Structures via Reasoning: Applying Artificial Intelligence to cryo-EM to Reveal Structural Variability

PI: James Evans
Cryo-electron microscopy (EM) is a powerful technology in structural biology capable of characterizing small biomolecules to large complexes with many interacting proteins with atomic to sub-nanometer resolution. Small internal variances in structure are critical to understanding the function of proteins. However, the latest cryo-EM reconstruction and modeling approaches sometimes fail in resolving subclasses within populations of 2D images, which differ only by small variances in contrast. Researchers at PNNL plan to overcome this limitation using deep learning, thereby helping with the identification of subtle structural differences within the protein density. Comparing between expected atomic structure and the 3D volume refined by an AI method will reveal whether local differences in structure are due to intrinsic molecular differences or just accumulated noise.
Towards this goal, researchers plan to implement a new cryo-EM reconstruction approach, building upon the success of current software packages that perform conventional alignments and classification. This output is leveraged as input to the ML method that will fractionate the population into the highest resolution 3D classes. The novelty of this method stems from using a custom supervised reasoning pipeline that evaluates multiple pixel locations within the protein density to decipher whether single pixel variances are correlated to changes in heterocomplex stoichiometry, subunit position, and ligand binding state.
PNNL’s approach will not only advance the cryo-EM field, but the ability to detect subtle inner structural differences of biological molecules will allow differentiation of virus strains, a much-needed capability today.