Pratyush Tiwary

Structure prediction tools using Generative Artificial Intelligence (AI) have significantly advanced, offering rapid predictions of the most stable structure for generic proteins/RNA and even generating ensembles with dynamics. This might suggest that molecular dynamics (MD) and statistical mechanics are now maybe obsolete. However, I will demonstrate why these methods remain critical for ensuring AI approaches are both predictive and reliable for chemistry and biology. I’ll discuss how current AI predictions can sometimes result from memorization or hallucination and show how integrating Generative AI with enhanced MD and statistical mechanics provides a more predictive, though slower, alternative to using AI alone. Examples will include kinases, crystal polymorphs and RNA, revealing thermodynamic and dynamic properties such as drug residence times, conformational populations, mutation effects, and melting curves derived from chemical identity and force fields. Lastly, I’ll illustrate how these integrated methods enable predictions of biomolecular properties under thermodynamic conditions far from training data, including temperature, pressure, and chemical potential.

Sponsorship of an event does not constitute institutional endorsement of external speakers or views presented.