Recently, I came across this news release from the FDA, where as of April 2025, the agency declared it would be phasing out animal testing for monoclonal antibodies (man-made proteins meant to stimulate the immune system) and other drugs. Instead of using in-vivo mouse models, researchers and companies will be encouraged to use advanced computer simulations to predict a drug’s behavior or human based lab models – which could potentially serve as a more accurate model for testing drugs. 

Working in a lab, I always feel a little bad when a mouse is euthanized for an experiment. In addition, the lab research process is lengthy and riddled with many setbacks; if a therapeutic is optimized using in-vitro cell cultures, it’s not guaranteed that it’ll work in an in-vivo mouse model. Due to the fact that there isn’t an unlimited supply of mice, and that breeding and sustaining colonies is resource-intensive, therapies can’t be optimized for in-vivo directly, either. Further, 90% of drugs that enter clinical trials fail for a variety of reasons, with common ones being unforeseen toxicity and a lack of efficacy.  

Just as we ask, “what can be better?”, when engineering the next generation of cancer therapeutics, we also have to ask, “what can be better?”, when validating them. In-vivo animal models have been the standard in biomedical research for decades. While we probably won’t move away from them completely, there has to be more accurate, human-representative models for understanding disease and testing therapeutics. 

Stem Cells & Organoids 

Alternative methods to in-vivo animal models are currently being developed, and include 3D cell culture models derived from stem cells (these are primitive, undifferentiated cells that are able to give rise to any stem cell type). 

Organoids are one of these models, and are organ-like structures derived from self-organizing stem cells. Because they are able to self-assemble, these organoids better mimic the in-vivo environment of organs in comparison to 2D cell cultures, and contain multiple cell types that are able to self-interact. They’re more complex to grow, and require a 3D scaffold consisting of extracellular matrix proteins (a diverse group of molecules providing structural support and allow for cell adhesion and tissue organization) to promote cellular differentiation into miniature organs. 

So far, organoids have been used to model specific diseases for basic science research. For instance, liver organoids were generated from patients with α1-antitrypsin deficiency (a genetic disorder that leads to buildup of a protein, causing liver and lung damage), and found that the A1AT aggregated in organoid cells – shedding light into the mechanism of the disease. Using human derived organoids as a disease model is more accurate than inducing the disease in a mouse and studying that. Furthermore, organoids grown from colorectal tumors were used to perform drug sensitivity assays, correlating better to how a human would respond to the compound.

A simpler version of organoid models are spheroids. These form sphere-like structures, and are created when different cell types aggregate in low-attachment cultureware. Often, spheroids are used for studying the tumor microenvironment (and can be established from tumor cell lines) since their structure allows for the investigation of cell to cell signaling and nutrient gradients, especially after treatment with small-molecule therapeutics. 

Nonetheless, organoids still have their downsides. They have a short lifespan due to incomplete vascularization (in other words: blood vessels and oxygen transport can’t be perfectly simulated). In addition, immune system interactions cannot be simulated either, as organoids are not models for a full organism; this creates a significant limitation of using organoids as a sole model for drug screening. Because of their short lifespan, organoids may not be sustainable in the long-term, especially when it comes to screening a lot of compounds. 

The Computational Side 

Just fifty years ago, using computational and machine learning methods to streamline drug discovery was unheard of. However, now, computational methods can be used to screen large libraries of small molecule compounds to determine which ones have affinity for a certain target. Furthermore, machine learning models – known as quantitative structure-property relationship (QSPR) models – can be implemented to predict the solubility and lipophilicity of compounds, as well as their bioavailability or their ability to cross the blood-brain barrier. These models are trained on large experimental datasets. 

I’m certainly not an expert in this computational side; however, I find it very cool that deep learning models can narrow down therapeutic candidates, saving time, resources (and mice) in the lab. 

Looking Ahead 

It is unlikely that in-vivo animal models will ever be replaced; they’ve been a hallmark of biomedical research for decades, and remain the only way of conducting tests in a full organism (before moving on to humans). 

Yet, better models are now being developed to understand disease and prototype treatments. These have the potential to reduce the number of in-vivo models used and the number of therapeutics that fail upon entering clinical trials – accelerating the pace of research.