Leading artificial intelligence tools used for predicting protein structures routinely generate results that are physically and chemically impossible, according to research published in the Proceedings of the National Academy of Sciences. The findings, led by Rensselaer Polytechnic Institute professor George I. Makhatadze, reveal critical blind spots in how deep learning models handle thermodynamic principles and ionizable residues in biological research.
Evaluating Deep Learning Models and Thermodynamic Rules
Protein structure prediction tools convert flat sequences of amino acids into the three-dimensional shapes that dictate a protein’s function. According to Makhatadze’s evaluation of widely used deep learning platforms, these tools frequently overlook underlying scientific rules of protein folding. Furthermore, every single model tested rated its own accuracy higher than the actual results warranted.
“The major conclusion of the paper essentially is: trust but verify,” Makhatadze explained, emphasizing that researchers must validate AI outputs using physics-based methods.
Did you know? Google’s DeepMind AI laboratory developed AlphaFold2, a pioneering platform that shared the 2024 Nobel Prize in Chemistry for its contributions to protein structure prediction. Despite this accolade, current models still face structural limitations when analyzing complex variant sequences.
Comparing AlphaFold2, RoseTTAFold2, and Transformer-Based Models
Widely adopted platforms like AlphaFold2 and RoseTTAFold2 rely heavily on evolutionary data and structural databases. According to the published research, these tools produce implausible structures for variant sequences by prioritizing statistical patterns over the underlying thermodynamic principles of folding.
“AlphaFold is considered the gospel of the field,” Makhatadze noted. “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.”
In contrast, transformer-based protein language models—such as OmegaFold and Meta’s ESMFold—rely directly on protein sequences rather than pre-existing structural databases. Makhatadze found fewer scientific impossibilities within this alternative class of tools. However, neither category performed well when handling proteins containing ionizable residues.
The Challenge of Ionizable Residues in Scientific Labs
Proteins featuring amino acid side chains that gain or lose protons based on their surrounding environment present a major hurdle for current algorithms. According to Makhatadze, predictive AI tools have not been adequately trained to account for these ionizable residues, leading directly to scientifically impossible predictions in laboratory settings.
To overcome these systemic shortcomings, the research paper recommends pairing AI-generated predictions with computer simulations of molecular dynamics. This dual approach strengthens confidence in generated structures and highlights the necessity of coupling machine learning with physics-based refinement.
Pro Tip: Computational biologists and laboratory researchers should never blindly trust algorithmic outputs. Pairing machine learning tools with physics-based molecular dynamics simulations is essential for confirming structural accuracy.
Frequently Asked Questions
Why do AI protein structure prediction tools generate impossible results?
Makhatadze at Rensselaer Polytechnic Institute, deep learning models often prioritize statistical patterns derived from evolutionary data and structural databases over the fundamental thermodynamic principles of protein folding.
How can scientists verify AI-generated protein structures?
Researchers recommend combining machine learning predictions with physics-based methods, such as computer simulations of molecular dynamics, to validate structural accuracy and account for complex elements like ionizable residues.
What is the difference between AlphaFold2 and transformer-based models like ESMFold?
AlphaFold2 and RoseTTAFold2 rely on structural databases and evolutionary data, whereas transformer-based models like OmegaFold and Meta’s ESMFold utilize raw protein sequences. While transformer models showed fewer thermodynamic impossibilities in testing, both categories struggle with ionizable residues.
Future Outlook for Computational Biology
With the publication of this study in the Proceedings of the National Academy of Sciences, developers of AI folding models face mounting pressure to incorporate physicochemical validation directly into their platforms. Makhatadze stresses that future iterations must rely on a balanced combination of pattern recognition and physical laws.
Until these updates take effect, the core takeaway for laboratory scientists remains straightforward: model outputs require rigorous, independent verification before they guide downstream experiments.
What is your experience with AI tools in biological research? Share your thoughts in the comments below, explore our related articles on computational biology, or subscribe to our newsletter for the latest peer-reviewed scientific updates.
Related reading