Biology subjects

Bouwmeester, R.

Publications and source records attributed to Bouwmeester, R..

At least 19 recordsLinked to original sources

bioRxiv2026

Evaluating the use of non-linear models in data-driven rescoring of peptide-spectrum matches

In mass spectrometry (MS)-based proteomics, computational tools match acquired tandem MS spectra to peptides from a sequence database. Machine learning increasingly supports this task through peptide-spectrum match (PSM) rescoring, in which a classifier, typically a linear semi-supervised model, refines the initial matching score. However, Mokapot allows the user to choose among different machine learning algorithms of increasing complexity, from the default linear support vector machine (LSVM) to random forest and XGBoost. Here, we use an entrapment approach to assess the effect of this increasing complexity on PSM identification and the accuracy of the estimated false discovery rate (FDR). We show that, while more complex models increase the number of identified PSMs at a fixed FDR threshold, this gain reflects a bias towards random matches from the target proteome database rather than genuine identifications. Indeed, for the most complex model, the entrapment FDR reaches 6.3% instead of the estimated 1% decoy FDR. This bias thus yields overly optimistic FDR estimates, indicating that model complexity in PSM rescoring must be carefully balanced against this overfitting risk.

bioinformatics↗

bioRxiv2026

Predicting and Elucidating Peptide Retention Mechanisms with Graph Attention Networks

Liquid chromatography (LC) is a key technology in bottom-up proteomics, separating proteolytic peptides to decrease sample complexity, enhance coverage, and increase the robustness of protein identification and quantification. Although high-resolution mass spectrometry has advanced significantly, comparable progress in LC has lagged, primarily due to a limited understanding of peptide-column interactions. To bridge this knowledge gap, we introduce a novel deep learning model (PeptideGNN) based on a Graph Neural Network (GNN) architecture to model and elucidate peptide behaviors across various separation conditions. Trained to accurately predict peptide retention times on ten diverse proteomic datasets, the model subsequently employed a saliency mapping technique to interpret the underlying retention mechanisms. Our model consistently outperformed existing retention-time predictors across multiple datasets, while the saliency mapping, importantly, revealed insights into peptide-stationary phase interactions, highlighting the effects of neighboring amino acids, post-translational modifications (PTMs), chromato-graphic columns, and mobile phase additives on peptide retention.

bioinformatics↗

bioRxiv2026

LFQ Benchmark Dataset - Generation Beta: Assessing Modern Proteomics Instruments and Acquisition Workflows with High-Throughput LC Gradients

Recent advances in liquid chromatography-mass spectrometry (LC-MS) have accelerated the adoption of high-throughput workflows that deliver deep proteome coverage using minimal sample amounts. This trend is largely driven by clinical and single-cell proteomics, where sensitivity and reproducibility are essential. Here, we extend our previous benchmark dataset (PXD028735) using next-generation LC-MS platforms optimized for rapid proteome analysis. We generated an extensive DDA/DIA dataset using a human-yeast-E. coli hybrid proteome. The proteome sample was distributed across multiple laboratories together with standardized analytical protocols specifying two short LC gradients (5 and 15 min) and low sample input amounts. This dataset includes data acquired on four different platforms, and features new scanning quadrupole-based implementations, extending coverage across different instruments and acquisition strategies. Our comprehensive evaluation highlights how technological advances and reduced LC gradients may affect proteome depth, quantitative precision, and cross-instrument consistency. The release of this benchmark dataset via ProteomeXchange (PXD070049 and PXD071205), allows for the acceleration of cross-platform algorithm development, enhance data mining strategies, and supports standardization of short-gradient, high-throughput LC-MS-based proteomics.

bioinformatics↗

bioRxiv2026

quantms-rescoring enables deep proteome coverage across protein quantification, immunopeptidomics, and post-translational modifications experiments.

The growing volume of public proteomics datasets and the advent of novel machine learning (ML)-based methods create unprecedented opportunities for discovery through large-scale reanalysis. However, traditional desktop tools are increasingly insufficient for processing and integrating data at this scale. To address this challenge, we present a novel package, quantms-rescoring, that extends the cloud-native quantms workflow with a machine learning-based rescoring module. Unlike prior tools that rescore single-engine outputs, quantms-rescoring seamlessly integrates multiple search engines (SAGE, COMET, and MSGF+), performs automatic model selection, model fine-tuning, and scales reproducibly on cloud infrastructures. In quantms-rescoring, we rely on multiple fragment-ion intensity (AlphaPeptDeep and MS2PIP) and retention-time prediction (DeepLC) methods to improve results from multiple peptide database search engines. It features automatic model selection, fine-tuning, and retraining for MS/MS intensity and retention time prediction to select the best model for a given dataset. We applied the novel workflow to five representative datasets spanning DDA label-free quantification, TMT 10-plex isobaric labelling of tumor proteomics data, immunopeptidomics, phospho-proteomics, and unseen lysine malonylation experiments. We achieved a 16-22.8% increase in identified spectra, along with the quantification of 2191 additional phosphorylated peptides and 1337 phosphosites. In the tandem mass tag (TMT)-labeled clear cell renal cell carcinoma dataset, 76 novel differentially expressed multiple search engines identified proteins with quantms-rescoring. Additionally, novel 11,688 HLA-II potential binders were detected in the immunopeptidomics dataset by multiple search engines with quantms-rescoring. For unseen malonylation data, we reported more than 58.8% malonylation PSMs and 30.5% modification sites than COMET alone. Together, these results show that integrating multi-engine searches with machine learning-derived features can be combined in a scalable workflow that enhances identification, PTM localization, and quantification performance.

bioinformatics↗

bioRxiv2025

iDeepLC: chemical structure information yields improved retention time prediction of peptides with unseen modifications

Deep learning has notably advanced the field of liquid chromatography-mass spectrometry-based proteomics. Accurate prediction of peptide retention times significantly enhances our ability to match LC-MS data with the correct peptides and proteins, especially for DIA data. While numerous models predict peptide LC retention times with high accuracy, few can accurately predict the retention times of chemically modified peptides, particularly those with modifications not encountered during model training. In our previously developed DeepLC model, accurate predictions could be made for unseen modifications by leveraging the chemical composition of (modified) residues. Here, however, we present a further enhancement of this model based on chemical structural information. The resulting model, called iDeepLC, shows overall more accurate predictions, and better generalization performance for predicting the retention time of unseen modifications than DeepLC. iDeepLC is freely available as open-source software under the Apache2 license and can be found at https://github.com/CompOmics/iDeepLC.

bioinformatics↗

bioRxiv2025

The E. coli PeptideAtlas Build: Characterizing the observed Escherichia coli pan-proteome and its post-translational modifications

Escherichia coli is a widely used model organism in molecular biology. Despite its pivotal role, a comprehensive proteome resource covering the E. coli pan-proteome and its post-translational modifications (PTMs) has been lacking. Here we present the E. coli PeptideAtlas build, the first comprehensive pan-proteome analysis of E. coli, generated from 40 high-quality public and in-house datasets spanning a broad diversity of strains, sample types, and experimental conditions, and comprising over 73 million MS/MS spectra. All datasets were reprocessed using both a closed search (Trans-Proteomic Pipeline using MSFragger) and an open search (ionbot). The E. coli PeptideAtlas build provides evidence for 4,755 proteins, including 1,410 previously lacking protein-level support in UniProt. The resource offers protein coverage, modification sites, raw spectra with matched peptides, and manually annotated metadata for the E. coli pan proteome. PTM profiling identified over 10,000 modification sites, including phosphorylation (3,806), acetylation (754), methylation (730), glutathionylation (352) and phosphoribosylation (226). Analysis of the glutathionylation sites revealed potential links to metal binding regulation. We also detected proteins likely associated with phages, underscoring the value of pan-proteomic approaches for studying host-phage interactions. All identifications are publicly accessible and traceable through the PeptideAtlas interface. We expect that the E. coli PeptideAtlas build will provide a useful resource for the community, which supports, for example, targeted MS experiment design, PTM enrichment method development, and strain typing. It allows straightforward lookups of protein and peptide identifications and facilitates comparative proteomic analyses by enabling the assessment of protein presence and variability across different E. coli strains. The build is available at https://db.systemsbiology.net/sbeams/cgi/PeptideAtlas/buildDetails?atlas_build_id=585.

bioinformatics↗

bioRxiv2025

Limitations of de novo sequencing in resolving sequence ambiguity

De novo peptide sequencing enables peptide identification from fragmentation spectra without relying on sequence databases. However, incomplete spectra create ambiguity, making unambiguous identification challenging. Recent deep learning advances have produced numerous de novo models that predict sequences and refine peptide-spectrum matches under such conditions. Yet, their relative strengths, weaknesses, and ability to handle spectrum ambiguity remain unclear. Here, we benchmark eight state-of-the-art models on three publicly available proteomics datasets, comparing performance using established metrics and quantifying inter-model agreement. We assess post-processing approaches, including iterative refinement, rescoring, and reranking, for their ability to improve identification accuracy, and perform an error analysis to identify common mispredictions and their causes. Model performance varied, with considerable overlap of correct identifications. Post-processing yielded no or only modest improvements. Most sequencing errors were model-independent and driven by limited fragment ion coverage, a limitation also observed in database searches with large search spaces.

bioinformatics↗

bioRxiv2025

DeepLC introduces transfer learning for accurate LC retention time prediction and adaptation to substantially different modifications and setups

While LC retention time prediction of peptides and their modifications has proven useful, widespread adoption and optimal performance are hindered by variations in experimental parameters. These variations can render retention time prediction models inaccurate and dramatically reduce the value of predictions for identification, validation, and DIA spectral library generation. To date, mitigation of these issues has been attempted through calibration or by training bespoke models for specific experimental setups, with only partial success. We here demonstrate that transfer learning can successfully overcome these limitations by leveraging pre-trained model parameters. Remarkably, this approach can even fit highly performant models to substantially different peptide modifications and LC conditions than those on which the model was originally trained. This impressive adaptability of transfer learning makes it a highly robust solution for accurate peptide retention time prediction across a very wide variety of imaginable proteomics workflows.

bioinformatics↗

bioRxiv2025

Collisional cross-section prediction for multiconformational peptide ions with IM2Deep

Peptide collisional cross-section (CCS) prediction is complicated by the tendency of peptide ions to exhibit multiple conformations in the gas phase. This adds further complexity to downstream analysis of proteomics data, for example for identification or quantification through feature finding. Here, we present an improved version of IM2Deep that is trained on a carefully curated dataset to predict CCS values of multiconformational peptides. The training data is derived from a large and comprehensive set of publicly available datasets. This comprehensive training dataset together with a tailored architecture allows for the accurate CCS prediction of multiple peptide conformational states. Furthermore, the enhanced IM2Deep model also retains high precision for peptides with a single observed conformation. IM2Deep is publicly available under a permissive open source license at https://github.com/compomics/IM2Deep.

bioinformatics↗

bioRxiv2024

Maximizing immunopeptidomics-based bacterial epitope discovery by multiple search engines and rescoring

Mass spectrometry-based discovery of bacterial immunopeptides presented by infected cells allows untargeted discovery of bacterial antigens that can serve as vaccine candidates. However, reliable identification of bacterial epitopes is challenged by their extreme low abundance. Here, we describe an optimized bioinformatical framework to enhance the confident identification of bacterial immunopeptides. Immunopeptidomics data of cell cultures infected with Listeria monocytogenes were searched by four different search engines, PEAKS, Comet, Sage and MSFragger, followed by data-driven rescoring with MS2Rescore. Compared to individual search engine results, this integrated workflow boosted immunopeptide identification by an average of 27% and led to the high-confidence detection of 18 additional bacterial peptides (+27%) matching 15 different Listeria proteins (+36%). Despite the strong agreement between the search engines, a small number of spectra (< 1%) had ambiguous matches to multiple peptides and were excluded to ensure high-confident identifications. Finally, we demonstrate our workflow with sensitive timsTOF SCP data acquisition and find that rescoring, now with inclusion of ion mobility features, identifies 76% more peptides compared to Q Exactive HF acquisition. Together, our results demonstrate how integration of multiple search engine results along with data-driven rescoring maximizes immunopeptide identification, boosting the detection of high-confidence bacterial epitopes for vaccine development.

molecular biology↗

bioRxiv2024

ClassiCOL: LC-MS/MS analysis for ancient species Classification via Collagen peptide ambiguation

LC-MS/MS extends on the MALDI-TOF ZooMS approach by providing fragmentation spectra for each peptide. However, ancient bone samples generate sparse datasets containing only a few collagen proteins, rendering target-decoy strategies unusable and increasing uncertainty in peptide annotation. ClassiCOL embraces and even extends this ambiguity using a novel isoBLAST approach. The exhaustive set of potential peptide candidates created in this way is then used to retain or reject different potential paths at each taxonomic branching point down to the taxonomic level attainable with the sample information, always allowing for potential mixtures in the process. As an end point, all considered ambiguity is graphically represented with a clear prioritization of the species in the sample. Using public as well as in-house data, we demonstrate the performance of this universal postprocessing approach on different instruments and explore the possibility of identifying genetic as well as sample mixtures. Diet reconstruction from 40,000 year old cave hyena coprolites illustrates the exciting potential of this approach. TeaserClassiCOL is a postprocessing tool that allows for more accurate species classification from LC-MS/MS measurements of collagen.

bioinformatics↗

bioRxiv2024

TIMS2Rescore: A DDA-PASEF optimized data-driven rescoring pipeline based on MS2Rescore

The high throughput analysis of proteins with mass spectrometry (MS) is highly valuable for understanding human biology, discovering disease biomarkers, identifying therapeutic targets, and exploring pathogen interactions. To achieve these goals, specialized proteomics subfields - such as plasma proteomics, immunopeptidomics, and metaproteomics - must tackle specific analytical challenges, such as an increased identification ambiguity compared to routine proteomics experiments. Technical advancements in MS instrumentation can counter these issues by acquiring more discerning information at higher sensitivity levels, as is exemplified by the incorporation of ion mobility and parallel accumulation - serial fragmentation (PASEF) technologies in timsTOF instruments. In addition, AI-based bioinformatics solutions can help overcome ambiguity issues by integrating more data into the identification workflow. Here, we introduce TIMS2Rescore, a data-driven rescoring workflow optimized for DDA-PASEF data from timsTOF instruments. This platform includes new timsTOF MS2PIP spectrum prediction models and IM2Deep, a new deep learning-based peptide ion mobility predictor. Furthermore, to fully streamline data throughput, TIMS2Rescore directly accepts Bruker raw mass spectrometry data, and search results from ProteoScape and many other search engines, including MS Amanda and PEAKS. We showcase TIMS2Rescore performance on plasma proteomics, immunopeptidomics (HLA class I and II), and metaproteomics data sets. TIMS2Rescore is open-source and freely available at https://github.com/compomics/tims2rescore.

bioinformatics↗

bioRxiv2024

Proteogenomics analysis of human tissues using pangenomes

The genomics landscape is evolving with the emergence of pangenomes, challenging the conventional single-reference genome model. The new human pangenome reference provides an extra dimension by incorporating variations observed in different human populations. However, the increasing use of pangenomes in human reference databases poses challenges for proteomics, which currently relies on UniProt canonical/isoform-based reference proteomics. Including more variant information in human proteomes, such as small and long open reading frames and pseudogenes, prompts the development of complex proteogenomics pipelines for analysis and validation. This study explores the advantages of pangenomes, particularly the human reference pangenome, on proteomics, and large-scale proteogenomics studies. We reanalyze two large human tissue datasets using the quantms workflow to identify novel peptides and variant proteins from the pangenome samples. Using three search engines SAGE, COMET, and MSGF+ followed by Percolator we analyzed 91,833,481 MS/MS spectra from more than 30 normal human tissues. We developed a robust deep-learning framework to validate the novel peptides based on DeepLC, MS2PIP and pyspectrumAI. The results yielded 170142 novel peptide spectrum matches, 4991 novel peptide sequences, and 3921 single amino acid variants, corresponding to 2367 genes across five population groups, demonstrating the effectiveness of our proteogenomics approach using the recent pangenome references.

bioinformatics↗

bioRxiv2023

Benchmarking DIA data analysis workflows

Data independent acquisition (DIA) has become a well-established method in LC-MS driven proteomics. Nonetheless, there are still a lot of possibilities at the data analysis level. By benchmarking different DIA analysis workflows using a ground-truth sample, consisting of a differential spike-in of UPS2 in a constant yeast background, we provide a roadmap for DIA data analysis of shotgun samples based on whether sensitivity, precision or accuracy is of the essence. Three different commonly used DIA software tools (DIA-NN, EncyclopeDIA and SpectronautTM) were tested in both spectral library mode and spectral library-free mode. In spectral library mode we used the independent spectral library prediction tools PROSIT and MS2PIP together with DeepLC, next to the classical DDA-based spectral libraries. In total we benchmarked 12 DIA workflows. DIA-NN in library-free mode or using in silico predicted libraries, together with Spectronaut in library-free mode, shows the highest sensitivity maintaining a high reproducibility and accuracy. In general, DIA-NN shows the best reproducibility, while the accuracy is comparable for all DIA workflows.

bioinformatics↗

bioRxiv2022

PhosphoLingo: protein language models for phosphorylation site prediction

Protein post-translational modifications (PTMs) play an important role in numerous biological processes by significantly affecting protein structure and dynamics. Effective computational methods that provide a sequence-based prediction of PTM sites are desirable to guide functional experiments. Whereas these methods typically train neural networks on one-hot encoded amino acid sequences, protein language models carry higher-level pattern information that may improve sequence based prediction performance and hence constitute the current edge of the field. In this study, we first evaluate the training of convolutional neural networks on top of various protein language models for sequence based PTM prediction. Our results show substantial prediction accuracy improvements for various PTMs with current procedures of dataset compilation and model performance evaluation. We then used model interpretation methods to study what these advanced models actually base their learning on. Importantly for the entire field of PTM site predictors trained on proteomics-derived data, our model interpretation and transferability experiments reveal that the current approach to compile training datasets based on proteomics data leads to an artefactual protease-specific training bias that is exploited by the prediction models. This results in an overly optimistic estimation of prediction accuracy, an important caveat in the application of advanced machine learning approaches to PTM prediction based on proteomics data. We suggest a partial solution to reduce this data bias by implementing negative sample filtering, only allowing candidate PTM sites in matched peptides that are present in the experimental metadata. Availability and implementationThe prediction tool, with training and evaluation code, trained models, datasets, and predictions for various PTMs are available at https://github.com/jasperzuallaert/PhosphoLingo. Contactsven.degroeve@vib-ugent.be and nico.callewaert@vib-ugent.be Supplementary informationSupplementary materials are available at bioRxiv.

bioinformatics↗

bioRxiv2022

Machine learning on large-scale proteomics data identifies tissue- and cell type-specific proteins

Using data from 183 public human data sets from PRIDE, a machine learning model was trained to identify tissue and cell-type specific protein patterns. PRIDE projects were searched with ionbot and tissue/cell type annotation was manually added. Data from physiological samples were used to train a Random Forest model on protein abundances to classify samples into tissues and cell types. Subsequently, a one-vs-all classification and feature importance were used to analyse the most discriminating protein abundances per class. Based on protein abundance alone, the model was able to predict tissues with 98% accuracy, and cell types with 99% accuracy. The F-scores describe a clear view on tissue-specific proteins and tissue-specific protein expression patterns. In-depth feature analysis shows slight confusion between physiologically similar tissues, demonstrating the capacity of the algorithm to detect biologically relevant patterns. These results can in turn inform downstream uses, from identification of the tissue of origin of proteins in complex samples such as liquid biopsies, to studying the proteome of tissue-like samples such as organoids and cell lines.

systems biology↗

bioRxiv2022

FAVA: High-quality functional association networks inferred from scRNA-seq and proteomics data

Protein networks are commonly used for understanding how proteins interact. However, they are typically biased by data availability, favoring well-studied proteins with more interactions. To uncover functions of understudied proteins, we must use data that are not affected by this literature bias, such as single-cell RNA-seq and proteomics. Due to data sparseness and redundancy, co-expression analysis becomes complex. To address this, we have developed FAVA (Functional Associations using Variational Autoencoders), which compresses high-dimensional data into a low-dimensional space. FAVA infers networks from high-dimensional omics data with much higher accuracy than existing methods, across a diverse collection of real as well as simulated datasets. FAVA can process large datasets with over 0.5 million conditions and has predicted 4,210 interactions between 1,039 understudied proteins. Our findings showcase FAVAs capability to offer novel perspectives on protein interactions. FAVA functions within the scverse ecosystem, employing AnnData as its input source.

bioinformatics↗

bioRxiv2021

MS2Rescore: Data-driven rescoring dramatically boosts immunopeptide identification rates

Immunopeptidomics aims to identify Major Histocompatibility Complex-presented peptides on every cell that can be used in anti-cancer vaccine development. However, existing immunopeptidomics data analysis pipelines suffer from the non-tryptic nature of immunopeptides, complicating their identification. Previously, peak intensity predictions by MS2PIP and retention time predictions by DeepLC, have been shown to improve tryptic peptide identifications when rescoring peptide-spectrum matches with Percolator. However, as MS2PIP was tailored towards tryptic peptides, we have here retrained MS2PIP to include non-tryptic peptides. Interestingly, the new models not only greatly improve predictions for immunopeptides, but also yield further improvements for tryptic peptides. We show that the integration of new MS2PIP models, DeepLC, and Percolator in one software package, MS2Rescore, increases spectrum identification rate and unique identified peptides with 46% and 36% compared to standard Percolator rescoring at 1% FDR. Moreover, MS2Rescore also outperforms the current state-of-the-art in immunopeptide-specific identification approaches. Integration of immunopeptide MS2PIP models, DeepLC, and Percolator into MS2Rescore thus allows substantial improved identification of novel epitopes from existing immunopeptidomics workflows.

bioinformatics↗