23rd Human Proteome Organization World Congress

MoDPA: investigating diseases through modification-dependent protein associations

Data Integration: With Bioinformatics to Biological Knowledge

Both protein-protein associations and Post-Translational Modifications (PTMs) are crucial for regulating various biological processes, but the relationship between these two regulatory mechanisms is not fully understood. To bridge this gap, I am currently developing MoDPA (Modification-dependent Protein Associations), a computational approach that aims to identify functional associations between PTMs.Co-expression analysis is a widely utilized method for exploring the function and regulation of genes. Various techniques exist for assessing the relationship between two proteins, with the simplest being based on co-occurrence. My research project seeks to employ a similar approach to identify consistently co-occurring clusters of PTMs across experiments.To obtain PTM identifications, public datasets from the online PRIDE repository were uniformly reprocessed using ionbot, a machine learning-based peptide search engine.To tackle the high dimensionality and sparsity of PTM data, I employed a Variational Autoencoder (VAE) deep-learning model to compress the dataset into a low-dimensional latent space while retaining its essential features. An adjacency matrix was then constructed by calculating correlations between PTMs in the latent space, resulting in a network of co-occurring PTMs. Each node in the network represents a modification on a specific protein residue, and two modifications are connected by an edge if they were consistently co-identified in the analyzed mass spectrometry data.A clustering analysis revealed that proteins within related pathways tend to bear similar PTMs. For instance, proteins involved in phagocytosis display co-occurring Lysine and Arginine methylation events, whereas proteins related to translation displayed co-occurring Lysine acetylation events. Intriguingly, clusters containing a combination of acetylation, phosphorylation and methylation also emerged from the analysis, offering an opportunity to study the cross-talk between these modifications.The PTM association network, generated by MoDPA and annotated with biological information, will be made public and will serve as a source of novel hypotheses for researchers in life sciences.

Oral presentation

A next-generation covariation map advances functional annotation of the human proteome

Most research aiming to understand the molecular foundations of life and disease has focused on a limited set of increasingly well-known proteins. By contrast, thousands of human proteins remain "understudied": their biological function is poorly understood and annotation of their molecular properties is scarce. This annotation inequality hinders biomedical progress because mechanistic investigations of gene–disease associations typically focus on proteins that are already well known, a phenomenon also known as the street-light effect (Kustatscher et al, Nat Methods 2022, PMID 35534633).We have previously shown that protein covariation (coexpression) analysis is a powerful proteomics approach to link uncharacterised proteins to known cellular processes (Messner et al, Cell 2023, PMID 37080200; Kustatscher et al, Nat Biotechnol 2019, PMID 31690884). Here, we present a new covariation map of the human proteome. To assemble it, we developed a proteomics data processing pipeline that is fast, scalable, and includes appropriate FDR control for very large datasets. Based on the Fragpipe platform, this pipeline enabled us to re-process 23,000 previously published MS runs in less than two weeks. This resulted in the ProteomeHD.2 dataset, which covers the abundance changes of 16,000 proteins in response to 2,500 biological perturbations, quantified using SILAC labelling. This includes many microproteins for which proteomic evidence had so far been lacking (Kourtis et al, in preparation).To determine which proteins have similar covariation patterns across ProteomeHD.2, and might thus be functionally related, we developed a machine-learning-based strategy. This strategy not only considers whether two proteins have correlated expression but also incorporates additional information from the proteomics data that reflect the robustness and reliability of the underlying protein quantitations. The resulting proteome covariation map reveals functional associations for about 10,000 human proteins. We show that it captures functional associations as well as more established techniques, such as affinity-purification MS experiments, allowing us to make predictions on the potential biological functions of many previously unidentified and understudied human proteins.

PTMeXchange species specific PTM builds: meta-analysis of datasets and dissemination of high-quality PTM data for community use

Post-translational modifications (PTMs), of which phosphorylation is the most studied, play an important role in biological functions. Mass spectrometry (MS) and database searching are commonly used to detect and localise modification sites on proteins, with confidence being governed by peptide-spectrum match (PSM) and PTM localisation statistics. The aim of the PTMeXchange project is to re-analyse public enriched PTM datasets, focusing on accurate PTM localisation, integrating data across studies and disseminating the data to UniProtKB, linking it to the original MS evidence in PRIDE and PeptideAtlas in order to make PTM data FAIR (Findable, Accessible, Interoperable and Reusable). Here we demonstrate the re-analysis workflow using publicly available mass spectrometry proteomics datasets.Publicly available mass spectrometry datasets from ProteomeXchange were selected and curated to identify relevant datasets for each investigated species and modification. An open data re-analysis pipeline using the Trans Proteomic Pipeline (TPP) and statistical methods to control PTM false localisation were used. The &quot;PTM build&quot; data from these are then integrated into public databases, including UniProtKB, PRIDE and PeptideAtlas. This enables detailed exploration of scores and visualisation of source mass spectra, as a full evidence trail. Here we demonstrate the pipeline for species specific PTM builds, including data from our first major build, phosphorylation in Asian rice (Oryza sativa).For our first build, Oryza sativa, we identified eight relevant phosphoproteomics datasets. We preformed a simple meta-analysis combining all datasets, and assigning sites to categories based on their PTM localisation scores and occurrences in datasets: Gold-Silver-Bronze. This resulted in identifying 15,565 high-quality phosphosites on serine, threonine and tyrosine residues on rice proteins. We incorporated the use of pAla decoy identifications, enabling validation of false reporting within these categories. Only two pAla hits were observed in the Gold set, with 475 sites across all categories, indicating the overall FLR is very low.We have now generated several phosphorylation builds for a variety of species including Plasmodium falciparum, Saccharomyces cerevisiae, Homo sapiens and Mus musculus. We are also covering other modifications, requiring more complex workflows, including SUMOylation, lysine acetylation and ubiquitination.The re-analysed data for the completed builds have been loaded into UniProt Knowledge-Base - enabling researchers to visualise sites alongside other data on relevant proteins e.g. structural models from AlphaFold2, PeptideAtlas and the PRIDE database - enabling visualisation of source evidence, including scores. We also demonstrate how Universal Spectrum Identifiers (USIs) can be used to visualise mass spectra supporting modification sites and can be a valuable tool to investigate the evidence supporting identified modifications.

Professor Dr

A comparison of peptide- and spectrum-centric search engines beyond bar charts

New Technology: AI and Bioinformatics in Mass Spectrometry

Background: Data-independent acquisition (DIA) offers reproducibility and proteome coverage but yields complex spectra. Various algorithms were developed for peptide identification from DIA data. Peptide-centric engines rely on fragment ion elution curves, while spectrum-centric methods analyze spectra individually. Algorithms aim to separate true from false identifications (IDs) through diverse scores, with q-values controlling false discovery rates (FDR). We compare DIA-NN, Spectronaut, and Chimerys examining their unique statistics and scores. Methods: Publicly available mixed species and single-cell Orbitrap Astral DIA raw files were downloaded from PRIDE. Datasets were searched library-free with DIA-NN 1.8.1, Spectronaut 18.5 and Chimerys 2.0 in Proteome Discoverer 3.1 against normal or entrapment databases. Chimerys deconvolutes spectra, distributing shared fragment ion intensities among co-isolated precursors. Peptides are quantified by aggregating contributions over retention time with mokapot controlling FDR. Peptide and protein FDR is calculated by Proteome Discoverer 3.1. Empirical FDR measurements were conducted via entrapment experiments. Results: Search engine comparisons typically focus on ID counts, yet FDR control and reporting levels differ among DIA-NN, Spectronaut, and Chimerys. Filtering is done at the peptide group (Chimerys) or precursor level (DIA-NN and Spectronaut), which affects the number of reported IDs. We investigated the effect of mismatching the level at which FDR is estimated and at which IDs are counted. We observe that the number of peptide groups artificially increases by ~10% if FDR is controlled at the precursor level, but IDs are counted at the peptide group level. Applying 1% precursor FDR to all search engines reveals that Chimerys detects the same number of precursors as Spectronaut, while DIA-NN seemingly identifies 17% more.The accuracy of FDR estimates significantly impacts the number of IDs. We empirically assessed FDR control using entrapment experiments. Chimerys and Spectronaut maintain precursor counts, while DIA-NN loses ~16% of IDs at 1% empirical FDR. Accurate FDR control equalizes the performance of all search engines in identifying precursors on this dataset. However, identifying spurious IDs from DIA-NN necessitates entrapment analyses.Subsequently, we investigated IDs unique to each specific tool. Unique IDs from each tool often rely on few data points or fragments, and measures of fragment ion interference can inflate ID numbers. In conclusion, close inspection of search engine results is vital for accurate peptide identification and quantification from DIA data.Conclusion: We demonstrate the importance of scrutinizing search engine scores and FDR for accurate peptide identification and quantification from DIA data.

Achieving quantitative accuracy in large-scale DIA proteomics

Data-independent acquisition (DIA) proteomics has recently seen rapid growth in popularity as a robust analytical tool in basic and translational research. New mass spectrometry instrumentation has pushed the boundaries of sensitivity and speed, and in tandem with novel acquisition schemes and advances in analysis software enables innovative applications, including spatial and single-cell proteomics. The high diversity of sample types and DIA methods calls for a reliable framework to evaluate and optimise the accuracy of protein quantification and ensure that the quantitative methods scale well to large experiments.In this work, we designed and recorded a large-scale benchmark dataset to thoroughly examine the performance of quantification algorithms in DIA proteomics. The dataset contains a series of multi-species mixes similar to the LFQbench approach, but with the crucial difference that our design introduces a variable background to mimic natural proteome variation in a controlled fashion. The digest mixes were recorded in a high-throughput setting with fast gradients (200 - 500 samples/day) on an Evosep One system coupled to the latest-generation Bruker timsTOF Ultra mass spectrometer using dia-PASEF and slice-PASEF. Injection amounts ranged from 15 ng down to 0.75 ng for the human background matrix. Our dataset thus permits to investigate the effects of injection amount, chromatography method, acquisition scheme and sample heterogeneity on accuracy, precision and reproducibility of quantification. It further allows to ensure that the data processing algorithms used can cope with the challenges posed by large experiment size and sample diversity.We use the new benchmark dataset to validate our machine learning-driven quantification framework QuantUMS (Quantification using an Uncertainty Minimising Solution), integrated in the DIA-NN software. We find that not only ultra-fast chromatographic gradients but also low sample amounts exasperate the inherent signal ratio compression in DIA proteomics. Next, we observe that QuantUMS uniquely alleviates the ratio compression, thus enabling comprehensive quantitative proteomics of low sample amounts at a throughput of hundreds of samples per day. Our results further show that QuantUMS enables reliable quantification irrespective of the dataset size or the presence of acquisitions with sample amounts spanning an order of magnitude within the same experiment, meeting the demands of the recent challenging applications of DIA proteomics.In summary, we present a large-scale and diverse controlled benchmark data set acquired using the latest-generation mass spectrometry platform. This dataset allows to gain insights into the quantitative performance of DIA at the limits of instrument speed and sensitivity, as well as permits to evaluate, optimise and validate the next-generation computational methods for protein quantification.

Session
AI and Bioinformatics Approaches

Appointment

Chairs

Program

SessionAI and Bioinformatics Approaches

Appointment

Chairs

Program

Session
AI and Bioinformatics Approaches