VP-Detector: A 3D multi-scale dense convolutional neural network for macromolecule localization and classification in cryo-electron tomograms

Abstract

Background and objective: Cryo-electron tomography (cryo-ET) with subtomogram averaging (STA) is indispensable when studying macromolecule structures and functions in their native environments. Due to the low signal-to-noise ratio, the missing wedge artifacts in tomographic reconstructions, and multiple macromolecules of varied shapes and sizes, macromolecule localization and classification remain challenging. To tackle this bottleneck problem for structural determination by STA, we design an accurate macromolecule localization and classification method named voxelwise particle detector (VP-Detector).

Methods: VP-Detector is a two-stage particle detection method based on a 3D multiscale dense convolutional neural network (3D MSDNet). The proposed network uses 3D hybrid dilated convolution (3D HDC) to avoid the resolution loss caused by scaling operations. Meanwhile, it uses 3D dense connectivity to encourage the reuse of feature maps to reduce trainable parameters. In addition, the weighted focal loss is proposed to focus more attention on difficult samples and rare classes, which relieves the class imbalance caused by multiple particles of various sizes. The performance of VP-Detector is evaluated on both simulated and real-world tomograms, and it shows that VP-Detector outperforms state-of-the-art methods.

Results: The experiments show that VP-Detector outperforms the state-of-the-art methods on particle localization with an F1-score of 0.951 and a precision of 0.978. In addition, VP-Detector can replace manual particle picking in experiment on the real-world tomograms. Furthermore, it performs well in classifying large-, medium-, and small-weight proteins with accuracies of 1, 0.95, and 0.82, respectively. Finally, ablation studies demonstrate the effectiveness of 3D HDC, 3D dense connectivity, weighted focal loss, and training on small training sets.

Conclusions: VP-Detector can achieve high accuracy in particle detection with few trainable parameters and support training on small datasets. It can also relieve the class imbalance caused by multiple particles with various shapes and sizes.

Introduction

Cryo-electron tomography (cryo-ET) is a promising imaging technique that allows three-dimensional (3D) visualization of protein or macromolecular complexes at molecular resolution in their native context. The tomographic reconstruction is generated from a tilt series imaged by rotating the biological sample. However, the resolution of tomographic reconstructions is limited by two factors: the low signal-to-noise ratio (SNR) due to the low electron dose and the missing wedge caused by the missing tilt angles. When numerous noisy copies of a complex of interest appear in a set of tomograms, subtomogram averaging (STA) can significantly increase their resolution by extracting, classifying, aligning, and averaging subtomograms in their native biological surroundings [17], [35], [36]. STA provides biological insights into the interaction and function of structures imaged under close-to-life conditions. Accurate subtomogram localization and classification, as the first two steps in the STA process, are critical for the subsequent subtomogram alignment and averaging to improve the final resolution of subtomograms. Nevertheless, extracting and distinguishing macromolecules from the complex and crowded native environment is challenging because of the low-resolution tomographic reconstruction and varied shapes and sizes of multiple macromolecules.

Traditional particle identification approaches such as template matching are employed to localize each type of macromolecule independently in tomographic reconstructions. However, due to the low SNR and missing wedge, template matching is only suitable for large complexes and ineffective for identifying small structures or macromolecular species with different states. Recently, the deep learning technique has gained revolutionary success in the fields of computer vision and image processing. In particular, convolutional neuron networks (CNNs) have been applied to object detection [24], classification [5], [10], [25], [28] and segmentation [8], [27], [31] in cryo-ET due to their natural ability to handle high noise data and achieve faster speed than template matching. However, the previously mentioned approaches for localizing and classifying subtomograms have the following deficiencies: (1) Downscaling and upscaling operations employed in CNNs can cause resolution loss, which makes particle detection in low SNR cryo-ET images difficult. (2) It is difficult to obtain adequate training data because collecting a large amount of high-quality cryo-ET images and labeling the ground truth are both time-consuming. However, these networks are typically designed with multiple trainable parameter which require large numbers of training data to prevent overfitting. (3) The number of distinct macromolecules varies significantly, resulting in the class imbalance problem that has been overlooked in previous studies.

To overcome the above limitations, we propose an efficient and automatic two-stage particle detection method called VP-Detector for localizing and classifying macromolecules in cryo-electron tomograms. VP-Detector uses a 3D multiscale dense convolutional neural network (3D MSDNet), taking advantage of the following strategies: (1) To avoid the loss of the resolution caused by the scaling operations in traditional CNNs, the 3D hybrid dilated convolution (3D HDC) module is proposed in our network, which can capture multiscale features without sacrificing resolution. (2) To train on small datasets without overfitting, 3D dense connectivity is proposed in our network, which encourages the reuse of feature maps to reduce the number of trainable parameters. (3) To address the class imbalance problem caused by the varying size of multiple macromolecules, the weighted focal loss is proposed in our network, which can reweight the loss of different classes based on prior knowledge and focus on difficult samples and rare classes.

In the experiments, VP-Detector is evaluated on two simulated datasets and one real-world dataset, and the results demonstrates that VP-Detector is accurate and effective. First, particle localization is assessed on precision, recall, miss rate, and F1-score metrics. The result shows that VP-Detector outperforms the state-of-the-art methods on the simulated tomograms with the highest precision and F1-score. Next, VP-Detector can replace manual particle picking approach on real-world tomograms. Then, the performance of particle classification is measured based on class accuracy and group accuracy. The results show that VP-Detector performs well in all groups of proteins. Finally, ablation studies are conducted to prove the effectiveness of 3D HDC, 3D dense connectivity, weighted focal loss, and training on small training sets.