CryoJAM: Automating Protein Homolog Fitting in Medium Resolution Cryo-EM Density Maps

Obtaining atomic structures of large protein complexes from medium-resolution cryogenic electron-microscopy (cryo-EM) density maps is a critical bottleneck in the cryo-EM workflow. CryoJAM aims to automate this process by using a 3D Convolutional Neural Network model within a U-Net architecture. This model is trained on a novel loss function that leverages Fourier-Shell Correlation (FSC), as a proxy for quality of fit, and Root Mean Squared Error (RMSE) to help optimize fits within real space. Capitalizing on the gold-standard status of FSC in cryo-EM, this method introduces an innovative implementation of FSC into cryo-EM model fitting software, enhancing the precision and efficiency of structural analysis. After 25 epochs, CryoJAM successfully reduced the RMSE in 21 out of 26 of the test cases, effectively fitting homologous protein structures into medium-resolution cryo-EM densities.