By Deborah Borfitz 

June 23, 2021 | In the not-too-distant future, physicians and pharmaceutical companies could have a tool to predict how well a drug will perform in patients and optimize its therapeutic efficacy. The enabler will be “virtual clinical trials”—employing digital representations of patients and running in parallel with actual clinical trials—to understand what separates responders and nonresponders to a drug, according to Himanshu Kaul, D.Phil. (Oxon), research fellow in the Royal Academy of Engineering at the University of Leicester.

Kaul says he is a huge fan of mathematical models that are “agent-based” and therefore capture the heterogeneity of biological systems, including what is happening at the cell and organ levels that contribute to health and disease. In the long run, he is intent on building a cellular interaction model that factors in the omics, including genomics and transcriptomics as well as molecular perturbations, and scales across space and time.

The model’s initial output will be a prediction of average, rather than patient-specific, outcomes following a therapeutic intervention, says Kaul, who has already succeeded in building a virtual asthma patient based solely on cellular interactions (DOI: 10.1126/scitranslmed.aao6451). The “digital patient” here was an agent-based computational model capturing a trio of pathological features of the airway consistent with moderate to severe asthma and provided an explanation for how a DP₂ antagonist (fevipiprant, Novartis) succeeded in reducing smooth muscle mass in the phase 2 trial. 

It was a task that equation-based models would have been challenged to meet, Kaul says. Such models are typically used to understand biological mechanisms and, while successful, “kill out heterogeneity” because they tend to make use of system-level observables from which diseases emerge. Machine learning models can be highly effective but require more data than is often readily available. 

Interactions between epithelial, mesenchymal, and inflammatory cell types were selected to capture asthma pathogenesis via the virtual asthma patient, based on historical data in the literature, he says. The model was able to show “what parameters going wrong would lead to which of the asthma phenotypes.”

Notably, predictions by the digital patient about the clinical effects of fevipiprant were “very consistent” with findings of the drug in clinical trials, says Kaul. Researchers now hope to expand its predictive capabilities from generalized patient trends to individual patients over the course of a disease.

Building A Lung

Development of the digital asthma patient was an initiative of the European Commission-funded Airway Disease Predicting Outcomes through Patient Specific Computational Modelling (AirPROM) consortium that launched in 2011 with a focus on asthma and chronic obstructive pulmonary disease (COPD). Although funding technically ended in 2016, various strands of research that emerged from the consortium continue, says Kaul.

He is now leading a new research project, called “The Lung Pharmacome,” whose initial aim is to produce a prototype of a working in silico lung by 2026, he says. Patient-specific virtual clinical trials could begin as soon as a year thereafter.

The next step is to build the “omic interactions” into the cellular interaction model to capture patient-specific and, hence, inter-patient trends, says Kaul. In addition to colleagues at the University of Leicester’s schools of engineering and mathematics and department of respiratory sciences, he is working on the project with collaborators from the University of Auckland in New Zealand and Arizona State University, he says. 

An additional partnership with the Virtual Physiological Human Institute is dedicated to creating policy and regulation around this paradigm of computation-aided healthcare, Kaul says, which will “ultimately lead to the delivery of standardized, reproducible, and efficacious precision healthcare.” 

Consider the organ as hardware regulated by cells, the software, which are in turn governed by the internal program, he offers as an analogy. “Whether the program works optimally or malfunctions, the other two are affected.” The team will create mathematical models of the program interlinked with the software and hardware. 

Understanding where issues in one would lead to issues in the other could be the basis for developing therapies to “debug the code and improve the functioning of the system bottom up,” Kaul says. As envisioned, the digital patient would ultimately be useful to clinicians by simulating disease progression and helping them select the best intervention tailored to a patient’s profile—and the rationale for making that recommendation. 

In a study published last year in the European Respiratory Journal (DOI: 10.1183/13993003.00930-2019), Kaul and his colleagues used computational modeling to predict the impact of benralizumab (MedImmune, AstraZeneca) on airway smooth muscle mass in asthma, as had previously been demonstrated in clinical trials, he notes. Sponsor companies therefore might find the model to be a good companion to their current in vitro pharmacology program in deciding which drugs have a higher probability of succeeding when tested in humans. 

Currently, "more than 90% of drugs fail to reach the market,” says Kaul. “This is because we lack the capacity to predict the impact of drugs at the systems level, and this comes at a huge cost to pharmaceutical companies.” He contrasts this with practices in the aeronautical industry, where companies run a plane’s design through rigorous mathematical models to predict and improve its real-world performance. 

Here To There 

Cell interactions and hallmarks of asthma are well captured by the current iteration of the model, Kaul says. Many models of lung function have separately been developed to measure performance at the human level.

“We now need to integrate the software [cells] with the hardware [organ]… to understand the gene registry networks and pathological network motifs that show up when cells go rogue,” he says. “We might have genes interacting at different levels but also time, from seconds to minutes to months, and we will be… integrating [all of] those scales.” 

Airway organoids are being developed to validate findings from the enhanced model, says Kaul, adding that the team will work with others in the field using lung-on-a-chip airway devices. The process of aligning the in-silico models with the airway organoids will take at least three years, after which the predictive accuracy of the digital patient will be tested against findings of published clinical trials. 

Virtual clinical trials could begin in another five or six years, he says. The model will capture the “universal interactions” within the airways that are predictive of outcomes from multiple other lung diseases beyond asthma, including COPD and cystic fibrosis.