Proceedings of a Workshop—in Brief

Convened November 18, 2025

Developing Regenerative Medicine Therapies with Artificial Intelligence

Artificial intelligence (AI) technologies enable the processing of large amounts of data and the identification of complex patterns and correlations. In the field of regenerative medicine this analytical power could be applied to accelerate the development and translation of safe and effective therapies and products. For example, AI tools are being developed with the potential to better elucidate cell and tissue regenerative mechanisms, investigate potential interactions with the cellular microenvironment, and predict long-term product safety and efficacy.

On November 18, 2025, a National Academies of Sciences, Engineering, and Medicine planning committee under the auspices of the Forum on Regenerative Medicine in collaboration with the Forum on Drug Discovery, Development, and Translation convened a workshop to explore the potential applications of AI in regenerative medicine throughout the product development pipeline.¹ The overall goal, said Nabiha Saklayen, chief executive officer and founder of Cellino, is to "spark new ideas together" and "think about how [to] create a better and healthier future for our patients." The goals of the workshop were to explore how AI and related technologies can be used to improve the discovery of regenerative medicine therapies, consider the role of AI in pre-clinical models and supplemental nonclinical data, examine the potential uses of AI to support regenerative medicine clinical trials and regulatory processes, explore opportunities to use AI in the manufacturing process, and discuss the ethical and legal implications for AI in regenerative medicine and the ways AI can improve safe and effective regenerative medicine therapies.

Keynote Addresses

Explainable and Transparent AI for Biomedicine

Many biomedical AI² problems share a similar structure, said Su-In Lee, Boeing Endowed Professor in the Paul G. Allen School of Computer Science and Engineering at University of Washington: "Given patient features X, we want to predict Y." However, predicting the outcome Y is not enough, she continued. In regenerative medicine, the goal is to formulate AI questions such that the predictions offer actionable insights about interventions. It is important to understand how a model arrived at its prediction, but this is difficult because AI models can be "black boxes" with regard to interpretability, Lee explained. She then discussed her research on explainable AI, a subfield of AI, which is designed

"to enhance the interpretability of complex AI models" so that the contribution from the model inputs (e.g., patient features) provides insight into the outcome or predictions made by the model (Lee et al., 2024). These models are generalizable and can be used across many biomedical fields.

Identifying Biomarkers and Drivers of Disease with Explainable AI

Explainable AI models can explain "which features are true biomarkers or disease drivers," Lee said, and she discussed Alzheimer's disease (AD) as a use case. In one study (Beebe-Wang et al., 2021), Lee and colleagues used available RNA sequencing data from postmortem brain tissue and "modeled six Alzheimer's disease neuropathological phenotypes based on gene expression levels by using a multi-output neural network model" (i.e., the X above was gene expression, and Y was phenotype). In this case, explainable AI methodology was able to elucidate complex nonlinear relationships, which she said revealed "previously unknown molecular mechanisms underlying sex-dependent neuropathology." In another example, Lee described how applying explainable AI to unsupervised modeling of gene expression data can "pinpoint important genes that explain the expression variation within the dataset," and could potentially identify driver genes and important molecular pathways (Janizek et al., 2023).

Auditing AI Models with Explainable AI

Explainable AI can also be used to audit AI models for clinical use, thereby enhancing transparency and increasing trust and confidence in model outputs, Lee said. During the COVID-19 pandemic, numerous AI models were developed to predict COVID-19 from chest X-rays. Models performed well in hospitals where their training data were derived from but failed in testing at other hospitals. Explainable AI identified features that the models used for predictions and found that many models had learned shortcuts associated with the data sources and were not relying on true pathological features in the imaging, Lee said. For example, the models came to associate the presence of laterality or other text markers on the X-ray with COVID-19 (DeGrave et al., 2021). "Any kind of language model can hallucinate," Lee said, and she stressed the importance of continuously auditing AI outputs. Lee also spoke about another auditing AI model for understanding which features of dermoscopic images were important for predicting melanoma from skin images (DeGrave et al., 2025).

Using AI in the Development of Regenerative Medicine Cell Therapies

The overall goal of regenerative medicine is simple, said Seth Ettenberg, the president and chief executive officer of BlueRock Therapeutics; it is to study human disease in order to understand whether a patient's loss of a foundational cell type is disease causing and, if it is, to develop and deliver a cell-based therapy to replace what has been lost in an effort to restore function to that patient. Part of the complexity in developing an effective therapy is bringing the infused cells to the right location and having them accepted by the patient. Because autologous cell therapies use a patient's own cells for reprogramming, the cells are readily accepted when returned to the patient; however, manufacturing a personalized therapy batch by batch is logistically complex and expensive. Allogeneic cell therapies, which use cells derived from a healthy donor, can be reliably manufactured at scale from master cell banks. However, the body sees the therapy cells as foreign, and successful integration generally requires immunosuppression for the cells to be accepted or else to be delivered directly to immune-privileged sites such as the brain and eye. Ettenberg discussed BlueRock Therapeutics' development of an allogeneic cell therapy designed to restore lost function in patients with Parkinson's disease and how the company is using AI to automate and streamline this work. He explained that "end-to-end" capabilities are needed to deliver regenerative therapies from the preclinical stages through to commercial applications, which will require advances in development and manufacturing, therapeutic platforms, and expertise and resources.

Allogeneic cell therapies begin, Ettenberg explained, with the creation of a master cell bank of "well annotated and characterized" pluripotent stem cells (PSCs) derived from healthy donor cells. PSCs are then differentiated into the cell type needed to restore the patient's loss of function. As an example, Ettenberg discussed bemdaneprocel, an allogeneic dopaminergic neuron progenitor cell therapy for restoring function in patients with moderate Parkinson's disease. Typically, by the time a Parkinson's patient is diagnosed, "they have lost anywhere from 50 to 70 percent of the midbrain dopaminergic cells within their putamen," he said.

Bemdaneprocel is surgically implanted into the patient's brain, and an open-label phase 1 clinical trial in 12 patients (Tabar et al., 2025) found that "the procedure was well

tolerated," Ettenberg said. At 36-months post implantation the data show a continued favorable safety profile, he said. "Secondary clinical endpoints related to motor outcomes remain stable compared to 24 months and continue to show positive trends from baseline," he added, and there is "demonstrated cell engraftment and survival after stopping immune suppression at 12 months." Based on these phase 1 clinical trial data, bemdaneprocel received a Regenerative Medicine Advanced Therapy designation from the U.S. Food and Drug Administration (FDA), which allowed bemdaneprocel to proceed directly to a phase 3 clinical trial.³ Previously, bemdenaprocel had received fast-track designation from the FDA.

Using AI to Advance Regenerative Medicine

BlueRock is using AI for automating electronic laboratory notebook entries and quality control review, for creating first drafts of regulatory documents and clinical trial designs, and for enhancing product quality and streamlining manufacturing processes, Ettenberg said. One example is an automated large language model (LLM) for single-cell RNA sequencing (RNAseq) data, which he said allows both research and process-development scientists to see batch-to-batch and in-batch variability (i.e., differences between and within production runs) and produce "fine-tuned maps" that indicate the cell types present in a batch. This facilitates the timely identification and assessment of failed or deviated batches and allows for more rapid process improvements.

Looking ahead, Ettenberg said that BlueRock aspires to use AI for "end-to-end authoring and submission of regulatory documents" and "AI-enabled clinical trial design." There are other potential applications in manufacturing, such as "AI-enabled process automation and predictive control." There is also a role for AI in discovery research to, for example, predict in vivo efficacy and streamline the development of preclinical assays and models required for regulatory filings, he said.

Keynote Reflections: Research, Ethics, and Regulatory Considerations

Christopher Hartshorn, the chief of the Digital and Mobile Technologies Section at the National Center for Advancing Translational Sciences at the National Institutes of Health (NIH), reflected on the keynote presentations and highlighted the potential for AI-enabled biomanufacturing as opportunities for "integrating interpretable AI models into cell differentiation schemas, potency prediction, or organoid design" and for "using AI for real-time process monitoring and quality analytics during cell manufacturing." He said that there are a host of cautions associated with AI in biomedical research and clinical care, many of which revolve around data issues that can bias models (e.g., data heterogeneity, lack of standardization, provenance, small sample sizes). Inappropriate training data can lead to models that are "confidently wrong," he said. Another area for attention, he said, echoing Lee's earlier remarks, is incorporating transparency and explainability, which are important for establishing regulatory confidence in AI models.

NIH is focused on "developing evidence-based practices for increasing transparency and generalizability of algorithms as well as appropriate bias mitigation and benchmarking standards that are needed to ensure trust in them," Hartshorn said. Some of the efforts at NIH to support this include the National Clinical Cohort Collaborative, the Bridge2AI initiative, the Nutrition for Precision Health study, and the AIM-AHEAD program. More work is needed to develop annotated datasets, and, along with robust regulatory science, this could expand the scope of clinical decision support tools, he said.

In the short term, AI-enabled biomanufacturing is an area of opportunity, shifting from "craft scale manufacturing to data-driven production," Hartshorn said. Additionally, AI can aid in the acceleration of discovery in regenerative medicine (e.g., by correlating high-dimensional omics and imaging data). Achieving these advances will require "well-annotated standardized datasets" for model training, which he said depend on appropriate sharing of high-quality reference data and metadata. He also emphasized the importance of developing "a workforce fluent in both computational modeling and cell biology."

Barbara Evans, a professor of law and the Stephen C. O'Connell Chair at the University of Florida Levin College of Law, drew attention to the challenges of defining a regulatory model for AI tools that affect clinical care, as opposed to AI tools used in developing a biologic therapy, which is reviewed by FDA in the context of the biologics approval process. From a legal and ethical perspective, Evans said, there is also a role for state medical practice regulators in establishing boundaries for the

use of AI-based clinical decision-support tools. This includes defining appropriate use and required training for users and establishing checks and balances so that the expert human users of these AI tools can intervene and prevent model errors from proliferating and potentially adversely impacting patients.

On the topic of the potential distribution of liability, Evans said the 2013 FDA policy that established the regulation of software as a medical device "opened software developers up to becoming targets for product liability lawsuits." In the context of AI-based tools, she suggested that, for the purposes of a liability lawsuit, an inadequate training dataset could be considered a design defect or that undisclosed model bias could be seen as a failure to warn. Liability will be distributed among developers, manufacturers, and users of an AI tool, Evans said. However, she predicted that "the health care system and the health care professionals who are using a clinical decision support tool will be the primary target of lawsuits if a patient is injured."

The Department of Justice Data Security Program,⁴ which places limitations on the use of bulk sensitive personal data, could potentially impact open access to the data that are so critical for advancing the use of AI models in clinical care, and Evans noted the need to monitor the implementation of this rule.

"CBER is committed to promoting responsible and ethical use of AI in medical product development and product lifecycle management," said Steven Oh, the acting director of the Office of Cellular Therapy and Human Tissue at the FDA Center for Biologics Evaluation and Research (CBER). He highlighted several FDA publications, including a discussion paper on the use of AI and machine learning (ML) in the development of drugs and biologics, a publication on how FDA centers and offices are working together to develop policy related to AI uses in medical products and their development, and a draft guidance document on the use of AI in regulatory decision making for drug and biological products.⁵ An increasing number of regulatory submissions for regenerative medicine therapies involve AI and ML in some capacity, he said, such as "to predict, to classify, to cluster, and to identify anomalies in the dataset." Appropriate training datasets are especially critical for AI models that are intended for clinical use, and a challenge in regenerative medicine is gathering sufficient relevant data suitable for model training. FDA is open to reviewing AI models using a risk-based framework tied to context of use, Oh said. He noted some challenges, however, such as understanding what happens inside the "black box" and ensuring that AI systems are trained on high-quality datasets.

AI in the Pre-Clinical Development of Regenerative Medicine Therapies

Toward a Holistic and Dynamic Stem Cell State Landscape

Susanne Rafelski, the deputy director of scientific programs at the Allen Institute for Cell Science, described the fundamental biological modeling the institute is developing "to understand how cells organize, communicate, and change." The goal is to develop a more holistic representation of cell state by linking spatiotemporal observations of physical structure and molecular organization of the cell, taking advantage of the availability of large-scale single cell image datasets, quantitative image data science methods, and AI/ML tools (see Rafelski and Theriot, 2024).

Rafelski proposed a conceptual holistic cell state framework which accounts for bidirectional feedback among cell state categories such as cell organization, cell function, cell environment, and molecular census (i.e., all of the molecules within a cell), each of which has numerous visible or measurable elements or "observables." For simplicity, she suggested one could visualize a spring-connected tetrahedron which flexes in response to feedback among the four cell state categories (the vertices) to maintain a stable holistic cell state.

Generalizable quantitative frameworks that are integrated across these categories of observables are needed to understand the mechanisms underlying normal, damaged, and diseased cell states, Rafelski said, emphasizing that such understanding also requires identifying the molecular observables present and the "spatiotemporal multiscale

arrangement of all the molecules inside the cell," including organelles and cellular structures. This is where AI can be used for "dimensionality reduction," identifying patterns in complex data and interpreting how these many observables interact to create a holistic cell state over time. Generative AI is used to "decode . . . the pattern that the AI sees" for the purposes of interpreting and validating AI insights, she said.

Developing Quantitative Frameworks for Cell Organization

Over the past decade the Allen Institute has been developing quantitative frameworks for cell organization. The choice was made to focus on intracellular structures and organelles, and human-induced PSCs (hiPSC) are used to derive the cell type of choice for imaging studies visualizing cellular structures. As a case example Rafelski described using hiPSC-derived endothelial cells to investigate cell response to shear stress, which was conducted in a 2-dimensional cell system. The institute's new CellScapes initiative is using human hiPSC models and AI to understand how "cells organize themselves across scales . . . to form complex cell communities and tissues," she said. In this model, sheets of cells grown in 2-dimensional sheets can transform into 3-dimensional lumenoids⁶ through the addition of Matrigel to the culture medium. In closing, Rafelski suggested that holistic cell representations and dynamic cell state landscapes could be thought of as new types of data that could inform research on mechanisms of action and treatment targets and provide a "fundamental understanding of what cells can do" for engineering purposes.

Programmable Cell Therapies: Using AI in the Development of New Technologies to Control Cells

"Cell therapies have a lot of promise for different diseases, but ultimately we need to be able to control what cells do to get the most out of them," said Kyle G. Daniels, an assistant professor of genetics and, by courtesy, of neurosurgery at Stanford University. Chimeric antigen receptor (CAR) T cell therapy, or genetically engineering a patient's T cells to express a synthetic receptor, has proven highly effective against certain blood cancers. However, poor T cell survival, low proliferation, and insufficient tumor killing result in CAR T cells having limited efficacy against solid tumors, he said. There are also emerging CAR macrophage therapies which polarize these cells to a pro-inflammatory state (Wei et al., 2025). However, this polarization is temporary, and CAR macrophages can be repolarized by tumors to an immunosuppressive anti-inflammatory state. Often phagocytosis of tumor cells by engineered macrophages is insufficient for the therapy to be effective.

Daniels described his work to develop synthetic cytokine receptors (SCRs). The approach is for the SCRs to provide continuous signaling (i.e., no exogenous ligand is required). The response can be programmed to activate the desired signaling cascade through the incorporation and recombination of select signaling motifs from natural signaling domains into the SCR signaling domains. The goal is to "identify optimal signals that improve immune cell function," he said, referring to Bell et al., 2024. The first step is designing and building libraries of SCRs which are then tested in primary human T cells and macrophages using high-throughput methods (e.g., flow cytometry, live cell imaging). Data from this testing are used to train neural networks to predict how the signaling motif combinations in the SCRs influence immune cell functions, including differentiation, survival, proliferation, and the ability to kill tumor cells (Capponi and Daniels, 2023).

Incorporating around 500 SCRs with varying signal domains into HER2-specific CAR T cells and testing them against HER2-positive cancer cells resulted in "a huge amount of functional diversity with minimal structural diversity," Daniels said, which facilitated the use of a relatively small dataset for training an ML model. Neural network training produces "ensembles of 10 models," and together the model outputs can be used to identify which motif combinations are associated with undesirable outcomes, such as autonomous proliferation, or desirable outcomes, such as enhanced tumor killing. In a similar approach, SCRs can be used to polarize macrophages to proinflammatory or anti-inflammatory cell states. "Neural networks trained on screening data can predict how changes to the SCR will affect immune cell function," Daniels said. This approach could aid in developing new therapies and could also potentially be applied to other cell types.

Applying AI to Power Genetic Agency⁷

There are many known genetics illnesses, yet few available treatments, said Sam Sinai, a cofounder and the head of machine learning at Dyno Therapeutics. Promising gene therapies are emerging, but not without challenges, and there is an inherent tension between the patients who are eager to try a new treatment and the regulators who are charged with ensuring that products are safe and efficacious, he said. Sinai shared an example of how Dyno is using AI to overcome bottlenecks and bring genetic treatments to patients faster.

Over the past decade Dyno has been working to develop the adeno-associated virus (AAV) protein as a delivery mechanism for genetic therapies. "AAV transduction requires completion of multiple complex steps," Sinai said, including protein folding, assembly around the genetic payload, delivery to the target tissue, and expression in that tissue. This process can be influenced by modifying the sequence for AAV capsid.

The first step in the effort was the development of the Low-shot Efficient Accelerated Performance (LEAP) technology for AAV capsid design.⁸ This approach involves using high-throughput assays to measure sequence effects and using in vivo data in non-human primates to train ML models to predict how a sequence will behave in a functional assay, Sinai said. LEAP is "high-performance in vivo black box design" incorporating "billions of in vivo capsid measurements." A head-to-head comparison found that capsids designed using LEAP outperformed other designs (e.g., rational designed by experts) on measures of packaging and performance (specifically, liver detargeting and brain transduction), said Sinai. As a result, using LEAP allows for bypassing in vivo discovery studies and proceeding directly to validation studies of the AI-designed capsids, he said.

The next step is connecting sequence to mechanism. The new, unpublished⁹ Dyno Psi-1 model is "a structure-based foundation model capable of designing delivery or payload proteins with high efficiency, scale and fidelity," Sinai said. The last element is AI-assisted hypothesis exploration for more efficient therapeutic design.

Sinai described recent results from studies in nonhuman primates of Dyno-bn8a, a new AAV vector for delivery of muscle gene therapies. The results demonstrate the efficiency and potential of modeling genetic therapeutics with AI, he said, with "therapeutic delivery to muscle at significantly lower doses and with improved liver detargeting" compared with a currently marketed product. He noted the challenges of grounding AI models and said that "translation to patients requires grounding in human clinical data."

FDA's Risk-Based Approach to using AI to Support Regulatory Decision Making

FDA centers began formally coordinating efforts on how to incorporate AI into the regulatory framework in 2019, said Johnny Lam, the associate director for policy at CBER. Public and stakeholder engagement workshops were held, and several key discussion papers and agency publications were issued, which informed the development of FDA's 2025 draft guidance on the use of AI in regulatory decision making. In addition to the publications mentioned by Oh above (see footnote 4), Lam highlighted a discussion paper on AI in drug manufacturing¹⁰ and a landscape analysis of the use of AI in regulatory submissions received by the Center for Drug Evaluation and Research (CDER) from 2016 through 2021 (Liu et al., 2023). While not necessarily representative of all FDA centers, the landscape analysis shows the "exponential growth in the number of submissions" received by CDER that applied AI in some form, he said.

The 2025 FDA draft guidance Considerations for the Use of AI to Support Regulatory Decision-Making for Drug and Biologic Products provides product sponsors with "recommendations on the use of AI to produce information or data to support regulatory decision making regarding safety, effectiveness, or quality for drugs," Lam said. A key element of the guidance is "a risk-based credibility assessment framework to establish and evaluate the credibility of an AI model for a particular context of use." The two main factors that determine model risk are model influence and decision consequence, he said. The guidance also addresses life-cycle management of the credibility of AI model outputs.

FDA encourages product sponsors to engage with FDA early in the product development process, Lam said, and he highlighted a range of informal and formal options for engaging with the agency. "FDA evidentiary standards are the same, independent of the technology," he said, and "a risk-based approach is critical to foster innovation and to help [FDA] protect the public."

Panel Discussion

The panelists discussed the promise and the complexity of integrating AI into regenerative medicine. Sinai said that models can appear highly accurate yet fail when "optimized against an objective that's slightly misaligned," demonstrating that careful objective selection and evaluation are important. Daniels agreed with a workshop participant's comment on the importance of addressing sources of noise, explaining that models may inadvertently learn from batch effects and other sources of noise unless those variables are accounted for. From a regulatory perspective, Lam said that adoption will depend on thoughtful "model risk assessment" and the ability to establish credibility for specific contexts of use. Understanding how cell-based therapies or drugs interact with cells is a "ways away" but an exciting application to track, Rafelski said. Correlative science using clinical trial data is already providing predictive power to understand how CAR T cell design and manufacturing affect clinical outcomes, Daniels said.

AI in Regenerative Medicine Clinical Trials and Manufacturing

Practical Uses of AI for Clinical Trial Design

Clinical trial protocols are large, highly detailed documents developed by multiple contributors, and their complexity can make them feel disjointed and "often self-contradictory," said Marshall Summar, the executive officer at Uncommon Cures. AI-assisted cross-analysis can flag errors in minutes, versus a couple of weeks of effort needed by a person. As an example, he described an AI analysis of a pediatric protocol which found that the number of blood draws required would result in participants requiring blood transfusions two to three times over the course of the study.

Another practical application is using generative AI to review the science and the volumes of FDA guidance relevant to a protocol and create first drafts of regulatory documents. In the rare disease field, where the workforce is limited, this can be particularly helpful. Summar emphasized that review by human content experts is always required, as generative AI outputs can "go awry."

Summar discussed several ways he has been using AI in the conduct of rare disease clinical trials. A range of sources suggest there are more than 16,000 rare diseases. There are "about 150 rare diseases that have thousands of patients," he said; 90 percent of them affect less than 150 patients, and 80 percent affect less than 50 patients. After an AI system ran various statistical models for different types of trial designs, it concluded that a double-blind placebo-controlled trial would require about 160 participants for clinical relevance, more than the number of people impacted by the majority of rare diseases. Pairing patients with natural history data would require about 60 participants, more than are impacted by 80 percent of rare diseases. A model using patients as their own control (before and after treatment) would require 10 to 15 participants, which could potentially result in a viable clinical trial for many rare diseases. This information is important when discussing trial design with FDA and making the case for why classic approaches might not be sufficient or successful, said Summar.

AI was also used to understand why about 75 percent of rare disease trials fail, which is important because many organizations conducting rare disease trials are self-funded and have enough capital for one trial, Summar said. Only about 15 to 20 percent of these failures can be attributed to the science. AI determined that the most common cause of failure in rare disease trials was recruitment and retention of clinical trial participants.

Another potential application for AI in rare disease clinical trials is for phase 4 studies, which can extend for 10 to 15 years post-market. AI or chatbots, perhaps in phone apps, could be used to prompt participants with questions and therefore retain them long term, Summar said. This could also allow for real-time data on unexpected outcomes. There is a role for AI in adaptive clinical trial design. Perhaps an AI monitoring system, blinded to investigators, could detect protocol deviations or other concerns and identify potential adaptations, potentially saving resources and reducing failure rates, he said.

Predicting Product Potency and Enabling Cost-Effective Manufacturing

Aspen Neuroscience's approach to treating Parkinson's disease starts with iPSCs derived from patient skin biopsies, which are then differentiated into dopaminergic neuronal precursor cells, said Andrés Bratt-Leal, a cofounder and the senior vice president of research at Aspen. These autologous cells can be transplanted back into the brain of the patient without immune suppression. The cells then differentiate into mature dopaminergic neurons which will produce dopamine. He described how Aspen uses AI for supporting product quality of autologous cell therapies and for autonomous manufacturing.

Following the release of PluriTest in 2011, researchers no longer had to use a mouse teratoma model to empirically demonstrate the pluripotency of new stem cell lines. PluriTest, a bioinformatic assay developed "using hundreds of gene expression profiles from pluripotent and somatic cells" (Müller et al., 2011), can predict pluripotency of a new stem cell line based on gene expression data, Bratt-Leal said. Building off this technology, Aspen developed GraftTest, an ML model to predict the ability of autologous dopaminergic neuronal precursor cells to engraft and function as intended following transplantation into the brain. GraftTest was developed by transplanting patient-derived dopaminergic neuronal precursor cells into rats and using ML to correlate graft quantification with the RNASeq data of the implanted cells to identify which genes are being expressed.

Using a diverse set of cell lines as reference data for model development could lead to more universal biological insights compared with reference data based on only one cell line, Bratt-Leal said. Overall, the potency of a cell therapy for Parkinson's disease depends on engraftment (i.e., survival), differentiation and innervation in patient tissue, and the ability to release dopamine, and Aspen is developing other ML-based predictive models as part of a potency assay matrix for assessment of autologous Parkinson's cell therapies.

Bratt-Leal also described how Aspen is using the Celligent platform to automate production of autologous iPSCs. This platform combines AI and robotics to accomplish the same feeding, cleaning, picking, and passaging steps of iPSC production usually done by skilled human operators. In comparison, the Celligent platform has been found to be more consistent in selecting high quality pluripotent stem cell lines, and he suggested that this autonomous approach could potentially reduce the cost of manufacturing autologous iPSCs by several orders of magnitude.

Use of AI to Predict Donor Variability and Manufacturability for Ex-vivo CAR T

Seed funding and venture capital are hard to come by, and startups that are funded often rush to collect phase 1 clinical data, said Ken Harris, the chief strategy officer and head of AI of OmniaBio. Harris said startup companies seeking funding need to show potential investors how they are using AI and data for prediction and analysis to increase the efficiency of their platform. He referred participants to a 2024 McKinsey white paper that highlights the potential for the application of AI in pharmaceutical research, development, commercialization, and operations to generate billions of dollars in value annually.¹¹

Harris discussed ex-vivo CAR T-cell therapy as a case example of the potential of AI to predict the manufacturability of an autologous therapy. Key challenges for broadly realizing the potential of CAR T-cell therapy include scale, affordability, and access. Clinicians prescribe CAR T cell treatment to only about 30 percent of eligible patients, which is primarily due to reimbursement barriers and patients' inability to access one of only 160 clinical facilities in the United States with CAR T treatment capabilities. A CAR T treatment costs between $600,000 and $1 million, and Medicare reimburses around $487,000, inclusive of drug product and clinical care, Harris said. The pharmaceutical manufacturer recovers its costs, while the health system bears the unreimbursed cost for CAR T treatment typically arising from adverse events or extended care. Adding to the burden for hospitals is that up to 25 percent of patient treatments for licensed CAR T or clinical trial research subjects are associated with manufacturing failures (Baguet et al., 2024), and the pharmaceutical company absorbs the costs associated with out-of-specification drug material regardless of a decision to proceed with treatment. In such cases, neither the drug license holder (i.e., pharma) or hospitals are able to cover costs of that patient irrespective of clinical response, he said.

"Failed batches consume manufacturing capacity, contribute to higher overall costs, and erode patient and clinician confidence," Harris said. OmniaBio is developing a multimodal AI model to predict whether a patient's native

and unique T-cell population tested prior to apheresis using molecular methods can expand into an in-specification drug dose post CAR transduction. The reference database incorporates both public and confidential datasets, including clinical, sequencing, and manufacturing data. A challenge, Harris said, is that many of the manufacturing data needed to enable production predictions are proprietary. Preliminary validation tests show that the model "can predict 88 percent of patients that will not manufacture a good drug product and should not be put into the program." Even if the prediction is that a patient's cells will likely not produce a suitable product, it does not mean the patient cannot be treated with autologous CAR T-cell therapy, he said. This prediction empowers the clinician, patient, and manufacturer to make a joint decision about going forward with cell production and treatment, with the understanding that an out-of-specification CAR T product is unlikely to be covered by insurance.

One approach to navigating the regulatory landscape would be to follow the investigational new drug pathway for pharmaceuticals and biologics. If the intent is to have the AI model integrated into the electronic health record (EHR) as a clinical decision support tool, a de novo pathway may be needed for approval as software as a medical device, Harris said. He noted that FDA has information on lifecycle management for a regulated AI model.¹² Furthermore, the Coalition for Health Artificial Intelligence (CHAI), a not-for-profit organization, provides a certificate for health AI models if they have been validated to CHAI standards. Beyond the model itself, there is a larger information technology infrastructure that also needs to be regulated and compliant, he concluded.

The Future of AI-Enabled Regenerative Medicine Clinical Trials

Regenerative medicine trials face biological uncertainty at every stage, more so than other trials, said Vera Mucaj, Mayo Venture Partner at Mayo Clinic, noting that these are living systems and it is hard to capture one biological endpoint. She also discussed opportunities for AI approaches to help address some of the challenges. Before a clinical trial, AI could be used to accelerate discovery and increase the probability of success. She described using biological data to create AI-enabled simulations of cells, tissues, and organs that could be used to "predict cell behavior, tissue integration, and long-term viability" in response to a treatment. These models could potentially predict the development of toxicities, which is a common reason that phase I trials fail. AI can also be deployed to analyze volumes of multi-omics data for biomarker and therapeutic target discovery and for optimizing trial design (e.g., enrollment criteria, cohort sizes).

During a clinical trial, there are opportunities to use AI to ensure product quality and to automate endpoint assessment. The manufacturing of a cell and gene therapy is where bottlenecks often occur, Mucaj said. AI-driven process analytical technology (PAT) can be used for "real-time monitoring of in-line manufacturing data." These data could be used to predict product quality, increase batch success rates, and reduce costs. There is also a potential role for AI computer vision and ML in standardizing the assessment of complex clinical endpoints, helping to counter both human and model biases. After a trial, natural language processing and ML could be used to extract and analyze real-world data (e.g., from EHRs, registries, and claims data) to support the long-term monitoring of safety and efficacy required for regenerative medicine products, Mucaj said.

Looking forward, Mucaj highlighted manufacturing optimization as an immediate opportunity for the application of AI. While data generated by a manufacturing process can help optimize that specific process, those data are generally proprietary, and she noted the need for precompetitive or federated sharing of data from regenerative medicine manufacturing to enable the accumulation of robust datasets for model training. The challenge is that these data are heterogeneous and not standardized. Over the next 5 years, AI-driven PAT will become standard protocol for predictive quality control in the manufacturing of cell and gene therapies, she predicted, and automated image analysis will become "the standard for evaluating tissue regeneration and functional recovery in clinical trials." She said that eventually there will be AI-driven virtual cell modeling to predict responses to an investigational treatment before the start of patient dosing.

Mucaj discussed several critical requirements for success moving forward. The collection and use of the volumes of data needed for model training will require a data infrastructure that incorporates common data standards and a federated data-sharing platform. The workforce will need scientists and leaders who are cross trained in biology and machine learning

techniques. Also needed are ethical frameworks with "clear guidelines on fairness, transparency, and privacy, especially for real-world data in post-market surveillance," she said. Clear, risk-based guidance from global regulatory agencies on the use of AI models in health care is also essential.

Panel Discussion

During the discussion, Harris said that AI can help identify biomarkers predicting nonresponse or toxicity, saying "more information [is] better" about cell quality so clinicians and patients can make informed decisions, weighed alongside cost considerations. AI remains "more of a tool that you have to use through a learned intermediate than . . . a standalone," Summar said, underscoring the need for human oversight. From a systems perspective, Mucaj highlighted federated learning approaches in the tech sector, but using those models in the clinical space is more complicated because of privacy considerations. Early-generation advanced therapies are inherently expensive, Bratt-Leal said, and initial reimbursement in high-resource settings may subsidize future efficiencies. These therapies are "potentially life altering and life changing," Ettenberg said, and innovation should not be undermined by cost pressures alone.

Laying the Data Groundwork for Regenerative Medicine AI Tools: A FIRESIDE CHAT

Data Collection and Sharing

Elaine Nsoesie, an associate professor in the School of Public Health at Boston University, said her approach to data collection always begins with establishing the research question. Next, she identifies other expertise and perspectives that should be involved in determining which datasets are needed to answer the question and to identify and address biases in those datasets. She then establishes metrics of success and considers how the AI model will be impacted by implementation and deployment in different settings. Even with a clear definition of the data elements to be captured, the dataset will need to be cleaned to ensure that the data retained meet those criteria, Nsoesie said, as there are "many ways that people talk about the same thing." For example, in her infectious disease surveillance research, a search of social media looking for mentions of fever included hay fever and "Bieber Fever," a pop music reference, in the dataset.

Klaus Romero, the chief executive officer at the Critical Path Institute (C-Path), said the organization establishes and leads public–private partnerships "to identify and remove bottlenecks from the drug development process." Identifying unmet needs among their pharmaceutical, regulatory, and advocacy partners; designing solutions; and then determining whether or not an AI-based strategy is needed for the solution will define the data required to build the solution, he said. C-Path has developed a data and analytics platform to facilitate data sharing by pharmaceutical, academic, patient group, and other stakeholders for the purposes of building solutions. The platform has 720,000 individual records from such sources as clinical trials, observational studies, patient registries, and EHRs, Romero said. C-Path also partners with other data sharing platforms. These data are used to build the solutions C-Path is known for.

If a drug development solution—including AI-based tools—is submitted to a regulatory authority, a context-of-use statement is required, Romero said. The context of use dictates the validation that must be done and the evidence required for the tool to be endorsed by regulators.¹³

As an example of the value of integrating different types of data sources, Romero pointed to the approval of the first disease-modifying drug for the rare disease Friedreich's ataxia, which was accomplished using the integration and analysis of C-Path's data and analytics platform. When the results of the clinical trial were statistically positive but evidence of benefit was not clear, data from the Friedreich's ataxia integrated database at C-Path were incorporated into the trial data as external controls, and the new analysis supported approval. Normally, another trial would need to be done, but this is generally not possible for a rare disease. This example also demonstrates the role of the patient community in research and the value of the data that patients contribute, Romero said.

Shawn Sweeney, a senior director at the American Association for Cancer Research (AACR), agreed that planning for data collection starts with the end in mind, with defining the question and what is needed for the solution. When working with patient data it is important to establish whether protected health information (PHI) will be needed, he said. If only de-identified data are needed for the solution, then the path is easier. Sweeny said that working with PHI is much more challenging, even in a federated sharing model. Romero

agreed and said "the value proposition becomes critical" when deciding whether to pursue a solution requiring PHI.

Eric Rubin, the editor-in-chief of the New England Journal of Medicine and an adjunct professor at the Harvard T.H. Chan School of Public Health, said that the standard for basic science is that primary data has to be made public and that this is enabled by "buy-in from the community" and the availability and accessibility of databases that store these data. However, there is no standard for sharing of clinical trial data and "there is not total buy-in from the community that data sharing is critical," he said. The International Committee of Medical Journal Editors (ICMJE) requires that clinical trials submitted for publication in ICMJE journals include the authors' data-sharing policy. There are no specific requirements for what the policy should be, and the policy could be that data will not be shared, Rubin said. Fostering broad acceptance for a data-sharing standard for clinical trials—and epidemiologic studies in particular—is challenging, he explained, because researchers have invested significant funding, time, and effort in collecting their data and want exclusive first access to using those data. The research community needs to establish a way to "value the contribution of those data generators and acknowledge it when . . . [using] their data," he said.

"The beauty of AI is it can find things that we didn't anticipate," said Rubin, however, in many cases it is not fully known what the datasets used for training AI contain. Primary data are best for the purposes of training AI, but they are often not accessible, especially in the case of industry-sponsored clinical trials. Alternatively, he said, AI can be trained on data in the published literature, but these data have been curated and cleaned to address a specific question, and much of the granularity is lost in the publication.

Data Quality

AI models are good at synthesizing the large multimodal datasets that accumulate over the course of a patient's health care journey, Sweeney said. Many of these data are collected and logged in the EHR in accordance with existing standards and norms (e.g., imaging, blood tests, treatment history). However, clinicians' assessments of patients' current conditions and near-term prognoses are entered in the record as text, often low-quality text that is not parsable, he said. EHR systems were developed for billing purposes and not for patient care or research, Rubin said. There will always be sources of variability in datasets, Romero said. The questions to ask, he continued, are, "What is the residual uncertainty, and is that residual uncertainty good enough to support the decision that you want to inform?" Nsoesie reiterated the need to identify and address biases in datasets prior to using them to train AI models, otherwise the models will incorporate those biases. There are frameworks to guide the curation and cleaning of training data, she said.

Ethical and Legal Considerations for Data Sharing

AACR (Project GENIE™) data is open access, Sweeney said, meaning that either patients have given consent to share their de-identified data or their organization's institutional review board has determined sharing the data to be low risk. Data users must agree to the terms of access, which include agreeing to not attempt to re-identify participants and to secure permission before redistributing data. Per the Health Insurance Portability and Accountability Act (HIPAA), AACR de-identifies data by the removal of 18 known identifiers. Another HIPAA path for de-identification of a dataset is via expert determination, benchmarking the dataset against publicly available data. Orthogonal datasets present a risk for re-identification, Sweeney said. For example, the U.S. Census contains very granular data that could be used to re-identify a participant. Genomic data can be in the public domain either intentionally or inadvertently. Ultimately, "nothing is anonymized anymore," he said. The goal is to implement agreed-upon safeguards and monitor data use. Rubin agreed and added that many patients are "enthusiastic about joining in studies" both to learn more about their disease and for the benefit of other patients, and many are similarly willing to share their data with the appropriate safeguards in place. De-identification is a greater risk under certain circumstances, such as for an individual with a rare disease living in a small zip code. There are successful models for sharing genetic data, such as the NIH All of Us Research Program and the UK Biobank. There is tension between regulations that strive to maximize data privacy and security and the patients' mandate for their "data to be valuable and to have maximum impact," Romero said. However, collaborative efforts have provided and will continue to provide avenues to reconcile these two important societal mandates.

Promoting Public Trust in Sharing Data for AI in Regenerative Medicine

The experience at C-Path is that patient participation is the best approach to building public trust, Romero said. For example, researchers should show patients what happens when they contribute their data for research and be transparent about the potential positives and negatives of participating, he said. "Showing, not telling" how AI is making an impact is important, Sweeney said. For example, researchers should show patients how AI was used to discover and develop a treatment for their condition. It is important to acknowledge that this technology is still in its early days, but AI in regenerative medicine is beyond theoretical and has great promise, he said. Nsoesie described her "end-to-end data science" approach to her work with communities in Africa. All projects begin and end with engaging the community "so they understand exactly how AI and data science methods can be useful for them." The same approach is relevant to helping patients understand the usefulness of AI in their health care journey, she said.

Building Trust in AI for Regenerative Medicine: A FIRESIDE CHAT

Peter Bajcsy, a project lead in the Information Technology Laboratory at the National Institute of Standards and Technology (NIST), said that from his perspective as a data scientist, two challenges impacting trust in AI are data quality and uncertainty. He used his work on an implant for the treatment of age-related macular degeneration as a case example. Terabytes of data, including imaging data, were collected over the months it takes to process the autologous cells for implantation, and numerous data science studies using many different approaches were done. Ultimately, "it is a binary decision," he said, and the technician and clinician must decide if the implant is ready for the patient or not. Establishing quality-control measures (in this case, a transepithelial resistance threshold) can aid in the decision making and increase trust in the clinical decision, Bajcsy said, but "at the end of the day, the patient . . . has to trust that we did everything that we could to assure the quality of the data." The other challenge for trust in AI is uncertainty. Scientific measurements generally have a calculated range of uncertainty, Bajcsy said, but currently there is no calculated measure of uncertainty for AI measurements or predictions.

Susan Ariel Aaronson, a research professor of international affairs at George Washington University and a co-principal investigator of the NSF-NIST Trustworthy AI Institute for Law and Society, said innovation is essential for a country's growth. Policy makers have a difficult balance; they want to nurture innovation while simultaneously protecting people from harm and ensuring public trust in AI-enabled technologies; studies of governance have found that "the best way to get trust is to ask people what risks they see and listen to them, be responsive to them." With the caveat that the terms "trust" and "trustworthiness" had not been defined for the purposes of the discussion, Aaronson said trust is "endemic to being human." Trust in others is a learned experience that begins in infancy. Trustworthiness is "how individuals discern if they should trust the system," Aaronson said. For the application of AI in different regenerative medicine systems, she said, "the trust relationships necessary are very different" depending on the system (e.g., a therapy to be implanted in the brain to control movement versus regenerating skin).

George Eastwood, the executive director of the Emily Whitehead Foundation, spoke from his perspective as a representative of patients and caregivers. Trust in AI needs to start with building trust in the person the patient is talking to, such as the clinician, he said. Building trust in the system requires honesty about how some diagnostic and treatment decisions are made (e.g., based on a threshold of probability). "There is a unique connection to the patient within regenerative medicine," he said, because it is the patient's own cells that start the process, and patients develop a unique relationship with their clinicians. When asked by the foundation, 81 percent¹⁴ of patients responded they would share their EHR data with the manufacturer of their CAR T therapy. This trust is related to the patient's personal connection to the process, Eastwood said. The task is to "bring the patient in at the very beginning of the process" for regenerative medicine and gene-based treatment modalities.

Shannon Eaker from Xcell Biosciences, Inc., said a recent International Society for Cell and Gene Therapy (ISCT) committee publication incorporating findings from a public survey identified validation, regulatory readiness, data quality, and trust as key barriers to AI adoption in cell and gene therapy manufacturing (Di Cerbo et al., 2025). Many researchers hesitate to deviate from established

manufacturing processes, he said. Conducting a risk assessment is common practice when incorporating a new tool or technology into a manufacturing process, Eaker said, and risk assessment protocols already exist. Treating AI like any other manufacturing tool and conducting a risk assessment could help to build confidence in its safe and effective use inin manufacturing.

Promoting Open Dialogue, Transparency, and Education About AI

Anne Plant, an emeritus fellow at NIST, suggested that one approach to ensuring confidence in AI would be for consortia with broad representation to come to consensus on the type and depth of testing needed to provide a certain level of confidence in an AI model. Even if initially imperfect, it would be an agreed-upon, tangible measure. A challenge, Aaronson said, is that these AI systems are used globally, and "confidence, like trust, is normative, and the indicators differ among societies." She also emphasized the need for specificity regarding which AI system is being discussed, as different levels of governance would be needed for models used for research versus models deployed in the provision of care. AI models could not mimic human decision making yet, she said. Thinking as a patient, Eastwood said, he would support the use of AI to provide "the fastest, best answer," and then have that reviewed by an expert human clinician.

Building trust requires bringing the whole community together to identify issues and develop solutions, Bajcsy said. The NIST AI Consortium includes members from industry, academic institutions, and federal government agencies who are working to enable the identification of interoperable measurements and methodologies for the use of AI, such as AI standards and evaluations.¹⁵

Eastwood described how he regularly uses AI tools for functional support. As an example, he said that the Emily Whitehead Foundation now records all meetings using a custom-built generative pre-trained transformer (GPT) to extract action items and hold staff and volunteers accountable to each other for these tasks. At first one of the board members was very hesitant to have everything recorded, but there was already trust established between this person and the organization that was built over many years of working together, and the recordings proceeded. Over time the process improved, and he said they were coincidentally educating their board and workforce and earning their trust that the AI outputs were accurate. This is now the foundation's standard practice. Organizations should build trust and connections with people and then start using AI and "educate and grow together," Eastwood said. People around the world are feeling both concerned about government and corporate use of AI and excited about the potential of AI and "are starting to self-educate," Aaronson said. Based on her research, she said, most governments are "shirking their responsibility" to engage their constituents on AI policy issues and the future of AI in their lives. AI will transform manufacturing, the conduct of science, and the job market overall, and "we have an obligation to talk with each other about it," she said. More training programs are needed for those who could implement AI in a technical setting, Eaker said.

Understanding the potential consequences for those who misuse AI could help to promote trust, Bratt-Leal said. For example, what are the consequences for companies using published information to train models without agreement from the copyright holder? What are the consequences for those who misuse private data? Aaronson noted there are many legal cases related to the misuse of data by platforms such as ChatGPT and Character.AI. "Governments were not enforcing their laws on copyright and on personal data protection," she said, "and the side effect was less sharing of data."

Harris emphasized the need to "keep reinforcing the value proposition" of AI in health care (e.g., potentially lower drug costs, increased diagnostic accuracy, and rapid AI-generated responses from care providers). Unlike other areas where AI is being applied, AI in health care is regulated and subject to civil liability laws, Harris said. Bajcsy agreed that the public needs to be made aware of the value of AI but reiterated that they also need to be aware of the uncertainty. See Box1 for opportunities for the application of AI across several areas in the development of regenerative medicine therapies.

Envisioning the Future of AI and Regenerative Medicine: A FIRESIDE CHAT

During the final panel, speakers reflected on the discussions and highlighted opportunities to realize the potential of AI in regenerative medicine. Ettenberg started with a call to participate in forums such as this workshop because of the "power of interdisciplinary conversation like we're having

. . . to make sure that I'm sitting in rooms where not just innovators, but also regulators, payers, and patients and patient advocates are in that conversation together." Many ideas emerging from academic research remain unfunded because they cannot secure the capital needed to conduct phase 1 clinical studies, Ettenberg said, adding that AI can contribute to "lowering the activation energy for innovation . . . [in] regenerative medicine" and could "significantly impact the cost of capital to get to [phase 1]." Summar described AI as a potential "force multiplier." Investment in rare disease research has declined, which he said is due to declining return on investment associated with "pushback from the payer community." AI-assisted workflows can help teams increase efficiencies, reduce failures, and reduce costs across the product trajectory, from research to development to clinical

practice. For example, he said, the capital cost for developing a new therapy decreases as the failure rate decreases.

There are many well established, robust, explainable AI techniques that tend to be applied in other fields outside biomedicine first, Lee said. Applying these in biology and medicine could provide more mechanistic insights about model predictions and build trust in the reliability of AI simulations. AI simulations of biological systems, such as digital twins, have the potential to enable regenerative and precision medicine where each person has his or her our own digital model, Bajcsy said. Sinai stressed the importance of building filters into simulation models to prevent inaccurate and potentially harmful outputs.

Data generation for AI model development is another area of opportunity. There should be greater focus on creating datasets that are computer-ready from the start, Bajcsy said.

Current researchers and clinicians need to educate themselves in the fundamentals of AI and ML, fields that are dynamic and rapidly changing, Lee said. AI also needs to be incorporated into the education of the next generation. Researchers with deep expertise in both biology and computer science are needed, she said, adding that some of her students in the Medical Scientist Training Program have pursued a doctorate in computer science along with their medical degree. Patients are actively using AI-assisted tools to research their diseases, and patient organizations are developing new AI-assisted tools for their community, Summar said. In some cases, the patient community is ahead of the research community, and patients are not waiting for the rest of the field to catch up. Engaging with "patients has taught us a lot more than we've been able to teach them," Sinai said.

As discussed, FDA is receiving an increasing number of regulatory submissions that incorporate AI in some form. Ron Bartek, the president and co-founder of the Friedreich's Ataxia Research Alliance, said there are lessons to be learned about why some submissions are successful while others fail. Engaging with FDA to explore ways to more quickly share lessons learned from AI-enabled regulatory submissions (e.g., what succeeds, what does not, and why) to better understand how AI is being evaluated in practice could help address uncertainty in the field, rather than relying on longer timelines associated with formal guidance development, Bartek said. Another topic for consideration, he said, is the FDA¹⁶ and NIH¹⁷ initiatives to replace preclinical data collection in animal models with new approach methodologies, including AI-based approaches.

Topics for Future Discussion: Workforce Considerations

Speakers highlighted several issues related to workforce that merit further discussion. A challenge for accelerating the pace of regenerative medicine innovations with AI tools is the retention of top AI researchers, Sinai said. U.S. biotechnology research organizations cannot compete with the compensation packages offered by large tech companies. It is also hard to retain AI researchers in biotechnology when they see they are far outpaced by the progress of biotechnology companies in other countries. Current graduate students have expressed concerns about what they should be studying now to ensure their career is still relevant in the age of AI, Ettenberg said. These early career scientists should be part of and contribute to the ongoing conversations about the implementation of AI in research that will shape the future. Similarly, research team members are being tasked with implementing AI tools that might make their jobs less relevant, and they will need to adapt and transition to other roles, Ettenberg continued. He noted that subject matter expertise will continue to be needed for curation of information. In this era of "AI co-scientist"¹⁸ systems, it will be necessary to consider what it means to be a doctoral student or a postdoctoral trainee and how the education system needs to evolve, Lee said.

Closing Reflections

At the end of the workshop, Plant and Saklayen summarized observations and suggestions made by individual speakers on opportunities to advance the application of AI in the development of regenerative medicine therapies.¹⁹ 

Examples shared during the workshop demonstrate that "AI models are improving efficiencies in documentation and regulatory compliance," Plant said. AI models are also being developed to support prediction and decision making. AI models require substantial amounts of data, which highlights the critical importance of data sharing. Issues impacting data collection, sharing, and use were prevalent topics throughout the discussions. "Investment in AI within the context of regenerative medicine has not been sufficiently robust," Plant said, and greater capital investment is needed. Finally, Plant noted, there was much discussion of the need for standards that enable data sharing (e.g., standard data formats, harmonized metadata, and the protection of patient health information) and of standards for the evaluation of AI model reliability and model qualities such as bias. Challenges for the development of model standards, she noted, are identifying what should serve as ground truth and also the need for human intervention as a safeguard for clinical decision making, maintaining oversight, and managing uncertainty. "Building public trust through active participation and dialogue," Saklayen said, is important because people need to work together across sectors to arrive at solutions for the challenges considered here today at the workshop.

References

Baguet, C., J. Larghero, and M. Mebarki. 2024. Early predictive factors of failure in autologous CAR T-cell manufacturing and/or efficacy in hematologic malignancies. Blood Advances 8(2):337–42.

Beebe-Wang, N., S. Celik, E. Weinberger, P. Sturmfels, P. L. De Jager, S. Mostafavi, and S. I. Lee. 2021. Unified AI framework to uncover deep interrelationships between gene expression and Alzheimer's disease neuropathologies. Nature Communications 12(1):5369.

Bell, M., S. Lange, B. I. Sejdiu, J. Ibanez, H. Shi, X. Sun, X. Meng, P. Nguyen, M. Sutton, J. Wagner, A. KC, D. Langfitt, S. L. Patil, H. Tan, R. V. Pandey, Y. Li, Z. F. Yuan, A. A. Anido, M. Ho, H. Sheppard, P. Vogel, J. Yu, J. Peng, H. Chi, M. M. Babu, G. Krenciute, and S. Gottschalk. 2024. Modular chimeric cytokine receptors with leucine zippers enhance the antitumour activity of CAR T cells via JAK/STAT signalling. Nature Biomedical Engineering 8(4):380–96.

Capponi, S., and K. G. Daniels. 2023. Harnessing the power of artificial intelligence to advance cell therapy. Immunological Reviews 320(1):147–65.

Di Cerbo, V., H. W. Song, L. Herbst, S. J. Hart, R. Ladi, S. Jiang, A. Madbouly, L. Redmond, S. Stickland, A. Srinivasan, K. Trinkle, A. Ting, S. S. Eaker, and D. Sethi. 2025. Artificial intelligence, machine learning, and digitalization systems in the cell and gene therapy sector: A guidance document from the ISCT industry committees. Cytotherapy 27:903–9.

DeGrave, A. J., J. D. Janizek, and S. I. Lee. 2021. AI for radiographic COVID-19 detection selects shortcuts over signal. Nature Machine Intelligence 3:610–9.

DeGrave, A. J., Z. R. Cai, J. D. Janizek, R. Daneshjou, and S. I. Lee. 2025. Auditing the inference processes of medical-image classifiers by leveraging generative AI and the expertise of physicians. Nature Biomedical Engineering 9(3):294–306.

Howell, M. D., G. S. Corrado, and K. B. DeSalvo. 2024. Three Epochs of Artificial Intelligence in Health Care. JAMA 331(3):242–4. https://doi.org/10.1001/jama.2023.25057​​​​​​​​.

Janizek, J. D., A. Spiro, S. Celik, B. W. Blue, J. C. Russell, T. I. Lee, M. Kaeberlin, and S. I. Lee. 2023. PAUSE: Principled feature attribution for unsupervised gene expression analysis. Genome Biology 24(1):81.

Lee, S. I., and E. J. Topol. 2024. The clinical potential of counterfactual AI models. Lancet 403(10428):717.

Liu, Q., R. Huang, J. Hsieh, H. Zhu, M. Tiwari, G. Liu, D. Jean, M. K. ElZarrad, T. Fakhouri, S. Berman, B. Dunn, M. C. Diamond, and S. M. Huang. 2023. Landscape analysis of the application of artificial intelligence and machine learning in regulatory submissions for drug development from 2016 to 2021. Clinical Pharmacology & Therapeutics 113(4):771–4.

Müller, F. J., B. M. Schuldt, R. Williams, D. Mason, G. Altun, E. P. Papapetrou, S. Danner, J. E. Goldmann, A. Herbst, N. O. Schmidt, J. B. Aldenhoff, L. C. Laurent, and J. F. Loring. 2011. A bioinformatic assay for pluripotency in human cells. Nature Methods 8(4):315–7.

Rafelski, S. M., and T. A. Theriot. 2024. Establishing a conceptual framework for holistic cell states and state transitions. Cell 187(11):2633–51.

Tabar, V., H. Sarva, A. M. Lozano, A. Fasano, S. K. Kalia, K. K. H. Yu, C. Brennan, Y. Ma, S. Peng, D. Eidelberg, M. Tomishima, S. Irion, W. Stemple, N. Abid, A. Lampron, L. Studer, and C. Henchcliffe. 2025. Phase I trial of hES cell-derived dopaminergic neurons for Parkinson's disease. Nature 641(8064):978–83.

Wei, D., L. Wang, Y. Liu, X. Zuo, X. Shen, and R. S. Bresalier. 2025. CAR-macrophage cell therapy: A new era of hope for pancreatic cancer. Clinical Cancer Research 31(19):4018–31.

Disclaimer: This Proceedings of a Workshop—in Brief was prepared by Dara Ancona, Theresa M. Wizemann, and Sarah H. Beachy as a factual summary of what occurred at the workshop. The statements made are those of the rapporteurs or individual workshop participants and do not necessarily represent the views of all workshop participants, the planning committee, or the National Academies of Sciences, Engineering, and Medicine.

Planning Committee:Anne Plant (Co-chair), retired, National Institute of Standards and Technology; Nabiha Saklayen (Co-chair), Cellino; Ronald Bartek, Friedreich's Ataxia Research Alliance; Kapil Bharti, National Eye Institute; Timothy Chan, Cleveland Clinic; Christopher Hartshorn, National Center for Advancing Translational Sciences; Rosario Isasi, University of Miami; Amritha Jaishankara, IQVIA; John Knighton, Johnson & Johnson; Jagdeep Podichetty, Critical Path Institute; Scott Steele, U.S. Food and Drug Administration; Claudia Zylberberg, Kosten Digital. The National Academies' planning committees are solely responsible for organizing the workshop, identifying topics, and choosing speakers. Responsibility for the final content rests entirely with the rapporteurs and the National Academies.

Reviewers: To ensure that it meets institutional standards for quality and objectivity, this Proceedings of a Workshop—in Brief was reviewed by George Eastwood, Emily Whitehead Foundation and Krishanu Saha, University of Wisconsin-Madison. Kirsten Sampson-Snyder, National Academies of Sciences, Engineering, and Medicine, served as the review monitor.

Sponsors: This activity was supported by agreements between the National Academy of Sciences and Advanced Regenerative Manufacturing Institute; Alliance for Regenerative Medicine; American Society of Gene & Cell Therapy; BlueRock Therapeutics; Burroughs Wellcome Fund (Grant No. 1496874); California Institute for Regenerative Medicine; Cellino Biotech, Inc.; Department of Veterans Affairs (Contract No. 36C24E21C0011); Food and Drug Administration: Center for Biologics Evaluation and Research (Contract No. 1R13FD008396-01); International Society for Cellular Therapy; International Society for Stem Cell Research; Johnson & Johnson; Kosten Digital; National Institutes of Health (including National Center for Advancing Translational Services; National Eye Institute; National Institute of Arthritis and Musculoskeletal and Skin Diseases; National Institute of Diabetes and Digestive and Kidney Diseases; National Institute of Biomedical Imaging and Bioengineering; National Institute on Aging; and National Heart, Lung, and Blood Institute [Contract No. 75N98024D00037, Task Order No. 75N98026F00007]); National Institute of Standards and Technology; and United Therapeutics Corporation. Any opinions, findings, conclusions, or recommendations expressed in this publication do not necessarily reflect the views of any organization or agency that provided support for the project.

Staff:Sarah H. Beachy, Forum on Regenerative Medicine Director, Michelle Drewry, Associate Program Officer (through November 2025), Dara Ancona, Associate Program Officer (from February 2026), Carolyn Shore, Forum on Drug Discovery, Development, and Translation Director, and Ashley Pitt, Senior Program Assistant.

Suggested citation: National Academies of Sciences, Engineering, and Medicine. 2026. Developing Regenerative Medicine Therapies with Artificial Intelligence: Proceedings of a Workshop—in Brief. Washington, DC: National Academies Press. https://doi.org/10.17226/29427.

For additional information regarding the workshop, visit https://www.nationalacademies.org/projects/HMD-HSP-24-22/event/45414.

Copyright 2026 by the National Academy of Sciences. All rights reserved.