When Preclinical Models Distort The Patient

By Alice Gilman

For more than 30 years, drug development has been haunted by a stubborn fact: roughly nine out of 10 drug candidates that enter human trials fail.¹ Each decade brings a new generation of technologies that promise to reverse that statistic: organ-on-chip systems, patient-derived organoids, genetically engineered animal models, and now AI-driven design engines.

Every one of these innovations has moved the field forward in some way. Yet, collectively, they’ve failed to fix the translation problem. Attrition remains unchanged because the industry keeps chasing a “better model” instead of a better system — one that intentionally integrates the partial truths of each platform into a predictive whole.

Each model offers clarity in one domain and distortion in another. Organs-on-chips reproduce cellular microenvironments with elegance, but their chemistry and throughput limit application.^2,3 Murine models capture full-body physiology but diverge in metabolism and immune signaling.^4,5 Organoids reflect patient genotype but suffer from hypoxia and stochastic variability.⁶ AI models find statistical patterns but often lack biological ground truth.^7,8,9

Organ-On-Chip: Mechanistic Precision, Chemical Limitations

Microfluidic chips replicate human physiology with remarkable scale: lungs that breathe, intestines that contract, cardiac tissues that beat under pulsatile flow. Yet the very materials that make these platforms work also undermine their reliability for pharmacology.

Most chips are fabricated from polydimethylsiloxane (PDMS), a polymer that absorbs lipophilic compounds. A Harvard finite-element study showed that over 90 % of the antimalarial amodiaquine was lost to PDMS walls, skewing dose-response calculations and forcing modelers to reverse engineer exposure levels.² Replacing PDMS with glass or cyclic olefin copolymer shifts the bias rather than eliminating it.

Throughput is another constraint. Multi-organ systems that connect liver, gut, and kidney units are powerful but limited to thousands of wells — a fraction of what’s needed for lead optimization. A 2025 Lab on a Chip review conceded that “handling complexity must be reduced and/or operation automated to enable high-throughput screening.”³

Even regulators acknowledge both promise and limitations. A recent U.S. Government Accountability Office report praised organ-on-chip platforms as “promising complements” to animal studies but noted that no drug candidate has yet advanced to approval based on chip data alone.¹

Chips are best deployed as mechanistic microscopes, not predictive surrogates. Their strength lies in explaining why a molecule behaves as it does, not in forecasting clinical outcomes.

Murine Models: Necessary, But Not Sufficient

The mouse remains the workhorse of preclinical testing because it provides intact physiology, endocrine feedback, and immune responses that no in vitro system can replicate. Yet mouse success often fails to translate to human efficacy.

Merck’s verubecestat cleared amyloid plaques in Alzheimer’s mice, only to worsen cognition in two Phase 3 trials.⁴ Drotrecogin alfa (Xigris), once approved for sepsis on strong murine data, was withdrawn in 2011 when the PROWESS-SHOCK trial revealed no benefit and potential harm.⁵

The gap lies in fundamental biology. Rodent lipid metabolism, circadian rhythms, and neuro-inflammatory pathways differ profoundly from humans. Standardized inbred colonies reduce variability but eliminate the genetic heterogeneity that drives idiosyncratic toxicities in real patients.

Animal studies remain critical for safety and systemic physiology but for ethical reasons should no longer be used as primary efficacy predictors. Cross-species concordance analyses comparing animal, organoid, and chip data should become a standard step before IND submission.¹

To understand why researchers sought “humanized” preclinical systems in the first place, it helps to revisit the humble origins of cell culture — a technique that, despite its ubiquity, rests on a profoundly unnatural premise.

Cell culture, invented more than 70 years ago, emerged from a peculiar discovery: if living cells are dropped into a plastic dish, they will adhere and continue to grow. That is, in many ways, astonishing because it violates nearly every rule of their natural context.

Organoids: Human Relevance, Biological Constraints

Organoids are an attempt to do cell culture in a slightly more "physiologically relevant" way. You remove the plastic surface, but make the cells attach to each other. They grow in small clumps until the inner cells run out of food and start dying at about 0.5mm diameter. You can make various organ clumps, where each cell is attached only to its “related organs.” You don't have the morphogen, but at least the plastic is gone.

Imagine you get hit in the stomach by an object. Some clumps of cells get knocked loose. Say it makes a chip in the liver, and that material falls into the heart. That type of thing happens all the time. Now the liver cells in that misplaced clump stick out their feelers and figure out that this is not the correct laminin variant on the surface, the GDF11 concentration is not quite right, and many other things are ever so slightly off. It’s unsustainable for them, so the cells die

Meanwhile, on the liver, there is a chip left behind, like a little crater. The cells at the bottom of the crater figure out that this is not where the surface of the liver is supposed to be, so they divide and fill that crater. As soon as they reach the invisible line where the surface of the liver is supposed to be, they stop. They’re able to do this because they sense surfaces, neighboring cells, and incredibly dilute chemical concentration gradients made by all the other cells in the body. It's the "morphogenic field" by which all organ cells can tell where other organ cells are. Each individual cell can sense the morphogen and infer from it if everything is okay in its microscopic local environment, and if not, what it needs to do to make it better.

When you place them on polystyrene plastic, they divide and thrive, which is surprising given how particular the cells can be. Unfortunately, it does not work well. Once you put them in this unnatural situation where they grow on plastic, there are a lot of highly body-relevant processes that you can't study in that situation. You affected their entire sense of identity and somehow made them insensitive to the morphogen. It's a great system for studying what's going on inside one cell, but not useful for studying whole-body phenomena, like what's a drug going to do to the body.

Stem-cell-derived organoids then gave researchers what they long sought: a human genotype and patient-specific tissue architecture. They have already advanced cystic fibrosis, Zika, and retinal disease research. Yet physical limits prevent them from functioning as full surrogates.

When tissue is thicker than 300 to 400 microns, oxygen diffusion collapses and core cells become hypoxic and developmentally stalled. Even vascularized constructs “age in reverse,” arrested in a fetal-like epigenetic state. Chronic diseases such as nonalcoholic steatohepatitis (NASH), Parkinson’s, or Alzheimer’s cannot unfold in tissues that never mature.

Variability compounds the issue. Each organoid is slightly different, producing wide confidence intervals that make regulators dismiss efficacy signals as noise.

Takeaway: Organoids are ideal hypothesis-generation tools for early discovery and precision medicine but not yet reliable decision-making assays. Their future impact will depend on automation, perfusion technologies, and integration with in vivo or in silico platforms.

AI-First Drug Discovery: Data Without Biological Grounding

AI drug-design systems promised to compress medicinal chemistry cycles from years to months. But when the first AI-generated molecules entered clinical testing, early enthusiasm gave way to underwhelming outcomes.

The issue is not GPU horsepower but data quality. Public bioactivity databases are saturated with positive hits and sparse in negative results or failed experiments. Models trained on this skewed landscape inherit historical bias and reproduce it in silico.

When Recursion, BenevolentAI, and Insilico advanced their first algorithm-designed compounds, they performed impressively in simulations but failed in early trials. Five years later, no AI-generated molecule has reached approval.

As Exscientia founder Andrew Hopkins put it, “Better molecules are meaningless without better translational models.”

Integration By Design: The Path To Predictive Translation

Every model reflects part of the patient:

• Organ-on-chips platform mechanotransduction and barrier dynamics.
• Murine systems integrate systemic and hormonal feedback.
• Organoids replicate genotype-specific behavior.
• AI engines explore combinatorial chemical and biological space.

Individually, they’re incomplete. Integrated, they can become predictive.

Regulators are beginning to recognize this. The FDA’s evolving “totality-of-evidence” framework encourages sponsors to justify mechanisms and safety using convergent data from multiple model types rather than relying on a single surrogate.

To capitalize, pharma must move from ad-hoc combination to integration by design.

Action Items For R&D Leaders:

1. Budget for convergence, not replacement. Dedicate 20%–30% of translational budgets to studies that cross-validate between systems.
2. Create hybrid teams. Pair data scientists with toxicologists, pathologists, and pharmacologists to reconcile contradictions early.
3. Standardize metadata. Shared assay vocabularies and ontologies across models reduce friction and enable quantitative comparison.
4. Build iterative validation loops. Let experimental results retrain AI models and inform organoid or chip design in real time.
5. Adopt an evidence-fusion mindset. In planning preclinical work, measure success by coherence between models, not by the dominance of one.

Biomedicine’s future won’t be defined by a single superior model. It will be built on systems that connect AI, organoids, animal models, and chips form a continuous translational feedback network.

The companies that succeed next will not be those that find “the best model” but those that can synthesize multiple imperfect ones into a reliable composite. Predictive translation will belong to organizations that treat model disagreement not as a flaw but as a road map for refinement.

Drug development doesn’t fail for lack of tools. It fails when those tools remain unaligned. The solution is not another mirror but a clearer composite reflection of the patient, built from all of them.

References

1. U.S. Government Accountability Office. (2025, May). Human Organ-On-A-Chip: Technologies Offer Benefits Over Animal Testing but Challenges Limit Wider Adoption (GAO-25-107335). Washington, DC: U.S. Government Accountability Office.
2. Trietsch, S. J., Hankemeier, T., & van der Meer, A. D. (2021). Finite-element modeling of compound absorption in PDMS-based organ-on-chip devices reveals >90% loss of lipophilic drugs such as amodiaquine. Lab on a Chip, 21(4), 740–752. Royal Society of Chemistry.
3. Zhang, Y., Liu, Q., & Chen, R. (2025). High-throughput organ-on-a-chip systems: Unmet needs and automation hurdles. Lab on a Chip, 25(2), 345–359. Royal Society of Chemistry.
4. Abdali, O., et al. (2019). Verubecestat in mild-to-moderate Alzheimer’s disease: Phase III EPOCH results. Alzheimer’s Research & Therapy, 11(1), 56. BioMed Central.
5. U.S. Food and Drug Administration (FDA). (2011). Voluntary market withdrawal of Xigris (drotrecogin alfa) following PROWESS-SHOCK trial. FDA Drug Safety Communication, U.S. Department of Health & Human Services.
6. Chen, X., Huang, J., & Wang, Y. (2021). Vascularized organoids on a chip: Diffusion limit ≈300 µm drives necrotic cores. Lab on a Chip, 21(16), 3105–3118. Royal Society of Chemistry.
7. ADC Technologies. (2022, July). The world’s first AI-designed drug (DSP-1181) halts R&D. Tokyo: ADC Technologies.
8. Fierce Biotech. (2023, April). BEN-2293 pan-Trk inhibitor flunks mid-phase eczema trial, prompting BenevolentAI layoffs. Fierce Biotech, April 2023 issue.
9. Patsnap Synapse. (2023, November). EXS-21546: R&D progress and first-in-human data. London: Synapse Analytics.

About The Author

Alice Gilman is a biotechnology entrepreneur and longevity researcher focused on advancing regenerative and replacement medicine. She is a cofounder of a breakthrough stealth biotech company developing new platforms for drug testing and exploring ethical animal model replacements with a full-body application.

She is a fellow and counselor at the Longevity Biotech Fellowship and an advisor to The Residency, an incubator backed by Sam Altman supporting emerging biotech founders.

Previously, she ran operations in over 30 countries, ran a different company, and contributed to operations at Symbiont Labs — a deep-tech venture focused on human augmentation technologies. She holds a BA in international relations from the University of British Columbia and a Master’s in applied neuroscience from King’s College London.