measurements, outcomes, or events that are used to objectively determine the efficacy and safety of a study. AI could be used to mine the data and determine which endpoints are value-adds and which have limited or zero value in predicting a drug’s safety. This could go a long way toward improving study designs.
Most drug development programmes require two species to be used. AI could be used to mine data to determine if both species are necessary. If the data showed that one species could provide the necessary safety data to protect patient safety, the use of a second species may be avoided. This would not only save time and money, but would also reduce animal usage in the development of new therapies.
AI can also modernise the arduous and manual process of study preparation by optimising lab schedules. AI can help, for instance, streamline which animal studies should be conducted in which rooms and which studies should be prioritised; in doing so it can help generate protocols faster for a single drug. For large laboratories with thousands of animal rooms, having an AI model around to manage certain operations would be phenomenally helpful. A well-designed AI system could also greatly reduce the data entry required for a single study plan, accelerating data generation and enabling a study to start sooner.
It goes without saying that AI models are not foolproof. Like humans, they make mistakes and sometimes really big ones. There are well-documented examples of AI models generating false or misleading information and presenting it as fact; what the industry refers to as “hallucinations”. These AI models can erode trust in the research and the labs conducting it. For AI to become a reliable scientific partner, it will be important to address these glitches early on before damage is done. AI tools must be held to higher standards of factual accuracy and source traceability, and be trained on high-quality, scientifically relevant data. Wider AI adoption will ultimately depend on making tools more accessible to non-specialist users.
AI-powered digital pathology
Tissue analysis is another major area being accelerated by AI. In nonclinical drug development, pathologists have long been the primary interpreters and identifiers of adverse toxicities, but AI-powered image analysis is meaningfully expanding what’s possible when a programme needs more from its tissue data than a traditional histopathology evaluation can provide. When a programme needs to measure subtle morphometric changes across an entire tissue section, AI-powered image analysis can count specific cell populations with precision, generate data points suitable for standard statistical tests, and deliver results that visual scoring alone cannot. Whole-slide imaging (WSI), a foundational component of AI-driven technology, is a growing player in drug development today. Pathologists at Charles River, for instance, have evaluated more than 300 toxicology studies using WSI, and 45 New Drug Application submissions have included WSI-generated pathology data.
On the clinical side, AI‑enabled biomarkers are supporting earlier patient stratification, reducing biological and operational noise, and enabling more accurate response prediction across oncology trials, particularly those testing immuno-oncology drugs; and AI-enabled tissue quantification can overcome issues with tissue variability by reducing subjectivity, delivering more reproducible endpoints, and generating cleaner data earlier in drug development.
Virtual controls in animal toxicology studies
AI is also behind a new method that will significantly reduce animal use and improve experimental design. Virtual control groups (VCGs) leverage historical control data and machine learning algorithms to replicate the statistical power of traditional control groups and address preanalytical and analytical variations. A growing number of companies and contract research laboratories have been experimenting with and pioneering the use of VCGs across nonclinical toxicology. This year, Charles River published a re-analysis (3) of 20 studies using VCGs that compared findings with concurrent control groups. The study found no overall differences in the observed adverse or NOAEL, defined as the highest tested dose that causes no adverse effects, compared to the original study that used live animals.
The science behind VCGs is also gaining support from regulatory agencies. For instance, in March, a European Medicines Agency committee issued a draft qualification for VCGs, a crucial first step towards reducing animal testing for medicines. The EMA’s actions have been bolstered by the EU-funded VICT3R project, which includes nearly 40 organisations across 12 countries, and contributes valuable data and decades of experience toward the development of VCGs.
Don’t discount those AI-related “shiny objects”
Though a cautionary approach makes sense in the short-term with AI-powered research, it doesn’t mean that more complex AI models capable of rapidly identifying and vetting tomorrow’s blockbuster drugs is off the table. There is a tremendous amount of promising research using AI-powered tools to computationally identify and design millions of potential compounds, and predicting the ones destined for success. We are a long way off from turning the job once held by chemists over to computers. So, let’s celebrate the very real results that AI is delivering right now, and make sure the hype doesn’t overtake science.
References
2: Sibbald B. Death but one unintended consequence of gene-therapy trial. CMAJ. 2001 May 29;164(11):1612.
3: Lotfi L, O’Neill TW, Wilcox A, Bulera S. Virtual control groups in nonclinical research: A Re-analysis of 20 studies using virtual control group data. Regul Toxicol Pharmacol. 2026 Jun;168:106077.
From DDW Volume 27 – Issue 3, Summer 2026 – Read the digital issue here