UNSUPERVISED MEMBRANE SUBTRACTION IN CRYOGENIC ELECTRON MICROSCOPY IMAGES

ICLR 2026 Conference Submission13659 Authors

Keywords: cryo-EM, membrane, proteins, image processing, unsupervised, membrane representation

TL;DR: We present a novel method for unsupervised membrane subtraction from cryogenic-electron microscopy images of membrane proteins.

Abstract: Cryogenic electron microscopy (cryo-EM) of membrane proteins often requires extracting them from their membrane to simplify downstream image processing. While this step reduces the influence of membranes on 3D reconstruction, it also prevents proteins from being observed in their natural state. To overcome this limitation, we propose a two-step machine learning framework that avoids protein extraction: (1) membrane detection, which identifies the bilayer membrane, and (2) membrane subtraction, which digitally removes the detected membrane from the cryo-EM micrograph. Recent work has introduced supervised algorithms for membrane detection, but membrane subtraction remains relatively underexplored. Here, we present a novel unsupervised approach to membrane subtraction that models membranes using a general representation and computes a smooth estimate, which can then be subtracted from the original cryo-EM micrograph. Experimental results show that our method outperforms existing membrane subtraction alternatives and enables reliable 3D reconstruction of membrane proteins using cryo-EM without protein extraction.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 13659