Toxicity by Consortium Revisited: Structural Neighbors and Cross-Family Reach

Drug development is a long, expensive journey where most candidates fail before they ever reach a pharmacy shelf. A major reason for these failures is toxicity: a compound might cure a disease in its intended target but accidentally harm the body by interacting with the wrong proteins or causing damage in the liver, heart, or other organs. Scientists have long known that a molecule's physical shape and chemical properties often hint at how it will behave in the body. If a new drug looks very similar to a known poison, it is likely to be toxic itself. This principle of similarity has been used for decades to predict safety, but the tools and data available to make those predictions have changed drastically since the early days of the field.

Twenty-three years ago, researchers proposed a simple idea: if you have a new chemical, you can guess its toxicity by looking at the known toxicities of the chemicals it resembles most closely. They built a method where a new compound would gather a "consortium" of similar neighbors from a reference library, and those neighbors would vote on the new compound's safety. However, the original library they used has since become private and unavailable, leaving the method stuck in the past. Now, a team of researchers has rebuilt this approach using modern, public data to see if the old idea still holds up. They also asked a second, deeper question: does the sheer number of different types of protein targets a chemical resembles tell us something new about its safety, beyond just its size or how oily it is?

The researchers started by reconstructing their original voting system using a fresh, public collection of nearly 12,000 measured toxicity values for rats. They tested this method on thousands of new compounds to see if it could predict their lethal dose, a standard measure of how much of a substance is needed to kill half of a test population. The results were encouraging. For about two-thirds of the new compounds, the system could provide a specific prediction. Of those predictions, more than 90 percent were accurate enough to be within a single order of magnitude of the actual measured value. This performance is comparable to what the team achieved in 2003, proving that the core idea of using similar neighbors to predict toxicity remains valid even as the data landscape has shifted.

But the team did not stop at just predicting how lethal a substance might be. They expanded their test to twenty-eight different types of toxicity, ranging from whether a chemical damages the heart or causes mutations in DNA, to whether it blocks specific enzymes in the liver. They applied the same voting logic to these diverse endpoints. The method worked well across the board, correctly identifying active toxic compounds about 73 percent of the time and correctly identifying safe compounds about 96 percent of the time. Crucially, the system only makes a prediction when it finds enough close neighbors; if a compound is too unique, the method simply declines to answer, avoiding the trap of guessing blindly.

The more novel part of the study involved a different kind of search. Instead of just looking for similar chemicals, the researchers asked how many different families of proteins a new chemical resembles. They used a massive, public index of over 27,000 drug-like molecules that have been solved inside protein structures. For every new chemical, they counted how many distinct protein families its structural matches belonged to. They found a striking pattern: chemicals that resemble binders for many different protein families are far more likely to be toxic. For example, in the case of a specific liver enzyme called CYP3A4, compounds that matched no protein families were inhibitors only 2.5 percent of the time. In contrast, compounds whose matches reached sixteen or more different protein families inhibited that same enzyme 51.2 percent of the time. This suggests that a chemical's "promiscuity"—its tendency to fit into many different biological pockets—is a strong warning sign.

To ensure this pattern was not simply a result of the data, the researchers ruled out several alternative explanations. They showed that the result was not simply because larger, more complex molecules tend to hit more targets; the signal held true even after accounting for molecular size and oiliness. They also proved it was not because the chemicals were just close copies of well-studied compounds; even when they removed all known analogs from the search, the link between reaching many protein families and being toxic remained strong. Furthermore, they tested this finding on a completely separate set of 2,335 compounds measured by a different consortium, and the same trend appeared: compounds reaching many protein families were significantly more likely to be active inhibitors.

The researchers combined these two signals—the voting consortium and the count of protein families reached—into a single scoring tool designed to screen virtual libraries of chemicals before they are ever made in a lab. When they ran this tool on a large set of compounds, it successfully flagged the riskiest ones. By setting aside the worst-scoring fifth of the library, they could remove two-thirds of the compounds that bind to the androgen receptor and about half of the CYP3A4 inhibitors. This kind of early filtering is valuable because it allows scientists to discard dangerous paths before spending time and money synthesizing them. The tool requires only a two-dimensional drawing of a molecule and runs in seconds, needing no complex 3D modeling or docking simulations.

The study concludes that the old method of using similar neighbors is still a powerful predictor of toxicity, but it gains new strength when paired with a measure of how broadly a chemical's structure reaches across the protein world. The authors argue that the best way to improve these predictions further is through collaboration. They propose that pharmaceutical companies could share encoded data about their failed compounds—data that reveals toxicity without revealing the specific chemical structures—into a shared pool. Just as a larger library of neighbors improves the voting system, a larger shared pool of toxicity data would allow the method to answer for more compounds with greater confidence. The researchers have made their software and data available to the public, inviting others to contribute and use these tools to make drug discovery safer and more efficient.