The majority of lead compounds fail before they become valid candidates for preclinical testing.
Of the candidates that do make it to the preclinical stage, most will never reach clinical trials.
And of those that enter human studies, only one in ten makes it through all stages of clinical development.
While AI may seem like a new concept to many, it is far from novel in the world of drug discovery and development. Reportedly, protein structure prediction algorithms were introduced in the 1960s, but it was AI breakthroughs in the 2010s, such as deep learning and neural networks, that helped to propel the field forward.
Today, the pharmaceutical industry is at an inflection point. Where AI was once considered hype, it has now delivered vaccines and drugs to clinical trials.
This article reviews key lessons learned from a series of expert interviews exploring the role of AI across the drug development pipeline, covering drug discovery, virtual control groups, preclinical study conduct, and clinical trials.
From trial and error to a smarter search for medicines
Before development comes discovery. Traditionally, this follows one of two routes:
- Target-based drug discovery: Biological targets, including genes, proteins, and other molecules implicated in disease pathophysiology are identified. Researchers then confirm the target of interest and screen for compounds that alter its activity by modulating disease directly or by producing effects expected to provide symptomatic benefit.
- Phenotypic screening: Cells, isolated tissues, or animal models are exposed to compounds to identify whether they produce a beneficial change in phenotype. Promising compounds can then be investigated further to understand their mechanism of action, which may include identifying the target responsible for the observed effect (target validation).
Regardless of the approach taken, drug discovery requires significant trial-and-error experimentation that is labor-intensive, costly, and time-consuming. Once a compound has been identified and established as a lead (hit-to-lead), it will be further optimized (lead optimization) and taken forward into preclinical studies.
Dr. Jean-Philippe Vert, co-founder and chief executive officer at Bioptimus, shared his insights on how AI is accelerating drug discovery. “The most impactful effect of AI so far has been to reduce the time required,” he said.
“It’s not as if we couldn’t do these things before, but companies say it used to take five to seven years—now it’s less than two.” — Dr. Jean-Philippe Vert.
Vert attributed much of this efficiency gain to protein‒protein interaction predictions and target-based molecular design.
“In terms of the techniques used, it’s quite similar to those in large language models (LLMs),” he explained. “For example, a few years ago, AlphaFold (an LLM for proteins) was developed by DeepMind to predict the structure of proteins from their sequence. Now we can give it a sequence, and it can predict protein structure and interactions, and even design new proteins with target structures.”
AI in drug discovery
Some AI drug-design approaches combine LLMs with diffusion models. An LLM is used to analyze and interpret information about a drug target and translate it into instructions (i.e., physical or chemical constraints or requirements) for a candidate molecule. These instructions are given to the diffusion model, which can be thought of as a sculptor. Starting with a lump of clay, the model carves a 3D structure that is designed to meet the given requirements, such as fitting within the target’s binding pocket.
AI-developed drugs are beginning to reach more advanced stages of development. For example, Rentosertib—a first-in-class AI-discovered drug—is moving toward Phase 3 trials. Rentosertib was developed for idiopathic pulmonary fibrosis (IPF), a progressive, irreversible lung disease. Both the drug and its target were identified using AI approaches, based on multiomics data, clinical records, and scientific literature.
Moving from initial discovery to candidate selection into preclinical work can take several years, but in this case, it took just 18 months.
Existing IPF treatments can slow progression, but developers propose that Rentosertib could halt or reverse disease processes. Peer-reviewed studies have yet to demonstrate this, though the drug improved lung function and reduced IPF-related markers in a Phase 2a study.
While the discovery process speeds up, Vert cautioned that it doesn’t reduce the robust testing required before a candidate can progress: “The computational aspect is faster, but the rest largely stays the same because we still need to confront the AI predictions in the real world.”
Preclinical work is grounded in proof, not prediction
The start of preclinical work marks the transition from concept to evaluation, during which the ground truth is established. Researchers must define objectives, select appropriate disease models, evaluate pharmacological activity, and ensure studies comply with ethical and regulatory requirements.
Given this, insights from Dr. Steven Bulera, DABT, chief scientific officer of Discovery and Safety Assessment at Charles River Laboratories, seem logical. As the person responsible for the scientific conduct of preclinical studies, Bulera feels that AI can deliver operational gains but is not set to replace traditional preclinical testing.
“At last count, we have 15 in vivo sites worldwide,” said Bulera. “From an administrative perspective, we are evaluating the use of AI to schedule and make the best use of our resources across these sites.”
“We are also looking at AI for reporting, including production and interpretation, but this aspect is very much in early development,” he continued. “Subject matter experts still need to review the material and make sure the data is interpreted appropriately.”
AI has demonstrated potential beyond experimental design and administration, in areas including model selection and optimization, digital pathology, and biomarker discovery and analysis.
For example, histopathology requires visual interpretation by pathologists, with both inter-observer (between different pathologists) and intra-observer (evaluation of tissues by the same pathologist for different purposes) variability. But AI has the potential to transform images into quantitative data with greater result consistency than human interpretation.
However, Bulera highlighted data gaps as one of the factors limiting the potential of AI on the scientific side: “The data that’s being fed into systems is limited, whether it’s just that we don’t know enough about mechanisms and signal transduction cascades, or whether it’s pharmaceutical companies that are only feeding data into their own systems.”
“On that side, nobody is pulling all of the data together. How can they, without losing their competitive advantage?” — Dr. Steven Bulera.
Animal studies vs AI algorithms
Laura Lotfi, director of digital product management, scientific innovation, and global technology solutions at Charles River, echoed this sentiment when reflecting on the potential of virtual control groups (VCGs) in preclinical trials.
“Currently, the most significant limitation of VCGs is the lack of data and the way it’s standardized,” she said. “Is the data of good quality? Is it fit for the purpose of being used as a replacement?”
As part of the digital innovation team, Lotfi is heading up the investigation and implementation of VCGs in preclinical trials.
VCGs in preclinical studies
A VCG aims to replace the traditional in vivo control arm. With AI, historical control data are used to curate a dataset that matches key study parameters, such as subject characteristics, disease severity, and outcome measures. After the trial, comparative analyses are performed between the treatment and VCG groups.
Lotfi outlined the limitations of traditional control groups that have prompted this shift: “They consume significant financial and biological resources and raise important ethical questions, particularly when conducting nonclinical experiments. In that type of experiment, biological complexity in humans is not always reflected in controls. Therefore, there is an inherent limitation between the controls and the real-world questions you’re trying to address.”
“Without an appropriate control, a hypothesis remains a hypothesis. It cannot be addressed, as it cannot be untethered from unreliable evidence.” — Laura Lotfi.
As well as a push away from traditional methods, there has also been a pull toward VCGs. “VCGs represent one of the most promising tools that I’ve seen so far that honors these principles [the 3Rs: replacement, reduction, and refinement] in an evidence-based way,” noted Lotfi. “In practical terms, they significantly reduce the number of animals assigned to a study.”
Lotfi explained that VCGs are currently used alongside in vivo data; they do not yet replace a control group. “It strengthens your conclusions of a safety assessment,” she added, aligning with Bulera’s view that AI has a place in preclinical work, but cannot replace traditional approaches.
Both experts agreed that regulatory guidance has been limited. “While the US Food and Drug Administration (FDA) and European Medicines Agency, have signaled openness in terms of new approach methodologies, they’re not yet globally harmonized,” said Lotfi.
Bulera added: “I can list an entire alphabet of regulatory agencies, each of which may have different opinions.”
“Regulatory agencies don’t talk to each other very well; they may do similar things, but there’s no global standard. It’s kind of the wild west.” — Dr. Steve Bulera.
He noted that this is where contract research organizations can step in: “We are a neutral party, and we work with everybody, so we can be an impartial gatherer of information.”
“From our findings, we can go to regulators, saying, ‘Look, we have the support of these global pharmaceutical companies, here’s what we’re thinking. What do you think?’ And work from their feedback,” Bulera continued.
Vert concluded with one final benefit of utilizing AI within preclinical studies: “The field is trying to find faster ways to iterate, rather than waiting the 10 years it takes for a clinical trial to succeed or fail … We have learned that this can be done during the preclinical period, by feeding the LLMs data collected in cell cultures, organoids, animal models, and so on.”
With this, AI itself could help to overcome one of its key limitations—the data gap.
Separating signals from noise in clinical trials
It is evident that AI is being implemented to varying degrees across pharmaceutical development, from target identification through to Investigational New Drug application. This is the final regulatory step before clinical trials can commence, and the utilization of AI doesn’t stop here.
In fact, the use of AI in clinical trials is critical to addressing the data gaps highlighted by experts; access to and analysis of trial results enables the system to feed predictors of success and failure back to the earliest stages of development.
“This helps improve decision-making earlier in the process and enables more accurate AI predictions,” said Vert. “When we send a drug to clinical trials, even with the best minds and biggest budget, we still have a 90% failure rate … Biology is a complex domain, and when you have a lot of complex data, AI is good at making connections.”
Dr. Dominique Demolle, co-founder and chief executive officer of Cognivia, is passionate about identifying signals that are often mistaken for “noise.” In particular, she highlighted that patient-related variability is an untapped source of information.
“While biological variability has long been studied and increasingly controlled through stratification and biomarkers, behavioral variability remains significantly under-characterized,” said Demolle.
In a clinical trial, the control group mirrors the intervention group in terms of demographics and disease severity, but are they matched in motivation, adherence, and beliefs? When pharmacology is not separated from patient experience, researchers risk missing key insights: did the patient truly tolerate the medication well, or did a lack of adherence mean that side effects never came to light?
The dynamic, subjective nature of human behavior has historically made this type of variability difficult to measure.
Demolle stated that this is where AI can step in, creating three main opportunities:
- AI can identify and quantify sources of variability that would not typically be modeled.
- It can predict patient behavior, supporting proactive risk management, e.g., mitigating the risk of disengagement.
- Data interpretation could become more comprehensive, as AI “disentangles signals from noise.”
Vert agreed: “While AI has already revolutionized target discovery to preclinical work, the next phase is really, how can AI de-risk, optimize, and accelerate clinical trials?”
As with preclinical studies, he highlighted the opportunity for AI iteration during clinical trials.
Despite the potential, Demolle noted that there are limitations that could impact uptake.
“In practice, adoption may remain a barrier. AI-driven approaches typically require changes to established practices, and change is often perceived as carrying additional risk.” — Dr. Dominique Demolle.
Without clearly defined use cases, validation, and operational integration, opportunities will likely remain unrealized.
Although, in the world of clinical trials, regulators are supportive of AI-driven approaches. “A key reference point is the draft FDA guidance on the use of AI to support regulatory decision-making,” said Demolle. “The guidance outlines how AI-generated evidence can support regulatory decisions related to safety, efficacy, and quality.
“Models must be explainable, appropriately validated, and integrated into the overall evidentiary package in a way that can be reviewed and understood.” — Dr. Dominique Demolle.
“Overall, the direction is clear: regulators are not only open to AI but are actively defining the frameworks that will enable its safe and credible adoption in clinical development,” she concluded.
The future of drug development is AI-assisted, not AI-led
AI is already demonstrating its potential to improve pharmaceutical development. Experts are actively implementing AI throughout discovery, preclinical studies, and clinical trials, with tangible benefits. But they are doing so cautiously.
Lotfi summarized a key concern, one likely shared by AI users across industries: “What are the limitations that are not yet known?”
There are still plenty of questions around AI, but the evidence so far suggests it is more likely to complement traditional drug development than replace it entirely.
