When Less (Patients) is More - Digital Twins for More Efficient Trials

In James Cameron's "Avatar," humans pilot synthetic Na'vi bodies through neural interfaces, creating living proxies that traverse the alien world of Pandora. While such a technology, for the most part, is still a futurist’s dream, it mirrors a fast growing field of digital tools for use in R&D and healthcare - digital twins. 

Virtual Proxies for Real-Life Patients 

Instead of consciousness flowing into alien bodies, we are now capable of mapping human biology into complex digital models - creating virtual versions of patients and their diseases. These in-silico twins have potential beyond roaming jungles, returning information about realistic responses to clinical treatments or simply the progression of disease through accurate computational predictions. This can allow us to simulate disease progression, test treatments, flag patients at early risk of adverse effects and glimpse at future states in clinical trials without risking actual patient outcomes. Their use-case in the clinical space can be considered across two frontiers: 

1. while not universally considered equivalent to digital twins, synthetic control arms represent a significant focus for computational modeling companies - these approaches aim to augment or replace trial control arms with digital equivalents, generating predictive data based on historical or patient level input

2. a future state in which digital twins are deployed to augment active arms of clinical trials, in which the AI models have been trained sufficiently to extrapolate responses for novel drug candidates and/or targets. Moreover, patient-level models can also continuously be updated with real-life progress or clinical events. Much like Jake Sully's avatar enables him to experience a world otherwise inaccessible, these digital twins open windows into biological systems that would be impossible or unethical to assess in human subjects.

A Technological Response to Eroom's Law

The use of digital twins in clinical trials emerges against the backdrop of an unfortunate phenomenon in pharmaceutical R&D. Moore’s law - a rule of thumb whereby the cost of processing power falls by half roughly every two years - inspired the formulation of Eroom’s law - a parody converse of Moore’s Law (literally “Moore” spelled backwards), which characterises the declining efficiency of pharmaceutical R&D efforts, with exponentially increasing investment required for the commercialisation of a new therapeutic (Fig. 1). Digital twins represent a direct technological retort to Eroom’s law, offering an opportunity to tackle one of the biggest drains in the R&D pipeline using AI.

Figure 1. The pharmaceutical ‘productivity gap’, demonstrating the continuous increase in R&D expenditure with a stagnant level of novel drug introductions (Fernald et al., 2024)
https://doi.org/10.1016/j.drudis.2024.104160  

The scale of the challenge encountered by pharmaceutical companies, CROs and sponsors in the patient recruitment and retention space is significant. 80% of clinical trials fail to recruit enough patients as set in trial protocols or contractual agreements with third parties, with  37% of individual clinical trial sites failing to recruit enough patients. 11% of all clinical trial sites fail to recruit patients at all. Given that the clinical phase of drug development represents 30% of total development time, corresponding to $1.2B in costs, the fundamental value proposition for digital twin vendors seems to create itself. 

Multi-faceted Benefits 

The potential benefits of implementing digital twins in clinical development stretch beyond strictly fiscal considerations. As healthcare transitions to a more value-based care model in which patients increasingly demand meaningful participation in clinical trials, the use of digital twins offers a way to valorise patients' time and contributions. Firstly, computational modelling is an evidence-based method of optimising trial design, prospectively allowing clinical development teams to test protocol hypotheses and gain insight into patients before the first dose has been administered. This approach can both quantify the likelihood of protocol success and eliminate costly amendments. Secondly, there are many rare disease areas in which the use of digital twins can help to ‘widen’ constrained populations when it comes to recruitment. Real patients in search of improved therapeutics for their condition are averse to the risk of being placed in a placebo-arm or receiving the standard of care. However, synthetic control arms powered by digital twins create an ethical alternative that protects patient interests, allowing an individual to be recruited to the active arm whilst their Na’vi (an individualised model trained both on specifically that patient’s multimodal clinical information and historical trial and observational studies, if you’re being pedantic) coexists in the control arm. 

Technical and Ethical Considerations

Innovators who are bullish on the use of digital twins should be aware of some technical and ethical challenges with the technology. 

Firstly, data access and quality is one of the most important elements to address. It has been estimated that in 2025, the compound annual growth rate (CAGR) of healthcare data will be 36%. The proliferation of such data has, in theory, the potential to feed the development of tools that utilise that data, however, accessing that data is complicated - institutional data silos, stringent patient privacy protections and governance/sharing mechanisms all restrict utilisation. The challenge of data quality is also present across the various different kinds of clinical data on which companies will seek to train their AI models. Multi-modal data consisting of -omics, imaging modalities, and Real World Data (RWD) are increasingly valuable as training datasets. However, the interoperability of said datasets, i.e., the ability to integrate disparate data sources across different formats and collection methods, remains a significant technological hurdle. Harmonised formats such as OMOP serve as a bridge  for organisations to share datasets. Devising a digital twin AI architecture that can capture the underlying pathways and disease processes coherently is another technological milestone developers must first meet. 

Secondly, a fundamental challenge of modelling human biology gets left out of the equation when representing pharmacology at an individual or population level. AI models are defined according to the subset of measurable biomarkers and physiological parameters available to us. This process is by its very nature imperfect, and it must be acknowledged that it performs an oversimplification of complex biological behaviours. As a result, while AI can enhance our understanding and predictions, its limitations must be carefully considered on a case-by-case basis to ensure responsible and effective application.

The ultimate test for using a reductive set of biological features in-place of real patients should be one based on accuracy and clinical relevance. Indeed, this seems a measure some companies in the space have achieved to some extent already, with QuantHealth reporting 86% accuracy on the binary endpoint using their trial simulation platform. Predictive bias, and the potential for digital twins to proliferate existing inequities in historical trial data, should also be of high concern to healthcare leaders looking to utilise the technology. 

Regulatory Perspective with UNLEARN 

Although there is no specific digital twin guidance issued by FDA and EMA, both agencies have shown the appetite to assess their use through established regulatory gateways such as the FDA’s Medical Device Development Tool’s (MDDT) program and technical deep dives as part of the EMAs Artificial intelligence Workplan. There are several key players in the space which offer digital twin capabilities for synthetic control arms -  aiming to maintain statistical power whilst alleviating the demand for patient participation. Companies such as Nova In Silico, UNLEARN, and InSilicoTrials all develop digital twins to enhance trial efficiency. From a regulatory point of view, UNLEARN has reached a significant milestone toward regulatory acceptance through their TwinRCT™ solution, used primarily across neurodegenerative disease, immunology and others, as it is the only digital twin solution currently on the market that has received qualification status by the EMA, and also aligns with FDA guidance. 

Final Thoughts 

If we take digital twins to be a genuine shift in the way we define patient participation, evidence generation, and clinical validation, then we have a lot to be hopeful for. They may not just contribute to an inflection point in Eroom’s law as an incremental improvement in R&D, but as a rewriting of the clinical development process. Like most AI technologies however, the only way for stakeholders not to become jaded is to show real-world value front and center: incremental improvements; real demonstrations of accuracy, validation and ethical consideration; and regulatory assessment commensurate with the level of risk.