Why NAMs Are Replacing Animal Studies: 9 Examples That Show It Can Work

Introduction

The NIH recently updated its guidelines: grant proposals using animal studies must now justify why New Approach Methodologies (NAMs) are not used. In many domains-especially toxicology, pharmacology, and disease modeling-NAMs are becoming the expected default.

While the limitations of animal studies and ethical concerns are widely acknowledged, many classically trained researchers remain skeptical about how computational and human-based models can replace animal testing.

At AccuraScience, we’ve supported this shift in practical ways. We’ve helped researchers integrate NAM strategies into competitive grants and high-impact publications. From deep learning and digital twins to multi-omics models, we understand both the technical and regulatory complexities of this transformation.

NAMs simulate biological processes using human-relevant, often multi-modal data-replacing animal models with machine learning and mechanistic frameworks. Below, we share 9 concrete examples showing how NAMs can effectively substitute for animal experiments in diverse biomedical areas.

Throughout this discussion, we categorize human data sources using the following labels:

- Publicly available

- Available but restricted

- Private (collaborator/internal)

- Not yet available (must be generated)

1. Drug Development & Toxicology

Example: Predicting Liver Toxicity Without Animal Models

Scenario: You’re developing a new compound. Traditionally, you would give it to mice or rats to assess hepatotoxicity.

NAM Alternative: Use in vitro liver organoid data and human liver toxicity datasets (e.g., ToxCast, SIDER) to train a machine learning model that predicts hepatocyte damage.

Human Data Used:

- Gene expression profiles from human liver organoids exposed to test compounds - Available but restricted

- Known adverse event data (e.g., FAERS, SIDER) - Publicly available

- HepaRG transcriptomics - Available but restricted

Why It Works: These models predict drug-induced liver injury (DILI) mechanisms, such as ROS production or steatosis, using human-relevant input - bypassing rodent uncertainty.

2. Cancer & Disease Modeling

Example: Modeling Tumor Immune Microenvironment to Predict Therapy Response

Scenario: Instead of mouse xenografts, you want to simulate immune checkpoint inhibitor response in a human tumor.

NAM Alternative: Use scRNA-seq and spatial transcriptomics to model immune infiltration, exhaustion, and therapeutic outcome.

Human Data Used:

- scRNA-seq and spatial transcriptomics from human tumors - Publicly available

- Imaging mass cytometry (CODEX, MIBI) - Available but restricted

- Clinical response data (e.g., RECIST) - Available but restricted

Why It Works: You can model the tumor–immune landscape at single-cell resolution and simulate therapy effect - making predictions more human-relevant than in mice.

3. Personalized Medicine

Example: Digital Twin for Predicting Seizure Control in Epilepsy

Scenario: A patient with refractory epilepsy is facing multiple drug options. Instead of trial-and-error, can we predict response?

NAM Alternative: Build a digital twin using EEG, genomics, and prior treatment response to simulate effects of anticonvulsants.

Human Data Used:

- Whole-exome sequencing showing ion channel mutations - Publicly available

- Pharmacogenomics data (e.g., CYP2C9, SCN1A) - Publicly available

- EEG under prior treatments - Private

- Clinical seizure diaries and drug outcomes - Private

Why It Works: Digital twins simulate network-level brain excitability and drug action mechanisms - improving outcomes without subjecting animals to repeated seizure induction.

4. Organ-on-a-Chip & Microphysiology

Example: AI Analysis of Cardiotoxicity in Heart-on-Chip Systems

Scenario: You're screening compounds for pro-arrhythmic effects without relying on ECGs from guinea pigs or dogs.

NAM Alternative: Apply deep learning to video microscopy from human iPSC-derived cardiomyocytes in heart-on-chip devices.

Human Data Used:

- Beating video recordings of iPSC-cardiomyocytes - Available but restricted

- Reference cardiotoxic drug sets (e.g., QT-prolongers) - Publicly available

- Human clinical trial ECG/QT data - Publicly available

Why It Works: Deep neural networks can detect arrhythmic patterns or repolarization abnormalities in beating human cells, offering early-stage toxicity prediction.

5. Inflammatory Disease Simulation

Example: Modeling Cytokine Storm in Viral Infection

Scenario: You want to understand systemic inflammation caused by a virus (e.g., COVID-19) - without using ferret or mouse infection models.

NAM Alternative: Build a computational model of cytokine signaling using human PBMCs and patient-derived plasma data.

Human Data Used:

- Cytokine levels over time from hospitalized COVID-19 patients - Publicly available

- scRNA-seq and proteomics from LPS-stimulated PBMCs - Publicly available

- Clinical metadata (e.g., ICU admission, mortality) - Available but restricted

Why It Works: By modeling feedback loops and activation thresholds, researchers can simulate escalation to cytokine storm and test virtual interventions.

6. Neurotoxicity Evaluation in CNS Drug Discovery

Example: Predicting Neurotoxicity Using Human Brain Organoids

Scenario: Screening CNS-targeted compounds for off-target effects - traditionally done in rodents.

NAM Alternative: Use transcriptomic and electrophysiological features from human brain organoids exposed to known and novel compounds.

Human Data Used:

- Gene expression after drug exposure in organoids - Available but restricted

- Electrophysiological recordings (MEA, patch clamp) - Private

- Reference neurotoxic compound outcomes - Publicly available

Why It Works: You can identify synaptic dysfunction or toxicity signatures in human-relevant 3D neural models - improving prediction of neurodevelopmental or neurodegenerative risks.

7. Autoimmunity Post-Vaccination

Example: Predicting Autoimmune Flares After Vaccination

Scenario: Concern that vaccination may trigger lupus or RA flare in vulnerable individuals - usually tested in autoimmune-prone mice.

NAM Alternative: Model post-vaccine cytokine and immune cell activation based on human datasets from vaccinated autoimmune cohorts.

Human Data Used:

- Pre/post-vaccine cytokine and B/T cell panel data - Available but restricted

- Autoimmune disease transcriptomes (RA, SLE, MS) - Publicly available

- Flare incidence from vaccinated autoimmune patients - Private

Why It Works: Allows you to simulate immunologic risk in real patients using mechanistic immune models - without triggering flares in live animals.

8. Endocrine Disruption

Example: Predicting Estrogenic Activity Without Rodent Uterotrophic Assays

Scenario: Regulatory need to evaluate endocrine disruption - traditionally tested in rat models.

NAM Alternative: Combine transcriptomic response in MCF-7 cells and computational docking models to predict estrogenic or anti-estrogenic effects.

Human Data Used:

- Dose–response gene expression in ER+ cell lines - Publicly available

- Ligand–ER binding affinities via in silico docking - Publicly available

- Reference compounds with known hormonal effects - Publicly available

Why It Works: Enables mechanism-specific, high-throughput prediction of estrogen receptor modulation, aligned with human pathways.

9. Skin Wound Healing & Fibrosis

Example: Modeling Fibrosis in Human Skin Without Rabbit Models

Scenario: Testing anti-fibrotic therapies or wound healing compounds - commonly using rabbit or pig skin.

NAM Alternative: Analyze tissue remodeling in ex vivo human skin explants using image analysis, transcriptomics, and machine learning.

Human Data Used:

- Histology images of skin after compound exposure - Private

- Gene expression of fibrosis markers (COL1A1, ACTA2, TGFB1) - Publicly available

- Mechanical contraction or stiffness measurements - Available but restricted

Why It Works: Enables modeling of collagen deposition, epithelialization, and fibroblast activity in human-relevant systems - reducing reliance on large animal surgery.

Final Thoughts: What These Examples Teach Us

What we’ve tried to highlight here is not just theoretical. These are real and grounded examples from work we and others have already done.

In this NAM transformation, the same kinds of measurements - for both input variables and output outcomes - generally stay the same. We’re still working with cytokine levels, gene expression, organoid behavior, clinical outcomes. What actually replaces animal testing is the construction of a mechanistic or machine learning model trained on human-relevant datasets.

These datasets are often multi-modal: genomics, proteomics, imaging, clinical metadata - sometimes messy, often complex. The main practical disadvantage is clear: it takes a lot of carefully collected human data to build something that works. But once a model works and is validated, it becomes a lot easier to test another drug, another variant, another hypothesis. You don’t go back to square one. You simulate and refine.

We understand this change is not trivial. But this is where we bring real value. If you’re working on a grant proposal and wondering how to adapt to this shift - you’re not alone. And we’re ready to help.

This blog article was co-authored by Zack Tu, Ph.D., Lead Bioinformatician and Justin Li, Ph.D., Lead Bioinformatician. To learn more about AccuraScience's Lead Bioinformaticians, visit https://www.accurascience.com/our_team.html.