Protein-Folding AI Tools Generate Impossible Results, RPI Study Warns

Artificial intelligence tools designed to predict protein structures are producing results that violate fundamental laws of physics and chemistry, according to new research from Rensselaer Polytechnic Institute.

The study, published in the Proceedings of the National Academy of Sciences, examined widely deployed deep learning platforms and found they consistently generate structures that are scientifically implausible—while simultaneously overestimating their own accuracy.

Critical gaps in leading platforms

George I. Makhatadze, professor of biological sciences at RPI and the study's author, evaluated tools that predict how amino acid sequences fold into three-dimensional protein structures. His analysis revealed systematic failures across the field's most prominent platforms.

AlphaFold2, developed by Google's DeepMind laboratory and recognized with a share of the 2024 Nobel Prize in Chemistry, produced implausible structures for variant sequences by prioritizing statistical patterns over thermodynamic principles. RoseTTAFold2, a similar platform from the University of Washington, exhibited comparable problems. Both systems are trained on evolutionary data and structural databases.

"AlphaFold is considered the gospel of the field," Makhatadze noted, according to details first reported by RPI. "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."

Transformer-based protein language models—including OmegaFold and Meta's ESMFold, which rely on sequence data rather than structural information—showed fewer scientific impossibilities. However, no category of model performed adequately when proteins contained ionizable residues, amino acid side chains that can gain or lose protons based on their environment. Current AI tools lack training to account for these residues.

Why it matters

Protein structure prediction underpins drug discovery, enzyme engineering, and fundamental biological research. As AI becomes standard infrastructure in scientific labs, physically impossible predictions that go undetected can waste resources, delay discoveries, and undermine confidence in computational methods. The finding that every tested tool overestimated its own accuracy compounds the risk, since researchers may trust confidence scores that don't reflect actual reliability.

Path forward requires hybrid approach

Makhatadze recommends combining AI predictions with molecular dynamics simulations to validate structures before use. This physics-based refinement, the paper argues, strengthens confidence in machine learning outputs and addresses current blind spots.

The researcher expects platform developers to incorporate additional physicochemical validation layers in response to these findings. "It has to be a combination of pattern recognition and physics," he said.

His immediate guidance for scientists using AI tools is direct: "You cannot blindly believe everything the model predicts." The study's core message, Makhatadze explained, is "trust but verify"—AI outputs must be validated using physics-based methods before application in laboratory settings.

The research was first reported by Rensselaer Polytechnic Institute.