Human Intestinal Absorption Revisited: The 1998 Descriptors Survive, and the Method Extends to Permeability
This paper revisits a 1998 model for human intestinal absorption by demonstrating that its original performance relied on a fortunate data split, and subsequently establishes more robust, consensus-based predictive models for intestinal absorption and various permeability assays using genetic algorithm-selected descriptor sets and random forests across 250 stratified extractions.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer
When a person swallows a pill, the medicine must pass through a complex barrier: the wall of the intestine. If it cannot cross this wall, the drug never reaches the bloodstream to do its work. Scientists have long tried to predict whether a new chemical compound will make this journey successfully, using computer models that look at the shape and electrical properties of the molecule. For decades, a specific set of six mathematical measurements derived from a molecule's structure has been the standard tool for making these guesses. These measurements, which capture things like how the molecule is charged and how much surface area it has, were established in a famous study from 1998. That study suggested these six numbers were enough to predict absorption with high accuracy, a claim that has influenced drug development ever since. However, the reliability of such predictions depends heavily on how the data is divided and tested, and the original results relied on a single, specific split of the data that might have been unusually lucky.
A team of researchers recently revisited this classic problem to see if those original six measurements still hold up when tested more rigorously. They took the original 86 compounds from the 1998 study and, instead of testing them just once, ran the experiment 250 times, each time shuffling the data into different groups to see how the model performed on average. They found that the original success story was indeed a fortunate fluke. When the same six measurements were tested across these many different shuffles, the average error was nearly double what the original paper had reported. The model worked well on the data it had seen before, but it struggled to predict the unseen compounds as reliably as once thought. This revealed that the original result was not a robust rule, but rather a specific outcome of one particular arrangement of the data.
Undeterred, the researchers asked if these six measurements could still be useful if applied to a much larger collection of compounds. They took the same six numbers and applied them to a dataset of 780 different drug-like molecules. The results were better than on the small set, but the error rate remained higher than desired. The team then decided to let a computer search for a better combination of measurements. Using a method that mimics natural selection, the computer tested thousands of different combinations of molecular descriptors, looking for the set that produced the most accurate predictions across the 250 different data shuffles. This process identified a new, fixed set of twenty-five measurements that consistently outperformed the original six. This new model, built on a more robust foundation, reduced the prediction error significantly, bringing the average mistake down to a level that is much more useful for real-world drug design.
The researchers did not stop at just human absorption. They applied this same rigorous testing and selection process to three other difficult problems in drug delivery: how small molecules and peptides move through artificial cell layers, and how they pass through synthetic membranes. For each of these four distinct challenges, they built a dedicated model using a specific set of descriptors chosen by the computer. The results showed that while the old six measurements were not enough, a carefully selected, larger set of properties could predict these behaviors with much greater confidence. The new models for small molecules and peptides achieved high accuracy, correctly predicting the movement of compounds across cell barriers in the vast majority of cases. These findings suggest that the principles behind the old model were correct, but the specific tools needed to be updated and expanded to handle the complexity of modern drug discovery.
All four of these new models are now available for public use, allowing scientists to test new chemical structures against these refined standards. The work demonstrates that while the fundamental chemistry of absorption has not changed, the way we measure and predict it must evolve. By moving away from a single, lucky data split and embracing a method that tests predictions across hundreds of different scenarios, the researchers have provided a more honest and reliable map for navigating the journey of a drug from a pill to the bloodstream. The original six measurements were not wrong, but they were incomplete; the new twenty-five measurement sets offer a clearer, more complete picture of how drugs interact with the human body.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.