Virtual Controls in PDX Tumor Studies

Can we apply this non-animal, 3Rs-centric tool effectively in Discovery’s unregulated space? Our lab launched an experiment to find out with the help of a Digital Twins expert

Virtual control groups are to research what historical data from control groups is to replacing concurrent control groups in animal toxicity studies. 

OK, that’s a mouthful. But in plainer language, virtual controls, fed by historical data from years of studies, can make studies more efficient while helping reduce the use of animals. Consider my field of oncology. In the course of a year our laboratory site in Germany uses dozens of PDX mice—defined as models where human tissue or cells from a patient are implanted into an immunodeficient mouse or a humanized mouse—to vet hundreds of cancer drugs, mostly from within the booming field of immuno-oncology. Our PDX models are a vital pathway for clients in determining if their drug candidates are robust enough to move forward.

One of the inherent challenges with studies employing PDX models or other in vivo models is that they require a lot of animals. In our lab, for instance, we typically need 20-30 animals per study.  This is where virtual control models might help.

Recently our lab in partnership with Aitia, a Boston-based leader in the application of Causal AI and Digital Twins, began looking into whether virtual control groups could be applied in the early-stage unregulated space that our Discovery Oncology research occupies. The use of VCGs is hardly a new venture for Charles River. About a year ago, in partnership with several large pharma companies, our Safety Assessment teams began looking into replacing control animals with virtual controls in nonclinical safety studies. The pioneering work is leveraging years of legacy data and seeing how these “virtual” controls stand up to their “live” animal counterparts in toxicology studies.

In the Discovery Oncology realm, the scope of our work is a bit different. With the help of Aitia, we are hoping to probe our historical data for years of PDX studies to analyze how a tumor (breast, lung, colon etc.) changed over time, and potentially use this information to predict outcomes or interventions. Put another way, we are trying to develop “virtual” groups based on our PDX models by looking at the growth curves of those models. 

This is complex work, but here are a few things that have stood out thus far:

1. Data Annotation is key! Though we work in a nonregulated space, and therefore not subject to the same strict regulations, guidelines, and oversight as later-stage preclinical studies are, classifying data is still a key determining factor in well-run studies. The naming of the groups, naming of the vehicles, tumor ID, that all has to be harmonized to best analyze the data. This presented a dilemma for us because some of our data dates back 30 years, when data was, quite literally, entered by pen to paper. So, to overcome this challenge, a hardworking member of my team recently spent three to four months cleaning that data to ensure consistency. 
2. Genes do drive tumor growth! Certain extrinsic factors or mouse strains can impact a tumor’s growth, but at the end of the day what really drives it are the genetic mutations within the tumor itself. This was an important observation in our quest to develop a virtual control group because it meant we did not have to use the same vehicle, the same mouse strain, the exact same conditions to create a virtual control group for the exact same condition. In other words, no matter what control vehicle you are using you can always use the virtual animals and reduce animals in the process. Still …..  
3.  … A third-party vetting is required for that final seal of VCG approval. On the bioinformatics side we did the heavy lifting to make sure the data sets were clean, but once we finished our study, we needed an outside statistician to confirm our data was correct. That’s where Aitia came in. They are doing the work now, running the data again to confirm our approach was solid and that we did not make any mistakes. 
4. VCGs’ wider reach in Discovery. While our project was confined to the oncology space, it should theoretically be easy to transfer this principle to the central nervous system arena, where you typically have a similar set up per model that is run over and over again. Same for other in vivo models within Discovery.

So, what will the future be for VCGs in our oncology studies? We are awaiting results from Aitia to determine exactly how many years of data are needed to build a reliable virtual control. While we wait we are excited for the possibilities of employing a new tool the doesn't require the need for live animal controls. The fact that it appears capable of providing consistently good results in the Safety space is reassuring. Given that an estimated 20-30 animals are required per PDX drug study, VCGs could be a huge win for research and for animals.