Code to Cure is a newsletter co-hosted by the Wyss Institute at Harvard University and Milad Alucozai on the intersection of AI, biology, and healthcare transforming medicine.
By Milad Alucozai, Abishek Kulshreshtha, Megan Sperry, and Klaus Romero
The average new drug requires over 15 years and $2 billion in funding to traverse the journey from discovery to full approval. The first portion of the journey, among the riskiest and most scientifically complex portions, involves extensive testing of the drug in in silico, in vitro, and in vivo settings. Current standards all but ensure that new medicines are tested extensively in vivo animal models for safety before being permitted in patients.
But a tidal shift is underway. In vivo models are being complemented by, and in some cases supplanted by, more sophisticated in vitro and computational preclinical models that can substantially reduce the cost and timeline of early-stage drug development. These advanced models, including optimized human-relevant wet lab models and predictive computational/in silico tools, have the potential to dramatically improve clinical success rates in drug development. Importantly, adoption of these tools to advance new medicines along the critical path is nascent, but increasingly critical to modern drug discovery and development. Leveraging wet lab and computational/in silico preclinical models that are well-aligned with effects and outcomes in patients is critical to identifying today’s most promising drug candidates and bringing them to patients efficiently and safely.
“Traditional” Preclinical Development
Since the Federal Food, Drug, and Cosmetic Act was passed in 1938, the US Food and Drug Administration (FDA) has required products to be assessed for scientific validity and safety in animals before dosing human participants in Phase I trials. Most often, preclinical validation involves rodent and higher species models that are meant to mimic the genotype and/or pathophysiology of human disease. Some adverse events, such as cardiac effects, demonstrate general predictivity of human safety from animal observations. However, the lack of toxicity in animals has very low predictivity for lack of adverse events in humans for some organs and animal species, meaning that preclinical tests in vivo do not always translate well to humans. Unfortunately, these models often fail to accurately predict pharmacokinetics, pharmacodynamics, or safety in humans, leading to Phase I and II clinical trials failing due to lack of efficacy (60% of trials) or toxicity (30% of trials).
So, what’s the problem with these models? Although the full scope of issues is beyond this article, key distinctions between humans and animals include differences in how drugs are broken down and cleared from the body, difficulties in replicating the relevant disease pathophysiology, as well as a lack of human-representative genetic diversity in rodents. Humans exhibit a huge variation in drug metabolism, interactions, targets, and pathophysiology and disease progression, contributing to the complexity of their disease presentation. Scientists, regulators, and patient advocates alike recognize that while animal models currently provide a feasible route to test safety and efficacy before clinical tests in humans, the development of non-animal models that demonstrate stronger clinical translation will be essential to lead to faster, less expensive, and more reliable drug development.
Several organizations are addressing the complex scientific and regulatory challenges required to increase adoption of non-animal models. For example, the Critical Path Institute (C-Path) works with the FDA alongside industry and university researchers to advance non-animal models. C-Path was established under the auspices of the FDA’s Critical Path Initiative in 2005 and has a 20-year track record as an independent, public-private partnership. Since its founding, C-Path has been unique in its ability to develop actionable solutions through open, precompetitive collaboration with sharing of data and expertise. C-Path builds consensus among participating scientists from industry and academia with regulatory participation and iterative feedback. Such consensus provides the mechanism to generate the necessary confidence to assure the adoption of the medical product development solutions by sponsors and regulators. Examples of pathways through which this confidence is achieved include informal and formal regulatory pathways. Through these various mechanisms, sponsors can confidently adopt the solutions generated through C-Path’s collaborative approach, thus ensuring the continuous optimization of the medical product development process.
The Critical Path Institute (C-Path) hosted a series of public workshops, bringing together key stakeholders like the Food and Drug Administration (FDA), pharmaceutical and biotech leaders, model developers, and academia. These workshops aimed to establish initial performance and validation considerations for the Qualification of Computational In Vitro Models (CIVMs). These initiatives reflect a concerted effort to create a clear pathway for the regulatory acceptance of New Approach Methodologies (NAMs), ultimately supporting more human-relevant approaches in biomedical research and drug development. Innovations for which they have helped pave the way are included here.
Emerging Shifts in Preclinical and Translational Development
Acknowledgement of these challenges has resulted in efforts by researchers to develop alternative, more effective methods for preclinical and translational evaluation of drug candidates. Organs-on-a-Chip, organoids, human induced pluripotent stem cells, and more in vitro approaches are reviewed here. Notably, some Liver Chip models were recently found to outperform conventional models in predicting drug-induced liver injury. However, further refinement of these systems may be required due to the technical skills required to use such specialized microphysiological systems, difficulties replicating some whole organs, and limited long-term functionality.
An example of a complex in vitro approach comes from Emulate, which designs and builds Organ Chip microfluidic devices lined with living human cells for drug development, disease modeling, and personalized medicine. These microdevices are composed of a clear, flexible polymer about the size of a USB memory stick that contains hollow microfluidic channels lined with living human organ cells and human blood vessel cells. By mimicking human organs in vitro, Emulate aims to provide faster, better, and cheaper drug development and insights into human health compared to more simplistic cell culture systems or animal models that are physiologically different from humans. Their current Organ Chip products include brain, colon, duodenum, kidney, liver, and lung systems. Of note, the Emulate Liver Chip better predicted drug-induced liver injury than animal and hepatic spheroid models. Emulate was launched in 2014 from the Wyss Institute at Harvard University.
New methods for preclinical testing are not just limited to human-relevant in vitro systems. They can also include quantitative computational/in sililco models to predict drug metabolism, toxicities, and off-target effects. These models may incorporate multiple quantitative methods, from quantitative systems modeling, to AI-based tools. First-in-human trials could also be performed in perfused, donated organs that are unable to be used for transplant.
In December 2022, these changes in thinking resulted in the FDA Modernization Act 2.0 being signed into law. While alternative models were possible on a case-by-case basis previously, this Act specifically states the intent to utilize alternatives to animal testing for Investigational New Drug applications to enable clinical trials for novel drugs, or previously approved products expanding to a new indication or patient population. In September 2024, the FDA’s Center for Drug Evaluation and Research (CDER) accepted its first letter of intent for an organ-on-a-chip technology as a drug development tool. This is the first step in a three-part process that may qualify the Liver Chip as a drug development tool that can be used in any drug development program. Additionally, FDA CDER has the Fit-for-Purpose Initiative for regulatory acceptance of computational/in-silico drug development tools.
Digital Twins and Other In silico Preclinical Tools
Although AI is often strongly associated with the drug design stage, these tools can support multiple aspects of the drug development process, including preclinical, translational, clinical, manufacturing, and marketing phases. In fact, the FDA noted a substantial uptick in AI-driven applications in 2024.
Combined with evidence from human-relevant in vitro models, computational/in-silico tools offer a bridge from preclinical to clinical development. These tools can be used on multiple levels – either simulating the impact of a therapeutic on a disease itself or modeling clinical outcome measures. For the clinical stages of drug development, clinical trial simulation tools that allow the generation of virtual populations, through methods that vary from pharmacometrics, to quantitative systems pharmacology, to AI-based “digital twins” are also crucial quantitative tools that can optimize clinical trial design and conduct. And to generate these quantitative solutions, data sharing is key. To truly unlock the potential of these computational approaches, they must be deeply integrated with rich, high-fidelity human data. This necessitates a new infrastructure and standards for capturing and utilizing data, which is precisely the problem companies like Revalia Bio are addressing.
Revalia Bio is creating the home for human data by building the world’s first truly integrated Human Data StackTM to power a new category of Phase 0 Human TrialsTM. Revalia’s Phase 0 Human Trial platform integrates the various sources of human data to ask and answer questions that are impossible with Phase I/II clinical trials or current preclinical models alone. The key to the platform is Revalia’s “Human Organ Data Layer” made possible by a unique network of partnerships assembled to support organ donation for research when those organs are not suitable for clinical transplant. The Revalia team brings those organs back to life on proprietary organ perfusion technology to create a “Rosetta Stone” for human data that serves to integrate and contextualize all other sources of human data from patient medical records to cells in a petri dish. Revalia provides access to the Phase 0 Human Trial through a Platform as a Service model that enables biomedical developers to design, track, and interpret Phase 0 Human Trials at the click of a button guided by Rio, Revalia’s software companion. Through this new platform, Revalia aims to enable the global biomedical community—from academic scientists and hospitals to large biopharma—to effectively collaborate on creating better medicines faster. The team at Revalia believe that biomedical innovators should compete on delivering the best innovations to patients as fast as possible, not on access to critical human data.
In January 2025, the FDA released guidance entitled, “Considerations for the Use of Artificial Intelligence to Support Regulatory Decision Making for Drug and Biological Products,” which includes a framework for risk-based credibility assessment, considerations for maintenance of the algorithm across the product life cycle, and recommendations for early engagement with the FDA. Although these are currently nonbinding recommendations, their release suggests that in silico tools are starting to have a meaningful role in drug development.
Challenges and Areas for Innovation
Despite the promise of human-relevant in vitro and computational/in-silico models, there remain multiple opportunities for development and implementation of these tools in FDA applications. In vitro models with human relevance have proven successful for modeling safety and toxicity within individual cells/organs as well as modeling diseases linked to specific genetic mutations. However, accurately modeling complex disease pathophysiology and clinical outcomes remains difficult. Diseases characterized by behavioral symptoms and reliant on patient reporting, such as mental health and pain, are especially tough to capture in vitro. Additionally, scale-up of in vitro experiments to capture relevant genetic diversity in human populations remains a challenge. Pooled cell line approaches, also known as cell villages, have been proposed as a solution for achieving this at scale.
Genetically complex and poorly understood pathology, like neurobehavioral diseases and some cancers, are also more challenging to model in silico. Deep learning models could play a role in improving digital twin tools by forecasting unknowns, including using generative AI to fill in sparse data from real-world patients. The pairing of digital twins and generative AI is being explored across industries and could play a powerful role in drug discovery and clinical trial preparation.
There are many diseases with poorly understood pathophysiology, including those with variable mechanisms of onset across the population. Most existing datasets are incomplete or restricted for use. They are also incredibly noisy because the multi-omic and pathology datasets do not have matching donors across data types. From existing datasets, ML/AI algorithms detect very large signals and more subtle trends are completely obscured. To overcome these challenges, integrated approaches are starting to be used to collect complete, multi-dimensional datasets for use in AI/ML pipelines. One such example is the UK company Avatrial, which brings together legal and regulatory approvals, commercial licenses, hospital partnerships, state-of-the-art multi-omic technology, AI/ML analysis pipeline, and advanced experimental models.
Future Perspective
Expert organizations, including the FDA and C-Path, are exploring opportunities to better incorporate non-animal preclinical and translational models into the regulatory review and approval journey for new medicines. Through efforts such as C-Path’s Predictive Safety Testing Consortium, regulatory scientists are leveraging nascent regulatory pathways to obtain acceptance of non-clinical platforms, which can then be applied broadly by scientists and startups advancing novel drugs. The FDA’s support of these models via new regulatory channels could ultimately provide the pathway to rapid adoption of these in vitro, in chemico, and computational/in-silico models to replace certain animal tests.
The impact of future preclinical models could be groundbreaking. Current preclinical tests require several years and up to tens of millions of dollars for new drug candidates that often fail in the clinic, creating a risk-return profile that even traditional venture investors are increasingly reluctant to undertake. The incorporation of more translatable, faster, and less expensive preclinical models could ultimately disrupt the current financial landscape for drug development, potentially bringing previously unfundable new medicines to the fore. This opportunity highlights the importance of advancing computational/in-silico models that capture the complexity and diversity of human patients, a soon-to-be essential tool for drug developers of all types.