Predicting What the Body Will Do to a Molecule | Visual AI Learning Guide

The structure on the left is the only input. From it, separate models each answer a different question: how much crosses the gut wall, where it distributes in the body, how fast the liver breaks it down, how quickly it is cleared, and whether it is toxic or blocks the heart channel. Each model was trained on compounds whose behavior was actually measured in the lab, so it has learned which structural features tend to go with which outcomes. The result is a panel of estimates. Notice that no single number decides anything — the panel is read together, and it lets chemists set aside likely failures before spending weeks synthesizing them.

0:00 / 0:00

Auto-play

Before a compound is synthesized, AI models can estimate what the body will do to it. These predictions are grouped under the label ADMET: absorption, distribution, metabolism, excretion, and toxicity.

Absorption asks how much of an oral dose crosses from the gut into the bloodstream. Distribution asks where the compound goes once in circulation — whether it reaches the tissue of interest or is largely bound to blood proteins. Metabolism asks how quickly liver enzymes chemically modify and deactivate it. Excretion asks how fast the body removes it. Toxicity asks whether the compound or its breakdown products harm cells, and whether it interferes with proteins such as the hERG ion channel, whose blockage is associated with dangerous heart-rhythm effects.

Each of these is predicted from the same input the model already uses for binding: the molecular structure. A trained model maps structural features — polarity, size, the presence of particular chemical groups — onto measured outcomes from thousands of compounds whose ADMET behavior was determined experimentally. The output is a panel of estimates attached to the structure, letting chemists discard compounds that are likely to fail on absorption or safety before spending weeks making them.