Deep Semi-Supervised Learning Improves Universal Peptide Identification of Shotgun Proteomics Data

Abstract

In proteomic analysis pipelines, machine learning post-processors play a critical role in improving the accuracy of shotgun proteomics analysis. Most often performed in a semi-supervised manner, such post-processors accept the peptide-spectrum matches (PSMs) and corresponding feature vectors resulting from a database search, train a machine learning classifier, and recalibrate PSM scores based on the resulting trained parameters, often leading to significantly more identified peptides across q-value thresholds. However, current state-of-the-art post-processors rely on shallow machine learning methods, such as SVMs, gradient boosted decision trees, and linear discriminant analysis. In contrast, the powerful learning capabilities of deep models have displayed superior performance to shallow models in an ever-growing number of other fields. In this work, we show that deep neural networks (DNNs) significantly improve the recalibration of shotgun proteomics data compared to the most accurate and widely used post-processors, such as Percolator and PeptideProphet. Furthermore, we show that DNNs are able to adaptively analyze complex datasets and features for more accurate universal post-processing, leading to both improved Prosit analysis and markedly better recalibration of recently developed p-value scoring functions.

Deep Convolutional Neural Networks Help Scoring Tandem Mass Spectrometry Data in Database-Searching Approaches

“…Recently, several deep learning methods have been proposed, such as pDeep, Prosit, DeepMass, and ProteoTorchDNN, to improve the number of annotations. Prosit extracts nearly 60 features for every spectrum annotation obtained with the Andromeda/MaxQuant system, including the application of the intensities of the theoretical spectra predicted by deep LSTM networks.…”

“…Recently, several deep learning methods have been proposed, such as pDeep, 46 Prosit, 23 DeepMass, 47 and Proteo-TorchDNN, 25 to improve the number of annotations. Prosit extracts nearly 60 features for every spectrum annotation obtained with the Andromeda/MaxQuant system, 7 We have performed a grid search over the convolutional kernel number and its width to find the best architecture for Slider using the HumVar data set.…”

“…Several new score functions and database searching tools have also been introduced, including Mascot, HyperScore of X!Tandem, Morpheus, and MS Amanda; however, these traditional methods are based on manually constructed score functions and have resulted in only minor improvements as compared with SEQUEST . Recent studies have focused on (1) score calibration methods to provide a well-defined accurate semantics so that the spectrum annotations can be compared with each other ,,− or (2) postprocessing methods, such as Percolator, Prosit, PeptideProphet, ProteoTorchDNN, and DeepRescore to re-rank database search results to improve the number of spectrum annotations at various false discovery rates (FDRs).…”

Semi-supervised machine learning for sensitive open modification spectral library searching

“…For example, PSM rescoring in ANN-SoLo can trivially switch between the standard linear SVM, as used by Percolator, 6 or a random forest (Supplementary figure S5), which can learn a non-linear decision function and which has previously been shown to have beneficial performance during PSM rescoring. 8,33 Additionally, other machine learning models, such as neural networks, 9 could effortlessly be integrated as well if they support the popular Scikit-Learn application programming interface (API). 55 Furthermore, being able to directly manipulate the PSM rescoring functionality allowed us to get deeper insights into the workings of the machine learning model, for example, to perform a detailed analysis of the feature importances.…”

“…This makes it possible to improve the spectrum identification results by accepting more target PSMs at a specific FDR threshold. Since the original description of Percolator, 6 several approaches that extend upon this idea have been proposed, including tools that use different machine learning models instead of a linear SVM, [7][8][9] integration with various search engines, [10][11][12] and additional features that the classifier can use, often powered by deep learning. [13][14][15][16][17][18][19][20] Another recent development is the emergence of novel search engines that can efficiently perform so-called open modification searching (OMS).…”