How Preclinical Models Are Failing Oncology— And Why Patient‑Derived Organoids Might Finally Fix It
Experts warn that traditional animal models are masking critical drug-response signals, highlighting the need for patient-derived organoids and AI-driven prediction to bridge the translational gap in oncology.
As oncology pipelines grow more complex and more personalized, drug developers are confronting an uncomfortable truth: preclinical models that have underpinned cancer research for decades often fail to predict how a drug will perform in real patients. The field’s reliance on a limited panel of animal model systems—most notably patient‑derived xenografts (PDXs) propagated in immunodeficient mice has contributed to a translational gap that obscures subpopulation effects and leads to drug candidates being deprioritized too early.
With a career spanning over 20 years in drug discovery, regulatory science, and advanced in vitro model development, Dr. Madhu Lal‑Nag, chief scientific officer at InSphero and former director of the Trans‑NIH RNAi Facility at the National Center for Advancing Translational Sciences and program director at the US Food and Drug Administration (FDA), has seen how the mismatch between model systems and real patient biology can distort decision making when it comes to therapeutic development in oncology.
In this article, Dr. Lal‑Nag explains why traditional systems fall short, how patient‑derived organoids and other complex in vitro models (CIVM) offer a more faithful window into real tumor biology, and why the future of cancer drug development may increasingly depend on pairing non‑animal biology with AI‑driven prediction.
The translational gap
Despite rapid innovation, oncology is being held back by a reliance on simplified animal systems. Dr. Lal‑Nag explains that while PDXs are widely used, they remain “homogeneous” tools that cannot reflect the biological heterogeneity of real-world patients.
“The go‑to for regulatory submissions is still PDX… even though now in the in vitro space we have the capability to go straight to the patient resection. Patient tumors propagated in immunodeficient, homogenous animal models, lose the immune context of the disease and hence cannot recreate that heterogeneity and complexity that you see in the clinical trial population.” — Dr. Madhu Lal‑Nag
This mismatch creates a significant business and scientific risk: deprioritization of strong candidates and programs. When developers use models that “smooth out” important and relevant biological differences, they risk deprioritizing viable candidates that might have worked in certain specific human demographics, simply because the signal was lost in a uniform mouse population.
Figure 1. Homogeneous mouse models average out
biological diversity, causing drug‑response signals to disappear—whereas
patient‑derived, heterogeneous human models preserve distinct responder
subgroups. Credit: AI-generated image created using Microsoft Copilot
(2026).
The consequences of relying on immunodeficient animal models:
- Depending on the mechanism of action of a drug, propagating patient tissue in an immunodeficient animal may not show efficacy since the response to the drug may involve the interplay of the tumor and its microenvironment.
- Once a human tumor is placed into a mouse, the model no longer reflects the diversity of patient responses that developers need to see
- Potentially viable therapies are abandoned early because the animal model provided a false negative for the reasons stated above.
A new regulatory reality
The need to bridge this gap is now a regulatory reality. Dr. Lal‑Nag recalls a pivotal 2019 shift at the FDA under then‑Center for Drug Evaluation and Research director Dr. Janet Woodcock, which changed the logic of clinical‑trial analysis. The goal was to ensure that a drug’s success in a specific group of patients was no longer “hidden” by its failure in the general population.
Instead of just looking at overall survival across an entire trial, the agency began looking for subpopulations—specific cohorts where a drug might be life-saving.
“A clinical trial might show no benefit overall, but a specific subgroup—for example, women with breast cancer who carry a particular co‑occurring mutation—could respond beautifully. You’re never going to see that in a mouse model.” — Dr. Madhu Lal‑Nag
This shift carries major implications: if regulators expect sponsors to identify responder subgroups, developers cannot rely on genetically uniform animal models that are incapable of revealing this variability. Developers require a model that can more accurately mirror human diversity.
The impact of regulatory stratification:
- Modern regulatory analysis focuses on identifying specific patient cohorts where a drug is effective rather than relying on broad population averages.
- Standardized animal models create a technical bottleneck because they cannot mirror the human diversity required to find these specific responders.
Organoids and CIVMs: a closer approximation to real patient biology
To bridge the gap between the lab and the clinic, Dr. Lal‑Nag advocates for moving towards a tiered approach that ultimately involves going “straight to the patient resection” (the biopsy). By using organoids and spheroids—3D models grown directly from human tumor tissue—researchers can preserve the natural heterogeneity required to identify how different subpopulations will respond to a drug.
“You’re trying to get as close to the patient as possible without being in the patient, and I don’t think you can get more physiologically relevant than that.” — Dr. Madhu Lal‑Nag
Because organoids retain the unique biological signatures of the original patient, they enable researchers to observe responses in specific human tissues that would otherwise be lost in an average. This directly supports the modern clinical shift toward identifying successful subpopulations rather than just looking at a broad, non-specific average. Importantly, organoids are not a complete recapitulation of the patient: standard tumor organoids reconstruct the epithelial compartment with high fidelity but capture the immune and stromal microenvironment less completely, and clonal selection can occur in culture. The field is rapidly advancing co-culture and immune-competent formats to close that remaining gap—an active frontier that strengthens, rather than undermines, the case for human-relevant models.
What organoids enable:
- Testing therapies on 3D tissue that retains the key genomic alterations and tissue architecture of the donor tumor.
- Identifying specific patient profiles that respond to a drug to prevent successful signals from being “averaged out” or deprioritized.
- Generating data that better reflects real‑world clinical diversity before a trial begins.
Non-animal models meet AI-enabled prediction
The final piece of the puzzle is the integration of AI and machine learning. Dr. Lal‑Nag is enthusiastic about AI’s potential to deepen these insights, but she warns that the biological assays must be built for accuracy and reproducibility from day one.
Advancing Biomarker Discovery With Liquid Biopsy Research
This webinar explores how liquid biopsy research is advancing biomarker discovery through extracellular vesicle (EV)-associated biomarkers and circulating tumor DNA (ctDNA).View Webinar
She warns against retrofitting fragile assays into high‑throughput formats and then attempting to train algorithms on inconsistent data.
“If you want a predictive algorithm, you need to design the assay with that goal from the start. If you retrofit an assay to make it high‑throughput, the data may be physiologically relevant, but it won’t be robust or reproducible.” — Dr. Madhu Lal‑Nag
The goal is to co-develop the biology and the algorithms. When high‑quality patient‑derived data from organoids are paired with AI, developers can begin mapping which patient groups are most likely to benefit—a shift from testing to predicting.
The future of AI in new approach methodologies (NAMs):
- Co-developing high-quality organoid data with AI creates a predictive engine to map clinical success for diverse patient groups.
- Assays must be designed for AI from day one to avoid the retrofitting of fragile lab processes that produce inconsistent data.
As cancer drug development becomes increasingly personalized, the limitations of conventional mouse‑based systems are becoming more pronounced. Patient‑derived organoids offer a more faithful representation of human biology, enabling early detection of subpopulations, more meaningful preclinical decisions, and a clearer path toward AI‑supported trial design. This direction is increasingly reflected in regulatory policy: the FDA’s April 2025 Roadmap to Reducing Animal Testing in Preclinical Safety Studies signals broad regulatory momentum toward human-relevant NAMs, of which patient-derived organoids are a leading example.
“Using non predictive, immunodeficient homogeneous animal populations, you deprioritize drugs that may actually be working in a certain demographic, just because you’re not looking at it the right way.” — Dr. Madhu Lal‑Nag
Key takeaways:
- Unlike animal models, organoids and other CIVMs preserve the complex heterogeneity of a patient’s tumor, capturing drug responses that are routinely masked in uniform animal models.
- The analytical shift at the FDA toward identifying specific responders aligns with the strengths of patient-derived models, which allow for stratification earlier in the pipeline.
- The next frontier of oncology R&D lies in co-developing models with AI frameworks to move from “testing” to “predicting” patient outcomes.
This content includes text that has been created with the assistance of generative AI and has undergone editorial review before publishing. Technology Networks' AI policy can be found here.
Add Technology Networks as a preferred Google source to see more of our trusted coverage.
