DeepIso: A Deep Learning Model for Peptide Feature Detection

TL;DR

DeepIso introduces a deep learning model using CNNs for peptide feature detection in LC-MS/MS proteomics data, outperforming existing tools by learning complex data representations and adapting to new data.

Contribution

We propose a novel CNN-based model, DeepIso, for peptide feature detection in LC-MS/MS data, capable of learning from high-dimensional data and evolving with new proteomic datasets.

Findings

Achieved 93.21% sensitivity and 99.44% specificity on antibody dataset

Demonstrated deep learning's potential to improve peptide detection accuracy

Outperformed traditional feature detection tools in proteomics workflows

Abstract

Liquid chromatography with tandem mass spectrometry (LC-MS/MS) based proteomics is a well-established research field with major applications such as identification of disease biomarkers, drug discovery, drug design and development. In proteomics, protein identification and quantification is a fundamental task, which is done by first enzymatically digesting it into peptides, and then analyzing peptides by LC-MS/MS instruments. The peptide feature detection and quantification from an LC-MS map is the first step in typical analysis workflows. In this paper we propose a novel deep learning based model, DeepIso, that uses Convolutional Neural Networks (CNNs) to scan an LC-MS map to detect peptide features and estimate their abundance. Existing tools are often designed with limited engineered features based on domain knowledge, and depend on pretrained parameters which are hardly updated despite huge amount of new coming proteomic data. Our proposed model, on the other hand, is capable of learning multiple levels of representation of high dimensional data through its many layers of neurons and continuously evolving with newly acquired data. To evaluate our proposed model, we use an antibody dataset including a heavy and a light chain, each digested by Asp-N, Chymotrypsin, Trypsin, thus giving six LC-MS maps for the experiment. Our model achieves 93.21% sensitivity with specificity of 99.44% on this dataset. Our results demonstrate that novel deep learning tools are desirable to advance the state-of-the-art in protein identification and quantification.