# GTpick: A deep neural network for Cryo-EM particle detection

## TL;DR

GTpick is a new deep learning model for detecting particles in Cryo-EM images, improving accuracy and 3D reconstruction quality.

## Contribution

GTpick introduces cross-attention and Focal Loss to enhance particle detection in Cryo-EM images.

## Key findings

- GTpick outperforms existing methods in Recall and F1 scores for Cryo-EM particle detection.
- GTpick improves the resolution of 3D density maps reconstructed from detected particles.
- GTpick shows limitations in performance on very large datasets.

## Abstract

Accurate identification of protein particles in cryo-electron microscopy (Cryo-EM) images is crucial for achieving high-resolution three-dimensional (3D) structural reconstruction. However, this task faces multiple challenges, including low signal-to-noise ratios, densely distributed particles, and class imbalance. To address these issues, this study proposes a target detection algorithm named GTpick, built upon the DETR framework. GTpick introduces a cross-attention mechanism to enhance the interaction between target queries and specific image features. In addition, a grouped one-to-many label assignment strategy is employed to improve recall in densely populated regions, and a Focal Loss function is incorporated to mitigate the adverse effects of background noise and class imbalance on detection accuracy. Experiments on large-scale Cryo-EM datasets demonstrate that GTpick outperforms existing machine learning-based particle-picking methods in terms of the resolution of 3D density maps reconstructed from detected particles and achieves superior Recall and F1 scores, particularly excelling in the Recall metric.

•GTpick is a Cryo-EM particle detection model built upon the DETR framework.•GTpick achieves higher recall and true positive counts, demonstrating strong capability in Cryo-EM particle detection.•Detection results from GTpick lead to improved 3D reconstruction quality when processed with CryoSPARC.•GTpick still shows room for improvement in performance when applied to very large datasets.

GTpick is a Cryo-EM particle detection model built upon the DETR framework.

GTpick achieves higher recall and true positive counts, demonstrating strong capability in Cryo-EM particle detection.

Detection results from GTpick lead to improved 3D reconstruction quality when processed with CryoSPARC.

GTpick still shows room for improvement in performance when applied to very large datasets.