Model vs. Model: How Humanized Mice Compare to Organoids, Chips, and AI

‍

Preclinical Disease Models That Deliver Better Predictive Power‍

‍

In part 3 we explored how multi-cytokine and dual-organ humanized mouse platforms, such as MISTRG and NSG-SGM3, are advancing disease modeling by enabling human-like hematopoiesis, immune function, and organ-level integration for more physiologically relevant studies of cancer, infection, and inflammation. In this part 4 we tackle the rapid evolution of biomedical model systems, which has created a rich but fragmented landscape: organoids and 3D cultures provide genetically precise, patient-specific platforms; microphysiological systems (organs-on-chips) deliver unprecedented control over cellular environments; and in silico/AI simulations offer predictive power at unmatched speed and scale. Yet each of these approaches faces fundamental limitations in capturing the full complexity of a living organism. Humanized mice remain the only models capable of integrating immune, vascular, and systemic physiology over time, making them indispensable for validation. Increasingly, these models are viewed not as competitors but as collaborators in a sequential and hybrid workflow, one that regulatory agencies such as the FDA are beginning to formally recognize as the future of translational research.

‍

Organoids and Organoid–Mouse Hybrids

‍

Organoid cultures offer a high-throughput, genetically tractable platform that recapitulates many aspects of human organ architecture. Complex genetic modifications can be achieved in human organoids within weeks, compared to the many months often required to breed transgenic mice. This rapid turnaround makes organoids invaluable for initial mechanistic studies and large-scale drug screens where multiple genetic or pharmacological perturbations need evaluation. Notably, organoids derived from human tissues capture unique cell types and functions that are absent in murine models [1,2]. For example, human intestinal organoids contain motilin-producing enteroendocrine cells and BEST4/OTOP2-expressing absorptive cells that do not exist in the mouse gut, underscoring species-specific physiology that only human-based models can reveal [2]. The ability to establish organoids from patient biopsies also enables personalized disease models, allowing direct study of patient-specific mutations and therapeutic responses in vitro [1]. These strengths position organoids as powerful complements to in vivo studies, filling gaps in human relevance and experimental throughput that traditional mouse models cannot easily address [1].

‍

Despite their advantages, organoids have fundamental limitations in recapitulating the full complexity of living organs. By nature, organoids lack the vascular perfusion, immune cell repertoire, neural innervation, and mechanical forces present in vivo. This means processes that depend on multicellular interactions across tissues – immune responses, inflammation, endocrinology, or drug pharmacokinetics – are difficult or impossible to model fully in stand-alone organoids. For instance, organoid cultures cannot yet mimic an intact circulatory system or the recruitment of immune cells to a site of infection or tumor. These gaps constrain the utility of organoids for studying systemic diseases or interventions. Recent advances are beginning to bridge these gaps: co-culture systems that add specific stromal or immune components, and organoid–mouse hybrid approaches, are emerging to provide more physiological context. Transplanting human organoids into immunodeficient mice is one such strategy. In an in vivo environment, human organoids can vascularize and integrate into host tissue, allowing maturation and cell–cell interactions that approximate normal organ function. Notably, human colonic organoids engrafted in mice have been shown to develop crypt-villus structures more closely resembling native human intestines than the surrounding mouse epithelium [1]. By performing CRISPR gene edits in organoids prior to transplantation, researchers can interrogate human gene functions within a living organism,  something not feasible with standard mouse models alone.

‍

Humanized mouse models remain irreplaceable as the in vivo component providing systemic context to complement organoid findings. Immune-humanized mice, for example, enable studies of human immune–organ interactions and longitudinal disease dynamics that organoids alone cannot achieve. The MISTRG humanized mouse model of myelodysplastic syndrome (MDS) illustrates this advantage: it supports engraftment of patient-derived hematopoietic cells and maintains their human genetic complexity, while allowing multi-lineage interactions in an organism over time. Such mice can be serially sampled for human cytokines or immune responses and can model disease progression and therapy response over months of experimental windows far beyond organoid cultures [1]. In sum, organoids excel at rapid, reductionist exploration of human-specific biology, whereas humanized mice provide the integrative, whole-body validation. Increasingly, hybrid approaches (like organoid engraftment or “avatars”) combine these strengths, underscoring that organoids and humanized mice are more effective as partners than as competitors in modeling human disease.

‍

Microphysiological Systems (Organs-on-Chips)

‍

Precision and Throughput: Microphysiological systems – often called organ-on-a-chip devices – are engineered microfluidic platforms that aim to recreate critical aspects of organ-level function in a controlled setting. These devices incorporate living cells on biomimetic scaffolds with perfusion of media, allowing precise control of the cells’ microenvironment (such as fluid flow, stretch, or biochemical gradients). The result is a miniature “organ” that can be observed in real time under tightly defined conditions. Organs-on-chips have demonstrated an extraordinary ability to dissect mechanical and chemical cues in physiology. A landmark example is the lung-on-a-chip, which consists of human alveolar epithelial cells on one side of a porous membrane and capillary endothelial cells on the other, subjected to cyclic stretching to simulate breathing [3]. This system successfully replicated the alveolar–capillary barrier and even showed that physiological breathing motions accentuate nanoparticle absorption and inflammatory responses in the lung. Such biomimetic chips have provided insights into organ function and pathophysiology that are difficult to attain in vivo – for instance, identifying how mechanical forces contribute to lung edema and drug toxicity, or how shear stress influences vascular inflammation. Moreover, because chips are amenable to parallelization, they are useful for high-throughput drug screening under organ-specific conditions. Researchers can test many compounds on identical “organs” arrayed on chips, with automated readouts of barrier integrity, electrophysiology, or other functional metrics. This level of precision and throughput makes MPS platforms valuable for mechanistic studies and initial safety screens in drug development [3].

‍

Limitations in Systemic Biology: Despite their sophisticated design, organ-on-chip systems cannot yet capture the full physiological complexity of living organisms [4]. Each chip typically models a narrow slice of biology – e.g. the lung’s air-blood barrier or the liver’s metabolic functions – in isolation. Crucially, they often lack immune cells, endocrine feedback, multi-organ communication, and the adaptive capacity of real organs. For example, a liver-on-chip can mimic enzyme-mediated drug metabolism, but it will not account for an immune-mediated liver injury or the neuroendocrine regulation of metabolism. Likewise, a heart-on-chip can beat and respond to drugs affecting contractility, but it does not experience circulating immune cells or hormonal surges that a heart in a living organism would encounter. This reductionism means MPS models may miss emergent phenomena that arise only when multiple organ systems interact. To address this, researchers are developing multi-organ chip platforms that connect several organ modules into one circuit, permitting inter-organ crosstalk via a shared perfusion medium [4]. Examples include two-organ partnerships like a gut–liver chip to study oral drug absorption and first-pass metabolism, or more complex “body-on-a-chip” setups linking liver, heart, lung, kidney, and others in physiologically scaled proportions. A recent demonstration interconnected up to ten different human tissue organoids in a closed loop to simulate pharmacokinetics across the whole body [4]. These advances show promise in recapitulating systemic drug distribution and multi-organ toxicities in vitro. However, multi-organ MPS technology remains technically challenging: each tissue must survive and function stably, and the fluidic coupling must reproduce realistic timings and concentrations (e.g. metabolites from a “liver” chip reaching a “tumor” chip at relevant levels). Ensuring that such systems are physiologically faithful (for instance, matching human organ sizes and cell ratios) is an ongoing hurdle. Consequently, even the most advanced MPS cannot yet duplicate an immune system’s complexity or a fully orchestrated pathological response (such as fever or fibrosis involving multiple organs). Regulatory agencies acknowledge these limitations – the U.S. FDA recently noted that while organ-on-chip and other new approach methodologies are promising, they are “not ready to fully replace animal testing” without further validation [5].

‍

Humanized mice complement MPS by providing the integrated physiology that chips lack. Immune-humanized mice, for example, have been used to study infections like human-tropic respiratory viruses that organ chips cannot truly emulate (since chips lack a full immune system to mount antiviral responses). The BLT-L humanized mouse model – which implants human lung tissue along with bone marrow, liver, and thymus – can be infected by human pathogens and generate human immune responses in the lung, an achievement far beyond current in vitro models. Additionally, only in an in vivo model can one observe behaviors like immune cell trafficking between organs, drug metabolism with feedback loops, or chronic toxicity that develops over weeks to months. Humanized mice thus serve as a critical bridge from the controlled conditions of MPS to the messy but crucial reality of organismal biology. In practice, a sequential strategy is often most powerful: one can use organ-on-chips to identify organ-specific mechanisms or screen compounds under defined conditions, then employ humanized mice to test the most promising findings in an environment that includes all the systemic variables (immune system, neuroendocrine axes, etc.). This way, the strengths of MPS (precision and throughput) are combined with the strengths of animal models (complex integrative physiology), leading to more robust and translatable results.

‍

In Silico and AI Models

‍

Target Prediction and Virtual Trials: The rise of in silico modeling and artificial intelligence has added a new dimension to biomedical research. AI-driven computational models can analyze vast datasets – spanning genomics, chemical libraries, clinical records, and more – to generate hypotheses or predictions at a speed and scale no experimental system can match. In drug discovery, machine learning algorithms are now routinely used to predict drug–target interactions and optimize lead compounds. For instance, AI models can screen millions of chemical structures against a target protein in silico to prioritize those most likely to bind strongly, focusing experimental testing on only the top candidates. This greatly accelerates the early-phase discovery process. AI can also suggest modifications to improve a drug’s pharmacokinetics or reduce toxicity by learning from large datasets of previous compounds. Beyond discovery, in silico trials are emerging as a tool in situations where traditional clinical trials are challenging. In rare diseases – where patient numbers are limited and enrolling a placebo control arm may be unethical – researchers can use computational disease models and “virtual patients” as synthetic control arms. By leveraging real-world patient data and disease simulations, these models can predict how an untreated group would behave, allowing efficacy of a new therapy to be assessed without a conventional placebo group [6]. Regulatory authorities have begun to accept evidence from such approaches. Notably, the FDA has approved therapies for ultra-rare diseases based on studies that compared treated patients to historical or simulated controls when no feasible alternative existed. In one case, a gene therapy for Batten disease was approved by the FDA using a natural history cohort as the comparator, effectively a synthetic control arm, due to the impracticality of a randomized trial. During the COVID-19 pandemic, computational modeling proved its value by predicting immune responses and optimal vaccine dosing schedules before clinical data were available [5]. Indeed, the remarkable concordance between some in silico vaccine efficacy predictions and the eventual trial outcomes in humans helped validate the predictive power of these models [5]. Generative AI techniques are further expanding capabilities by creating synthetic data to fill knowledge gaps – for example, generating plausible patient datasets for rare conditions to train algorithms, or designing novel protein structures as starting points for drug or vaccine development. Overall, AI and in silico methods offer an unprecedented ability to explore “what-if” scenarios and iterate through possibilities rapidly, essentially conducting virtual experiments that guide real-world research.

‍

Gaps in Biological Complexity: For all their promise, purely computational models have important limitations. Biology is exceedingly complex, and not all of it can be captured in silico. Models are abstractions of reality – they rely on known parameters and data. Thus, in silico predictions are only as good as the data and assumptions underlying them. When abundant high-quality data exist (for example, in well-studied domains like small-molecule drug–enzyme interactions), AI predictions can be highly accurate. But in frontier areas – a novel disease mechanism, a first-in-class therapeutic modality, or a multifactorial immune response – the algorithms may be flying partially blind. Emergent properties of living systems, which arise from intricate cell–cell communications, feedback loops, and microenvironmental factors, often elude computational modeling. The immune system is a prime example: while AI can predict an antigen’s binding to an HLA molecule or T-cell receptor to some extent, predicting a full immune response (with co-stimulation, memory formation, cytokine networks) is far more complex. Additionally, AI models can inadvertently incorporate biases or errors present in the training data, leading to spurious predictions. In drug safety, an AI might predict low toxicity for a compound because no training data suggested a particular off-target risk – but in a real human, that compound might be metabolized into a toxic intermediate, a nuance the model didn’t “know.” Therefore, experimental validation remains essential. Humanized mice serve as an important proving ground for AI-generated hypotheses in a setting that approximates human physiology. For example, if a machine learning model flags a new oncology drug as effective against a human tumor based on in vitro and omic data, one can test this by engrafting the patient’s tumor into an immune-humanized mouse to see if the predicted response occurs in vivo. The MISTRG humanized mouse has been used in this way to evaluate candidate immunotherapies: it can sustain human myeloid and lymphoid cells, thus if an AI-designed bispecific antibody is predicted to activate T cells against a leukemia, the MISTRG mouse can reveal whether this actually happens in a living organism. Similarly, humanized mice with human HLA gene knock-ins allow vaccine concepts generated by computational immunology to be tested for real T-cell responses and immunogenicity that mirror human immune constraints. In each case, the discrepancies between in silico predictions and in vivo outcomes provide invaluable feedback to refine the models. This synergy between AI and animal models is increasingly recognized as the optimal path forward: use AI to narrow the search space and design candidates, then use human-relevant experimental models to verify and further explore those candidates. Indeed, regulatory agencies are crafting frameworks to integrate AI into the drug development pipeline alongside traditional testing. The FDA recently issued guidance on using AI for drug development and is funding initiatives to qualify in silico models as part of regulatory submissions [5]. Nonetheless, the consensus is that while AI can drastically streamline research, it cannot replace biological testing entirely, especially for capturing unanticipated effects. Humanized mice and other advanced preclinical models will remain critical to ensure that AI’s virtual predictions translate into real-world safety and efficacy.

‍

Why Integration Matters

‍

In the landscape of modern biomedical research, advanced models like organoids, microphysiological systems, and in silico simulations should not be viewed as competitors to humanized mice, but rather as complementary tools that operate at different scales [6]. Each model system has distinct strengths and blind spots. By strategically integrating these approaches, scientists can leverage the best of each while mitigating their individual limitations. For example, a logical workflow might proceed as follows: 

‍

1. Use in silico AI models to sift through millions of drug candidates or pinpoint key molecular targets using patient data and machine learning. 
2. Apply organoid or organ-on-chip models to test the top candidates or hypotheses in a controlled human-cell context, rapidly obtaining data on efficacy, mechanism, and organ-specific toxicity. 
3. Finally, take the most promising interventions into humanized mouse models to observe their effects in an integrated whole-body context, assessing pharmacokinetics, immune engagement, long-term outcomes, and cross-organ effects that neither organoids nor chips can fully predict. 

‍

This sequential, multi-model strategy accelerates discovery (by front-loading the quick, high-throughput tests) while preserving rigor (by back-loading the comprehensive in vivo validation). Indeed, many pharmaceutical companies are now adopting such “pipeline” approaches, where computational screening and in vitro human models act as filters before any compound ever enters animal testing. This not only saves time and resources but also aligns with ethical efforts to minimize animal use.

‍

Moreover, hybrid strategies are emerging that blur the lines between these model systems. Organoid-on-a-chip technology, for instance, marries the 3D cellular complexity of organoids with the fluidic perfusion and control of microchips [6-7]. These hybrids can reproduce aspects of tissue architecture (like lumen-forming epithelial tubes) while also providing vascular flow or mechanical forces, thus extending the physiological relevance of in vitro models. Another example is combining AI with laboratory models in real-time: machine learning algorithms can analyze live imaging data from organ chips to detect subtle phenotypic changes and adapt the experimental conditions on the fly (“AI-in-the-loop” experimentation). On the animal front, organoid-humanized mice – where human organoids are engrafted into mice – exemplify a fusion of in vitro and in vivo: they create chimeric animals that carry human tissues or tumors (derived from organoids) alongside a human immune system, thereby enabling drug testing on “human” organs within a living host. As these hybrid and integrative methodologies develop, the boundaries between silico, in vitro, and in vivo will continue to soften, ideally leading to models that are both highly human-relevant and holistically complex [6-7].

‍

A collaborative, integrated model ecosystem is also aligned with regulatory trends. Recognizing the scientific and ethical imperative, regulators have begun encouraging the use of New Approach Methodologies (NAMs) – which include organoids, organ-on-chips, and computational models – to supplement and in some cases replace traditional animal studies. In 2022, the U.S. Congress passed the FDA Modernization Act 2.0, which removed the historical mandate for animal testing in drug approval if valid alternative methods are available. This policy shift, along with the FDA’s April 2025 roadmap, envisions that animal tests will become the exception rather than the norm in preclinical safety studies within the next few years [5]. However, regulators are clear that alternative models must demonstrate their reliability and relevance before displacing animal models entirely. In practice, this means an integrated approach: using organoids and chips for what they do best (human-specific and mechanistic insights) and using humanized mice as the anchor for complex validation that ties everything together. By combining evidence from AI simulations, in vitro human models, and in vivo humanized models, researchers can present a convergent case for a drug’s efficacy and safety that is more robust than any single approach alone. Such a strategy not only improves confidence in translational success when moving to human trials, but also adheres to the 3Rs principles (Replacement, Reduction, Refinement of animal use) in a scientifically grounded way.

‍

Conclusion

‍

Alternative model systems – organoids, organ-on-chip microphysiological systems, and in silico simulations – have each expanded what is possible in biomedical research, offering insights at speeds and resolutions that traditional animal models alone could never achieve. Yet, rather than displacing animal models, these technologies are becoming collaborators with them. Humanized mice remain a cornerstone for modeling the integrated physiology of human diseases, especially when immune, endocrine, and multi-organ interactions are involved. The consensus emerging in the field is that no single model is sufficient on its own; true understanding and successful drug development arise from the synthesis of evidence across models. Organoids and chips can provide early-stage, human-specific data, and AI can guide decision-making, but humanized mice supply the ultimate test of whether those findings hold in a living, complex organism. As the FDA’s recent actions highlight, the future of preclinical research will increasingly rely on a toolkit of complementary models. Rather than a competition of “model vs. model,” the paradigm is shifting to cooperation: in silico models generate hypotheses, in vitro models refine and test them under defined conditions, and in vivo humanized models confirm and expand the results in a system as close to a human patient as possible. Embracing this integrated approach will improve translational predictability — getting therapies from bench to bedside more efficiently — while also reducing redundant animal usage. In summary, humanized mice, organoids, chips, and AI should be seen as synergistic parts of a single research ecosystem. When used in concert, they can accelerate discovery and development in ways none of them could accomplish alone, ultimately benefitting patients by delivering new treatments that have been rigorously vetted from every angle.

References

1. Kim, J., Koo, B.K. & Knoblich, J.A. Human organoids: model systems for human biology and medicine. Nat. Rev. Mol. Cell Biol. 21, 571–584 (2020).
2. Beumer, J. et al. High-resolution mRNA and secretome atlas of human enteroendocrine cells. Cell 182, 1062–1064 (2020).
3. Huh, D. et al. Reconstituting organ-level lung functions on a chip. Science 328, 1662–1668 (2010).
4. Edington, C.D. et al. Interconnected microphysiological systems for quantitative biology and pharmacology studies. Sci. Rep. 8, 4530 (2018).
5. FDA pushes to replace animal testing. Nat. Biotechnol. 43, 655 (2025).
6. Yildirim, Z., Swanson, K., Wu, X., Zou, J. & Wu, J. Next-gen therapeutics: pioneering drug discovery with iPSCs, genomics, AI, and clinical trials in a dish. Annu. Rev. Pharmacol. Toxicol. 65, 71–90 (2025).
7. Huang, Y., Liu, T., Huang, Q. & Wang, Y. From organ-on-a-chip to human-on-a-chip: a review of research progress and latest applications. ACS Sens. 9, 3466–3488 (2024).

‍