Meditations

He also described how AlphaFold’s ability to predict 3D protein structures may enable drug developments, and briefly shared an ongoing work called Isomorphic which focuses on the ‘chemistry space’ . By this, he appeared to mean that this model tries to use chemical dynamics to predict suitable binding spots on protein surfaces for medicinal compounds. This project seems very interesting as it appears to be potentially enhancing our understanding of various drug-disease relationships. 

I was hoping that he could say something more about whether AlphaFold’s predictions of the 200 million protein structures shed some light to the physical/biological principles by which protein-folding occurs in nature. Can one draw insights into biophysical mechanisms of protein-folding by examining aspects of the trained AlphaFold architecture? A quick reading of the related Wiki-pages and the Nature paper did not seem to reveal anything in this aspect though. The nice interview ended with Demis commenting that he would love to see A.I. being used to elucidate the ‘fundamental nature of reality, like performing Planck-scale experiments’. In physics, our current understanding based on experimentally proven theories will break down at Planck scale. This is fundamentally because of a mathematical inconsistency that arises when one simply treats the gravitational field as a quantum field in the same way that we understand the three other fundamental forces. We are limited experimentally as there is a limit to the energy of the particle colliders that we can make on Earth. So unlike the case of protein structures, it’s not quite about the complexity and sheer magnitude of data that allows A.I. to be a revolutionary tool. But nonetheless, it is still interesting to see whether and how our future developments in A.I. may help particle physicists harness data, especially from astrophysical observations, to figure out answers to these deep questions!