July 24, 2026
Researchers at Rensselaer Polytechnic Institute (RPI) have found that today's leading artificial intelligence tools for predicting protein structures routinely generate results that are physically and chemically impossible, exposing critical blind spots in how AI is being applied across scientific research. The work, published in the Proceedings of the National Academy of Sciences (PNAS), serves as a cautionary reminder that AI still requires human oversight and physics-based verification to produce reliable results in the lab.
The paper, authored by George I. Makhatadze, professor of biological sciences and Constellation Endowed Chair at RPI, evaluated widely used deep learning tools for predicting how flat sequences of amino acids fold into the three-dimensional structures that determine a protein's function. Makhatadze found that these tools frequently overlook the underlying scientific rules of protein folding – and, notably, that every tool tested rated its own accuracy higher than the results warranted.
“The major conclusion of the paper essentially is: trust but verify,” Makhatadze explained. “You have to verify [AI outputs] using physics-based methods.”
AI has become indispensable for analyzing the massive data sets used to predict protein folds. For example, Google’s DeepMind AI laboratory – called AlphaFold2 – shared the 2024 Nobel Prize in Chemistry for its contributions to protein structure prediction.
But according to Makhatadze's work, AlphaFold2 and RoseTTAFold2 — a similar deep learning-based prediction platform developed at the University of Washington — both produced “implausible structures for variant sequences” by “[prioritizing] statistical patterns over the underlying thermodynamic principles of folding.” Both tools are trained on evolutionary data and structural databases.
“AlphaFold is considered the gospel of the field,” Makhatadze said. “It is very good, and it does many things well. But occasionally it makes mistakes, because there simply isn’t enough of the right kind of data in the model yet.”
Makhatadze found fewer scientific impossibilities in a different class of tools – “transformer-based protein language models” that rely on protein sequences rather than structural data. Tools in this class included OmegaFold and the Meta-developed ESMFold. However, neither category of model performed well when proteins contained ionizable residues, meaning amino acid side chains that can gain or lose a proton depending on their surrounding environment.
According to Makhatadze, current predictive AI tools simply haven’t been trained to account for ionizable residues, sometimes leading to scientifically impossible predictions.
To address these shortcomings, Makhatadze recommends pairing AI-generated predictions with computer simulations of molecular dynamics to validate protein structures. That combination, the paper notes, “strengthens confidence in AI-generated structures and underscores the need to couple machine learning with physics-based refinement.”
With the paper’s publication, Makhatadze expects the developers of AI predictive folding models to address this blind spot in existing platforms by incorporating additional levels of physicochemical validation. “It has to be a combination of pattern recognition and physics,” he said.
In the meantime, Makhatadze offers one piece of advice for AI users in scientific labs and beyond: “You cannot blindly believe everything the model predicts.”