Virtual Control Groups: The Next Chapter

From hybrid models to full validation, a look at where the field is at with this New Approach Methodology

It was just five years ago that laboratories, including Charles River, began exploring whether virtual control groups in nonclinical toxicology settings were a feasible strategy for reducing the use of research animals. The move was not without precedent. VCGs have been well studied in clinical trial settings to improve efficiency, reduce costs, and enhance ethical standards by replacing or supplementing traditional, concurrent control groups (CCGs) with historical data.

But in the nonclinical realm, where animals have for decades played a critical role in ensuring that drugs are both safe and effective before being administered to people, the adoption of VCGs is another thing entirely. They need to be supported by pristine historical control data collected and curated over many years. They depend on large, well-structured, and harmonized historical control datasets. And they need reliable biomarkers that provide objective, quantitative, and reproducible metrics, enabling researchers to accurately compare treatment effects against historical benchmarks without compromising study safety or sensitivity.

So, where do we stand on the journey toward replacing animal controls? Is the data compiled so far strong enough to support the use of VCGs commercially, particularly in the heavily regulated GLP space? And are drug developers on board with using VCGs in their studies? Here is what we know so far:

• Hybrid models appear to be a bridge to the future. While replacing animals with VCGs is a reasonable long-term goal, the industry favors studies that combine CCGs and VCGs for study interpretation. Building a perfect VCG to replace animal controls may not be possible at this time. Adopting hybrid models can also help alleviate the looming problem of data drift that would inevitably arise if the industry went full VCG and then found itself short of usable data.
• VCG hybrids are inching toward the market: While VCGs remain an active area of exploration by contract research organizations and large pharma,  hybrid models are being offered commercially in limited situations.
• Fully-validated VCG models are necessary for regulatory acceptance: As with any software, the data that feed a VCG model must undergo rigorous validation so they can be used widely with confidence, particularly in more regulated areas of toxicology.
• The scope of VCGs will eventually expand: While most research on VCGs has been confined to general toxicology studies, virtual controls might also be useful in more specialized safety studies, such as the two-year carcinogenicity study, and expand beyond the handful of large and small animal species now being targeted.

A retrospective study published this year in Regulatory Toxicology and Pathology by scientists from Charles River, a pioneer in the drive to move VCGs into prime time, provides some of the clearest guidance yet on how to leverage VCGs effectively using minimal selection criteria. But the study rightly points out that for VCGs to truly transform how we conduct routine GLP-tox studies, input from a much wider tent will be needed. Regulatory authorities, specialty organizations relating to anatomic pathology, clinical pathology, and toxicology, as well as academia and the pharmaceutical industry, will be required to determine the appropriate use of VCGs in the future, the study notes.

Most importantly, perhaps, the paper highlighted proof-of-concept findings dating back to the launch of the VCG project at Charles River. Researchers applied VCG data from multiple animal species retrospectively to 20 pilot studies and compared them with CCGs to assess statistical alignment and biological relevance. They found there were no changes in the no observed adverse effect level or NOAEL—defined as the highest tested dose or concentration of a substance that causes no significant adverse effects on test subjects—compared to the CCG animals. Body weight, body weight gain, food consumption (in rats only), and clinical observation conclusions were also unchanged. However, there were differences in interpretation between clinical pathology and anatomic pathology findings.

Because test article-related effects in clinical and anatomic pathology affect clinical trial development, monitoring plans, and the interpretation of unexpected changes, the slight variability they observed between VCGs and CCGs highlighted the importance of including additional experimental criteria when evaluating clinical pathology parameters, the study noted.

Laura Lotfi, the lead author of the study, said the differences in pathology were not entirely unexpected. “We knew there would likely be threshold differences, especially on the pathology side.

However, the VCGs just make it much more visible because you are relying on previously collected datasets,” said Lotfi. “So that was a very big finding, and in our conclusions, we recommended that host site images or the pathology slides be available for pathologists using the virtual control group animal as any other endpoint of that database.”

Looking ahead, Lotfi says how we use VCGs will always come down to clear selection criteria. “How do we best use VCGs? What are the selection criteria that we really need to consider to be non-negotiable, absolutely critical, versus things that the data science shows us are really not that important?”

Lotfi said they have conducted seven other retrospective analyses, in collaboration with different clients that use refined sets of selection criteria; they are eyeing one for publication later this year or early next year. They are also actively working with 15 sponsors on VCGs, including three clients that are close to implementing virtual control groups in a dual-design format, loosely defined as the assessment of VCGs and full control groups in parallel over a reasonable time frame to confirm there are no differences in outcomes.

All this work is going a long way toward filling the gaps that, for now, prevent full replacement of animal controls with VCGs. With the value of VCGs growing by the day, Lotfi is often asked why the industry isn’t prepared to make the full switch to VCGs now, and her answer goes something like this. “While, I do think that we understand a lot more than what we did back in 2022, on the operational side there’s still some unknowns in the process; we need to ensure the right questions are being asked, such as is it feasible to do the virtual control group on a specific study, what type of pathway is it, and what type of molecule is it. Is it a first-in-human enabling study, and if so, what are the risks?  There is a clear risk assessment and a decision tree that we are building right now to make sure that the right VCG is available from a context of use and selection attribute perspective.”

The encouraging news is that VCGs are starting to find their way into commercial drug development in certain capacities. A few of the Sponsors Charles River works with are already using a hybrid model, and Charles River is eyeing a soft launch the second half of this year of their hybrid VCG model.

With the help of an internal IT team, Lotfi says Charles River is also hoping to validate its VCG model for use in a GLP setting, which will go a long way toward building trust with regulatory authorities. “This is a humongous effort that needs to happen. While I  know from a Study Director's perspective that we can use non-validated methods on a GLP study, I would only want to do that under very select circumstances because the second you have an exception to the protocol, the FDA is going to ask a lot of questions.”