A new AI model could speed up drug development by reducing the number of parameters needed to predict how molecules bind to proteins — a vital step in designing new drugs.
The Predicting Affinity Through Homology (
AI model slashes drug discovery parameters
It was created by Bruce Donald, professor of computer science and biochemistry at Duke University, and Yuxi (Jaden) Long, a former undergraduate in Donald’s lab who is now a graduate student at Memorial Sloan Kettering Cancer Center.
Donald said: “Previous models, which used deep learning, use billions of parameters and tens of thousands of feature.
“They report good correlations, but we don’t know why.”
PATH combines interpretable machine learning with algebraic topology — a field of mathematics that studies shapes — to help researchers trace predictions back to specific atomic interactions.
Long said: “We can now understand how the algorithm made these predictions.
“We can see exactly how much each atom contributed.”
The model addresses a major challenge in drug discovery: distinguishing molecules that will truly bind to target proteins from those that will not.
Binding is essential for most drugs to function, but many models overpredict binding because they are trained primarily on positive examples.
Donald said: “If you look at a million small molecules and a protein target, only two or three will actually bind.
“Most previous models predict binding because they have only seen positive examples, so they are ‘trained to please.’
“PATH incorporates a second module that specifically discriminates between binders and non-binders.”
PATH can be used with proteins, small molecules, peptides and antibodies.
The Donald Lab now plans to apply it in developing cancer drugs and HIV antibodies — especially targeting kinases, which are enzymes that modify proteins, and transcription factors, which regulate gene expression.
The goal is to design drugs that bind precisely to their intended targets, improving effectiveness while reducing side effects.
