Digital twin for ex-VIVO organ

Abstract

Provided herein are methods and systems for generating a digital twin of an isolated organ using machine learning models trained on data relating to the ex vivo organ perfusion platform. Various uses of the digital twin disclosed herein are also provided.

Description

TITLE: DIGITAL TWIN FOR EX-VIVO ORGAN CROSS REFERENCE:

[0001] This PCT application claims the benefit to U.S. Provisional Application No. 63 / 603,060, filed November 27, 2023, the contents of which are incorporated herein by reference. FIELD:

[0002] The present disclosure relates generally to digital organs, and more specifically, to digital ex vivo organs. BACKGROUND:

[0003] Digital twins have recently emerged in the healthcare setting. Digital twin studies have been reported by researchers in cardiology, immunology, diabetes, hepatology, drug discovery, and oncology. In-silico models of the heart and lung have only been used to understand the mechanics of these organs. SUMMARY:

[0004] In one aspect, disclosed herein is a method for generating a digital twin of an isolated organ, comprising: receiving, by a computing device, a plurality of ex vivo organ perfusion time series data entries from a first assessment period; generating, by the computing device, one or more organ parameters for a first forecast period at least in part by processing the time series data entries from the first assessment period with a trained machine learning model, wherein the first forecast period occurs after the first assessment period; and displaying, by the computing device, using a graphical user interface, a simulation output comprising the generated one or more organ parameters for the first forecast period, thereby generating the digital twin of the isolated organ.

[0005] In another aspect, disclosed herein is a method for generating a digital twin of an isolated organ, comprising: receiving, by a computing device, a plurality of donor chart data; generating, by the computing device, one or more organ parameters during ex vivo organ perfusion at least in part by processing the donor chart data entries with a trained machine learning model; and displaying, by the computing device, using a graphical userinterface, a simulation output comprising the generated one or more organ parameters, thereby generating the digital twin of the isolated organ.

[0006] In another aspect, disclosed herein is a system for generating a digital twin of an isolated organ, comprising: (a) a memory to store executable components; (b) a processor, operably linked to the memory to implement the executable components, the executable components comprising: (i) a model-receiving platform for receiving a trained machine learning model generated from training data comprising a plurality of ex vivo organ perfusion time series data entries from a first assessment period; (ii) a simulation platform that simulates one or more organ parameters for a first forecast period at least in part by processing the time series data entries with the trained machine learning model, wherein the first forecast period occurs after the first assessment period; and (iii) an output-generating component to generate a simulation output comprising the generated one or more organ parameters from the first forecast period; and (c) a user interface to display the simulation output.

[0007] In another aspect, disclosed herein is a system for generating a digital twin of an isolated organ, comprising: (a) a memory to store executable components; (b) a processor, operably linked to the memory to implement the executable components, the executable components comprising: (i) a model-receiving platform for receiving a trained machine learning model generated from training data comprising a plurality of donor chart data entries; (ii) a simulation platform that simulates one or more organ parameters for a first forecast period during ex vivo organ perfusion at least in part by processing the donor chart data entries with the trained machine learning model; and (iii) an output-generating component to generate a simulation output comprising the generated one or more organ parameters from the first forecast period; and (c) a user interface to display the simulation output.

[0008] In another aspect, disclosed herein is a computer-readable storage medium with instructions stored thereon that causes a system to: (a) receive a trained machine learning model generated from training data comprising a plurality of ex vivo organ perfusion time series data entries from a first assessment period; (b) simulate one or more organ parameters for a first forecast period, wherein the first forecast period occurs after the firstassessment period; and (c) generate a simulation output comprising the simulated one or more organ parameters for the first forecast period.

[0009] In another aspect, disclosed herein is a computer-readable storage medium with instructions stored thereon that causes a system to: (a) receive a trained machine learning model generated from training data comprising a plurality of donor chart data entries; (b) simulate one or more organ parameters for a first forecast period during ex vivo organ perfusion; and (c) generate a simulation output comprising the simulated one or more organ parameters for the first forecast period.

[0010] In yet another aspect, disclosed herein is a method for training a machine learning system to produce a digital twin of a lung, the method comprising: collecting a set of input data for a first time period, the input data comprising high-resolution time-series data, physiology and biochemistry data, metabolomic and protein biomarkers, lung image data, and transcriptomics data; pre-processing the set of input data, the pre-processing comprising generating breath-by-breath parameters from the high-resolution time series ventilator data, hourly parameter values from the physiology and biochemistry data, generating hourly biomarker values from the metabolomic and protein biomarkers, performing principal components analysis (PCA) on the lung image data to generate principal components; and generating gene enrichment scores from the transcriptomics data; training a first model to predict breath parameters for a second time period by processing the breath-by-breath parameters, the hourly parameter values, the hourly biomarker values, the principal components, and the gene enrichment scores; and training a second model to predict hourly values, image principal components, and transcriptomic changes for the second time period, by processing the hourly parameter values, the hourly biomarker values, the principal components, and the gene enrichment scores.

[0011] Also disclosed herein is a method for training a machine learning system to produce a digital twin of an organ, the method comprising: collecting a first set of input data, the input data comprising high-resolution time-series ex vivo organ function data; collecting a second set of input data, the data comprising organ function data retrieved from one or more electronic sources; training, in a first stage, a machine learning model to predict one or more organ parameters from the second set of input data; and training, in a second stage,the machine learning model to predict the one or more organ parameters from the first set of input data.

[0012] In a further aspect, disclosed herein is a use of the digital twin generated by the method or the system of the present disclosure in a study, optionally a clinical study.

[0013] Other features and advantages of the present disclosure will become apparent from the following detailed description. It should be understood, however, that the detailed description and the specific examples while indicating preferred embodiments of the disclosure are given by way of illustration only, since various changes and modifications within the spirit and scope of the disclosure will become apparent to those skilled in the art from this detailed description. BRIEF DESCRIPTION OF THE DRAWINGS:

[0014] The embodiments of the application will now be described in greater detail with reference to the attached drawings in which:

[0015] FIGs. 1A-1C is an illustration of the Toronto Ex Vivo Lung Perfusion (EVLP) Platform. (A) Image of a human lung on the EVLP platform. (B) Schematic drawing of the individual circuit components for the EVLP system including: ICU ventilator, pressure monitor, CDI monitor, and perfusate reservoir for cytokine testing. (C) Examples of the multi- modal and organ-specific data (blue boxes) generated during EVLP, which can be analyzed using machine learning to create digital twins of human lungs (center). (D)Modeling strategies for developing human lung digital twins: static digital lungs (purple) – forecasting models based solely on baseline functional lung data; and dynamic digital twins (orange) – forecasting models that are continuously updated and recalibrated with ongoing functional lung data. (Images: UHN, Biorender)

[0016] FIG.2 is a graph showing real-time ventilator data capture during human EVLP. Breath-by-breath recording and analysis of dynamic compliance measurements (line) compared to the data derived from the traditional approach of hourly recording (dots).

[0017] FIG. 3 is a graph showing real-time data recording in lung perfusate using a porcine model of EVLP. Real-time data extraction of EVLP perfusate features using theCDI550 monitor to quantify: partial pressure of oxygen (pO2) and carbon dioxide (pCO2), and perfusate pH.

[0018] FIG. 4 is a graph showing an embodiment forecasting results of pO2levels during EVLP using high-resolution data. Real-time pO2data measured in EVLP perfusate using the CDI550 monitor compared to the predicted pO2values using a KNN model. Using high-resolution data, predicted pO2values were within 2% of the recorded values.

[0019] FIG.5 are graphs showing prediction of lung function during EVLP using deep learning in an embodiment. The CNN-based model displayed precise predictive profiles for: (A) dynamic compliance (mean absolute error (MAE)=10.3 mL / cmH2O vs. clinical benchmark=15.8mL / cmH2O); (B) stress index (MAE=0.07); (C) mean pressure (MAE=0.23 cmH2O vs. clinical benchmark=0.33cmH2O); and (D) peak pressure (MAE=0.9 cmH2O vs. clinical benchmark=1.6cmH2O). Data shown are representative EVLP cases in the study cohort. The predicted data and observed data are shown on each graph.

[0020] FIG.6 are graphs showing examples of prediction of lung function during EVLP using deep learning additional to the example of Fig.5.

[0021] FIG.7 is a diagram illustrating the study population for digital twin modeling.

[0022] FIG. 8 is a diagram showing modeling strategies for developing human lung digital twins: static digital lungs – forecasting models based solely on baseline functional lung data; and dynamic digital twins – forecasting models that are continuously updated and recalibrated with ongoing functional lung data.

[0023] FIG.9 is a diagram showing digital twin forecasting pipeline of X-ray derived image principal component features. CNN=convolutional neural network; PC = principal component.

[0024] FIG. 10 is a graph showing high-resolution time-series trace of breath -by- breath dynamic compliance.

[0025] FIGs.11A-11Q are graphs showing multi-modal digital twins of human lungs can accurately forecast a full panel of parameters of lung physiology, biochemistry, -omics, and image features. Doted lines: Observed; darker solid line: static digital lung; lighter solid line: dynamic digital lung. Static and dynamic multi-modal forecasting digital twins of asimulated human lung on EVLP, using baseline functional data (dashed black lines) for: lung physiology (a, dynamic compliance; b, static compliance; c, edema; d, plateau airway pressure; e, peak and mean airway pressures; partial pressure of CO2 (f) and O2 (g); h, pulmonary arterial and left atrial vascular pressure; i, expiratory volume); lung biochemistry (j, calcium and chloride; k, sodium and potassium; l, bicarbonate and base excess; m, pH); lung –omics (n, metabolomic biomarker (glucose and lactate); o, proteomic biomarker (IL-6, IL-8, IL-10, IL-1β)); p, transcriptomics (gene enrichment scores for lung disease-related pathways); and lung imaging (q, image principal component values (normalized) derived from the 3rd hour X-ray image). Abbreviations: DT-digital twin; PC-principal component; IL- interleukin; AU-arbitrary units.

[0026] FIGs. 12A-12L are images and graphs showing enhanced therapeutic evaluation using human lung digital twins. FIGs. 12A-12B are representative images of human lungs with suspicion with pulmonary embolism. FIG.12A is an image of human lungs with evidence of possible vascular obstructions on the posterior surface. FIG. 12B is an image of a blood clot obtained during retrograde flushing of the lungs in FIG. 12A. Representative data from measurements of the observed (black) pulmonary arterial pressure (FIG. 12C) and edema formation (FIG. 12D) in an alteplase treated human lung vs the corresponding digital twin for that lung. Evaluation of the efficacy of Alteplase to modify pulmonary arterial pressure one (FIGs.12E and 12F) and two (FIGs.12G and 12H) hours after administration using a conventional control cohort (FIGs.12E and 12G) or digital twin cohort (FIGs.12F and 12H). Evaluation of the safety of Alteplase to reduce edema formation one (FIGs. 12I and 12J) and two (FIGs. 12K and 12L) hours after administration using a conventional control cohort (FIGs.12I and 12K) or digital twin cohort (FIGs.12J and 12L). Box plots depict median, upper and lower quartiles, and range. Legend: a-Two-sided Mann- Whitney test; b-Two-sided Wilcoxon matched-pairs signed rank test.

[0027] FIGs. 13A-13C are graphs showing time-series forecasting of liver functional parameters. PHA: pressure of hepatic artery; PPV: pressure of portal vein; PIVC: pressure of inferior vena cava; AST: aspartate aminotransferase; ALT: alanine aminotransferase.

[0028] FIG. 14 is a graph showing pre- and post-transformation of the PV curve by clockwise (IPF) and counterclockwise 10 degrees (COPD). Black - Original curve; darker grey - clockwise 10 degree rotation; lighter grey - counterclockwise 10 degree rotation.

[0029] FIGs. 15A-15B are graphs showing example flow (FIG. 15A) and pressure (FIG. 15B) changes upon PV curve transformation. Black - Original curve; darker grey - clockwise 10 degree rotation; lighter grey - counterclockwise 10 degree rotation.

[0030] FIGs. 16A-16B are graphs showing digital twin of ex vivo lung model forecasting performance on transformed breath parameter data. Black - Observed; darker grey - Static digital lung; lighter grey - Dynamic digital lung.

[0031] FIG. 17 is a diagram illustrating generation of a lung digital twin according to certain embodiments.

[0032] FIG. 18 is a diagram illustrating digital twin of a generic organ pipeline according to certain embodiments.

[0033] FIG. 19 is a diagram illustrating a disease-specific digital twin pipeline according to certain embodiments.

[0034] FIG. 20 is a diagram illustrating generation of a liver digital twin according to certain embodiments. DETAILED DESCRIPTION OF THE DISCLOSURE:

[0035] The following is a detailed description provided to aid those skilled in the art in practicing the present disclosure. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs. The terminology used in the description herein is for describing particular embodiments only and is not intended to be limiting of the disclosure. All publications, patent applications, patents, figures and other references mentioned herein are expressly incorporated by reference in their entirety. I. Method

[0036] It is demonstrated herein that a machine learning model trained on data generated from an ex vivo organ perfusion platform can be used to predict various organ parameters. Accordingly, in one aspect, provided herein is a method for generating a digital twin of an isolated organ, the method comprises: receiving, by a computing device, a plurality of ex vivo perfusion time series data entries from a first assessment period; generating, by the computing device, one or more organ parameters for a first forecast period at least in part by processing the time series data entries from the first assessment period with a trained machine learning model, wherein the first forecast period occurs after the first assessment period; displaying, by the computing device, using a graphical user interface, a simulation output comprising the generated one or more organ parameters for the first forecast period, thereby generating the digital twin of the isolated organ. In one aspect, provided herein is a method of generating a digital twin. The digital twin may be a digital twin of an isolated organ. The method may comprise receiving a plurality of ex vivo organ perfusion time series data entries. The plurality of ex vivo organ perfusion time series data entries may be from a first assessment period. The receiving of the plurality of ex vivo organ perfusion time series data entries may be receiving by a computing device. The method may comprise generating one or more organ parameters. The one or more organ parameters may be for a first forecast period. The generating of the one or more organ parameters may be generating at least in part by processing the time series data entries from the first assessment period. The generating of the one or more organ parameters may be by a computing device. The processing of the time series data may be processing with a trained machine learning model. The first forecast period may occur after the first assessment period. The method may comprise displaying a simulation output. The simulation output may comprise the generated one or more organ parameters for the first forecast period. The displaying may be displaying by the computing device. The displaying may be displaying by a graphical user interface. The method may thereby generate the digital twin of the isolated organ.

[0037] In another aspect, provided herein is a method for generating a digital twin of an isolated organ, the method comprises: receiving, by a computing device, a plurality of donor chart data; generating, by the computing device, one or more organ parameters duringex vivo organ perfusion at least in part by processing the donor chart data entries with a trained machine learning model; and displaying, by the computing device, using a graphical user interface, a simulation output comprising the generated one or more organ parameters, thereby generating the digital twin of the isolated organ. In one aspect, provided herein is a method for generating a digital twin of an isolated organ. The method may comprise receiving a plurality of donor chart data. The method may comprise generating one or more organ parameters. The one or more organ parameters may be organ parameters during ex vivo organ perfusion. The generating of the one or more organ parameters may be generating at least in part by processing the donor chart data entries. The generating of the one or more organ parameters may be generating by a computing device. The processing of the donor chart data entries may be processing with a trained machine learning model. The method may comprise displaying a simulation output. The simulation output may comprise the generated one or more organ parameters. The displaying may be displaying by the computing device. The displaying may be displaying by a graphical user interface. The method may thereby generate the digital twin of the isolated organ.

[0038] As used herein, the term “isolated organ” refers to an organ isolated from the body in an ex-vivo perfusion system. The term encompasses a partial organ, for example, a partial lung, for example, a lung lobe.

[0039] The isolated organ can be of any animal origin. The animal can be a mammal, optionally human.

[0040] The isolated organ can be an organ of any condition, such as a healthy organ, a diseased organ, or an injured organ.

[0041] As used herein, the term “data entry” or “data entries” refers to data relating to an ex vivo organ perfusion platform that can be used as training data for a machine-learning model. The data can comprise, for example, data from the ex vivo organ perfusion platform during perfusion, and / or information relating to a donor or the donated organ, which may be obtained from a donor chart.

[0042] In some embodiments, the data entries comprise information obtainable from donor charts. In some embodiments, the data entries comprise one or more of donor age, donor sex, donor height, donor weight, donor smoking history (positive or negative history),donor cause of death, donor type (cardiac death or brain death), lung on EVLP (Left, Right, Both Right and Left or at least one lobe), and EVLP cold ischemic time (CIT).

[0043] In some embodiments, the method further comprises receiving, by the computing device, a plurality of donor chart data entries. In some embodiments, the generating of the one or more organ parameters for the first forecast period further comprises processing the time series data entries from the at least one additional assessment period with a trained machine learning model.

[0044] It is disclosed herein that the forecasting model can be updated and recalibrated with ongoing functional data.

[0045] In some embodiments, the method further comprises receiving, by the computer device, a plurality of ex vivo organ perfusion time series data entries from at least one additional assessment period. In some embodiments, the generating of the one or more organ parameters for the first forecast period further comprises processing the time series data entries from the at least one additional assessment period with a trained machine learning model.

[0046] In some embodiments, the at least one additional assessment period occurs before the first forecast period.

[0047] In some embodiments, the method further comprises generating, by the computing device, one or more organ parameters for at least one additional forecast period at least in part by processing the time series data entries from the first assessment period with a trained machine learning model, wherein the at least one additional forecast period occurs after the first assessment period and the first forecast period.

[0048] In some embodiments, the simulation output further comprises the generated one or more organ parameters for the at least one additional forecast period.

[0049] In some embodiments, generating the one or more organ parameters for the at least one additional forecast period further comprises processing the one or more organ parameters for the first forecast period.

[0050] The machine learning model can process time series data entries from one or more assessment periods. An assessment period can refer to a period when data from anorgan undergoing ex vivo organ perfusion are obtained. The data from an assessment period can be used by the computing device to forecast parameters for a later time period (a forecast period). The data from an assessment period may be transformed and the transformed data can then be used by the computing device to forecast paraments for a forecast period.

[0051] It is disclosed herein that the forecasting model can be updated and recalibrated with ongoing functional data.

[0052] In some embodiments, the at least one additional assessment period is one additional assessment period. In some embodiments, the at least one additional assessment period is two additional assessment periods. In some embodiments, the at least one additional assessment period is three additional assessment periods. In some embodiments, the at least one additional assessment period is four additional assessment periods. In some embodiments, the at least one additional assessment period is five additional assessment periods. In some embodiments, the at least one additional assessment period is six additional assessment periods. In some embodiments, the at least one additional assessment period is more than six additional assessment periods.

[0053] In some embodiments, the at least one additional assessment period occurs before the first forecast period. In some embodiments, at least one of the additional assessment periods occurs before the first forecast period. In some embodiments, one of the additional assessment periods occurs before the first forecast period. In some embodiments, two of the additional assessment periods occur before the first forecast period. In some embodiments, three of the additional assessment periods occur before the first forecast period. In some embodiments, four of the additional assessment periods occur before the first forecast period. In some embodiments, five of the additional assessment periods occur before the first forecast period. In some embodiments, six of the additional assessment periods occur before the first forecast period. In some embodiments, more than six of the additional assessment periods occur before the first forecast period. In some embodiments, all of the additional assessment periods occur before the first forecast period.

[0054] In some embodiments, the at least one additional assessment period occurs after the first forecast period. In some embodiments, at least one of the additional assessment periods occurs after the first forecast period. In some embodiments, one of theadditional assessment periods occurs after the first forecast period. In some embodiments, two of the additional assessment periods occur after the first forecast period. In some embodiments, three of the additional assessment periods occur after the first forecast period. In some embodiments, four of the additional assessment periods occur after the first forecast period. In some embodiments, five of the additional assessment periods occur after the first forecast period. In some embodiments, six of the additional assessment periods occur after the first forecast period. In some embodiments, more than six of the additional assessment periods occur after the first forecast period. In some embodiments, all of the additional assessment periods occur after the first forecast period.

[0055] In some embodiments, the method further comprises generating, by the computing device, one or more lung parameters for at least one additional forecast period at least in part by processing the time series data entries from the first assessment period with a trained machine learning model, wherein the at least one additional forecast period occurs after the first assessment period and the first forecast period.

[0056] In some embodiments, the simulation output further comprises the generated one or more lung parameters for the at least one additional forecast period.

[0057] In some embodiments, the at least one additional forecast period is one additional forecast period. In some embodiments, the at least one additional forecast period is two additional forecast periods. In some embodiments, the at least one additional forecast period is three additional forecast periods. In some embodiments, the at least one additional forecast period is four additional forecast periods. In some embodiments, the at least one additional forecast period is more than four additional forecast periods.

[0058] The forecasting model can forecast based on both the time entries from the one or more assessment periods and the one or more organ parameters generated from the one or more forecast periods. For example, the method can generate one or more organ parameters for a second forecast period by processing the time series data entries from the first assessment period and the one or more organ parameters for the first forecast period.

[0059] The forecasting model can forecast based on one or more organ parameters generated from one or more forecast periods. For example, the model can forecast a fifth forecast period base on one or more organ parameters generated from one or more of theprevious forecast periods, such as the fourth forecast period, the third forecast period, the second forecast period, and / or the first forecast period. Such sliding window approach can allow continuous forecasting.

[0060] In some embodiments, the method further comprises generating, by the computing device, one or more organ parameters for a plurality of additional forecast periods at least in part by processing the time series data entries from at least one previous forecast period.

[0061] In some embodiments, the simulation output further comprises the generated one or more organ parameters for the plurality of additional forecast periods.

[0062] Any suitable assessment period and forecast period can be selected. For example, data entries can be collected at about one hour after the organ is subject to ex vivo organ perfusion, and the model can forecast the organ parameters at about two hours after the organ is subject to ex vivo organ perfusion. The model can further forecast the organ parameters at about three hours after the organ is subject to ex vivo organ perfusion. If data entries are collected for at least one additional assessment period, then data entries can be collected at about one hour and at about two hours after the organ is subject to ex vivo organ perfusion, and the model can forecast the organ parameters at about three hours after the organ is subject to ex vivo organ perfusion. It can be appreciated that other assessment periods and forecast periods can be used, and they may or may not be at regular intervals.

[0063] In some embodiments, the first assessment period is at about 15 min of ex vivo organ perfusion. In some embodiments, the first assessment period is at about 30 min of ex vivo organ perfusion. In some embodiments, the first assessment period is at about 45 min of ex vivo organ perfusion. In some embodiments, the first assessment period is at about 1 hour of ex vivo organ perfusion. In some embodiments, the first assessment period is at about 75 min of ex vivo organ perfusion. In some embodiments, the first assessment period is at about 90 min of ex ex vivo organ perfusion. In some embodiments, the first assessment period is at about 105 min of ex vivo organ perfusion. In some embodiments, the first assessment period is at about 2 hours of ex vivo organ perfusion. In some embodiments, the first assessment period is after more than 2 hours of ex vivo organ perfusion.

[0064] In some embodiments, the at least one additional assessment period is a second assessment period. In some embodiments, the second assessment period is at about 15 min after the first assessment period. In some embodiments, the second assessment period is at about 30 min after the first assessment period. In some embodiments, the second assessment period is at about 45 min after the first assessment period. In some embodiments, the second assessment period is at about 1 hour after the first assessment period. In some embodiments, the second assessment period is at about 75 min after the first assessment period. In some embodiments, the second assessment period is at about 90 min after the first assessment period. In some embodiments, the second assessment period is at about 105 min after the first assessment period. In some embodiments, the second assessment period is at about 2 hours after the first assessment period. In some embodiments, the second assessment period is more than 2 hours after the first assessment period.

[0065] In some embodiments, the first forecast period is at about 15 min after the first assessment period. In some embodiments, the first forecast period is at about 30 min after the first assessment period. In some embodiments, the first forecast period is at about 45 min after the first assessment period. In some embodiments, the first forecast period is at about one hour after the first assessment period. In some embodiments, the first forecast period is at about 75 min after the first assessment period. In some embodiments, the first forecast period is at about 90 min after the first assessment period. In some embodiments, the first forecast period is at about 105 min after the first assessment period. In some embodiments, the first forecast period is at about 2 hours after the first assessment period. In some embodiments, the first forecast period is at more than 2 hours after the first assessment period.

[0066] In some embodiments, the at least one additional forecast period is a second forecast period. In some embodiments, the second forecast period is about 15 min after the first forecast period. In some embodiments, the second forecast period is about 30 min after the first forecast period. In some embodiments, the second forecast period is about 45 min after the first forecast period. In some embodiments, the second forecast period is about one hour after the first forecast period. In some embodiments, the second forecast period is about75 min after the first forecast period. In some embodiments, the second forecast period is about 90 min after the first forecast period. In some embodiments, the second forecast period is about 105 min after the first forecast period. In some embodiments, the second forecast period is about 2 hours after the first forecast period. In some embodiments, the second forecast period is at more than 2 hours after the first forecast period.

[0067] Any suitable data entries can be used with the methods disclosed herein. In some embodiments, the data entries comprise time series data entries. In some embodiments, the data entries comprise non-time series data entries. In some embodiments, the time series data entries comprise hourly data entries. In some embodiments, the time series data entries comprise one or more selected from the group consisting of: physiological data, biochemical data, transcriptomic data, metabolomic biomarker data, proteomic biomarker data, and imaging data. In some embodiments, the time series data entries comprise physiological data. In some embodiments, the time series data entries comprise biochemical data. In some embodiments, the time series data entries comprise omics data. In some embodiments, the time series data entries comprise transcriptomic data. In some embodiments, the time series data entries comprise metabolomic biomarker data. In some embodiments, the time series data entries comprise proteomic biomarker data. In some embodiments, the time series data entries comprise imaging data.

[0068] Any suitable machine-learning model can be used. In some embodiments, the trained machine learning model comprises a neural network. In some embodiments, the neural network is a CNN. In some embodiments, the neural network is an RNN.

[0069] The ex vivo organ perfusion time series data entries can comprise synthetic data. Synthetic data refer to data that are not obtained directly from a real organ. For example, synthetic data can be data transformed from real data, such as data obtained from a real organ. The synthetic data can, for example, capture characteristics of a diseased organ.

[0070] In some embodiments, the plurality of ex vivo organ perfusion time series data entries comprise synthetic data. In some embodiments, the synthetic data comprise synthetic data transformed from ex vivo organ perfusion time series data entries obtained from a real organ. In some embodiments, the synthetic data comprise disease synthetic data.

[0071] The method disclosed herein can be used to generate a digital twin of any organ. In some embodiments, the organ is a lung, a liver, a heart, a kidney, or a pancreas.

[0072] In some embodiments, the digital twin is a digital twin of an isolated mammalian organ. In some embodiments, the isolated mammalian organ is an isolated human organ.

[0073] In some embodiments, the digital twin is a digital twin of a healthy organ, a diseased organ, or an injured organ. In some embodiments, the digital twin is a digital twin of a healthy organ. In some embodiments, the digital twin is a digital twin of a diseased organ. In some embodiments, the digital twin is a digital twin of an injured organ. Lung digital twins

[0074] In some embodiments, the organ is a lung.

[0075] Any suitable machine-learning model can be used. For example, if a model is trained to take in features of donated lungs placed on EVLP to predict physiological parameters that are measured hourly during EVLP, machine-learning algorithms such as a a K-Nearest Neighbour (KNN) or an Extreme Gradient Boosting (XGBoost) model may be used.

[0076] An ex vivo lung perfusion (EVLP) platform allows ex vivo assessment of a lung. For example, with EVLP, a donor lung can be maintained in a normothermic environment (37°C), perfused with an acellular solution, and monitored for a period of time, for example up to six hours. During this time, the active metabolic state of the ex vivo donor lung can be monitored, for example, by recording parameters related to organ quality and performance, including gas exchange (e.g., pO2, pCO2), lung compliance (static & dynamic), airway pressure, acid-base chemistry (e.g., pH), electrolytes (e.g. K+), and inflammatory proteins (e.g., cytokines). A data entry of the disclosed method can relate to any of the aforementioned parameters.

[0077] The data from the EVLP platform during perfusion that can be used as training data can be any data from lung assessment on EVLP. For example, the data can be physiological data, biochemical data, omics data such as transcriptomic data, metabolomic biomarker data, and / or proteomic biomarker data, and / or imaging data. Physiological data obtainable during EVLP can include, but not limited to, for example pO2, pCO2, dynamiccompliance, static compliance, PA & LA pressure, vascular resistance, airway pressure (peak, mean, plateau), positive end expiratory pressure (PEEP), and / or volume of perfusate (loss and exchange). Biochemical data can include but not limited to, for example Ca2+, Cl-, K+, Na+, base excess, HCO3-, and pH. Transcriptomic biomarker data can relate to one or more lung disease-related pathways, such as hypoxia, inflammatory response, apoptosis, TP53 signaling, TNF-a signaling, interleukin-s signaling, PI3K / AKT / mTOR signaling, interleukin-6 signaling, TGF-b signaling, and / or oxidative phosphorylation. Metabolomic biomarker data, can include but not limited to, for example glucose, and / or lactate. Proteomic data obtainable during EVLP can be for example biomarker levels and can relate to for example inflammation (such as IL-10, IL-1β, IL-6, IL-8). Biochemical data and omics data may be measured from the perfusate. Imaging data can include, but not limited to, for example X-ray data.

[0078] In some embodiments, the data entries comprise physiological data, biochemical data, transcriptomic data, metabolomic biomarker data, proteomic biomarker data, and / or imaging data from lung assessment on EVLP.

[0079] In some embodiments, the physiological data comprises one or more selected from the group consisting of: air pressure, vascular pressure, partial pressure of O2, partial pressure of CO2, lung compliance, expiratory volume, and estimates of edema, or surrogates thereof such as perfusate loss. In some embodiments, the surrogate of an estimate of edema comprises perfusate loss.

[0080] In some embodiments, the biochemical data comprises one or more selected from the group consisting of: calcium, chloride, sodium, potassium; bicarbonate, base excess, and pH.

[0081] In some embodiments, the transcriptomic data comprises gene enrichment scores for one or more lung disease-related pathways. In some embodiments, the one or more lung disease related pathways are hypoxia, inflammatory response, apoptosis, TP53 signaling, TNF-a signaling, interleukin-s signaling, PI3K / AKT / mTOR signaling, interleukin-6 signaling, TGF-b signaling, and oxidative phosphorylation.

[0082] In some embodiments, the metabolomic biomarker data comprises at least one selected from the group consisting of: glucose and lactate.

[0083] In some embodiments, the proteomic biomarker data comprises at least one selected from the group consisting of: IL-6, IL-8, IL-10, and IL-1β.

[0084] In some embodiments, the imaging data comprises X-ray image features.

[0085] The data entries used to train models to forecast the one or more lung parameters can comprise donor chart data entries. Donor chart data entries can relate to the donor of the lung, such as age, sex, height, weight, smoking history (positive or negative history), cause of death, and / or type of death (e.g. cardiac death, brain death). Donor chart data entries can relate to characteristics of the donated lung, such as the lung on EVLP (Left, Right, Both Left or Right, or at least one lobe), and / or EVLP cold ischemic time (CIT).

[0086] In some embodiments, the method further comprises receiving, by the computing device, a plurality of donor chart data entries. In some embodiments, the donor chart data entries comprise one or more of donor age, donor sex, donor height, donor weight, donor smoking history, donor cause of death, type of death, lung on EVLP, and EVLP cold ischemic time (CIT).

[0087] It is disclosed herein that various data can be collected during ex vivo lung perfusion (EVLP) and used to train a machine-training model to forecast various lung parameters at a later time point. For example, data can be collected at a first assessment period to forecast parameters for a first forecast period. Data from the first assessment period can also be used to forecast parameters for at least one additional forecast period in addition to the first forecast period.

[0088] In some embodiments, the one or more lung parameters generated comprise physiological parameter, biochemical parameter, transcriptomic parameter, metabolomic biomarker parameter, proteomic biomarker parameter, and / or imaging parameter.

[0089] In some embodiments, the physiological parameter comprises one or more selected from the group consisting of: air pressure, vascular pressure, partial pressure of O2, partial pressure of CO2, lung compliance, expiratory volume, and estimates of edema, or surrogates thereof such as perfusate loss. In some embodiments, the surrogate of an estimate of edema comprises perfusate loss.

[0090] In some embodiments, the biochemical parameter comprises one or more selected from the group consisting of: calcium, chloride, sodium, potassium; bicarbonate, base excess, and pH.

[0091] In some embodiments, the transcriptomic parameter comprises gene enrichment scores for one or more lung disease-related pathways.

[0092] In some embodiments, the one or more lung disease related pathways are hypoxia, inflammatory response, apoptosis, TP53 signaling, TNF-a signaling, interleukin-s signaling, PI3K / AKT / mTOR signaling, interleukin-6 signaling, TGF-b signaling, and oxidative phosphorylation.

[0093] In some embodiments, the metabolomic parameter comprises at least one selected from the group consisting of: glucose and lactate.

[0094] In some embodiments, the proteomic biomarker parameter comprises at least one selected from the group consisting of: IL-6, IL-8, IL-10, and IL-1β.

[0095] In some embodiments, the imaging parameter comprises X-ray image features.

[0096] The time series data entries can be data that are collected at specific time interval, such as hourly data entries. The time series data entries can be data that are collected at each breath.

[0097] In some embodiments, the time series data entries are hourly data entries. In some embodiments, the time series data entries are breath-by-breathe data entries.

[0098] In some embodiments, the one or more lung parameter comprises an hourly parameter. In some embodiments, the one or more lung parameter comprises a per-breath parameter.

[0099] It is demonstrated that the method disclosed herein can accurately forecast various lung parameters, as measured by, for example, mean absolute percentage error (MAPE).

[0100] The term “mean absolute percentage error” or “MAPE” as used herein refers to the percentage difference between the forecasted and observed values.

[0101] In some embodiments, the one or more lung parameters has a mean absolute percentage error (MAPE) of less than about 1%. In some embodiments, the one or more lung parameters has a mean absolute percentage error (MAPE) of about 1%. In some embodiments, the one or more lung parameters has a a MAPE of about 2%. In some embodiments, the one or more lung parameters has a MAPE of about 3%. In some embodiments, the one or more lung parameters has a MAPE of about 4%. In some embodiments, the one or more lung parameters has a MAPE of about 5%. In some embodiments, the one or more lung parameters has a MAPE of about 6%. In some embodiments, the one or more lung parameters has a MAPE of about 7%. In some embodiments, the one or more lung parameters has a MAPE of about 8%. In some embodiments, the one or more lung parameters has a MAPE of about 9%. In some embodiments, the one or more lung parameters has a MAPE of about 10%. In some embodiments, the one or more lung parameters has a MAPE of about 11%. In some embodiments, the one or more lung parameters has a MAPE of about 12%. In some embodiments, the one or more lung parameters has a MAPE of about 13%. In some embodiments, the one or more lung parameters has a mean a MAPE of about 14%. In some embodiments, the one or more lung parameters has a MAPE of about 15%. In some embodiments, the one or more lung parameters has a MAPE of about 16%. In some embodiments, the one or more lung parameters has a MAPE of about 17%. In some embodiments, the one or more lung parameters has a MAPE of about 18%. In some embodiments, the one or more lung parameters has a MAPE of about 19%. In some embodiments, the one or more lung parameters has a MAPE of about 20%.

[0102] In some embodiments, the one or more lung parameters is a per-breath parameter. In some embodiments, the per-breath parameter comprises air pressure. In some embodiments, the per-breath parameter comprises dynamic compliance. In some embodiments, the per-breath parameter comprises expiratory volume. In some embodiments, the per-breath parameter has a mean absolute percentage error (MAPE) of less than about 10%. In some embodiments, the per-breath parameter has a MAPE of less than about 9%. In some embodiments, the per-breath parameter has a MAPE of less than about 8%. In some embodiments, the per-breath parameter has a MAPE of less than about 7%. In some embodiments, the per-breath parameter has a MAPE of less than about 6%.In some embodiments, the per-breath parameter has a MAPE of less than about 5%. In some embodiments, the per-breath parameter has a MAPE of less than about 4%. In some embodiments, the per-breath parameter has a MAPE of less than about 3%. In some embodiments, the per-breath parameter has a MAPE of less than about 2%. In some embodiments, the per-breath parameter has a MAPE of between about 1% and about 10%. In some embodiments, the per-breath parameter has a MAPE of between about 2% and about 9%. In some embodiments, the per-breath parameter has a MAPE of between about 2% and about 8%. In some embodiments, the per-breath parameter has a MAPE of between about 2% and about 7%.

[0103] In some embodiments, the one or more lung parameters is an hourly parameter. In some embodiments, the hourly parameter comprises vascular pressure. In some embodiments, the hourly parameter comprises partial pressure of O2. In some embodiments, the hourly parameter comprises partial pressure of CO2. In some embodiments, the hourly parameter comprises static compliance. In some embodiments, the hourly parameter has a MAPE of less than about 15%. In some embodiments, the hourly parameter has a MAPE of less than about 14%. In some embodiments, the hourly parameter has a MAPE of less than about 13%. In some embodiments, the hourly parameter has a MAPE of less than about 12%. In some embodiments, the hourly parameter has a MAPE of less than about 11%. In some embodiments, the hourly parameter has a MAPE of less than about 10%. In some embodiments, the hourly parameter has a MAPE of less than about 9%. In some embodiments, the hourly parameter has a MAPE of less than about 8%. In some embodiments, the hourly parameter has a MAPE of less than about 7%. In some embodiments, the hourly parameter has a MAPE of less than about 6%. In some embodiments, the hourly parameter has a MAPE of less than about 5%. In some embodiments, the hourly parameter has a MAPE of less than about 4%. In some embodiments, the hourly parameter has a MAPE of less than about 3%. In some embodiments, the hourly parameter has a mean absolute percentage error (MAPE) of between about 3% and 15%. In some embodiments, the hourly parameter has a MAPE of between about 3% and 14%. In some embodiments, the hourly parameter has a MAPE of between about 4% and 13%. In some embodiments, the hourly parameter has a MAPE of between about 4% and 12%. In some embodiments, the hourly parameter has a MAPE ofbetween about 4% and 11%. In some embodiments, the hourly parameter has a MAPE of between about 4% and 10%.

[0104] In some embodiments, the one or more lung parameters is a biochemical parameter. In some embodiments, the biochemical parameter comprises calcium. In some embodiments, the biochemical parameter comprises potassium. In some embodiments, the biochemical parameter comprises chloride. In some embodiments, the biochemical parameter comprises sodium. In some embodiments, the biochemical parameter comprises bicarbonate. In some embodiments, the biochemical parameter comprises base excess. In some embodiments, the biochemical parameter comprises pH. In some embodiments, the biochemical parameter has a mean absolute percentage error (MAPE) of less than about 15%. In some embodiments, the biochemical parameter has a mean absolute percentage error (MAPE) of less than about 15%. In some embodiments, the biochemical parameter has a MAPE of less than about 14%. In some embodiments, the biochemical parameter has a MAPE of less than about 13%. In some embodiments, the biochemical parameter has a MAPE of less than about 12%. In some embodiments, the biochemical parameter has a MAPE of less than about 11%. In some embodiments, the biochemical parameter has a MAPE of less than about 10%. In some embodiments, the biochemical parameter has a MAPE of less than about 9%. In some embodiments, the biochemical parameter has a MAPE of less than about 8%. In some embodiments, the biochemical parameter has a MAPE of less than about 7%. In some embodiments, the biochemical parameter has a MAPE of less than about 6%. In some embodiments, the biochemical parameter has a MAPE of less than about 5%. In some embodiments, the biochemical parameter has a MAPE of less than about 4%. In some embodiments, the biochemical parameter has a MAPE of less than about 3%. In some embodiments, the biochemical parameter has a MAPE of less than about 2%. In some embodiments, the biochemical parameter has a MAPE of less than about 1%. In some embodiments, the biochemical parameter has a MAPE of between about 0.1% and about 13%. In some embodiments, the biochemical parameter has a MAPE of between about 0.1% and about 10%. In some embodiments, the biochemical parameter has a MAPE of between about 0.1% and about 9%. In some embodiments, the biochemical parameter has a MAPE of between about 0.1% and about 8%.

[0105] In some embodiments, the one or more lung parameters is a transcriptomic parameter. In some embodiments, the transcriptomic parameter comprises a post-EVLP transcriptomic level change of a gene. In some embodiments, the gene comprises a gene involved in a lung disease-related pathway. In some embodiments, the transcriptomic parameter has a mean absolute percentage error (MAPE) of less than about 5%. In some embodiments, the transcriptomic parameter has a MAPE of less than about 4%. In some embodiments, the transcriptomic parameter has a MAPE of less than about 3%. In some embodiments, the transcriptomic parameter has a MAPE of less than about 2%. In some embodiments, the transcriptomic parameter has a MAPE of between about 0.1% and 5%. In some embodiments, the transcriptomic parameter has a MAPE of between about 0.5% and 4%. In some embodiments, the transcriptomic parameter has a MAPE of between about 1% and 3%.

[0106] In some embodiments, the one or more lung parameters is a metabolic parameter. In some embodiments, the metabolic parameter comprises lactate. In some embodiments, the metabolic parameter comprises glucose. In some embodiments, the metabolic parameter has a mean absolute percentage error (MAPE) of less than about 15%. In some embodiments, the metabolic parameter has a MAPE of less than about 14%. In some embodiments, the metabolic parameter has a MAPE of less than about 13%. In some embodiments, the metabolic parameter has a MAPE of less than about 12%. In some embodiments, the metabolic parameter has a MAPE of less than about 11%. In some embodiments, the metabolic parameter has a MAPE of less than about 10%. In some embodiments, the metabolic parameter has a MAPE of less than about 9%. In some embodiments, the metabolic parameter has a MAPE of less than about 8%. In some embodiments, the metabolic parameter has a MAPE of less than about 7%. In some embodiments, the metabolic parameter has a MAPE of less than about 6%. In some embodiments, the metabolic parameter has a MAPE of less than about 5%. In some embodiments, the metabolic parameter has a MAPE of less than about 4%. In some embodiments, the metabolic parameter has a MAPE of less than about 3%. In some embodiments, the metabolic parameter has a MAPE of less than about 2%. In some embodiments, the metabolic parameter has a MAPE of between about 4% and 14%. In some embodiments, the metabolic parameter has a MAPE of between about 4% and 13%.

[0107] In some embodiments, the plurality of ex vivo organ perfusion time series data entries comprise synthetic lung disease data. In some embodiments, the lung disease comprises chronic obstructive pulmonary disease (COPD). In some embodiments, the lung disease comprises idiopathic pulmonary fibrosis (IPF). In some embodiments, the lung disease comprises acute respiratory distress syndrome (ARDS). In some embodiments, the lung disease comprises pulmonary hypertension (PAH). In some embodiments, the lung disease comprises asthma. In some embodiments, the lung disease comprises a disease with underlying pathological explanations that alter the mechanism of breathing.

[0108] In some embodiments, the one or more lung parameters generated comprise one or more of dynamic compliance, peak pressure, mean pressure, or stress index.

[0109] During EVLP, measurements can be taken at multiple times, for example, every hour or every 30 minutes. Accordingly, a data entry can include a time component and the training data can comprise a sequence of parameters as a time-series data.

[0110] In some embodiments, the data entries comprise EVLP time series data entries.

[0111] As demonstrated herein, lung parameters measured at an initial stage of a EVLP case can predict lung parameters at a later stage of the EVLP case.

[0112] In some embodiments, the data entries comprise EVLP time series data entries from a first assessment period. The trained machine learning model can generate one or more lung parameters for a second assessment period. In some embodiments, the second assessment period occurs after the first assessment period.

[0113] The present method can be used to forecast one or more lung parameters at one or more time points in the second assessment period and is not restricted to any time interval represented by the data entries from the first assessment period. For example, the data entries from the first assessment period can be hourly parameters, and the generated one or more lung parameters for the second assessment period can be any specified second, minute and / or hour after initiation of perfusion. The second assessment period can be longer than the first assessment period. Accordingly, the output can, for example, be represented by a graph showing values over a period of time.

[0114] Any suitable machine-learning model can be used. For example, if a model is trained on time-series data collected from the initial stage of a clinical EVLP case to forecast parameters of the rest of the EVLP case, machine-learning algorithms such as a CNN-based encoder-decoder model may be used.

[0115] The data entries can comprise data from one of more lung lobes. The data entries can be of any animal original, optionally a mammal origin, optionally a human origin.

[0116] In some embodiments, the data entries comprise data from one or more lung lobes.

[0117] In some embodiments, the data entries comprise data from animal, optionally from mammal, optionally from human.

[0118] The method disclosed herein encompasses combining data of more than one lung simulation into a single model or output. Liver digital twins

[0119] In some embodiments, the organ is a liver.

[0120] In some embodiments, the one or more organ parameters are one or more liver parameters.

[0121] In some embodiments, the one or more liver parameters comprise one or more selected from the group consisting of: a physiological parameter, a biochemical parameter, a transcriptomic parameter, a metabolomic biomarker parameter, a proteomic biomarker parameter, and an imaging parameter. In some embodiments, the one or more liver parameters comprise a physiological parameter. In some embodiments, the one or more liver parameters comprise a biochemical parameter. In some embodiments, the one or more liver parameters comprise an omics parameter. In some embodiments, the one or more liver parameters comprise a transcriptomic parameter. In some embodiments, the one or more liver parameters comprise a metabolomic biomarker parameter. In some embodiments, the one or more liver parameters comprise a proteomic biomarker parameter. In some embodiments, the one or more liver parameters comprise an imaging parameter.

[0122] In some embodiments, the physiological parameter comprises one or more selected from the group consisting of: Hepatic Artery Flow, Portal Vein Flow, Pressure of Inferior Vena Cava, Pressure of Hepatic Artery, Pressure of Portal Vein, and Temperature.

[0123] In some embodiments, the biochemical parameter comprises one or more selected from the group consisting of: chloride, sodium, potassium, and pH.

[0124] In some embodiments, the metabolomic biomarker parameter comprises one or more selected from the group consisting of: glucose, lactate, AST, and ALT.

[0125] In some embodiments, the plurality of ex vivo organ perfusion time series data entries comprise synthetic liver disease data. II. System and Computer-Readable Medium

[0126] In another aspect, provided herein is a system for generating a digital twin of an isolated organ, comprising: (a) a memory to store executable components; (b) a processor, operably linked to the memory to implement the executable components, the executable components comprising: (i) a model-receiving platform for receiving a trained machine learning model generated from training data comprising a plurality of ex vivo organ perfusion time series data entries from a first assessment period; (ii) a simulation platform that simulates one or more organ parameters for a first forecast period at least in part by processing the time series data entries with the trained machine learning model, wherein the first forecast period occurs after the first assessment period; and (iii) an output-generating component to generate a simulation output comprising the generated one or more organ parameters from the first forecast period; and (c) a user interface to display the simulation output.

[0127] In another aspect, provided herein is a system for generating a digital twin of an isolated organ, comprising: (a) a memory to store executable components; (b) a processor, operably linked to the memory to implement the executable components, the executable components comprising: (i) a model-receiving platform for receiving a trained machine learning model generated from training data comprising a plurality of donor chart data entries; (ii) a simulation platform that simulates one or more organ parameters for a first forecast period during ex vivo organ perfusion at least in part by processing the data entrieswith the trained machine learning model; and (iii) an output-generating component to generate a simulation output comprising the generated one or more organ parameters for the first forecast period; and (c) a user interface to display the simulation output.

[0128] As used herein, the term “system” may include a hardware and / or software system that operates to perform one or more functions.

[0129] In yet another aspect, provided herein is a computer-readable storage medium with instructions stored thereon that causes a system to: (a) receive a trained machine learning model generated from training data comprising a plurality of ex vivo organ perfusion time series data entries from a first assessment period; (b) simulate one or more organ parameters for a first forecast period, wherein the first forecast period occurs after the first assessment period; and (c) generate a simulation output comprising the simulated one or more organ parameters for the first forecast period.

[0130] In yet another aspect, provided herein is a computer-readable storage medium with instructions stored thereon that causes a system to: (a) receive a trained machine learning model generated from training data comprising a donor chart plurality of data entries; (b) simulate one or more organ parameters for a first forecast period during ex vivo organ perfusion; and (c) generate a simulation output comprising the simulated one or more organ parameters for the first forecast period.

[0131] A “computer-readable storage medium” refers to any available medium that can be accessed by one or more local or remote computing devices and may be random access memory (RAM), read only memory (ROM), electrically erasable programmable read only memory (EEPROM), flash memory or other memory technology, compact disk read only memory (CD-ROM), digital versatile disk (DVD), Blu-ray disc (BD) or other optical disk storage, magnetic storage devices, solid state drives or other solid state storage devices, and / or any other storage device or storage disk in which information is stored for any duration. A computer-readable storage memory can include any non-transitory computer readable storage medium. The term “non-transitory” is expressly defined to include any type of computer readable storage device and / or storage disk and to exclude propagating signals and transmission media.

[0132] The computer-readable storage medium of the present disclosure can be stand-alone, or can be incorporated into other computer-readable storage media, for example, a software for ex vivo organ perfusion.

[0133] The computer-readable storage medium can be distributed physically (e.g. through the use of a physical drive) and / or over the internet (e.g. via cloud storage). III. Use

[0134] The digital twin of the present disclosure can be used in a variety of applications.

[0135] For example, cohorts of control organs in a clinical study for organ transplantation can be replaced or supplemented by the digital twin of the present disclosure.

[0136] The digital twin can be used to simulate organs of any condition, such as healthy organs, diseased organs, injured organs, etc.

[0137] The digital twin can be used in determining effects of therapies on rehabilitating organs. For example, simulation of a healthy organ and a diseased organ can be compared to actual data from a diseased organ challenged with a therapy intervention, and the data can allow a user to discern the relative improvement of the organ in the treatment data compared to baseline disease data and normal health data for a given donor profile parameters. This can allow determination of when to insert or return the treated organ to the recipient, and or to determine the safety and efficacy of a therapeutic.

[0138] The digital twin can be used in applications relating to transplantation. For example, by understanding organ function in the hours ahead, clinical teams can intervene and plan patient logistics.

[0139] The digital twin can be used in out-of-body treatment (e.g. targeted gene therapy).

[0140] The digital twin can be used outside of a clinical setting, for example as a research tool.

[0141] The digital twin can also be used in veterinary settings.

[0142] Accordingly, in a further aspect, provided herein is use of the digital twin generated by the disclosed method or the disclosed system in a study, optionally a clinical study. Also provided is use of the digital twin as a research tool.

[0143] In some embodiments, the study is a therapeutic study or a transplantation study.

[0144] In some embodiments, the digital twin is used as a control, optionally a paired- control.

[0145] In some embodiments, the study is a human clinical study. In some embodiments, the study is a veterinary clinical study. IV. Definitions

[0146] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains.

[0147] The term “about” or “approximately” when used in reference to a particular recited value, means the value may vary from the recited value by no more than 10%, 5%, 2% or 1%. Where a particular value is recited, it can be understood that the value is modified by the term “about” or “approximately”, unless indicated otherwise.

[0148] The term “comprise”, “comprising”, or the like means additional elements or components other than those recited may be present. Other terms such as “include”, “contain”, “have” and the like have similar meaning.

[0149] The term “consist of”, “consisting of” or the like means no additional component is present.

[0150] As used herein, the singular forms “a,” “an,” and “the” include plural references, unless indicated otherwise. For example, a reference to “a molecule” can be a reference to more than one molecule.

[0151] The term “or” is used to mean “and / or”, unless it is indicated explicitly to refer to alternatives only. It should also be noted that the term “or” is generally employed in its sense including “and / or” unless the content clearly dictates otherwise. These terms canconvey that any combination is specifically contemplated. Solely for illustrative purposes, the expression “A, B, and / or C” can mean A individually; B individually; C individually; A and B; B and C; A and C; and A, B, and C

[0152] The above disclosure generally describes the present application. A more complete understanding can be obtained by reference to the following specific examples. These examples are described solely for the purpose of illustration and are not intended to limit the scope of the application. Changes in form and substitution of equivalents are contemplated as circumstances might suggest or render expedient. Although specific terms have been employed herein, such terms are intended in a descriptive sense and not for purposes of limitation. EXAMPLES Example 1

[0153] Digital twins are comprehensive computer-based representations of physical objects. A true digital twin in healthcare and medical research remains unrealized (1,2). Current reports of digital twins in medicine are generally extensions of prognostic disease models that do not fully depict the systems-level nature of digital twin technology (3–5). Using advanced engineering, transplant surgeons have developed biomedical platforms for ex vivo (outside the body) organ perfusion (EVOP), which enables the comprehensive study of isolated human organs under physiological conditions outside the body (6). Ex vivo systems exist for human lungs, livers, hearts, pancreases, and kidneys and, historically, have been used to assess the condition of donor organs for transplant suitability prior to surgery (7-11). However, true multi-modal digital twins of human organs remain to be developed.

[0154] A method for creating detailed digital twins of human lungs using the ex vivo lung perfusion (EVLP) platform is described herein (Fig. 1a). EVLP is a self-contained, closed circuit that perfuses an isolated human lung at 37°C (body temperature) while allowing the organ to ‘breathe’ and perform its physiological function of gas exchange using an ICU- grade mechanical ventilator (7) (Fig. 1b). During EVLP, each monitor in the system generates organ-specific data in real-time, enabling a detailed, comprehensive, multi-modal assessment of lung function (Fig.1c).

[0155] We have accumulated experience with over 1,000 clinical EVLP cases and have used this data to improve the precision of decisions about organ suitability for transplantation (12).

[0156] EVLP has been used as a platform to evaluate therapeutics, including preclinical studies on: CRISPR gene editing of lungs (13), stem cell therapeutics (14), blood group modification prior to surgery (15), and immunomodulation (16). However, the current approach to evaluating these new therapeutics involves lengthy and costly preclinical studies. While EVLP has significantly advanced the assessment of novel therapeutics, two central problems remain for clinical research: (1) low study enrollment (i.e., the scarcity of available organs); and (2) poorly defined control groups – the inherent variance in the baseline function of human lungs limits the ability to efficiently conduct case-matched or pooled cohort analyses.

[0157] These challenges are mirrored in preclinical pharmaceutical drug development pipelines that normally require around 12-15 years to complete, involve significant resource investment (around $2.8 billion) (17), and may only offer modest assurance of therapeutic efficacy in humans. Ultimately, the major bottleneck in research of novel therapies stems from the reliance on poorly representative preclinical animal models and two-arm study designs that require a separate control group for baseline comparisons.

[0158] To date, the approach of leveraging ex vivo and machine learning to build digital organ twins has not been demonstrated.

[0159] Herein, we show that the EVLP system provides access to unparalleled biospecimens and data that enable multi-modal computational models of human lungs at a systems-level that can be used to forecast physiological (i.e., lung mechanics, gas exchange), biochemical (i.e., pH, electrolytes), proteomic biomarkers, metabolomic biomarkers, transcriptomic, and radiomic (i.e., imaging) data from human lungs (Fig. 1 c). We provide evidence that the ex vivo approach using digital twins can result in highly accurate machine-learning models, capable of precisely twinning the organ. Importantly, we validated the power of the digital twin approach to advance therapeutic research programs by demonstrating drug safety and efficacy of a pharmacological intervention applied to human lungs during EVLP.Example 2

[0160] Real-Time Data Acquisition and Analysis during EVLP

[0161] Rationale: Current EVLP protocols involve hourly monitoring of: lung physiology (i.e., gas exchange (pO2and pCO2), static and dynamic compliance, airway pressure), biochemistry (i.e., glucose, pH, electrolytes), and biology (i.e., cytokine, chemokines) (Table 1). These protocols simplify functional organ assessments to only 2-4 data points per feature and do not realize the full breadth of time series data that is available (i.e., 100 to >1000 data points). As such, the ability to generate digital twins and forecast lung function is limited by the current approach to EVLP data collection. This study aims to build a real-time data acquisition system that is compatible with the EVLP platform to facilitate real- time data monitoring and recording. Table 1: Standard donor characteristics and lung function assessments during EVLPLegend: Donor features are extracted from patient records; Physiological measurements are obtained during EVLP; Biochemical and biological features were measured in EVLP perfusate solution. pO2= oxygen partial pressure; pCO2= carbon dioxide partial pressure; LA=left atrial; PA=pulmonary artery; IL-8=interleukin-8; IL-6=interleukin-6; IL-10=interleukin- 10; IL-1β=interleukin-1beta; BMI=body mass index; DBD=donation after brain death; DCD=donation after cardiac death; PEEP=positive end-expiratory pressure.

[0162] High-resolution flow and pressure data from an ICU-grade ventilator (Maquet Servo-i, Siemens Healthineers) were recorded at 100Hz during a human EVLP case. Using this highly resolved time series data, software code was developed to construct the ventilation waveform and perform continuous, breath-by-breath analysis of critical lung function parameters, such as dynamic compliance (Fig.2). Moreover, feasibility testing for real-time data monitoring in EVLP perfusate using a porcine model of EVLP was performed. High-resolution data representative of lung physiology (i.e., pO2, pCO2) and biochemistry (i.e., pH values, electrolytes) were acquired using a blood parameter monitor designed for cardiopulmonary bypass surgery (CDI Blood Parameter Monitoring System 550, Terumo Cardiovascular) (Fig.3).

[0163] The data suggests: (1) real-time data acquisition and recording of human lungs during EVLP is completely feasible using conventional equipment developed for patient monitoring, and; (2) the EVLP platform can be leveraged to generate high-resolution time- series data that can be aligned for complex, multi-modal analysis of lung function. Example 3

[0164] Establishing a real-time data acquisition system to capture high- resolution profiles of lung function during EVLP.

[0165] The ventilator-specific code developed in Example 2 was validated. To do this, data (i.e., static and dynamic compliance, etc.) from n=20 EVLP cases were recorded and analyzed by data extraction algorithm, and compared to manual values derived by a highly- trained lung intensivist. Clinical Laboratory Standards Institute method comparison statistical techniques (i.e., accuracy, difference plots, bias plots, correlation plots) were be conducted to test for any differences in the values obtained from the automated and manual approaches.

[0166] The feasibility of real-time data recording in human EVLP cases using the CDI Blood Parameter Monitor was tested. Real-time perfusate data (i.e., pO2, pCO2, pH, K+, etc.) was compared to values obtained using a standard blood gas analyzer (RADIPPoint 500e Blood Gas System, Siemens Healthineers) using the statistical techniques noted above. Perfusate cytokines were measured using a rapid cytokine assay (TORdx LUNG, SQIDiagnostics Inc.), and additional acid-base assessments (i.e., HCO3- and base excess) were derived from pH and similarly analyzed.

[0167] Algorithms for the automated analysis of real-time EVLP data were developed. Using Python and R programming, automatic analysis looping or time-series analysis code (i.e., tsfresh in Python) on lung waveform data described above was constructed. This approach extracts all of the relevant data features (i.e., min / max values, trends, etc.) for downstream analysis by the digital twin model. With highly resolved time series data across multiple modalities, it was possible to align and compute novel functional assessments of human lungs on EVLP such as: stress index (i.e., lung overdistention vs. atelectasis) and shunt fraction (i.e., perfusate that bypasses oxygenation). The analysis of each new feature was automated alongside the standard lung function parameters.

[0168] All real-time inputs for united multi-modal data storage and analysis were be combined. Using the Microsoft Azure cloud-computing platform, a central storage and processing system for real-time EVLP dataset were developed. All of the data monitoring systems (i.e., ventilator and perfusate monitors) were linked and sent to a central network that ran the code developed in above to extract, analyze, and store EVLP data in real-time via the Azure cloud. Example 4

[0169] Development and Validation of Digital Twins for Ex Vivo Donor Lungs

[0170] Rationale: The concept of a digital twin (in-silico model) is based on an accurate statistical simulation that represents a physical object and predicts its behaviour and performance. Recently, deep neural networks have become accepted approaches to analysis for time series data―similar to the EVLP data that will be generated in Example 3. In contrast to simpler models such as regression and KNN, deep learning can pick up subtle and non-linear relationships from large-scale longitudinal data. Successful applications of deep neural networks outperforming traditional machine learning methods have been published across many fields of healthcare and medicine. A functional digital twin of the EVLP lung will be developed using convolutional neural networks (CNN) and recurrent neural networks (RNN), which have been shown to capture the temporal sequence of multiple parallel time-series parameters. Briefly CNN and RNN are deep learning approaches toprediction where CNNs use filters within convolution layers to transform data and RNNs learn from and use other data points in a sequence to generate the next output in a series. Moreover, CNNs have been applied to recognize time-series data as 1-dimensional arrays effectively. The DT-EVLP model can be used to forecast key functional parameters of a lung on EVLP, including biophysical / biomechanical, biochemical, and biological features. With real-time monitoring of lung function (Examples 2 and 3), thousands of time series data points across different monitoring systems can be available as input data for the DT-EVLP forecasts.

[0171] Experiments on data forecasting using a retrospective cohort of n=476 historical EVLP cases with tabular (hourly) data were conducted. Briefly, data was split 80:20 for training (n=381) and testing (n=95) using a k-nearest neighbor (KNN) model. Univariate forecasting predictions were conducted for: lung compliance, pO2, pCO2, pH and K+measurements—key features that are essential to determining lung quality. Data collected at 1, 2, and 3 hours of perfusion were used to predict the 4thhour values recorded during EVLP. Strong agreement of predicted and observe values was observed―within 10% of the true value for: 83% of dynamic compliance (R2=0.94), 79% of static compliance (R2=0.94), 86% of pO2(R2=0.66), 81% of PCO2(R2=0.82), and 100% of pH (R2=0.80) and K+(R2=0.91) measurements in the test set. pO2is a critical functional parameter and, importantly, the standard deviation of predicted pO2measurements was only 24mmHg. Moreover, a KNN forecasting model was built using the waveform pO2data collected in the porcine experiment in Example 2. Fig. 4 illustrates improved forecasting using real-time data obtained during EVLP with predicted pO2values within 2mmHg (<2%) of observed values.

[0172] The data suggests that accurate forecasting of EVLP data is possible. Using a limited dataset (i.e., only 3 datapoints per feature) and a simple KNN forecast model, it was possible to predict future hourly EVLP clinical parameters. As expected, the ability to accurately forecast will depend on the time-series behaviour of a given feature (i.e., the waveform data of lung compliance versus pH). Example 5

[0173] Developing a DT-EVLP model

[0174] The DT-EVLP model of a human lung can be constructed using the high- resolution time series data obtained from n=248 clinical EVLP cases (training dataset = 80%), which can provide an opportunity for time-series forecasting using deep learning models. Input features can include donor characteristics (i.e., age, sex, BMI, cause of death, lung function before organ recovery), standard EVLP assessments of biophysical / biomechanical, biochemical, and biological features (Table 1) measured in real-time, as well as any new functional assessments obtained in Example 3. The time-series data for each EVLP assessment feature is expected to be unique (i.e., Figs.2 and 3) and different deep learning approaches for the DT-EVLP model will be investigated (Summarized in Table 2). Different RNNs (i.e., long short-term memory (LSTM) models, gated recurrent units,), transformers and Gaussian process approaches can be tested. Similarly, several CNN approaches, such as WaveNet, T-WaveNet, Smooth-CNN can be tested. Table 2: Characteristics of deep learning approaches for the development of DT-EVLP

[0175] A reductionist approach can be used for modeling strategy. First, all time series data collected up until 3.5 hours of EVLP can be used to predict the 4thhour measurements. The data window can be decreased by 30 minutes (i.e., use all data up to 3 hours) to forecast the remainder of the case. This process can be repeated until the first hour is reached, as the EVLP circuit reaches 37°C and full flow at this time and any intervention can be administered thereafter. Fine tuning of DT-EVLP hyperparameters can be performed and the model that reports the highest degree of accuracy can be selected for validation.

[0176] To validate the DT-EVLP model, an equivalence study can be conducted to compare observed and predicted (DT-EVLP) lung function parameters in n=62 clinical EVLP cases (test dataset = 20%). The DT-EVLP model performance can be considered validated if all of the predicted EVLP features have an R2>0.95 and a mean absolute error of less than 5% from the observed values. Example 6

[0177] Utility of DT-EVLP: Optimizing Lung Recruitment during EVLP

[0178] Rationale: Lung recruitment is a technique whereby an increase in pressure is applied to the airways in an attempt to open up any alveoli that have collapsed in order to improve oxygenation. Donor lungs can often have significant portions that are derecruited during the organ recovery, transport and preservation processes. Lung injury is exacerbated by poor recruitment and, if transplanted, this can lead to poor recipient outcomes, such as prolonged mechanical ventilation and / or early mortality. Thus, there is a need to improve donor lung recruitment. If successful, this will lead to better patient post-transplant outcomes and more lungs available for transplantation that would otherwise have been discarded due to poor oxygenation results. Importantly, the EVLP platform provides an opportunity to improve oxygenation during tidal ventilation; however, current EVLP protocols do not titrate PEEP (positive end expiratory pressure) to optimize lung recruitment, but rather follow a fixed PEEP strategy (i.e., 5cmH20). Studies have suggested that a tailored approach to PEEP can improve lung recruitment and oxygenation.

[0179] As noted earlier, the central challenges in conducing a study investigating optimal PEEP for EVLP involves obtaining and defining a control group. The classical design for similar studies would be a two-arm trial where one cohort of donor lungs receive thestandard EVLP protocol and the other arm would be the optimized PEEP group. As such, a single-arm study can be conducted where an optimized PEEP protocol is applied to all donor lungs and the digital twins of the study lungs can represent lung function under standard conditions.

[0180] For each lung, real-time data can be collected and used to generate a digital twin with the DT-EVLP model. During EVLP, a decremental PEEP strategy can be employed from the 2ndhour onwards to identify the PEEP level associated with the highest lung compliance. Briefly, during each hourly assessment, a pressure-volume curve can be derived prior to conducting a recruitment maneuver. PEEP is then set to a high level (14cmH2O) and decreased by 2cmH2O every 60 seconds (10 breaths) (i.e., PEEP settings = 14-12-10-8- 5cmH2O). The PEEP setting that results in the highest compliance value can be identified as the best PEEP during the titration. PEEP can then be set to the selected value until the next lung assessment. This PEEP selection procedure can be repeated for the duration of EVLP.

[0181] An optimized PEEP strategy can result in improved lung oxygenation and this can be demonstrated by a significant increase in observed pO2compared to the levels predicted by the DT-EVLP model. The primary endpoint can be pO2levels in EVLP perfusate. Secondary outcomescan be assessed against the predicted DT-EVLP counterpart in a similar manner and will include measurements of: pCO2, pH, cytokine levels, glucose, acid- base chemistry, and electrolyte levels plus, any features arising from Example 3. Additional secondary analyses can compare the predicted patient outcome using data derived from DT- EVLP to the actual patient outcome for: (1) transplant suitability; and (2) duration of mechanical ventilation post-transplant.

[0182] Example 7

[0183] Two background steps to verify the feasibility of the digital lung program are described below.

[0184] First, high-resolution flow and pressure data were recorded from an ICU-grade ventilator during a human EVLP case. Using this highly resolved time series data, software code was developed to construct the ventilation waveform and perform continuous, breath- by-breath analysis of critical lung function parameters, such as dynamic compliance. Moreover, feasibility testing for real-time data monitoring in EVLP perfusate was conductedusing a porcine model of EVLP. High resolution data representative of lung physiology (i.e., pO2, pCO2) and biochemistry (i.e., pH values, electrolytes) were acquired using a specialized monitor.

[0185] Secondly, using a retrospective cohort of n=476 historical EVLP cases with tabular (hourly) data, experiments on data forecasting were conducted. Briefly, data was split 80:20 for training (n=381) and testing (n=95) using a k-nearest neighbor (KNN) model. Univariate forecasting predictions were conducted for: lung compliance, pO2, pCO2, pH and K+measurements—key features that are essential to determining lung quality. Data collected at 1, 2, and 3 hours of perfusion were used to predict the 4thhour values recorded during EVLP. Strong agreement of predicted and observed values was observed―within 10% of the true value for: 83% of dynamic compliance (R2=0.94), 79% of static compliance (R2=0.94), 86% of pO2(R2=0.66), 81% of pCO2(R2=0.82), and 100% of pH (R2=0.80) and K+(R2=0.91) measurements in the test set. The standard deviation of predicted pO2measurements was 24mmHg. A KNN forecasting model was built using the waveform pO2data collected in the porcine experiment that demonstrated improved forecasting using real- time data obtained during EVLP with predicted pO2values within 2mmHg (<2%) of observed values.

[0186] Studies were performed to demonstrate the digital organ program’s feasibility.

[0187] Briefly, a neural network-based machine learning model accurately predicted future lung function during EVLP assessments for: dynamic compliance (mean absolute error (MAE)=10.3 mL / cmH2O vs. clinical benchmark=15.8mL / cmH2O); stress index (MAE=0.07); mean pressure (MAE=0.23cmH2O vs. clinical benchmark=0.33cmH2O); and peak pressure (MAE=0.9cmH2O vs. clinical benchmark=1.6cmH2O). Example 8

[0188] This example describes the development of a deep learning approach to simulate lung function by leveraging high-resolution time-series lung ventilation data and demonstrate its predictive utility during EVLP.

[0189] Methods: Flow and pressure data were recorded at a resolution of 100Hz using the Servo-i ventilator (Maquet, USA) during n=50 clinical EVLP cases performed at ourinstitution from 2019-2022, generating n>1.2x106breath datapoints during EVLP. A convolutional neural network (CNN)-based encoder-decoder model was trained using data from a previous assessment period to forecast data for the next EVLP assessment period (i.e., 3h simulation based on 2h assessment data) for: lung compliance, peak and mean airway pressures, and stress index profiles. A long-short term memory (LSTM) model was then trained to predict transplant suitability using either the observed (actual) or forecasted (simulated) lung function data.

[0190] Results: The CNN-based model accurately predicted future lung function during EVLP assessments (Figure 5). Importantly, there was no difference in the predicted EVLP outcome using the actual or simulated multivariate ventilation data as inputs to the LSTM model (p=0.88).

[0191] Conclusion: This study demonstrates a deep learning method for analyzing EVLP ventilator waveform data, enabling precise simulation of isolated donor lung function. This approach offers an efficient means to expedite clinical research by replacing the control arm of clinical trials with accurate digital lung simulations. Example 9

[0192] Section 1: Predicting hourly EVLP parameters using EVLP tabular data

[0193] Section 1 mainly demonstrates that we can use the information in donor charts (i.e. donor height, donor weight, donor sex, etc.) to forecast hourly EVLP physiological parameters. This serves as one of the approaches to build the digital twin model.

[0194] Here, we describe two classes of Machine Learning Models trained to take in features of donated lungs placed on Ex-Vivo Lung Perfusion (EVLP) to predict physiological parameters that are measured hourly during EVLP.

[0195] Donor information that serves as the input for these models includes donor age, donor sex, donor height, donor weight, donor smoking history (positive or negative history), donor cause of death, donor type (cardiac death or brain death), lung on EVLP (Left, Right, or Both), and EVLP cold ischemic time (CIT). EVLP CIT is a phase of donor lung preservation following organ retrieval and preceding EVLP, when the lung is kept in a hypothermic state without a blood supply.

[0196] The outputs are physiological parameters that serve as markers of lung function during EVLP. They include lung dynamic compliance (cDyn), lung peak pressure (pPeak), lung mean pressure (pMean), and lung plateau pressure (pPlat). Each of these parameters are measured hourly during EVLP for 4 hours.

[0197] The first class of model is a K-Nearest Neighbour (KNN), a simple non- parametric Machine Learning algorithm. In total, we devised 16 unique KNN models for the purpose of predicting specific physiological parameters for a given hour of EVLP from donor features of the lung (4 physiological parameters × 4 hourly measurements = 16 different KNN models). The Mean Absolute Error, Mean Squared Error, training set size, and testing set size for each KNN model are reported in Table 1. Available data for each model was divided into training and testing sets in an 80 / 20 train / test split. As an EVLP procedure progresses, they may be terminated short of 4 hours, resulting in smaller training and testing sets for the machine learning models as the predict parameters later in the EVLP.

[0198] The second class of model is an Extreme Gradient Boosting (XGBoost) model, a more advanced Machine Learning algorithm. Like the KNN, we devised 16 different XGBoost models to predict the same hourly physiological parameters over the course of EVLP. The Mean Absolute Error, Mean Squared Error, training set size, and testing set size for each XGBoost model are reported in Table 2. These models were trained on the same data sets as the KNN, with an 80 / 20 split to create the training and testing seats for model development. Table 3. Performance and training details on the K-Nearest Neighbour models that use donor features of donated lungs to predict hourly EVLP physiological parametersTable 4. Performance and training details on the XGBoost models that use donor features of donated lungs to predict hourly EVLP physiological parameters

[0199] Forecasting time-series EVLP ventilation parameters during assessment period

[0200] This example demonstrates that the time-series data collected from the initial stage of a clinical EVLP case can be used to forecast the key physiological parameters of the rest of the EVLP case. This serves as another approach to build the digital twin model.

[0201] 1. Data structure

[0202] The model takes in per-breath ventilation parameter value sequence from an assessment period to forecast the per-breath ventilation parameters of the next assessment for an EVLP case. In total, there were 94 input-output pairs from n = 60 clinical EVLP cases. The ventilation parameters include dynamic compliance, peak pressure, mean pressure, and stress index. The length of the input and output sequence can be determined as a parameter that can be fine tuned or adjusted based on the EVLP protocol.

[0203] 2. Models

[0204] The CNN-based encoder-decoder model is built for univariate sequence-to- sequence prediction of EVLP ventilation parameters. The encoder portion of the model takes in the input sequence (a sequence of ventilation parameters as a time-series data), generates a latent representation of the input sequence, then projects it to map with the output sequence. The model achieved a mean absolute error (MAE)=10.3 mL / cmH2O vs. clinical benchmark=15.8mL / cmH2O for dynamic compliance; an MAE=0.23 cmH2O vs.clinical benchmark=0.33cmH2O for mean pressure; an MAE=0.9 cmH2O vs. clinical benchmark=1.6cmH2O for peak pressure; and an MAE=0.07 for stress index.

[0205] Representative graphs are shown in Figure 6.

[0206] Other example methodologies to build digital organs using ex vivo systems can include: (1) MLP (multi-layer perceptron) models: We have currently tested this approach: The MLP-based mixing model handles the input through N number of layers of MLP. The input data is mixed along the time and feature domains as it passes through each layer of MLP. N is a parameter that can be finetuned during training. The model achieved an MAE of 6.5 mL / cmH2O for dynamic compliance. (2) Recurrent neural network (RNN)-based encoder-decoder models (3) Transformer-based models such as: • PatchTST • FEDformer • Inverted Transformer • Foundation model: TimeGPT • Autoformer • Large Language Models as time-series forecasters (4) Regression models such as: Lasso, Ridge, Elastic Net Example 10 Methods

[0207] Study Population:

[0208] Inclusion and ethics: A total of n=1,000 consecutive EVLP cases conducted at Toronto General Hospital from 2008-2024 were reviewed for inclusion in the study cohort. In accordance with the Declaration of Helsinki, University Health Network (UHN) Research Ethics Board (REB) and institutional approval was obtained for the collection, storage, and analyses of the biospecimens and data used in this study (UHN REB#12-5488-13 and UHN REB#11-0170-AE); informed patient consent was obtained from study participants. Afterreview of each EVLP case file, n=49 cases were excluded from analyses for data missingness. A total of n=951 cases of isolated human lung perfusion were included for digital twin modeling and analyses in this study. During the study period, a total of n=16 EVLP cases were treated with Alteplase and, therefore, left out of digital twin model training. Sample sizes for each of the data modalities are described in Fig.7.

[0209] EVLP Protocol:

[0210] All ex vivo lung perfusion procedures were conducted according to the previously reported Toronto EVLP protocol (7). Briefly, after lung retrieval, the left atrium and pulmonary artery were cannulated and connected to the EVLP circuit for perfusion with an acellular low potassium, dextran solution. Target flow of the circuit was set at 40% of predicted cardiac output based on the size of the human lung donors. Circuit flow started as 10% of the target flow and incrementally increased to reach full target flow within the first hour of lung perfusion. Circuit temperature started at 20°C and gradually increased to reach a physiological temperature of 37°C within the first 30 minutes of perfusion to avoid injuring the cold-stored lungs. Lung ventilation started after 30 minutes of perfusion. An ICU-grade ventilator (Servo-i (Maquet, NJ, USA) or Bellavista (Vyarie Medical, IL, USA)) was connected via an endotracheal tube and set to flow-controlled ventilation. Under baseline breathing setting, ventilation settings were set at a tidal volume (TV) of 7 mL / kg of predicted body weight, a constant positive end-expiratory pressure (PEEP) of 5 cmH2O, a respiratory rate of 7 breaths / minute and fraction of inspired oxygen (FiO2) of 21%. During all EVLP cases, an evaluation of lung function was performed hourly. During this assessment period, the ventilation settings were set to a TV of 10 mL / kg of predicted body weight, PEEP of 5 cmH2O, a respiratory rate of 10 breaths / minute and an FiO2 of 100%.

[0211] Biospecimen and Data Collection:

[0212] For all EVLP cases, the lung function data collection methodology was consistent with the Toronto EVLP Protocol (7). A complete list of all human lung data obtained during EVLP can be found in Table 5. Physiological parameters were recorded hourly and at high-frequency (1 data point / 10 ms) using an ICU-grade ventilator (Servo-i (Maquet, NJ, USA) or Bellavista (Vyarie Medical, IL, USA)). Lung perfusate samples were extracted from the sampling port of the reservoir on the EVLP system. Perfusate sampleswere used for immediate testing using arterial blood gas analyzers (GEM Premier 5000 (Werfen, Catalonia, Spain) and RAPIDPoint 500 (Siemens, Munich, Germany)) for gas exchange, metabolomic, and biochemical data. Additional perfusate aliquots were frozen and stored at -80°C for downstream proteomic biomarker measurement using a multiplexed enzyme-linked immunosorbent assay (ELISA) (Ella, Protein Simple, CA, USA) for four proteomic biomarkers: interleukin (IL)-6, IL-8, IL-1ß, and IL-10. Pulmonary arterial and left atrial pressure data were recorded with a continuous pressure monitor (IntelliVue MX450 (Philips, MA, USA)). Pre- and post-EVLP lung tissue biopsy samples of approximately 1cm x 3cm were collected using a mechanical stapler and snap-frozen in liquid nitrogen for microarray analysis (Clariom D Assay (Thermo Fisher Scientific, MA, USA)) (42,43). Hourly lung edema was estimated as the volume of perfusate solution not returned to the EVLP circuit reservoir. Radiographic (X-Ray) images were taken at 1 and 3 hours of perfusion using standard portable radiography (DRX-Revolution, Carestream Health, Rochester, NY, USA) by medical radiology technicians (27). The clinical history of each lung was recorded at the time of organ donation and accessed using the Toronto Lung Transplant Program (TLTP) database. Table 5. Human lung cohort characteristicsLegend: SD=standard deviation. %=percentage of each category of the total n cases

[0213] Data Preprocessing:

[0214] An automated data preprocessing pipeline was coded using Python to preprocess each of the five main data modalities for a given EVLP case: 1) high-resolution (10 ms) time-series ventilator data; 2) hourly lung function parameters; 3) lung X-ray images; 4) transcriptomics; and 5) proteomic biomarker data.

[0215] For high-resolution ventilator flow and pressure data, breath segmentation was first performed to segment the entire ventilator waveform into individual breaths with identified inspiratory and expiratory phases. Next, a list of four physiological parameters (dynamic compliance, peak airway pressure, mean airway pressure, and expiratory volume) were extracted for every breath identified in the ventilator waveform for each case. For this study, a total of n=211,240 individual breaths were labelled using a semi-supervised approach guided by clinical specialists. One of six clinically relevant labels were assigned to each breath: “Normal”, “Assessment”, “Bronchoscopy”, “Recruitment”, “Inspiratory Pause”, and “Noise”. Normal and assessment breath labels were used in forecasting analysis tasks, while the rest of the labels were included in a detailed data review to derive clinical feature annotations: clinical intervention status of bronchoscopy, lung recruitment, respiratory pauses as yes / no indicator variables to denote if certain clinical intervention happened between each pair of two consecutive hourly assessments; and the number of inspiratory pauses that happened during each of the hourly assessments as count values. These clinical intervention features were used to enable subsequent multi-modal forecasting tasks.

[0216] Hourly lung function features were retrieved for all historical cases from the TLTP database. X-ray image features were extracted using latent features from the last convolutional layer of a convolutional neural network (CNN)-based approach (ResNet-50) according to previous reports (26).

[0217] Principal component analysis was used to fit the image feature vectors from the last convolutional layer before the fully connected layer to extract the top ten principal components (PC). Microarray data from lung tissue biopsies were analyzed using Gene Set Enrichment Analysis (GSEA, version 4.3.3). Single-sample GSEA (ssGSEA), an extension of GSEA, was used to calculate separate enrichment scores for the Hallmark gene sets (44) for each of the paired human lung samples. Protein concentrations were measured usingthe Ella multiplex ELISA platform and exported from the analysis software (Simple Plex Runner, version 4.0.0.28).

[0218] Digital Twins of Human Lungs Model Development:

[0219] For every assessment metric generated by an ex vivo human lung, a specific multi-modal time-series forecasting model was developed, optimized, and validated for that parameter. All code pipelines were developed using the following libraries: Pandas – 2.2.2; NumPy – 1.26.3; PyTorch – 2.3.1; XGBoost – 2.10; Scikit-learn – 1.5.1; SHAP – 0.46.0; SciPy – 1.13.0 in Python (version 3.11.8). Each model included the optimal multi-modal set of other lung features to accurately predict the target feature.

[0220] All forecasting models employed two digital twin approaches: ‘static’ and ‘dynamic’ digital twins of human lungs (Fig.8). The static digital lung used only the baseline (i.e., 1st hour) multi-modal lung function data as input to forecast multiple future time points (i.e., 2nd and 3rd hour lung function), whereas the dynamic digital lung received continuous updates from recently available data (i.e., 1st and 2nd hour) and recalibrated to forecast future time points (i.e., 3rd hour). Mean Absolute Error (MAE), which measures the difference between the forecasted and observed values, was used as the primary evaluation metric for loss function optimization during model training. Model performance reported throughout was the MAE and Mean Absolute Percentage Error (MAPE, %) derived from the validation cohort from k-fold cross-validation for a given feature and model.

[0221] Tabular Data Forecasting: XGBoost regressor models (45) were used to perform multi-modal forecasting for hourly lung physiological parameters, biochemical parameters, metabolomic biomarkers, image PC features, transcriptomics, and proteomic biomarkers.

[0222] For lung physiology, biochemistry, and metabolomic biomarker forecasting, all models were designed using a multivariate input approach where all available physiological and biochemical parameter values were used to forecast each of the 17 hourly features for both the static and dynamic digital lung forecasting approaches. A total of 68 XGBoost regressor models were constructed for 17 parameters. For each parameter, four XGBoost models were built to forecast each of the following: static approach to forecast: 1) the 2nd hour parameter and 2) the 3rd hour parameter based on the baseline multi-modal input data;3) static approach to forecast the 3rd hour parameter based on the 1st baseline and forecasted 2nd hour data; and 4) dynamic approach to forecast the 3rd hour parameter based on the observed baseline and 2nd hour data. All 68 models were trained and fine-tuned using 20-fold cross-validation on either n=754 EVLP cases for 2nd parameter forecasting or n=700 cases for 3rd hour parameter forecasting. Using a similar approach, an additional 4 XGBoost regressor models were trained for hourly edema forecasting on n=608 EVLP cases for 2nd parameter forecasting or n=586 cases for 3rd hour parameter forecasting.

[0223] The XGBoost regressor model was used to forecast 10 clinically relevant X-ray features (top 10 PCs) at the third hour using 10-fold cross-validation. A total of 20 models were trained and finetuned – for each PC values, two models were built to forecast the 3rd hour PC values using either the observed 2nd hour lung physiology, biochemistry and metabolomic data (dynamic approach) or the forecasted 2nd hour values noted above (static approach) (Fig.9). To evaluate the clinical relevance of image PC features, we selected four radiological labels that were significantly correlated with observed image PC features at the 3rd hour (27): the degree of two clinical radiology findings: consolidation and infiltrate, (for each lung lobe: 0-none, 1-minimal, 2-moderate, 3-pronouced), and the likelihood of each of the two diagnoses: pneumonia and aspiration (1-very unlikely, 2-unlikely, 3-intermediate, 4- likely, 5-very likely). Pearson correlation analyses were conducted between each of the clinical labels and the observed image PC values, PC values forecasted by static digital lung, and PC values forecasted by dynamic digital lung, respectively.

[0224] The XGBoost regressor model was further used to predict post-EVLP transcriptomic-level changes of the 50 Hallmark gene pathway enrichment scores (44). A total of 50 models were constructed and fine-tuned with 20-fold cross-validation on n=88 cases to predict each of the 50 gene set enrichment scores using observed baseline, 2nd, and 3rd hour lung physiology, biochemistry, metabolomic biomarkers and 3rd hour image features as multi-modal input, representing the dynamic digital lung approach. To establish the static approach, an additional 50 models were trained to forecast post-EVLP transcriptomic changes for each of the 50 gene sets using forecasted 2nd and 3rd lung physiology, biochemistry, metabolomic biomarkers, and 3rd hour image PC features.

[0225] The proteomic biomarker forecasting model was trained using an XGBoost regressor model to forecast each of the four proteomic markers at the 2nd and 3rd hour timepoints. Using a similar approach, the dynamic digital lung model combined observed lung physiology, biochemistry, metabolomic biomarkers, and image features as model input, whereas the static digital lung approach utilized the forecasted values. In addition, the proteomic biomarker XGBoost model also incorporated linear trend values of each of the proteomic biomarkers up to the timepoint to be forecasted. A total of 82 cases were included in model training with 20-fold cross-validation.

[0226] High-Resolution Data Forecasting: Gated recurrent unit (GRU) models (46) were trained on high-resolution breath-by-breath parameter data of dynamic compliance, peak airway pressure, mean airway pressure, and expiratory volume. An example time- series trace of dynamic compliance of a single case is shown in Fig.10. The model used a time-series segmental approach to focus on forecasting breaths in selected windows which contain different lengths of continuous breaths from lung assessment periods. In this study, for the 1st hour assessment (A1), three different sizes for input selected windows were selected: 1) first 50 breaths of A1; 2) first 50 breaths and last 50 breaths of A1; and 3) 20 normal breaths before A1 combined with first 50 breaths and last 50 breaths of A1. These three different input sequence set-ups were each used to build a static digital lung model to forecast all four breath parameter values for every breath of the first 50 breaths from the 2nd hour lung assessment period (A2). The same process was repeated to build a static digital lung model to forecast all four breath parameter values for every breath of the first 50 breaths from the 3rd hour lung assessment period (A3). Then, each of the three selected window set- ups were further combined with the first 50 breaths of observed A2 to forecast all four parameter values for every breath of the first 50 breaths from the 3rd hour lung assessment period (A3), demonstrating the dynamic digital lung approach. Lastly, for the static digital lung approach to forecast 3rd hour breath parameter values, the above-mentioned static and dynamic models were stacked sequentially to derive forecasted 2nd hour breath parameters values, which were then further concatenated with the observed 1st hour baseline data as input sequence to forecast 3rd hour breath parameters. In total, 48 different models based on the combination of 4 parameters x 3 selected window set-ups x 4 digital lung approaches(3 static and 1 dynamic) were tested. Each model was trained and fine-tuned with 20-fold cross-validation and repeated over 10 seed values to ensure model stability.

[0227] Breath traces outside of lung assessment periods were reviewed to derive additional static features to aid time-series forecasting of breath parameters (refer to ‘Data Preprocessing’ for more details). These additional static features together with hourly lung function data were passed as separate channels to the GRU model for the forecasting tasks. Following the same set-ups, for the GRU model with static feature channels, a total of 48 models were trained using 20-fold cross-validation and repeated over 10 seed values, resulting in a total of 96 models and 960 rounds of training for high-resolution data forecasting.

[0228] The simulated lung profile used to demonstrate the full utility of the digital lung model in Figs.11A-11Q was generated by building a k-nearest neighbor model (KNN) based on historical cases in the TLTP database matched for: age, sex, height, weight, brain or cardiac death, and total lung capacity (size). The lung profile was first generated by randomly sampling for age, sex, height, weight, brain / cardiac death and size. Then, these anchoring features were used as input features of the KNN model to find n=3 closest neighbors (cases) in the historical database for each data modality. Next, for each parameter to be synthesized in hourly tabular data, image features and omics, the average value of the 3 neighboring cases was taken as the synthetic value for each corresponding parameter at each timepoint. For time-series per-breath parameter values, the same procedures were repeated except for the average parameter values of every breath that were calculated based on n=2 neighboring cases to preserve as much inter-breath parameter variability as possible.

[0229] Clinical utility of Digital Twin Model in Preclinical Therapeutic Development:

[0230] During clinical EVLP, a 20 mg dose of Alteplase (tissue-type plasminogen activator (tPA)) was administered to human lungs with clinical suspicion of pulmonary emboli (Figs. 12A-12B). Alteplase was administered via the perfusate reservoir (pulmonary vasculature) after the first (baseline) assessment of the ex vivo lung. A total of 16 EVLP cases received Alteplase treatment in the study cohort. Two cases were excluded from the analysis due to incomplete data arising from <2 h of perfusion on EVLP. The Alteplase- treated lungs were used as a real-world digital twin utility validation cohort in this study. Themechanism of action of Alteplase is thrombolysis that leads to reduced pulmonary vascular resistance (PVR), a measurement derived using the following formula:

[0231]

[0232] In the study cohort, the left atrial pressure was constantly maintained at 4 mmHg; therefore, pulmonary arterial pressure was the key modulator of PVR and identified as the primary endpoint to evaluate the therapeutic efficacy of Alteplase. To further evaluate the therapeutic safety profile, hourly lung edema (mL) was measured and compared in Alteplase-treated lungs. The conventional control cohort was generated by randomly sampling the historical population of untreated lungs on EVLP. A p-value of 0.05 was considered significant for this comparison. Example 11

[0233] Human Lung Characteristics for Digital Twin Development

[0234] The clinical profiles of the human lungs used to develop the digital twin models are summarized in Table 6. A total of 951 isolated lungs assessed for transplant suitability using the ex vivo circuit were included in this study. Multi-modal lung function was evaluated on the EVLP circuit for at least three hours. The study population represented lungs from disease-free organ donors (i.e., no lung-specific diseases or chronic illnesses), with a mean age of 46 ± 16 years and 64% of the cohort was male. The average lung volume was 6.5 ± 1.3 L, which was representative of a typical adult population. Donor cause of death was primarily due to anoxia (40%) or cerebrovascular injury (32%). Table 7 provides a comprehensive list of all lung assessments, including sampling modalities and measurement frequencies, incorporated into the digital twin model. Table 6. Human lung cohort characteristicsTable 7. Full panel of parameters included in digital twin model developmentExample 12

[0235] Digital Twins of Human Lungs Accurately Forecast Human Lung Function

[0236] Lung Physiology

[0237] A total of thirteen different physiological parameters were accurately forecasted by static and dynamic digital lungs using a multi-modal machine learning approach (Table 8). The lung's capacity to stretch and expand during ventilation was assessed using measurements of dynamic compliance, mean and peak airway pressure, and expiratory volumes. These measurements were derived for every single breath of the lung using high- resolution ventilator flow and pressure data (10 ms) (Fig.10) and accurately predicted using a multi-modal GRU approach. The MAE and MAPE values reported for each of the per- breath parameters in Table 8 represented the average MAE and MAPE across 50 forecasted breaths. The MAPE values were within 2% to 7% of the observed values for each metric (Table 8). Other physiological measurements, such as gas exchange ability (e.g., partial pressures of O2 and CO2), were observed hourly and forecasted using a multi-modal XGBoost approach, achieving an MAPE of 4% to 13% (Table 8). Notably, the digital twin approach predicted lung edema within approximately 30 mL of observed values (Table 8) – a key factor in the safety profile of many therapeutics. Table 8: Digital twins of human lungs accurately forecast physiological and biochemical lung function

[0238] For all physiological parameters, the dynamic digital lung approach showed similar or improved accuracy compared to the static digital lung approach (Table 8). However, the static digital lung approach provided multiple pathways to predict future lung function: directly from baseline or via predicted values at recently available timepoints that were used to constantly recalibrate the model. Table 9 shows that a static digital twin model can also directly forecast 2nd and 3rd hour assessments from the baseline with high accuracy. Table 9. Complete performance characteristics of static and dynamic digital lung strategies for digital twins of human lungs on lung physiology, biochemistry, metabolomics, and proteomics

[0239] In addition to improved forecasting accuracy, high-resolution physiological parameters enabled deep-learning-based forecasting of parameter values for every breath in selected forecasting windows with varying look-back window sizes (e.g., number of breaths = 50, 100, 120, or 170) for all static and dynamic digital lung forecasting set-ups (Table 10). The GRU model forecasting results showed substantial improvement compared to a non-ML approach (e.g., linear models) for both static and dynamic approaches. Furthermore, for assessments such as dynamic compliance, the addition of clinical static features improved prediction MAEs (Table 10). Table 10. Complete performance characteristics of static and dynamic digital lung strategies for digital twins of human lungs on high resolution breath parameter forecasting

[0240] Lung Biochemistry

[0241] Lung biochemistry was assessed in the perfusate solution during EVLP for several electrolytes (e.g., [sodium] and [potassium]) and acid-base indicators (e.g., pH and [bicarbonate]). The static and dynamic digital lung approach correctly forecasted all seven biochemical evaluations with an MAPE of 0.83-13% and 0.63-7.2%, respectively (Table 8). Similar to physiological assessments, the dynamic digital lung approach demonstrated improved forecasting precision across all parameters. Notably, the dynamic digital lung modeling approach dramatically improved the acid-base prediction MAPE – pH (from 0.85% to 0.46%), bicarbonate (from 5.5% to 2.7%) and base excess (from 13% to 7.2%) – compared to the static digital lung results (Table 8).

[0242] Lung –Omics

[0243] Metabolomic Biomarkers

[0244] Cellular metabolism was modelled by the digital twin using glucose and lactate levels in the EVLP perfusate. Overall, both the static and the dynamic digital lungs demonstrated good performance in forecasting glucose (within 0.3-0.5mM) and lactate (within 0.5-0.9mM) levels (Table 11). As previously observed, the addition of further data via the dynamic digital lung models predicted glucose and lactate levels with improved accuracy, resulting in MAPE values of 6.9% for lactate and 4.9% for glucose (Table 11). Table 11: Digital twins of human lungs accurately forecast lung -omics

[0245] Transcriptomics

[0246] The ability to forecast transcriptomic changes in a lung was enabled by tissue biopsies collected before and after EVLP. Thus, using the pre-EVLP biopsy as a transcriptomic baseline, a multi-modal approach incorporating additional lung evaluations enabled the digital twin to predict the post-EVLP transcriptome. Table 11 highlights that the ex vivo digital lung can accurately forecast transcriptomic changes in lung disease-related pathways18–25 with high accuracy (MAPE: 1-3%). For example, both the static and dynamic digital lung models accurately forecasted the gene enrichment score of the TGF- β signaling pathway, hypoxia pathway, and TP53 pathway within 1% of the observed values. Additional gene sets were observed to have similar forecasting results (Table 12). Unlike the lung physiology and biochemistry parameters, for transcriptomic changes prediction, the dynamic model did not always perform better than the static digital lung (Table 12). Table 12. Digital twins of human lungs can accurately forecast transcriptomics changes of Hallmark gene pathways

[0247] Enrichment scores are unitless measurements that reflect the activity level of the biological processes associated with genes in a particular gene set and MAPE values can be less clinically representative and interpretable. Therefore, a key aspect of transcriptomic analysis, beyond knowing the gene set enrichment score, is identifying temporal changes in gene set enrichment for a given lung. Importantly, the ex vivo digital twin approach demonstrated an accuracy of 79-100% to correctly identify the temporal changes in gene enrichment (i.e., upregulation) (Table 13). Notably, the per-case prediction of the hypoxia pathway reached 100% accuracy, showing both the static and dynamic digital lungs can perfectly predict the directional change of the hypoxia pathway for individual cases (Table 13). Table 13. Pre-case prediction accuracy of transcriptomic changes of lung disease-related pathways using static and dynamic digital twin strategies

[0248] Proteomic Biomarkers

[0249] EVLP enables the study of circulating proteins in the perfusate solution. Using an ELISA-based approach, we observed that the dynamic digital lung approach can forecast the level of IL-10 within 23 pg / mL, IL-1J3 within 2.5 pg / mL, IL-6 within 6.2 ng / mL, and IL-8 within 0.53 ng / mL of the observed values (Table 11). In the study cohort, we observed larger than expected MAPE values due to the extreme range of protein concentrations with some approaching zero and causing the average percentage error values to be skewed and less representative of the forecasting performance. Importantly, the model showed a medianMAPE of 23% for IL-10, 26% for IL-1J3, 22% for IL-6, and 37% for IL-8, which was aligned with the forecasting accuracy for other parameters.

[0250] Lung Imaging

[0251] The EVLP platform provides a unique opportunity to image the lung in isolation, without the many confounding factors present in a traditional chest X-ray, such as the heart, chest wall, and ribs. X-ray images taken at one and three hours of perfusion were acquired and studied for digital twin development. A convolutional-neural network approach was used to derive the top ten principal components as previously described (26). A multi-modal XGBoost model was able to precisely forecast the third hour image PCs (Table 14). Since PC values are unitless, MAE and MAPE are less clinically intuitive for interpreting model results. It has been previously reported that the image PCs are strongly correlated to radiographic findings and diagnoses (27). Importantly, both static and dynamic digital lung forecasts of image PC values maintained strong correlations with important clinical radiology findings (e.g., consolidation and infiltration) and diagnoses (e.g., aspiration and pneumonia), underscoring the accuracy of the digital twin forecasting approach and its ability to identify clinically significant image features (Table 15). Table 14. Digital twins of human lungs accurately forecast X-ray image-derived featuresTable 15: Digital twins of human lungs accurately forecast X-ray image features and radiological findingsExample 13

[0252] A Digital Twin of a Human Lung using EVLP

[0253] EVLP provides an opportunity to gather complex and multi-modal data of an isolated human lung for digital twin testing. Figs.11A-11Q illustrates a static and dynamic model for a comprehensive, multi- modal digital twin using synthetic data. A k-nearest neighbour approach (described in Methods) was used to construct a complete baseline profile of lung physiology, biochemistry, -omics, and imaging for a human lung on the ex vivo platform (Figs. 11A-11Q). Consistent with model performance metrics, each functional assessment of the test case was accurately forecasted after 2 and 3 hours of perfusion. Example 14

[0254] Digital Twins of Human Lungs Enhance the Evaluation of Therapeutics

[0255] To expand on the findings of more efficient clinical studies using digital twins, real-world evidence was evaluated for human lungs treated with the therapeutic Alteplase. Alteplase, a tissue-type plasminogen activator that is often used clinically to lyse clots, has thrombolytic effects and reduces vascular resistance (28,29). During the multiorgan recovery process, lungs are at a heightened risk of injury related to pulmonary emboli formation and EVLP offers an opportunity to evaluate and treat this risk using Alteplase (28).

[0256] Based on our clinical protocol, Alteplase was administered after baseline data (i.e., 1st hour data) was acquired. Using the 1st hour baseline data as input to the digitaltwin model allowed for the generation of untreated counterfactual outcomes of Alteplase, further enabling the use of causal inference to control for confounding factors and identifying individual treatment effects based on the digital twin-based forecasted values and observed data. Thus, the therapeutic efficacy and safety profile of Alteplase-treated lungs were assessed using a paired digital twin approach (i.e., each treated lung was compared to its corresponding untreated digital twin). Pulmonary arterial pressure (PAP) was evaluated as the primary efficacy endpoint for therapeutic benefit, while an estimate of edema was used as a metric to assess the drug's safety profile. Fig.12A and 12B illustrate gross macroscopic evidence of pulmonary emboli in a human lung. Fig. 12C and Fig. 12D demonstrate therapeutic benefit (lowered PAP) and safety (no additional edema formation) of Alteplase for a treated lung vs. its corresponding digital twin.

[0257] To further evaluate the benefits of the digital twin in preclinical research, pooled statistical analyses were performed (Figs. 12E-12L, Table 16). For lungs with function suitable for transplantation at the end of EVLP (n=6) (Table 16), we observed a similar safety and efficacy response to Alteplase. For lungs that did not appear to respond to Alteplase (n=8) (Table 16), lung function was not suitable for transplantation and was associated with poor-patient outcomes. Using a conventional cohort approach, an independent, two-sample comparison between the Alteplase-treatment group (i.e., responders) versus the conventional control cohort revealed no significant differences in PAP or edema at 1 or 2 hours after treatment (Figs.12E, 12G, 12I, 12K). Importantly, using digital twins as controls allowed for paired statistical analyses of the observed vs. forecasted PAP and edema values. The digital twin approach identified significantly lower PAP two hours after treatment (Fig. 12h and Table 16). Correspondingly, the effects of Alteplase were only observed in lungs that were deemed suitable for transplantation with good post-operative outcomes (Table 16). In lungs where there was no observed impact of Alteplase, lung function did not improve to the point of the organ being deemed suitable for transplantation with good patient outcomes (Table 16). Moreover, while edema values were not significantly different 120 minutes post- treatment in the suitable lung cohort, there was a trend towards the treated lungs showing improved lung function, contributing to the observed lower edema values 120 minutes post- treatment (Fig.12l).Table 16. Digital lungs demonstrated clinical utility in evaluating preclinical therapeutic efficacy and safety of Alteplase

[0258] The digital twin framework described herein can effectively reflects the concept of a true digital twin. Since the ex vivo model was built using disease-free lungs, the digital twins created can be representative of general lung health. The ex vivo platform, for example, can be leveraged to study damaged or diseased lungs, enabling the digital twin approach disclosed herein to be precisely tailored to specific disease models, such as those for idiopathic pulmonary fibrosis (IPF) and chronic obstructive pulmonary disease. The ex vivo approach can enable the collection of a greater volume, frequency and variety of biospecimens (e.g., tissue biopsies and perfusate) and data (e.g., mechanical ventilation measurements and imaging), which would otherwise be too invasive to gather directly from patients. This comprehensive data collection can enable multi-modal parameter forecasting and a complete systems-level representation of the organ. Additionally, the ex vivo approach can allow for real-time data collection and monitoring. Currently, ventilation data is recorded with a frequency of 1 data point / 10 ms. Example 15

[0259] FIG.17 is a diagram illustrating generation of a lung digital twin according to certain embodiments.

[0260] The input data to the machine learning system can include high-resolution time- series data. The input data to the machine learning system can include physiology (e.g., i.e., pO2, pCO2) and biochemistry (e.g., pH values, electrolytes) data. The input data to themachine learning system can include metabolomic biomarkers (e.g., glucose and lactate) and protomic (e.g., IL-6, IL-8, IL-10, IL-1β) biomarkers. The input data to the machine learning system can include image data. The image data can include X-ray images of lung tissue. The input data to the machine learning system can include transcriptomics (e.g., gene enrichment scores).

[0261] The input data can be pre-processed before it is provided to the machine learning system. Pre-processing data may include adjusting or reducing the dimensionality of the data and / or can include generating an alternative representation of the data. In some implementations, pre-processing can include generating additional data that can be provided to the machine learning system.

[0262] The high-resolution time-series data can be pre-processed to generate breath- by-breath parameters. The physiology and biochemistry data can be pre-processed to generate hourly parameter values. The metabolomic and protein biomarkers can be pre- processed to generate hourly biomarker values. The image data can be processed using principal components analysis (PCA) to generate a subset of highly predictive image features (e.g., ten features). The transcriptomics data can be processed to generate gene enrichment scores using single-sample Gene Set Enrichment Analysis (GSEA).

[0263] One or more machine learning models can process the input data and / or any intermediate products derived from pre-processed input data to generate forecasted parameters for a lung digital twin. For example, a machine learning system can include a gated recurrent unit (GRU) to process the breath-by-breath parameters to forecast breath parameters for the digital twin. The machine learning system can also include an XGBoost regressor to process the hourly parameter values, hourly biomarker values, the principal components, and the gene enrichment scores to produce forecasted values of these parameters.

[0264] FIG. 18 is a diagram illustrating a digital twin of a generic organ pipeline according to certain embodiments. FIG.18 shows that many of the same sub-systems that could be used to generate the digital twin of the lung (e.g., from FIG.17) can be re-purposed to generate a digital twin or another organ.

[0265] The ex vivo perfusion system generates multi-modal, organ specific data. Ex vivo perfusion system data can be used to generate digital twins of other organs, such as the heart, kidney, and the pancreas. Example parameters can include, e.g., for a heart, time- series electrocardiogram data coming from a heartbeat. The parameters may be characterized by descriptive statistics, such as averages (means, medians, or modes), expected values, variances, covariances, standard deviations, distributions (e.g., a normal or Gaussian distribution, or a binomial distribution), or other descriptive statistics.

[0266] First, the system can perform data interpretation to place the structure and format of the data in a medical context in accordance with a clinical perfusion procedure. For example, organ assessment might introduce some specific patterns to the data that can be difficult to understand without understanding the clinical workflow first.

[0267] The model can be pretrained on literature-informed synthetic data. A computer program or model can process literature (e.g., public datasets, scientific publications (journal articles)) to generate synthetic values for parameters related to organ functioning (e.g., within feasible physiological ranges).

[0268] The input data can be pre-processed before it is provided to the machine learning system. Pre-processing data may include adjusting or reducing the dimensionality of the data and / or can include generating an alternative representation of the data. In some implementations, pre-processing can include generating additional data that can be provided to the machine learning system.

[0269] Once the model is pre-trained, the pre-processed ex vivo data can be input to the model to fine-tune the model. After the fine-tuning is complete, the model can be used to forecast parameters for an organ during a second time period when given a set of input parameters for the organ in a first time period, hence serving as a digital twin for the organ.

[0270] FIG. 19 is a diagram illustrating a disease-specific digital twin pipeline according to certain embodiments.

[0271] An organ-specific digital twin model of a healthy organ (e.g., the organ-specific model of FIG.18) can be maintained in computer memory (e.g., in a database). This organ- specific digital twin model can be used to generate a digital twin of a diseased organ.

[0272] First, the model is pretrained on disease-specific data. This data can be synthetic data. Synthetic data can be generated from a literature review. A machine learning system can analyze large volumes of text to determine example hourly parameters indicating organ function. This data can be collected and used in a dataset. Then, pretraining can be performed on the disease-specific synthetic data.

[0273] Pre-processing data may include adjusting or reducing the dimensionality of the data and / or can include generating an alternative representation of the data. In some implementations, pre-processing can include generating additional data that can be provided to the machine learning system.

[0274] Then, the model can be fine-tuned on real data. Real data can be collected from hospitals and labs and input to the model over several iterations to improve the model’s accuracy. The real data can be real ex vivo data on diseased organs or ex vivo perfusion of diseased organs. The real data can be provided to the model that has been pretrained on the pre-processed data for fine-tuning.

[0275] Once the model has been fine-tuned, it can be provided with data to generate a disease-specific digital twin of the organ for a future time period.

[0276] FIG.20 is a diagram illustrating generation of a liver digital twin according to certain embodiments.

[0277] Data may be generated using a combination of synthetic and real ex vivo liver data. The data may be ex vivo perfusion of a diseased liver. The synthetic data may be received from an extensive literature review of time series parameters of livers.

[0278] Pre-processing data may include adjusting or reducing the dimensionality of the data generating an alternative representation of the data, or otherwise modifying the data to convert that data to a format that a machine learning model can understand and process. In some implementations, pre-processing can include generating additional data that can be provided to the machine learning system.

[0279] Once the data has been pre-processed, the model can be pre-trained on the synthetic data. Then, the model may be fine-tuned using ex vivo liver data. Multiple modelscan be used to forecast different liver parameters. Hence, a digital twin model of the liver is generated. Example 16

[0280] Digital Twin Model of Diseased Lungs

[0281] Ex vivo lung perfusion (EVLP) provides a unique opportunity to preserve, stabilize and assess donor lungs outside of the body. EVLP also acts as a therapeutic platform to treat donor lungs to improve lung function and repair lung injuries. Previously demonstrated digital twins of ex vivo lungs can be further extended to create digital twins of diseased lungs - human lungs with characteristics of lung diseases such as Chronic obstructive pulmonary disease (COPD) and idiopathic pulmonary fibrosis (IPF).

[0282] Time-series ventilator flow, pressure, and volume data obtained during EVLP can be used as the foundation to model lung disease characteristics. Here is shown a method to transform breath pressure-volume (PV) loop to build lung disease synthetic data that can be used to model the functional behaviour of a diseased lung on ex vivo perfusion. Firstly, the method involves sigmoidal curve fitting on PV curves of individual breaths to derive four parameters that describe the shape of the curve. Two sigmoidal curves are fitted to the PV curve - one for the inspiratory portion of the curve, and the other for the expiratory portion of the curve. By applying rotation transformation of every data point relative to the first data point at the start of a breath, one can derive PV curves that capture COPD / IPF characteristics based on ventilator data from real non-diseased ex vivo lungs (Figure 14). In the example, it is shown that a counterclockwise 10 degree rotation to represent COPD lung characteristics and a clockwise 10 degree rotation to represent IPF characteristics. Then, the flow and pressure ventilator traces are derived based on the transformed curve (Figs.15A-15B). The derived flow and pressure data represent breaths associated with COPD / IPF characteristics. The generation of such datasets enabled an important step towards digital twins of diseased lungs by allowing for pre-training, fine-tuning, or directly running inference on our existing digital twin model. Here it is shown an example of inference testing on transformed lung data using the existing digital twin model and its forecasting performance for both the static digital lung and dynamic digital lung approaches (Figs.16A-16B).

[0283] Data may be generated using a combination of synthetic and real ex vivo data. The data may be ex vivo perfusion of a diseased lung. The synthetic data may be received from an extensive literature review of time series parameters of lungs.

[0284] Pre-processing data may include adjusting or reducing the dimensionality of the data generating an alternative representation of the data, or otherwise modifying the data to convert that data to a format that a machine learning model can understand and process. In some implementations, pre-processing can include generating additional data that can be provided to the machine learning system.

[0285] Once the data has been pre-processed, the model can be pre-trained on the synthetic data. Then, the model may be fine-tuned using ex vivo diseased lung data. Multiple models can be used to forecast different lung parameters. Hence, a digital twin model of the diseased lung is generated. Example 17

[0286] Digital Twin of Ex Vivo Human Liver

[0287] Background: Ex vivo liver perfusion provides the opportunity to preserve, stabilize and assess liver function outside the body. During liver perfusion, multi-modal liver function data is generated and can be modelled using artificial intelligence to construct digital twins of ex vivo livers. Briefly, a digital twin of an ex vivo liver is a machine learning-based time-series forecasting model that accurately predicts a full suite of liver functional data, such as hepatic artery flow, pH, temperature, bile production, and liver functional enzymes, etc.

[0288] Methods: Using liver function data from an ex vivo liver perfusion system, an extensive literature research was conducted to identify and verify data frequency, ranges, central tendency (i.e., mean, median), trends, and output format for each parameter of liver function.

[0289] Results: To demonstrate the forecasting ability of the ex vivo liver digital twin model, we applied a random forest model to predict future liver functional parameters shown in Figure 13. The prediction accuracy was measured using mean absolute error (MAE) and was reported in Table 17 for a list of parameters. In addition to a random forest approach to modeling, a digital twin of ex vivo human livers can be modeled by various machine learningmodels that have the capacity to perform time-series forecasting, including but not limited to: deep learning neural network such as recurrent neural network, transformer-based architectures, tree-based models, regression, representation learning, contrastive learning strategies. Table 17: Forecasting performance of the digital twin of ex vivo livers using a random forest model

[0290] Ex vivo liver data preprocessing

[0291] Each parameter in the list of liver functional data (Table 17) was 24 hours long. Due to different sampling frequency, the parameters with the frequency of 1 data point / hour would have 24 data points in raw data, and the parameters with the frequency of 1 data point / second would have 86400 data points in raw data.

[0292] In this example, a random forest model was used with a sliding window approach (sliding window size = n) to generate the training and test datasets: the first n data points would be used as input and the n+1th data point would be used as output. Following this approach, we formed input-output pairs to train the random forecast model. In this example, for a parameter with 24 data points and a sliding window size of n = 3, 21 input- output pairs were formed and used in the forecasting example. As Figure 13 shows, for example, the glucose trace has 24 observed data points and 21 predicted data points. On the other hand, due to the high dimension of parameters with frequency of 1 data point / second, we downsampled the 86400 data points to keep one data point every 1000 seconds, reducing the data size by 1000 fold. With a sliding window size of n = 10, the parameters were forecasted every 1000s starting at the 11th data point. The above mentioned example is showing one of the many approaches one can use to forecast the time-series traces of liver function data.24 hours time duration is also for example purpose, the data should not be limited by the time duration, it can be any time duration should the data is available for the specific organ. Machine Learning Models

[0293] The following sections generically describe some of the machine learning models to process organ data and should not be construed to limit any aspects or teachings of this disclosure. Random Forests

[0294] Random Forests are a type of machine learning algorithm used for classification, regression, and other predictive tasks. They operate by constructing a multitude of decision trees during training, with each tree learning from a random subset of the input data and features. The output of the Random Forest is determined by aggregating the predictions made by each individual tree. For classification tasks, this aggregation typically takes the form of a majority vote, while for regression tasks, the average of the individual tree predictions is used.

[0295] Each decision tree in the forest is grown by recursively splitting the data based on feature values, with the goal of maximizing the homogeneity of the target variable within each resulting subset. To ensure diversity among the trees, the Random Forest algorithmintroduces two main sources of randomness: (1) bootstrapping, where each tree is trained on a random subset of the data with replacement, and (2) random feature selection, where only a subset of the available features is considered for splitting at each node. XGBoost

[0296] XGBoost (Extreme Gradient Boosting) is a machine learning algorithm used for classification, regression, and other predictive tasks, based on an ensemble learning technique known as gradient boosting. XGBoost constructs a series of decision trees, where each successive tree is trained to correct the errors made by the previous trees in the sequence. The model is trained in an iterative manner, with each tree contributing to the overall prediction by adjusting the weights of previously misclassified data points.

[0297] In gradient boosting, the model is built by optimizing a loss function, which measures the discrepancy between the predicted and actual values. The optimization is performed using gradient descent, where the algorithm minimizes the loss by adjusting the weights of the model in the direction of the negative gradient. In XGBoost, this process is enhanced through a combination of techniques, such as regularization, which reduces overfitting by penalizing overly complex trees, and a more efficient tree-building process, which uses a novel split finding algorithm to accelerate training.

[0298] XGBoost incorporates additional enhancements, such as parallelization and support for sparse data, enabling faster model training and more efficient use of computational resources. The model's robustness, accuracy, and flexibility make it suitable for handling large datasets, high-dimensional feature spaces, and various types of data, including structured and unstructured forms. Gated Recurrent Unit (GRU)

[0299] A Gated Recurrent Unit (GRU) is a type of recurrent neural network (RNN) architecture used for sequence modeling tasks, such as time series prediction, speech recognition, and natural lan-guage processing. The GRU is designed to address the vanishing gradient problem commonly encountered in traditional RNNs, allowing it to learn long-term dependencies more effectively.

[0300] The GRU operates by utilizing two primary gates: the reset gate and the update gate, which control the flow of information at each time step.

[0301] Reset Gate: The reset gate determines how much of the previous hidden state should be "forgotten" when computing the new hidden state. It modulates the degree to which the previous information contributes to the current computation.

[0302] Update Gate: The update gate controls how much of the new information should be incorpo-rated into the current hidden state. It decides the balance between retaining the old state and incorporating the new information at each time step.

[0303] In the GRU model, the hidden state at each time step is updated based on a weighted combination of the previous hidden state and the candidate hidden state, which is computed using the reset and update gates. This mechanism enables the model to selectively retain and forget information as it processes sequences, providing flexibility and enabling it to capture long-term dependencies more effectively than traditional RNNs.

[0304] The GRU is typically trained using backpropagation through time (BPTT) or other gradient-based optimization techniques, and is computationally more efficient than more complex architectures like Long Short-Term Memory (LSTM) networks, due to its simpler structure with fewer gates. K-Nearest Neighbors

[0305] K-Nearest Neighbors (K-NN) is a non-parametric, instance-based machine learning algorithm that can be used for classification and regression tasks. The algorithm operates by analyzing the distances between a given query point and a set of labeled data points, then classifying the query point or predicting its value based on the labels or values of the 'k' nearest neighbors in the dataset.

[0306] In the context of classification, the class of the query point is determined by the majority class of its 'k' closest neighbors, where 'k' is a positive integer that determines the number of nearest neighbors to consider. These distances are typically computed using a distance metric, such as Euclidean distance or Manhattan distance, but other distance metrics may also be used depending on the application.

[0307] For regression tasks, the predicted value for the query point is derived by averaging the values of the 'k' nearest neighbors.

[0308] The K-NN algorithm does not require any explicit training phase, as it makes decisions based on the training data during the query phase. This makes it a highly intuitive and simple model, often utilized in scenarios where interpretability and adaptability to different data distributions are important. However, the efficiency of K-NN can be impacted by the size of the dataset and the choice of distance metric.

[0309] While the present application has been described with reference to what are presently considered to be the preferred examples, it is to be understood that the application is not limited to the disclosed examples. To the contrary, the application is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

[0310] All publications, patents and patent applications are herein incorporated by reference in their entirety to the same extent as if each individual publication, patent or patent application was specifically and individually indicated to be incorporated by reference in its entirety. Specifically, the sequences associated with each accession numbers provided herein including for example accession numbers and / or biomarker sequences (e.g. protein and / or nucleic acid) provided in the Tables or elsewhere, are incorporated by reference in its entirely.

[0311] The scope of the claims should not be limited by the preferred embodiments and examples, but should be given the broadest interpretation consistent with the description as a whole.CITATIONS FOR REFERENCES REFERRED TO IN THE SPECIFICATION 1. Sun, T., He, X. & Li, Z. Digital twin in healthcare: Recent updates and challenges. Digit. Health 9, 205520762211496 (2023). 2. Yao, J.-F., Yang, Y., Wang, X.-C. & Zhang, X.-P. Systematic review of digital twin technology and applications. Vis. Comput. Ind. Biomed. Art 6, 10 (2023). 3. Cen, S., Gebregziabher, M., Moazami, S., Azevedo, C. J. & Pelletier, D. Toward precision medicine using a “digital twin” approach: modeling the onset of disease-specific brain atrophy in individuals with multiple sclerosis. Sci. Rep.13, 16279 (2023). 4. Walsh, J. R., Roumpanis, S., Bertolini, D. & Delmar, P. Evaluating Digital Twins for Alzheimer’s Disease using Data from a Completed Phase 2 Clinical Trial. Alzheimers Dement.18, e065386 (2022). 5. Lv, Z., Guo, J. & Lv, H. Deep Learning-Empowered Clinical Big Data Analytics in Healthcare Digital Twins. IEEE / ACM Trans. Comput. Biol. Bioinform.21, 660–669 (2024). 6. Iske, J. et al. Pushing the boundaries of innovation: the potential of ex vivo organ perfusion from an interdisciplinary point of view. Front. Cardiovasc. Med.10, 1272945 (2023). 7. Cypel, M. et al. Technique for Prolonged Normothermic Ex Vivo Lung Perfusion. J. Heart Lung Transplant.27, 1319–1325 (2008). 8. Eshmuminov, D. et al. An integrated perfusion machine preserves injured human livers for 1 week. Nat. Biotechnol.38, 189–198 (2020). 9. Ardehali, A. et al. Ex-vivo perfusion of donor hearts for human heart transplantation (PROCEED II): a prospective, open-label, multicentre, randomised non-inferiority trial. The Lancet 385, 2577–2584 (2015). 10. Parmentier, C. et al. Normothermic Ex Vivo Machine Perfusion of Discarded Human Pancreas Allografts: A Feasibility Study. Transpl. Int.36, 10936 (2023). 11. Hamar, M. & Selzner, M. Ex-vivo machine perfusion for kidney preservation: Curr. Opin. Organ Transplant.23, 369–374 (2018). 12. Sage, A. T. et al. A machine-learning approach to human ex vivo lung perfusion predicts transplantation outcomes and promotes organ utilization. Nat. Commun.14, 4810 (2023). 13. 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).14. Nykänen, A. I. et al. Engineered mesenchymal stromal cell therapy during human lung ex vivo lung perfusion is compromised by acidic lung microenvironment. Mol. Ther. - Methods Clin. Dev.23, 184–197 (2021). 15. 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). 16. Nykänen, A. I. et al. Lung Transplant Immunomodulation with Genetically Engineered Mesenchymal Stromal Cells—Therapeutic Window for Interleukin-10. Cells 13, 859 (2024). 17. Singh, N. et al. Drug discovery and development: introduction to the general public and patient groups. Front. Drug Discov.3, 1201419 (2023). 18. Drakopanagiotakis, F., Xifteri, A., Polychronopoulos, V. & Bouros, D. Apoptosis in lung injury and fibrosis. Eur. Respir. J.32, 1631–1638 (2008). 19. Sauler, M., Bazan, I. S. & Lee, P. J. Cell Death in the Lung: The Apoptosis– Necroptosis Axis. Annu. Rev. Physiol.81, 375–402 (2019). 20. Kiszałkiewicz, J. M. et al. Evaluation of selected IL6 / STAT3 pathway molecules and miRNA expression in chronic obstructive pulmonary disease. Sci. Rep.11, 22756 (2021). 21. Evans, C. E. Hypoxia-Inducible Factor Signaling in Inflammatory Lung Injury and Repair. Cells 11, 183 (2022). 22. Pham-Danis, C. et al. Inflammation promotes aging-associated oncogenesis in the lung. Preprint at https: / / doi.org / 10.1101 / 2024.03.01.583044 (2024). 23. Wu, Q. et al. p53: A Key Protein That Regulates Pulmonary Fibrosis. Oxid. Med. Cell. Longev.2020, 1–13 (2020). 24. Ye, Z. & Hu, Y. TGF‑β1: Gentlemanly orchestrator in idiopathic pulmonary fibrosis (Review). Int. J. Mol. Med.48, 132 (2021). 25. Hou, J. et al. TNF‑α‑induced NF‑κB activation promotes myofibroblast differentiation of LR‑MSCs and exacerbates bleomycin‑induced pulmonary fibrosis. J. Cell. Physiol. 233, 2409–2419 (2018). 26. Chao, B. T. et al. A Multi-Modal Machine-Learning Approach that Combines Imaging and Functional Data Obtained During Ex Vivo Lung Perfusion Significantly Improves Prediction of Transplant Outcomes. J. Heart Lung Transplant.43, S81 (2024). 27. Chao, B. T. et al. A radiographic score for human donor lungs on ex vivo lung perfusion predicts transplant outcomes. J. Heart Lung Transplant.43, 797–805 (2024).28. Machuca, T. N. et al. Injury-Specific Ex Vivo Treatment of the Donor Lung: Pulmonary Thrombolysis Followed by Successful Lung Transplantation. Am. J. Respir. Crit. Care Med. 188, 878–880 (2013). 29. Collen, D. Molecular mechanism of action of newer thrombolytic agents. J. Am. Coll. Cardiol.10, 11B-15B (1987). 30. Gillette, K. et al. A Framework for the generation of digital twins of cardiac electrophysiology from clinical 12-leads ECGs. Med. Image Anal.71, 102080 (2021). 31. Lauzeral, N. et al. A model order reduction approach to create patient-specific mechanical models of human liver in computational medicine applications. Comput. Methods Programs Biomed.170, 95–106 (2019). 32. Dubs, L. et al. Assessment of extracranial carotid artery disease using digital twins – A pilot study. NeuroImage Clin.38, 103435 (2023). 33. Lai, M., Yang, H., Gu, J., Chen, X. & Jiang, Z. Digital-twin-based Online Parameter Personalization for Implantable Cardiac Defibrillators. in 2022 44th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC) 3007–3010 (IEEE, Glasgow, Scotland, United Kingdom, 2022). doi:10.1109 / EMBC48229.2022.9871142. 34. Ross, J. L. et al. Enhancing Longitudinal Clinical Trial Efficiency with Digital Twins and Prognostic Covariate-Adjusted Mixed Models for Repeated Measures (PROCOVA-MMRM). Preprint at https: / / doi.org / 10.48550 / ARXIV.2404.17576 (2024). 35. Zhu, T., Li, K., Herrero, P. & Georgiou, P. GluGAN: Generating Personalized Glucose Time Series Using Generative Adversarial Networks. IEEE J. Biomed. Health Inform. 27, 5122– 5133 (2023). 36. Susilo, M. E. et al. Systems-based digital twins to help characterize clinical dose- response and propose predictive biomarkers in a Phase I study of bispecific antibody, mosunetuzumab, in NHL. Clin. Transl. Sci.16, 1134–1148 (2023). 37. Ahmadian, H. et al. Toward an artificial intelligence-assisted framework for reconstructing the digital twin of vertebra and predicting its fracture response. Int. J. Numer. Methods Biomed. Eng.38, e3601 (2022). 38. Finlayson, S. G. et al. Adversarial attacks on medical machine learning. Science 363, 1287– 1289 (2019).39. Hough, O. et al. Improved ex Vivo Lung Perfusion (EVLP) with Dialysis and Nutrition to Achieve Successful 36h EVLP and Lung Transplantation. J. Heart Lung Transplant.42, S60 (2023). 40. Lau, N.-S. et al. Long-term ex situ normothermic perfusion of human split livers for more than 1 week. Nat. Commun.14, 4755 (2023). 41. Shin, J. et al. Use of metabolomics to identify strategies to improve and prolong ex vivo lung perfusion for lung transplants. J. Heart Lung Transplant.40, 525–535 (2021). 42. Baciu, C. et al. Transcriptomic investigation reveals donor-specific gene signatures in human lung transplants. Eur. Respir. J.57, 2000327 (2021). 43. Allen, J. et al. Ex vivo lung perfusion moderates gene expression differences between cardiac death and brain death donor lungs. JHLT Open 3, 100027 (2024). 44. Liberzon, A. et al. The Molecular Signatures Database Hallmark Gene Set Collection. Cell Syst.1, 417–425 (2015). 45. Chen, T. & Guestrin, C. XGBoost: A Scalable Tree Boosting System. in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining 785–794 (2016). doi:10.1145 / 2939672.2939785. 46. Bahdanau, D., Cho, K. & Bengio, Y. Neural Machine Translation by Jointly Learning to Align and Translate. Preprint at http: / / arxiv.org / abs / 1409.0473 (2016).

Claims

CLAIMS:

1. A method for generating a digital twin of an isolated organ, comprising: receiving, by a computing device, a plurality of ex vivo organ perfusion time series data entries from a first assessment period; generating, by the computing device, one or more organ parameters for a first forecast period at least in part by processing the time series data entries from the first assessment period with a trained machine learning model, wherein the first forecast period occurs after the first assessment period; displaying, by the computing device, using a graphical user interface, a simulation output comprising the generated one or more organ parameters for the first forecast period, thereby generating the digital twin of the isolated organ.

2. The method of claim 1, further comprises receiving, by the computing device, a plurality of donor chart data entries.

3. The method of claim 2, wherein the generating of the one or more organ parameters for the first forecast period further comprises processing the donor chart data entries with the trained machine learning model.

4. The method of claim any one of claims 1-3, further comprising receiving, by the computer device, a plurality of ex vivo organ perfusion time series data entries from at least one additional assessment period.

5. The method of claim 4, wherein the generating of the one or more organ parameters for the first forecast period further comprises processing the time series data entries from the at least one additional assessment period with a trained machine learning model.

6. The method of claim 5, wherein the at least one additional assessment period occurs before the first forecast period.

7. The method of any one of claims 1-6, further comprising generating, by the computing device, one or more organ parameters for at least one additional forecast period at least in part by processing the time series data entries from the firstassessment period with a trained machine learning model, wherein the at least one additional forecast period occurs after the first assessment period and the first forecast period.

8. The method of claim 7, wherein the simulation output further comprises the generated one or more organ parameters for the at least one additional forecast period.

9. The method of claim 8, wherein generating the one or more organ parameters for the at least one additional forecast period further comprises processing the one or more organ parameters for the first forecast period.

10. The method of claim 4, wherein the at least one additional assessment period is a second assessment period.

11. The method of claim 10, wherein the first assessment period is at about one hour of ex vivo organ perfusion.

12. The method of any one of claims 10-11, wherein the second assessment period is at about two hours of ex vivo organ perfusion.

13. The method of any one of claims 10-12, wherein the first forecast period is at about one hour after the second assessment period.

14. The method of any one of claims 1-13, wherein the first forecast period is at about one hour after the first assessment period.

15. The method of claim 8 or 9, wherein the at least one additional forecast period is a second forecast period.

16. The method of claim 15, wherein the second forecast period is at about one hour after the first forecast period.

17. The method of any one of claims 1-6, further comprising generating, by the computing device, one or more organ parameters for a plurality of additional forecast periods at least in part by processing the time series data entries from at least one previous forecast period.

18. The method of claim 17, wherein the simulation output further comprises the generated one or more organ parameters for the plurality of additional forecast periods.

19. The method of any one of claims 1-18, wherein the time series data entries comprise hourly data entries.

20. The method of any one of claims 1-19, wherein the time series data entries comprise breath-by-breath data entries.

21. The method of any one of claims 1-20, wherein the time series data entries comprise one or more selected from the group consisting of: physiological data, biochemical data, transcriptomic data, metabolomic biomarker data, proteomic biomarker data, and imaging data.

22. The method of any one of claims 1-21, wherein the trained machine learning model comprises a neural network.

23. The method of claim 22, wherein the neural network is a CNN or RNN.

24. The method of any one of claims 1-23, wherein plurality of ex vivo organ perfusion time series data entries comprise synthetic data.

25. The method of claim 24, wherein the synthetic data comprise synthetic data transformed from ex vivo organ perfusion time series data entries obtained from a real organ.

26. The method of claim 24 or 25, wherein the synthetic data comprises disease synthetic data.

27. The method of any one of claims 1-26, wherein the organ is a lung, a liver, a heart, a kidney, or a pancreas.

28. The method of any one of claims 1-27, wherein the organ is a lung.

29. The method of claim 28, wherein the data entries comprise data from one or more lung lobes.

30. The method of claim 29, wherein the one or more organ parameters are one or more lung parameters.

31. The method of claim 30, wherein the one or more lung parameters comprise one or more selected from the group consisting of: a physiological parameter, a biochemical parameter, a transcriptomic parameter, a metabolomic biomarker parameter, a proteomic biomarker parameter, and an imaging parameter.

32. The method of claim 31, wherein the physiological data and / or physiological parameter comprises one or more selected from the group consisting of: airpressure, vascular pressure, partial pressure of O2, partial pressure of CO2, lung compliance, expiratory volume, and an estimate of edema or surrogate thereof.

33. The method of claim 31, wherein the biochemical data and / or biochemical parameter comprises one or more selected from the group consisting of: calcium, chloride, sodium, potassium; bicarbonate, base excess, and pH.

34. The method of claim 31, wherein the transcriptomic data and / or transcriptomic parameter comprises gene enrichment scores for one or more lung disease-related pathways.

35. The method of claim 34, wherein the one or more lung disease related pathways are hypoxia, inflammatory response, apoptosis, TP53 signaling, TNF-a signaling, interleukin-s signaling, PI3K / AKT / mTOR signaling, interleukin-6 signaling, TGF-b signaling, and oxidative phosphorylation.

36. The method of claim 31, wherein the metabolomic biomarker data and / or metabolomic biomarker parameter comprises at least one selected from the group consisting of: glucose and lactate.

37. The method of claim 31, wherein the proteomic biomarker data and / or proteomic biomarker parameter comprises at least one selected from the group consisting of: IL-6, IL-8, IL-10, and IL-1β.

38. The method of claim 31, wherein the imaging data and / or imaging parameter comprises X-ray image features.

39. The method of any one of claims 30-38, wherein the one or more lung parameters comprises a mean absolute percentage error (MAPE) of about 15% of less.

40. The method of any one of claims 30=39, wherein the one or more lung parameters comprises one or more of dynamic compliance, peak pressure, mean pressure, or stress index.

41. The method of any one of claims 30-40, wherein the plurality of ex vivo organ perfusion time series data entries comprise synthetic lung disease data.

42. The method of claim 41, wherein the lung disease comprises chronic obstructive pulmonary disease (COPD).

43. The method of claim 41, wherein the lung disease comprises idiopathic pulmonary fibrosis (IPF).

44. The method of claim 41, wherein the lung disease comprises pulmonary hypertension (PAH).

45. The method of claim 41, wherein the lung disease comprises asthma.

46. The method of claim 41, wherein the lung disease comprises acute respiratory distress syndrome (ARDS).

47. The method of claim 41, wherein the lung disease comprises a disease with underlying pathological explanations that alter the mechanism of breathing.

48. The method of claim 2, wherein the organ is a lung, wherein the donor chart data entries comprise one or more of donor age, donor sex, donor height, donor weight, donor smoking history, donor cause of death, type of death, lung on ex vivo perfusion, and ex vivo perfusion cold ischemic time (CIT).

49. The method of any one of claims 1-27, wherein the organ is a liver.

50. The method of claim 49, wherein the one or more organ parameters are one or more liver parameters.

51. The method of claim 50, wherein the one or more liver parameters comprise one or more selected from the group consisting of: a physiological parameter, a biochemical parameter, a transcriptomic parameter, a metabolomic biomarker parameter, a proteomic biomarker parameter, and an imaging parameter.

52. The method of claim 51, wherein the physiological parameter comprises one or more selected from the group consisting of: Hepatic Artery Flow, Portal Vein Flow, Pressure of Inferior Vena Cava, Pressure of Hepatic Artery, Pressure of Portal Vein, and Temperature.

53. The method of claim 51, wherein the biochemical parameter comprises one or more selected from the group consisting of: chloride, sodium, potassium, and pH.

54. The method of claim 51, wherein the metabolomic biomarker parameter comprises one or more selected from the group consisting of: glucose, lactate, AST, and ALT.

55. A method for generating a digital twin of an isolated organ, comprising: receiving, by a computing device, a plurality of donor chart data; generating, by the computing device, one or more organ parameters during ex vivo organ perfusion at least in part by processing the donor chart data entries with a trained machine learning model; anddisplaying, by the computing device, using a graphical user interface, a simulation output comprising the generated one or more organ parameters, thereby generating the digital twin of the isolated organ.

56. The method of claim 55, wherein the donor chart data entries comprise one or more of donor age, donor sex, donor height, donor weight, donor smoking history, donor cause of death, type of death, lung on ex vivo perfusion, and ex vivo perfusion cold ischemic time (CIT).

57. The method of claim 55 or 56, wherein the trained machine learning model comprises a k-nearest neighbors (KNN) model or a decision tree model.

58. The method of claim 57, where the decision tree model is a gradient boosting model.

59. The method of claim 58, where the gradient boosting model is extreme gradient boosting (XGBoost).

60. The method of any one of claims 1-59, wherein the digital twin is a digital twin of a healthy organ, a diseased organ, or an injured organ.

61. The method of any one of claims 1-60, wherein the digital twin is a digital twin of an isolated mammalian organ.

62. The method of claim 61, wherein isolated mammalian organ is an isolated human organ.

63. A system for generating a digital twin of an isolated organ, comprising: a. a memory to store executable components; b. a processor, operably linked to the memory to implement the executable components, the executable components comprising: (i) a model-receiving platform for receiving a trained machine learning model generated from training data comprising a plurality of ex vivo organ perfusion time series data entries from a first assessment period; (ii) a simulation platform that simulates one or more organ parameters for a first forecast period at least in part by processing the time series data entries with the trained machine learningmodel, wherein the first forecast period occurs after the first assessment period; and (iii) an output-generating component to generate a simulation output comprising the generated one or more organ parameters from the first forecast period; and c. a user interface to display the simulation output.

64. The system of claim 63, wherein the training data further comprises a plurality of donor chart data entries.

65. The system of claim 63 or 64, wherein the training data further comprises a plurality of ex vivo organ perfusion time series data entries from at least one additional assessment period.

66. The system of claim 65, wherein the first forecast period occurs after at least one of the at least one additional assessment period.

67. The system of any one of claims 63-66, wherein the simulation platform further simulates one or more organ parameters for at least one additional forecast period.

68. The system of claim 67, wherein simulating one or more organ parameters for at least one additional forecast period comprises at least in part processing the time series data entries from the first assessment period and the one or more organ parameters generate from the first forecast period.

69. The system of any one of claims 63-68, wherein the simulation output further comprises the one or more organ parameters from the at least one additional forecast period.

70. The system of any one of claims 63-69, wherein the trained machine learning model comprises a neural network.

71. The system of claim 70, wherein the neural network is a CNN or RNN.

72. The system of any one of claims 63-71, wherein the plurality of ex vivo organ perfusion time series data entries comprise synthetic data.

73. The system of claim 72, wherein the synthetic data comprises synthetic data transformed from ex vivo organ perfusion time series data entries obtained from a real organ.

74. The system of claim 72 or 73, wherein the synthetic data comprises disease synthetic data.

75. A system for generating a digital twin of an isolated lung, comprising: a. a memory to store executable components; b. a processor, operably linked to the memory to implement the executable components, the executable components comprising: (i) a model-receiving platform for receiving a trained machine learning model generated from training data comprising a plurality of donor chart data entries; (ii) a simulation platform that simulates one or more organ parameters for a first forecast period during ex vivo organ perfusion at least in part by processing the donor chart data entries with the trained machine learning model; and (iii) an output-generating component to generate a simulation output comprising the generated one or more organ parameters from the first forecast period; and c. a user interface to display the simulation output.

76. The system of claim 75, wherein the simulation platform further simulates one or more organ parameters for at least one additional forecast period.

77. The system of any one of claims 75-76, wherein the simulation output further comprises the one or more organ parameters from the at least one additional forecast period.

78. The system of any one of claims 75-77, wherein the trained machine learning model comprises a k-nearest neighbors (KNN) model or a decision tree model.

79. The system of claim 78, where the decision tree model is a gradient boosting model.

80. The system of claim 79, wherein the gradient boosting model is extreme gradient boosting (XGBoost).

81. The system of any one of claims 63-80, wherein the digital twin is a digital twin of a healthy organ, a diseased organ, or an injured organ.

82. The system of any one of claims 63-81, wherein the digital twin is a digital twin of an isolated mammalian organ.

83. The system of claim 82, wherein the isolated mammalian organ is an isolated human organ.

84. The system of any one of claims 63-83, wherein the organ is a lung, a liver, a heart, a kidney, or a pancreas.

85. The system of any one of claims 63-84, wherein the organ is lung.

86. The system of any one of claims 63-84, wherein the organ is liver.

87. The system of claim 85, wherein the plurality of ex vivo organ perfusion time series data entries comprise synthetic lung disease data.

88. The system of claim 87, wherein the lung disease comprises chronic obstructive pulmonary disease (COPD).

89. The system of claim 87, wherein the lung disease comprises idiopathic pulmonary fibrosis (IPF).

90. The system of claim 87, wherein the lung disease comprises pulmonary hypertension (PAH).

91. The system of claim 87, wherein the lung disease comprises asthma.

92. The system of claim 87, wherein the lung disease comprises acute respiratory distress syndrome (ARDS).

93. The system of claim 87, wherein the lung disease comprises a disease with underlying pathological explanations that alter the mechanism of breathing.

94. A computer-readable storage medium with instructions stored thereon that causes a system to: a. receive a trained machine learning model generated from training data comprising a plurality of ex vivo organ perfusion time series data entries from a first assessment period; b. simulate one or more organ parameters for a first forecast period, wherein the first forecast period occurs after the first assessment period; and c. generate a simulation output comprising the simulated one or more organ parameters for the first forecast period.

95. The computer-readable storage medium of claim 94, wherein the training data further comprises a plurality of donor chart data entries.

96. The computer-readable storage medium of claim 94 or 95, wherein the training data further comprises a plurality of ex vivo organ perfusion time series data entries from at least one additional assessment period.

97. The computer-readable storage medium of claim 96, wherein the first forecast period occurs after at least one of the at least one additional assessment period.

98. The computer-readable storage medium of any one of claims 94-97, wherein the computer-readable storage medium causes the system to further simulate one or more organ parameters for at least one additional forecast period.

99. The computer-readable storage medium of claim 98, wherein simulating one or more organ parameters for at least one additional forecast period comprises at least in part processing the time series data entries from the first assessment period and the one or more organ parameters generate from the first forecast period.

100. The computer-readable storage medium of any one of claims 94-99, wherein the simulation output further comprises the one or more organ parameters from the at least one additional forecast period.

101. The computer-readable storage medium of any one of claims 94-100, wherein the trained machine learning model comprises a neural network.

102. The computer-readable storage medium of claim 101, wherein the neural network is a CNN or RNN.

103. The computer-readable storage medium of any one of claims 94-102, wherein the plurality of ex vivo organ perfusion time series data entries comprise synthetic data.

104. The computer-readable storage medium of any one of claims 94-103, wherein the synthetic data comprise synthetic data transformed from ex vivo organ perfusion time series data entries obtained from a real organ.

105. The computer-readable storage medium of any one of claims 94-104, wherein the synthetic data comprise disease synthetic data.

106. A computer-readable storage medium with instructions stored thereon that causes a system to: a. receive a trained machine learning model generated from training data comprising a plurality of donor chart data entries;b. simulate one or more organ parameters for a first forecast period during ex vivo organ perfusion; and c. generate a simulation output comprising the simulated one or more organ parameters for the first forecast period.

107. The computer-readable storage medium of claim 106, wherein the computer- readable storage medium causes the system to further simulate one or more organ parameters for at least one additional forecast period.

108. The computer-readable storage medium of any one of claims 106-107, wherein the simulation output further comprises the one or more organ parameters from the at least one additional forecast period.

109. The computer-readable storage medium of any one of claims 106-108, wherein the trained machine learning model comprises a k-nearest neighbors (KNN) model or a decision tree model.

110. The computer-readable storage medium of claim 109, where the decision tree model is a gradient boosting model.

111. The computer-readable storage medium of claim 110, wherein the gradient boosting model is extreme gradient boosting (XGBoost).

112. The computer-readable storage medium of any one of claims 96-111, wherein the digital twin is a digital twin of a healthy organ, a diseased organ, or an injured organ.

113. The computer-readable storage medium of any one of claims 96-112, wherein the digital twin is a digital twin of an isolated mammalian organ, 114. The computer-readable storage medium of claim 113, wherein the isolated mammalian organ is an isolated human organ.

115. The computer-readable storage medium of any one of claims 96-114, wherein the organ is a lung, a liver, a heart, a kidney, or a pancreas.

116. The computer-readable storage medium of any one of claims 96-115, wherein the organ is lung.

117. The computer-readable storage medium of any one of claims 96-115, wherein the organ is liver.

118. The computer-readable storage medium of claim 116, wherein the plurality of ex vivo organ perfusion time series data entries comprise synthetic lung disease data.

119. The computer-readable storage medium of claim 118, wherein the lung disease comprises chronic obstructive pulmonary disease (COPD).

120. The computer-readable storage medium of claim 118, wherein the lung disease comprises idiopathic pulmonary fibrosis (IPF).

121. The computer-readable storage medium of claim 118, wherein the lung disease comprises pulmonary hypertension (PAH).

122. The computer-readable storage medium of claim 118, wherein the lung disease comprises asthma.

123. The computer-readable storage medium of claim 118, wherein the lung disease comprises acute respiratory distress syndrome (ARDS).

124. The computer-readable storage medium of claim 118, wherein the lung disease comprises a disease with underlying pathological explanations that alter the mechanism of breathing.

125. Use of the digital twin generated according to the method of any one of claims 1-62 or by the system of any one of claims 63-93 in a study, optionally a clinical study.

126. The use of claim 125, wherein the study is a therapeutic study or a transplantation study.

127. The use of claim 125 or 126, wherein the digital twin is used as a control, optionally a paired-control.

128. The use of any one of claims 125-127, wherein the study is a human clinical study.

129. The use of any one of claims 125-127, wherein the study is a veterinary clinical study.

130. Use of the digital twin generated according to the method of any one of claims 1-62 or by the system of any one of claims 63-93 as a research tool.

131. A method for training a machine learning system to produce a digital twin of a lung, the method comprising: collecting a set of input data for a first time period, the input data comprising high-resolution time-series data, physiology and biochemistry data, metabolomic and protein biomarkers, lung image data, and transcriptomics data; pre-processing the set of input data, the pre-processing comprising generating breath-by-breath parameters from the high-resolution time seriesventilator data, hourly parameter values from the physiology and biochemistry data, generating hourly biomarker values from the metabolomic and protein biomarkers, performing principal components analysis (PCA) on the lung image data to generate principal components; and generating gene enrichment scores from the transcriptomics data; training a first model to predict breath parameters for a second time period by processing the breath-by-breath parameters, the hourly parameter values, the hourly biomarker values, the principal components, and the gene enrichment scores; and training a second model to predict hourly values, image principal components, and transcriptomic changes for the second time period, by processing the hourly parameter values, the hourly biomarker values, the principal components, and the gene enrichment scores.

132. The method of claim 131, wherein the first model comprises a gated recurrent unit (GRU).

133. The method of claim 131, wherein the second model comprises a regression model.

134. The method of claim 132, wherein the second model comprises a gradient boosting model.

135. The method of claim 133, wherein the second model comprises extreme gradient boosting (XGBoost).

136. The method of claim 131, wherein the lung-image data comprises X-ray images.

137. The method of claim 131, wherein generating breath-by-breath parameters comprises: segmenting a ventilator waveform into a plurality of breaths with identified inspiratory and expiratory phases; and extracting a list of physiological parameters for a breath of the plurality of breaths; and assigning a clinically relevant label to the breath.

138. A method for training a machine learning system to produce a digital twin of an organ, the method comprising: collecting a first set of input data, the input data comprising high-resolution time-series ex vivo organ function data; collecting a second set of input data, the data comprising organ function data retrieved from one or more electronic sourcestraining, in a first stage, a machine learning model to predict one or more organ parameters from the second set of input data; and training, in a second stage, the machine learning model to predict the one or more organ parameters from the first set of input data.

139. The method of claim 32, wherein the surrogate of the estimate of edema comprises perfusate loss.

Images