Rensselaer Polytechnic Institute (RPI) warns that today's leading AI tools for predicting protein structures can produce results that are scientifically impossible, highlighting the need for caution as AI becomes more deeply integrated into scientific research. Published in the Proceedings of the National Academy of Sciences (PNAS), the research found that widely used deep learning models sometimes generate protein structures that violate fundamental physical and chemical principles while simultaneously overestimating the accuracy of their own predictions.
The study, led by Professor George I. Makhatadze, evaluated popular protein-folding systems including AlphaFold2, RoseTTAFold2, OmegaFold, and ESMFold. While these AI models have significantly accelerated biological research, the analysis revealed that some struggle with proteins containing ionizable amino acid residues and can prioritize statistical patterns over the thermodynamic rules that govern real protein folding. According to Makhatadze, researchers should "trust but verify" AI-generated predictions rather than treating them as definitive scientific results.
To improve reliability, the researchers recommend combining AI-generated protein structures with physics-based molecular dynamics simulations, which can validate whether predicted structures are physically plausible. This hybrid approach leverages AI's speed while ensuring that predictions remain grounded in established scientific principles. The findings also encourage AI developers to incorporate more physicochemical knowledge into future protein prediction models to reduce these blind spots.
The study underscores a broader lesson for AI in scientific research: despite remarkable advances, human expertise and scientific validation remain essential. As AI becomes a standard tool in laboratories, researchers must continue to verify AI outputs through established experimental and computational methods rather than relying solely on machine-generated predictions. The authors conclude that AI should complement scientific judgment—not replace it—ensuring that breakthroughs are both efficient and scientifically sound.
