Targeted proteomics data interpretation with DeepMRM

Abstract

Targeted proteomics is widely utilized in clinical proteomics; however, researchers often devote substantial time to manual data interpretation, which hinders the transferability, reproducibility, and scalability of this approach. We introduce DeepMRM, a software package based on deep learning algorithms for object detection developed to minimize manual intervention in targeted proteomics data analysis. DeepMRM was evaluated on internal and public datasets, demonstrating superior accuracy compared with the community standard tool Skyline. To promote widespread adoption, we have incorporated a stand-alone graphical user interface for DeepMRM and integrated its algorithm into the Skyline software package as an external tool.

Keywords: Skyline; machine learning; multiple reaction monitoring; object detection; peak detection; quality control; quantification; selected reaction monitoring; targeted proteomics.

Graphical abstract

Figure 1

DeepMRM workflow for detecting peak groups of targeted peptides Given the target list and MRM/PRM/DIA data as input, transition chromatograms of both heavy and light peptides are fed into the model as 2-channel heatmap images. Multiscale 1D feature maps are extracted and processed by two subnetworks: a classifier for determining whether candidate peak groups are present and a regressor for detecting the boundary of candidate peak groups. Detected peak groups are then examined by a CNN-based transition classification model to select transitions unaffected by interference or noise. The abundances of target peptides are estimated based on peak areas of heavy and light peptides.

Figure 2

Comparison of DeepMRM and Skyline with or without mProphet Relative quantification and distribution of the abundance of heavy peptides for (A–D) noisy dataset and (E–H) complex background dataset. There is no difference between Skyline default and Skyline at FDR 5% in the complex background dataset because mProphet did not filter any results. Red dotted lines represent the theoretical values according to the heavy peptide abundance. In the boxplots, the centerline, edges, and whiskers represent the median, the first and third quartile, and 1.5× interquartile range, respectively. Outlier points outside of the whiskers are indicated by dot symbols. The table below the boxplot shows averaged correlation coefficients, arctangent absolute percentage error over 43 targeted peptides, and the number of peak groups in each quantile box (PCC, Pearson’s correlation; SPC, Spearman’s rank correlation; MAAPE, mean arctangent absolute percentage error). See also Figure S3.

Figure 3

DeepMRM performances on MRM, PRM, and DIA datasets (A and B) (A) Average precision (AP) and recall (RC) scores for MRM, PRM, and DIA datasets and (B) scatterplot comparing the results of light/heavy ratios calculated by peaks manually annotated and those by peaks detected by DeepMRM. The scale of axes is log2. The table below the scatterplot shows correlation coefficients (PCC, Pearson’s correlation; SPC, Spearman’s rank correlation) and mean arctangent absolute percentage error (MAAPE). See also Figure S2 and Table S2.