This review explores state of the art machine learning and deep learning models for peptide property prediction in mass spectrometry-based proteomics, including, but not limited to, models for predicting digestibility, retention time, charge state distribution, collisional cross section, fragmentation ion intensities, and detectability. The combination of these models enables not only the in silico generation of spectral libraries but also finds many additional use cases in the design of targeted assays or data-driven rescoring. This review serves as both an introduction for newcomers and an update for experienced researchers aiming to develop accessible and reproducible models for peptide property predictions. Key limitations of the current models, including difficulties in handling diverse post-translational modifications and instrument variability, highlight the need for large-scale, harmonized datasets, and standardized evaluation metrics for benchmarking.

Citrullination is a critical yet understudied post-translational modification (PTM) implicated in various biological processes. Exploring its role in health and disease requires a comprehensive understanding of the prevalence of this PTM at a proteome-wide scale. Although mass spectrometry has enabled the identification of citrullination sites in complex biological samples, it faces significant challenges, including limited enrichment tools and a high rate of false positives due to the identical mass with deamidation (+0.9840 Da) and errors in monoisotopic ion selection. These issues often necessitate manual spectrum inspection, reducing throughput in large-scale studies. In this work, we present a novel data analysis pipeline that incorporates the deep learning model Prosit-Cit into the MS database search workflow to improve both the sensitivity and the precision of citrullination site identification. Prosit-Cit, an extension of the existing Prosit model, has been trained on 53,000 spectra from 2500 synthetic citrullinated peptides and provides precise predictions for chromatographic retention time and fragment ion intensities of both citrullinated and deamidated peptides. This enhances the accuracy of identification and reduces false positives. Our pipeline demonstrated high precision on the evaluation dataset, recovering the majority of known citrullination sites in human tissue proteomes and improving sensitivity by identifying up to 14 times more citrullinated sites. Sequence motif analysis revealed consistency with previously reported findings, validating the reliability of our approach. Furthermore, extending the pipeline to a tissue proteome dataset of the model plant Arabidopsis thaliana enabled the identification of 200 citrullination sites across 169 proteins from 30 tissues, representing the first large-scale citrullination mapping in plants. This pipeline can be seamlessly applied to existing proteomics datasets, offering a robust tool for advancing biological discoveries and deepening our understanding of protein citrullination across species.

Recent developments in machine-learning (ML) and deep-learning (DL) have immense potential for applications in proteomics, such as generating spectral libraries, improving peptide identification, and optimizing targeted acquisition modes. Although new ML/DL models for various applications and peptide properties are frequently published, the rate at which these models are adopted by the community is slow, which is mostly due to technical challenges. We believe that, for the community to make better use of state-of-the-art models, more attention should be spent on making models easy to use and accessible by the community. To facilitate this, we developed Koina, an open-source containerized, decentralized and online-accessible high-performance prediction service that enables ML/DL model usage in any pipeline. Using the widely used FragPipe computational platform as example, we show how Koina can be easily integrated with existing proteomics software tools and how these integrations improve data analysis.

Proteomics, the study of proteins within biological systems, has seen remarkable advancements in recent years, with protein isoform detection emerging as one of the next major frontiers. One of the primary challenges is achieving the necessary peptide and protein coverage to confidently differentiate isoforms as a result of the protein inference problem and protein false discovery rate estimation challenge in large data. In this chapter, we describe the application of artificial intelligence-assisted peptide property prediction for database search engine rescoring by Oktoberfest, an approach that has proven effective, particularly for complex samples and extensive search spaces, which can greatly increase peptide coverage. Further, it illustrates a method for increasing isoform coverage by the PickedGroupFDR approach that is designed to excel when applied on large data. Real-world examples are provided to illustrate the utility of the tools in the context of rescoring, protein grouping, and false discovery rate estimation. By implementing these cutting-edge techniques, researchers can achieve a substantial increase in both peptide and isoform coverage, thus unlocking the potential of protein isoform detection in their studies and shedding light on their roles and functions in biological processes.

Immunopeptidomics is crucial for immunotherapy and vaccine development. Because the generation of immunopeptides from their parent proteins does not adhere to clear-cut rules, rather than being able to use known digestion patterns, every possible protein subsequence within human leukocyte antigen (HLA) class-specific length restrictions needs to be considered during sequence database searching. This leads to an inflation of the search space and results in lower spectrum annotation rates. Peptide-spectrum match (PSM) rescoring is a powerful enhancement of standard searching that boosts the spectrum annotation performance. We analyze 302,105 unique synthesized non-tryptic peptides from the ProteomeTools project on a timsTOF-Pro to generate a ground-truth dataset containing 93,227 MS/MS spectra of 74,847 unique peptides, that is used to fine-tune the deep learning-based fragment ion intensity prediction model Prosit. We demonstrate up to 3-fold improvement in the identification of immunopeptides, as well as increased detection of immunopeptides from low input samples. Immunopeptidomics is crucial for the discovery of potential immunotherapy and vaccine candidates. Here, the authors generate a ground truth timsTOF dataset to fine-tune the deep learning model Prosit, improving peptide-spectrum match rescoring by up to 3-fold during immunopeptide identification.

Liquid chromatography-coupled mass spectrometry (LC-MS/MS) is the primary method to obtain direct evidence for the presentation of disease- or patient-specific human leukocyte antigen (HLA). However, compared to the analysis of tryptic peptides in proteomics, the analysis of HLA peptides still poses computational and statistical challenges. Recently, fragment ion intensity-based matching scores assessing the similarity between predicted and observed spectra were shown to substantially increase the number of confidently identified peptides, particularly in use cases where non-tryptic peptides are analyzed. In this chapter, we describe in detail three procedures on how to benefit from state-of-the-art deep learning models to analyze and validate single spectra, single measurements, and multiple measurements in mass spectrometry-based immunopeptidomics. For this, we explain how to use the Universal Spectrum Explorer (USE), online Oktoberfest, and offline Oktoberfest. For intensity-based scoring, Oktoberfest uses fragment ion intensity and retention time predictions from the deep learning framework Prosit, a deep neural network trained on a very large number of synthetic peptides and tandem mass spectra generated within the ProteomeTools project. The examples shown highlight how deep learning-assisted analysis can increase the number of identified HLA peptides, facilitate the discovery of confidently identified neo-epitopes, or provide assistance in the assessment of the presence of cryptic peptides, such as spliced peptides.

Sample multiplexed quantitative proteomics assays haveproved tobe a highly versatile means to assay molecular phenotypes. Yet, stochasticprecursor selection and precursor coisolation can dramatically reducethe efficiency of data acquisition and quantitative accuracy. To addressthis, intelligent data acquisition (IDA) strategies have recentlybeen developed to improve instrument efficiency and quantitative accuracyfor both discovery and targeted methods. Toward this end, we soughtto develop and implement a new real-time spectral library searching(RTLS) workflow that could enable intelligent scan triggering andpeak selection within milliseconds of scan acquisition. To ensureease of use and general applicability, we built an application toread in diverse spectral libraries and file types from both empiricaland predicted spectral libraries. We demonstrate that RTLS methodsenable improved quantitation of multiplexed samples, particularlywith consideration for quantitation from chimeric fragment spectra.We used RTLS to profile proteome responses to small molecule perturbationsand were able to quantify up to 15% more significantly regulated proteinsin half the gradient time compared to traditional methods. Taken together,the development of RTLS expands the IDA toolbox to improve instrumentefficiency and quantitative accuracy for sample multiplexed analyses.

High-quality labeled datasets for applying machine and deep learning to several tasks in MS based proteomics.

Machine learning (ML) and deep learning (DL) models for peptide property prediction such as Prosit have enabled the creation of high quality in silico reference libraries. These libraries are used in various applications, ranging from data-independent acquisition (DIA) data analysis to data-driven rescoring of search engine results. Here, we present Oktoberfest, an open source Python package of our spectral library generation and rescoring pipeline originally only available online via ProteomicsDB. Oktoberfest is largely search engine agnostic and provides access to online peptide property predictions, promoting the adoption of state-of-the-art ML/DL models in proteomics analysis pipelines. We demonstrate its ability to reproduce and even improve our results from previously published rescoring analyses on two distinct use cases. Oktoberfest is freely available on GitHub (https://github.com/wilhelm-lab/oktoberfest) and can easily be installed locally through the cross-platform PyPI Python package.

Proteomics is the interdisciplinary field focusing on the large-scale study of proteins. Proteins essentially organize and execute all functions within organisms. Today, the bottom-up analysis approach is the most commonly used workflow, where proteins are digested into peptides and subsequently analyzed using Tandem Mass Spectrometry (MS/MS). MS-based proteomics has transformed various fields in life sciences, such as drug discovery and biomarker identification. Today, proteomics is entering a phase where it is helpful for clinical decision-making. Computational methods are vital in turning large amounts of acquired raw MS data into information and, ultimately, knowledge. Deep learning has proved its success in multiple domains as a robust framework for supervised and unsupervised machine learning problems. In proteomics, scientists are increasingly leveraging the potential of deep learning to predict the properties of peptides based on their sequence to improve their confident identification. However, a reference dataset is missing, covering several proteomics tasks, enabling performance comparison, and evaluating reproducibility and generalization. Here, we present a large labeled proteomics dataset spanning several tasks in the domain to address this challenge. We focus on two common applications: peptide retention time and MS/MS spectrum prediction. We review existing methods and task formulations from a machine learning perspective and recommend suitable evaluation metrics and visualizations. With an accessible dataset, we aim to lower the entry barrier and enable faster development in machine learning for proteomics.

Proteome-wide measurements of protein turnover have largely ignored the impact of post-translational modifications (PTMs). To address this gap, we employ stable isotope labeling and mass spectrometry to measure the turnover of >120,000 peptidoforms including >33,000 phosphorylated, acetylated, and ubiquitinated peptides for >9,000 native proteins. This site-resolved protein turnover (SPOT) profiling discloses global and site-specific differences in turnover associated with the presence or absence of PTMs. While causal relationships may not always be immediately apparent, we speculate that PTMs with diverging turnover may distinguish states of differential protein stability, structure, localization, enzymatic activity, or protein-protein interactions. We show examples of how the turnover data may give insights into unknown functions of PTMs and provide a freely accessible online tool that allows interrogation and visualisation of all turnover data. The SPOT methodology is applicable to many cell types and modifications, offering the potential to prioritize PTMs for future functional investigations.

Isobaric labeling increases the throughput of proteomics by enabling the parallel identification and quantification of peptides and proteins. Over the past decades, a variety of isobaric tags have been developed allowing the multiplexed analysis of up to 18 samples. However, experiments utilizing such tags often exhibit reduced identification rates and thus show decreased analytical depth. Re-scoring has been shown to rescue otherwise missed identifications but was not yet systematically applied on isobarically labeled data. Because iTRAQ 4/8-plex and the recently released TMTpro 16/18-plex share similar characteristics with TMT 6/10/11-plex, we hypothesized that Prosit-TMT, trained exclusively on 6/10/11-plex labeled peptides, may be applicable to these isobaric labeling strategies as well. To investigate this, we re-analyzed nine publicly available datasets covering iTRAQ and TMTpro labeling for samples with human and mouse origin. We highlight that Prosit-TMT shows remarkably good performance when comparing experimentally acquired and predicted fragmentation spectra (R of 0.84 - 0.9) and retention times (ΔRT95% of 3% - 10% gradient time) of peptides. Furthermore, re-scoring substantially increases the number of confidently identified spectra, peptides, and proteins.

Machine learning has been an integral part of interpreting data from mass spectrometry (MS)-based proteomics for a long time. Relatively recently, a machine-learning structure appeared successful in other areas of bioinformatics, Transformers. Furthermore, the implementation of Transformers within bioinformatics has become relatively convenient due to transfer learning, i.e., adapting a network trained for other tasks to new functionality. Transfer learning makes these relatively large networks more accessible as it generally requires less data, and the training time improves substantially. We implemented a Transformer based on the pretrained model TAPE to predict MS2 intensities. TAPE is a general model trained to predict missing residues from protein sequences. Despite being trained for a different task, we could modify its behavior by adding a prediction head at the end of the TAPE model and fine-tune it using the spectrum intensity from the training set to the well-known predictor Prosit. We demonstrate that the predictor, which we call Prosit Transformer, outperforms the recurrent neural-network-based predictor Prosit, increasing the median angular similarity on its hold-out set from 0.908 to 0.929. We believe that Transformers will significantly increase prediction accuracy for other types of predictions within MS-based proteomics.

The prediction of fragment ion intensities and retention time of peptides has gained significant attention over the past few years. However, the progress shown in the accurate prediction of such properties focused primarily on unlabeled peptides. Tandem mass tags (TMT) are chemical peptide labels that are coupled to free amine groups usually after protein digestion to enable the multiplexed analysis of multiple samples in bottom-up mass spectrometry. It is a standard workflow in proteomics ranging from single-cell to high-throughput proteomics. Particularly for TMT, increasing the number of confidently identified spectra is highly desirable as it provides identification and quantification information with every spectrum. Here, we report on the generation of an extensive resource of synthetic TMT-labeled peptides as part of the ProteomeTools project and present the extension of the deep learning model Prosit to accurately predict the retention time and fragment ion intensities of TMT-labeled peptides with high accuracy. Prosit-TMT supports CID and HCD fragmentation and ion trap and Orbitrap mass analyzers in a single model. Reanalysis of published TMT data sets show that this single model extracts substantial additional information. Applying Prosit-TMT, we discovered that the expression of many proteins in human breast milk follows a distinct daily cycle which may prime the newborn for nutritional or environmental cues.

The study of microbiomes has gained in importance over the past few years and has led to the emergence of the fields of metagenomics, metatranscriptomics, and metaproteomics. While initially focused on the study of biodiversity within these communities, the emphasis has increasingly shifted to the study of (changes in) the complete set of functions available in these communities. A key tool to study this functional complement of a microbiome is Gene Ontology (GO) term analysis. However, comparing large sets of GO terms is not an easy task due to the deeply branched nature of GO, which limits the utility of exact term matching. To solve this problem, we here present MegaGO, a user-friendly tool that relies on semantic similarity between GO terms to compute the functional similarity between multiple data sets. MegaGO is high performing: Each set can contain thousands of GO terms, and results are calculated in a matter of seconds. MegaGO is available as a web application at https://megago.ugent.be and is installable via pip as a standalone command line tool and reusable software library. All code is open source under the MIT license and is available at https://github.com/MEGA-GO/.