DiffFit: Visually-Guided Differentiable Fitting of Molecule Structures to Cryo-EM Map

We introduce DiffFit, a differentiable algorithm for fitting protein atomistic structures into experimental reconstructed Cryo-Electron Microscopy (cryo-EM) volume map. This process is essential in structural biology to semi-automatically reconstruct large meso-scale models of complex protein assemblies and complete cellular structures that are based on measured cryo-EM data. Current approaches require manual fitting in 3D that already results in approximately aligned structures followed by an automated fine-tuning of the alignment. With our DiffFit approach, we enable domain scientists to automatically fit new structures and visualize the fitting results for inspection and interactive revision. Our fitting begins with differentiable 3D rigid transformations of the protein atom coordinates, followed by sampling the density values at its atom coordinates from the target cryo-EM volume. To ensure a meaningful correlation between the sampled densities and the protein structure, we propose a novel loss function based on a multi-resolution volume-array approach and the exploitation of the negative space. Such loss function serves as a critical metric for assessing the fitting quality, ensuring both fitting accuracy and improved visualization of the results. We assessed the placement quality of DiffFit with several large, realistic datasets and found its quality to be superior to that of previous methods. We further evaluated our method in two use cases. First, we demonstrate its use in the process of automating the integration of known composite structures into larger protein complexes. Second, we show that it facilitates the fitting of predicted protein domains into volume densities to aid researchers in the identification of unknown proteins. We open-sourced (github.com/nanovis/DiffFitViewer) DiffFit as a plugin in ChimeraX. All supplemental materials are available at osf.io/5tx4q.