Why AI Can’t Yet Crack the Code of RNA and DNA Folding

Artificial intelligence (AI) has quickly become a widely discussed topic, as its reach is expanding to numerous fields beyond computer science. In structural biology, AI has been increasingly used to predict 3D structures of various biomolecules including proteins, carbohydrates and even the interaction between these molecules.

However, a study, “Critical Assessment of RNA and DNA Structure Predictions via Artificial Intelligence: The Imitation Game,” recently published in the Journal of Chemical Information and Modeling, reveals that AI still struggles in some instances to replicate the same structures as determined using experimental data. Selected as an Editor’s Choice by the American Chemical Society, a designation reserved for work of exceptional interest and influence, the paper presents a rigorous evaluation by IBBR Fellows Christina Bergonzo and Alexander Grishaev from the National Institute of Standards and Technology (NIST).  They used AlphaFold, an AI system developed by Google DeepMind, to predict structures for specific oligonucleotide sequences and then compared the predicted structures to those determined experimentally using nuclear magnetic resonance (NMR) techniques.

RNA folding is a complicated process in that it depends on many factors including the presence of metal ions, environmental conditions and the fidelity of the genetic code. More specifically, the presence of metal ions alter the ionic conditions and thus can greatly influence folding. The researchers introduced metal ions and examined the response of the AI structure prediction; the AI failed to fully capture the effects of a changing biophysical environment, often losing or changing several structural features.

In addition, the researchers looked at sequences with structures that would change significantly when a mutation was introduced, and did the same comparison with AI generated structures. The researchers observed a mixed performance; the AI tended to resolve more-common structures with higher accuracy, but did not perform as well for less common aspects. These results underscore the fact that AI-based models fail to understand the underlying mechanisms behind folding, and instead, mimic training data sets.

In a related interview with Chemical and Engineering News, Bergonzo and Grishaev again emphasize the importance of validating predictions generated by AI against experimental evidence. They also highlight the need for future models to better account for or incorporate experimental data and the complexities inherent in biomolecular structures.