The potential of AI for experiments in structural biology

Andrea Thorn

/Hamburg [DE]

Structural biology is key to understanding basic processes of life and a major driver for the development of new therapies. However, these structures do not directly result from the experiment, but are merely models which explain the observed data according to a priori knowledge. Consequently, our structures are only as good as our limited understanding of the underlying principles.

However, in crystallography, the gap between model and reality remains clearly evident in hard to interpret maps and large R values. In cryo-EM, model properties do not align well with the actual specimen. Machine learning methods, in particular convolutional neural networks, have been applied to a variety of problems in the electron cryo microscopic and crystallographic structure solution of biological macromolecules. However, until recently, their acceptance by the community was limited to tasks where they replaced repetitive work and visual checks were easy, such as particle picking, crystal centering or crystal recognition. With Artificial Intelligence (AI) based protein fold prediction now revolutionizing the field, it is clear that their scope could be much wider, including structure determination, prediction and validation. Experimental biomolecular structure determination could potentially profit immensely from AI, which may even pave the way to joint analyses of data from different labs and methods, which would significantly advance our understanding of the molecules that govern key biological processes.

Whether we will be able to exploit this potential fully will depend on the manner in which we use machine learning: training data must be well-formulated, methods need to utilize appropriate architectures, and outputs must be critically assessed, which may even require explaining AI decisions.

Go back