#32 · Structural bio & protein design

Deep learning automates particle picking, fusion, and event detection in microscopy

The dominant thread is learned models that convert raw experimental detections into discrete, interpretable entities, whether fused super-resolution particles, picked cryo-EM particles, or exocytic events. DeepSRFusion applies a self-supervised point-cloud framework to 3D particle fusion in single-molecule localization microscopy, reporting ~1.6 nm resolution. ExoFILT uses a transfer-learning deep classifier to identify exocytic events in single-particle tracking data, cutting manual annotation ten-fold. CryoSIP couples a multi-frequency adaptive U-Net with SAM for cryo-EM particle picking, aiming to reduce false positives. What is new is the shift from manual annotation toward learned pipelines that operate directly on localization or particle data. Only DeepSRFusion and ExoFILT report quantitative gains in the item text, while the CryoSIP summary gives no comparable figure.