AI drug discovery models fail to understand physics in protein-ligand binding
Researchers at the University of Basel found that state-of-the-art AI models for drug discovery, including AlphaFold and RosettaFold, memorize patterns from training data rather than understanding the underlying physical chemistry of protein-ligand interactions. The models failed to recognize when binding sites were altered or blocked, and struggled particularly with novel proteins dissimilar to training data.
What happened
Researchers at the University of Basel tested state-of-the-art AI models including AlphaFold and RosettaFold to verify whether they understood the physics underlying protein-ligand interactions. The models had shown apparently high success rates in predicting how proteins interact with small molecules such as drugs.
The researchers modified hundreds of sample proteins so that ligand binding sites had completely different charge distributions or were blocked entirely. In more than half of the cases, the AI models predicted the same complex structure as if binding were still possible, despite the modifications that should have prevented binding.
The researchers tested a similar approach with modified ligands designed to be unable to dock to proteins, and the models again failed to recognize the problem. The researchers concluded that the models memorize patterns from training data rather than learning the underlying principles of physical chemistry.
Reported impact
The models struggled particularly with novel proteins that did not show similarity to training data, which are precisely the proteins of greatest interest for developing innovative drugs.
According to the researchers, only around 100,000 protein structures together with their ligands were available for training the AI models, relatively few compared to other AI training datasets, yet the models achieved suspiciously high success rates.
Index analysis
The models' reliance on pattern matching rather than physical understanding means they may generate confident predictions for novel protein-ligand interactions that lack scientific validity, potentially misleading drug development efforts.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.