Cell Painting and machine learning for mechanistic profiling of hepatotoxic botanicals and their constituents

People have long turned to plants to support their health, from simple teas to complex herbal supplements. These botanical products are a massive part of the global wellness industry, yet they present a unique challenge for safety experts. Unlike a single chemical drug, a bottle of herbal supplement is a complex mixture containing hundreds of different compounds. While we know some plants can harm the liver, it is often impossible to tell which specific ingredient in the mix is responsible for the trouble or how it causes that damage. Traditional safety testing struggles with this complexity, often leaving regulators and consumers in the dark about the true risks hidden inside a natural remedy.

To solve this, researchers are turning to a method called Cell Painting. Imagine a microscope that doesn't just take a picture of a cell, but paints it with five different fluorescent colors, each highlighting a specific part of the cell's machinery, such as the nucleus, the energy factories, or the outer skin. When a chemical disturbs a cell, these painted parts change shape or arrangement in unique ways. By capturing these changes, scientists can see not just if a cell is dying, but how it is reacting. This visual data is then fed into computer models trained on thousands of known chemicals. These models act like a library of biological responses, allowing researchers to compare a new plant extract against a vast database of known behaviors to guess what is happening inside the cell.

In a recent study, a team of scientists applied this approach to thirteen different botanical extracts and eighty-five of their individual chemical ingredients. They tested these substances on primary human liver cells, which are the most relevant model for understanding liver safety. The researchers exposed the cells to eight different concentrations of each plant extract and its components, then used the Cell Painting technique to capture the cellular response. They also ran standard health checks to see if the cells were leaking or losing energy, providing a baseline for comparison. The goal was to see if this high-tech imaging could spot activity that traditional tests missed and, more importantly, to identify which specific ingredients within the complex plant mixtures were driving those effects.

The results showed that the Cell Painting method was far more sensitive than the standard health checks. It detected biological activity in eleven out of thirteen plant extracts and forty-six out of eighty-five individual ingredients, whereas the traditional tests only found activity in nine extracts and forty ingredients. Crucially, the imaging method often spotted these effects at much lower concentrations. For example, in nine out of the thirteen cases where the plant extract caused a reaction, at least one single ingredient within that extract caused an even stronger reaction at a lower dose. This suggests that the overall effect of a plant mixture is often driven by a small handful of powerful ingredients, rather than the sum of all its parts working together equally.

The study also revealed that the most potent plant extracts were Aristolochia and Goldenseal. Both of these showed strong signs of cellular stress at very low doses. The researchers were able to trace this back to specific ingredients, such as aristolochic acid I in Aristolochia and berberine in Goldenseal, which were themselves highly active. However, the relationship between the whole plant and its parts was not always straightforward. In some cases, like with Ashwagandha and Comfrey, the individual ingredients were highly active in the lab, yet the whole plant extract did not show the same strong reaction at the tested levels. Conversely, the Asian Ginseng extract showed a strong reaction even though its individual ingredients were less potent on their own. This highlights that the way ingredients interact inside a mixture can change the final outcome, making it difficult to predict safety just by looking at isolated chemicals.

Beyond simply detecting activity, the researchers used the visual data to generate hypotheses about how these plants might be affecting the liver. By comparing the cellular "paintings" of the botanicals to a massive reference library of over a thousand chemicals with known effects, the computer models suggested specific mechanisms of action. For instance, the models predicted that Goldenseal and Aristolochia might be interfering with the body's hormonal systems or damaging DNA, which aligns with what is known about their toxicity. The imaging also showed that different parts of the cell responded differently; for example, Aristolochic acid I caused a distinct change in the DNA channel, consistent with its known ability to cause genetic damage.

The study concludes that this combination of high-detail imaging and machine learning offers a powerful new way to screen complex plant mixtures. It allows scientists to move beyond asking simply "is this toxic?" to asking "what is this doing, and which part of the plant is doing it?" While the findings are based on lab-grown liver cells and do not directly prove that these plants will cause liver damage in people, the method successfully prioritizes the most active ingredients for further study. It provides a clear, scalable path to understanding the hidden drivers of toxicity in botanical products, helping to separate the beneficial from the harmful in the vast world of herbal remedies.