PickerOptimizer: A Deep Learning-Based Particle Optimizer for Cryo-Electron Microscopy Particle-Picking Algorithms

Abstract

Cryo-electron microscopy single particle analysis requires tens of thousands of particle projections for the structural determination of macromolecules. To free researchers from laborious particle picking work, a number of fully automatic and semi-automatic particle picking approaches have been proposed. However, due to the presence of carbon and different types of high-contrast contaminations, these approaches tend to select a non-negligible number of false-positive particles, which affects the subsequent 3D reconstruction.

In order to overcome this limitation, we present a deep learning-based particle pruning algorithm, PickerOptimizer, to separate erroneously picked particles from the correct ones. PickerOptimizer trained a convolutional neural network based on transfer learning techniques, where the pre-trained model maintains strong generalization ability and can be quickly adapted to the characteristics of the new dataset. Here, we build the first cryo-EM dataset for image classification pre-training which contains particles, carbon regions and high-contrast contaminations from 14 different EMPIAR entries. The PickerOptimizer works by fine-tuning the pre-trained model with only a few manually labeled samples from new datasets. The experiments carried out on several public datasets show that PickerOptimizer is a very efficient approach for particle post-processing, achieving F1 scores above 90%. Moreover, the method is able to identify false-positive particles more accurately than other pruning strategies. A case study further shows that PickerOptimizer can improve conventional particle pickers and complement deep-learning-based ones. The Source code, pre-trained models and datasets are available at https://github.com/LiHongjia-ict/PickerOptimizer/.