DeepTracer Diffusion: A Single‐Stage Diffusion Model for Accurate Cryo-EM Backbone Segmentation

Nathaniel Jewel, Haydn Tamura, Haowen Guan, Dong Si

Submission Track: Short papers presenting ongoing research or work submitted to other venues (up to 5 pages, excluding references)

Keywords: Cryo-EM, Denoising Diffusion Probabilistic Model (DDPM), Image segmentation, Deep learning for structural biology, Protein structure prediction, Single-stage diffusion

TL;DR: A single-stage DDPM directly segments cryo-EM voxels to produce improved backbone atom predictions and higher F1 scores.

Abstract: Precise atomic-level interpretation of macromolecular structures is vital for understanding biological mechanisms yet remains challenging due to the complex nature of cryo-electron microscopy (cryo-EM) data. Existing approaches have utilized either multiple convolutional neural networks or complex combinations of autoencoder and latent diffusion models to predict atom locations via image segmentation. We introduce DeepTracer Diffusion, a novel framework that leverages a single Denoising Diffusion Probabilistic Model (DDPM) to perform image segmentation, providing higher accuracy in terms of F1 score and predicted residues for predicted backbone atoms.

Submission Number: 42