SVcnn: an accurate deep learning-based method for detecting structural variation based on long-read data

Abstract

Background: Structural variations (SVs) refer to variations in an organism's chromosome structure that exceed a length of 50 base pairs. They play a significant role in genetic diseases and evolutionary mechanisms. While long-read sequencing technology has led to the development of numerous SV caller methods, their performance results have been suboptimal. Researchers have observed that current SV callers often miss true SVs and generate many false SVs, especially in repetitive regions and areas with multi-allelic SVs. These errors are due to the messy alignments of long-read data, which are affected by their high error rate. Therefore, there is a need for a more accurate SV caller method.

Result: We propose a new method-SVcnn, a more accurate deep learning-based method for detecting SVs by using long-read sequencing data. We run SVcnn and other SV callers in three real datasets and find that SVcnn improves the F1-score by 2-8% compared with the second-best method when the read depth is greater than 5×. More importantly, SVcnn has better performance for detecting multi-allelic SVs.

Conclusions: SVcnn is an accurate deep learning-based method to detect SVs. The program is available at https://github.com/nwpuzhengyan/SVcnn .

Keywords: Deep learning; Long-read sequencing data; SV caller; Structural variations.

Fig. 1

The figure shows the recall, precision, and F1-score of different SV callers using three ONT datasets for HG002, CHM13 and HG00733. a The recall of DELs in three datasets. b The recall of INSs in three datasets. c The precision of DELs in three datasets. d The precision of INSs in three datasets. e The F1-score of DELs in three datasets. f The F1-score of INSs in three datasets. From the histogram, we can clearly see that SVcnn achieved the best results in F1-score

Fig. 2

The figure shows the f1-score of different SV callers for SVs with different lengths. The histogram indicates that SVcnn has the highest f1-score when SVs are less than 500 bp. However, it performs poorly when SVs exceed 10k bp

Fig. 3

The F1-scores of SVcnn and SVision for calling a DELs and b INSs in HG002, CHM13, and HG00733. The main reason for the poor performance of SVision is that SVision outputs too many SVs (about 50,000), far exceeding the expected number (about 22,000) of SVs

Fig. 4

The figure shows the F1-score of different SV callers using the GIAB benchmark (HG002_SVs_Tier1_v0.6). From the histogram, we can see that SVcnn still has the best F1-score

Fig. 5

The F1-scores of SVcnn and other SV callers for calling a DELs and b INSs in HG002 under different sequencing depths. From the histograms, we can see that SVcnn has the best performance when the sequencing depth is greater than 5×. When sequencing depth is 5×, SVcnn is still the second-best method

Fig. 6

The overview of SVcnn. There are three main steps in SVcnn. (1) Detecting candidate SV regions, (2) Converting regions to images and building model, (3) Filtrating and outputting SVs

Fig. 7

In this figure, we show the alignments of 7 different reads. The reads in the green and blue boxes will be retained as reliable reads. The reads in the red box will be filtered as unreliable reads