(T1530-05-32) A Novel Approach to Preclinical Drug Development Using Digital Twins of Ex Vivo Human Lungs

Purpose: Current preclinical drug evaluation methods face several challenges, including the inability of animal and tissue organoid models to fully capture the dynamic complexity of the human body, leading to decreased translatability and an increase in R&D costs and timelines. Ex vivo lung perfusion (EVLP) was developed to assess isolated human lungs under physiological conditions (Fig. 1) (Cypel, 2008; Cypel, 2011). As such, EVLP has advanced preclinical research by enabling therapeutic testing directly in isolated human organs. EVLP has proven to be an effective platform for lung repair and therapeutic testing—such as gene therapy (Mesaki, 2023), stem cell therapy (Nakajima, 2019), creation of universal blood type organs (Wang, 2022), and immunomodulation (Nykänen, 2024); however, the conventional randomized two-arm study design presents several challenges, including limited availability of organs, which prolongs enrollment and increases costs, and inter-subject variability, which complicates result interpretation. Digital twins—computer simulations of physical objects—are an emerging concept in medicine, yet their potential in drug discovery remains unexplored. EVLP generates real-time, multi-modal data on lung function, offering a unique opportunity to train machine learning (ML) models to create digital twins of human lungs. Herein, we developed an ML-based approach that leverages >300 ML algorithms that work simultaneously to accurately simulate future lung function. As a proof of concept, we demonstrated the clinical utility of digital twins of ex vivo lungs for the preclinical evaluation of the safety and efficacy of an EVLP-targeted therapy.

Methods: Human lung physiology, biochemistry, proteomic and metabolomic biomarkers, transcriptomics, and medical imaging features were derived from n=1,000 EVLP clinical cases performed at the Toronto General Hospital (2008-2024). Multi-modal forecasting models were trained with k-fold cross-validation using a deep learning-based model called gated recurrent unit and a tree-based XGBoost model to forecast lung functional parameters at future time points with mean absolute error (MAE) and mean absolute percentage error (MAPE) as model evaluation metrics. To develop digital twins of human lungs, we deployed ‘static’ and ‘dynamic’ modeling strategies. The static digital twin takes single static baseline lung function data as input to forecast future lung function. In contrast, the dynamic digital twin gets recalibrated with additional functional data and forecasts the immediate next time point using the most recent lung function data. As a proof of concept to validate the clinical utility of digital lungs, we studied a cohort of 16 lungs that had been treated with tissue-type plasminogen activator (tPA), a thrombolytic therapy. Each lung was evaluated using the corresponding digital twin to assess therapeutic efficacy—measured by pulmonary arterial pressure (PAP)—and safety—measured by edema—on a personalized, case-by-case basis (Fig. 2a).

Results: A digital twin of an ex vivo human lung was successfully developed, accurately predicting lung function across more than 75 distinct functional parameters measured during EVLP. The digital twin approach achieved >90% accuracy in both short- and long-term forecasting of lung physiology, biochemistry, metabolites, and transcriptomics (Table 1). Moreover, we found that using a digital twin approach for preclinical therapeutic evaluation outperformed conventional preclinical study designs. Figure 2b shows a representative EVLP case treated with tPA, displaying a reduction in PAP compared to its corresponding digital twin, while no significant increase in edema was observed. Since digital twins successfully predicted the counterfactual outcome of the applied therapy, paired statistical analysis between the untreated digital twin and the corresponding treated lung was made possible. Next, in a cohort of n=16 human lungs treated with tPA, the digital twin paired analyses approach revealed a significant reduction in PAP within two hours of treatment (p-value = 0.031). In contrast, using a conventional two-arm study design, no significant differences in PAP were observed at 1- or 2-hours post-treatment. Thus, only a digital twin approach was able to identify preclinical treatment-induced reductions in PAP.<

Conclusion: Our digital twin approach creates a virtual model of human lungs, based on the world’s largest dataset of clinical ex vivo lung perfusion procedures. The digital twins in this study accurately predicted a comprehensive range of lung physiology parameters, as well as biochemical, transcriptomic, metabolomic, and proteomic biomarkers. Most notably, using real-world data, we validated the usefulness of digital twin models for preclinical therapeutic evaluation and showed their ability to reveal new insights into treatment effects. Digital twins allow direct comparison between human organs and their virtual counterparts, supporting advances in precision therapeutics and accelerating preclinical drug discovery.<

References: Cypel, M. et al. Technique for Prolonged Normothermic Ex Vivo Lung Perfusion. The Journal of Heart and Lung Transplantation 27, 1319–1325 (2008).
Cypel, M. et al. Normothermic ex vivo lung perfusion in clinical lung transplantation. N Engl J Med 364, 1431–1440 (2011).
Mesaki, K. et al. Immunomodulation of the donor lung with CRISPR-mediated activation of IL-10 expression. J Heart Lung Transplant 42, 1363–1377 (2023).
Nakajima, D. et al. Mesenchymal stromal cell therapy during ex vivo lung perfusion ameliorates ischemia-reperfusion injury in lung transplantation. J Heart Lung Transplant 38, 1214–1223 (2019).
Wang, A. et al. Ex vivo enzymatic treatment converts blood type A donor lungs into universal blood type lungs. Sci Transl Med 14, eabm7190 (2022).
Nykänen, A. I. et al. Lung Transplant Immunomodulation with Genetically Engineered Mesenchymal Stromal Cells-Therapeutic Window for Interleukin-10. Cells 13, 859 (2024).

Acknowledgements: The authors express their gratitude to the Toronto Lung Transplant Program Biobank for its support. This study was made possible with the efforts of the clinical teams past and present, including organ perfusion specialists, transplant fellows, nurses and students. In addition, the authors acknowledge the support of the AI HUB at University Health Network and the funding provided by the Canadian Institutes of Health Research (CIHR) [PJT-185944] and the J.P. Bickell Foundation Medical Research Grant. The funding organizations had no involvement in the study's design, data analysis, or manuscript preparation. Most importantly, we would like to thank the generosity of all organ donors that make this work possible.


Figure 1. Ex vivo lung perfusion (EVLP). a, EVLP system; b, Description of EVLP circuit and components.



Figure 2. a, Schematic of digital twin validation using real-world tPA treatment data. b, A representative tPA-treated human lung (black) alongside the corresponding digital twin (red) demonstrating therapeutic efficacy (PAP, squares) and safety (edema, circles). ML – machine learning; tPA – tissue-type plasminogen activator.


Table 1. Ex vivo digital twin ML model forecasting accuracy