Proteomic signatures improve risk prediction for common and rare diseases

    Multi-scale signaling and tumor evolution in high-grade gliomas

    quantms: a cloud-based pipeline for quantitative proteomics enables the reanalysis of public proteomics data

    Directional integration and pathway enrichment analysis for multi-omics data

    Variegated overexpression of chromosome 21 genes reveals molecular and immune subtypes of Down syndrome

    Imputation of label-free quantitative mass spectrometry-based proteomics data using self-supervised deep learning

    Host-derived protein profiles of human neonatal meconium across gestational ages

    Proteogenomic insights into the biology and treatment of pan-melanoma

    ADP-ribosylome analysis reveals homogeneous DNA-damage-induced serine ADP-ribosylation across wild-type and BRCA-mutant breast cancer cell lines

    Metabolic profiling reveals circulating biomarkers associated with incident and prevalent Parkinson's disease

Many diseases are diagnosed late because objective biomarkers are lacking. A research team from the MRC Epidemiology Unit at the University of Cambridge School of Clinical Medicine and GSK's UK R&D center collected data from more than 40,000 individuals in the UK Biobank Pharma Proteomics Project (UKB-PPP), and combined measurements of roughly 3,000 plasma proteins with clinical information to build sparse models predicting the 10-year incidence risk of 218 diseases. The results showed that sparse models using only 5 to 20 proteins predicted 67 different pathological conditions more accurately than models using basic clinical information alone or combined with 37 clinical assay data points. In addition, the sparse protein models outperformed the combination of basic information and clinical assay data in 52 diseases, including multiple myeloma, non-Hodgkin lymphoma, motor neuron disease, pulmonary fibrosis and dilated cardiomyopathy. Single-cell RNA sequencing of bone marrow samples from newly diagnosed patients revealed that 4 of the 5 predictive proteins were expressed exclusively in plasma cells, further confirming the strong predictive power of these proteins. External validation studies showed that the sparse protein models generalized well across the 6 diseases tested. The study shows that sparse plasma protein signatures—including disease-specific proteins and predictors shared across several diseases—can predict both common and rare diseases well.

Multi-scale signaling and tumor evolution in high-grade gliomas

To broaden our understanding of high-grade gliomas (both IDH-wildtype and IDH-mutant) across multiple molecular layers and to reveal the multifaceted regulatory mechanisms underlying their development and evolution, a research team led by Washington University in St. Louis and the Pacific Northwest National Laboratory integrated data from 14 proteomics, metabolomics, lipidomics and post-translational modification platforms together with genomic and transcriptomic data, analyzing 228 tumor samples (including 212 GBMs and 16 IDH-mutant astrocytomas) and comparing them with normal brain tissue and metastatic brain lesions. The results showed that upstream alterations from different sources converge on common downstream events, and that protein interactions and glycosylation site occupancy change at recurrence. In addition, recurrent genetic alterations and phosphorylation events were linked to the PTPN11 signaling pathway, suggesting that this pathway plays an important role in high-grade gliomas. The study provides new ideas and directions for the clinical treatment of neuro-oncology; future work may further improve treatment outcomes by investigating the mechanisms of the tumor microenvironment in greater depth, and may validate these new findings in additional experimental model systems to provide a more reliable basis for clinical treatment.

quantms: a cloud-based pipeline for quantitative proteomics enables the reanalysis of public proteomics data

With the rapid growth of public proteomics data, large-scale reanalysis faces computational challenges. The team of Yasset Perez-Riverol at the European Molecular Biology Laboratory developed quantms (https://quantms.org/), a cloud-based open-source platform designed for massively parallel proteomics data analysis. The platform supports three mainstream experimental types—DDA, isotope labeling and DIA—and is highly flexible and scalable. Beyond reanalyzing public proteomics data, quantms provides additional functions such as quality control reports and differential expression analysis. Validated through real use cases, quantms showed higher accuracy in the quantification of low-abundance proteins when reanalyzing large single-cell and tissue sample datasets. The platform adopts standardized file formats and reproducible workflows, supports parallel processing across multiple computers, and delivers markedly improved processing speed: when processing more than 1,000 mass spectrometry runs simultaneously, quantms executed about 40 times faster than MaxQuant. The team also innovatively proposed a method for integrating results across different datasets, successfully identifying 479 new proteins. All quantms modules are provided as BioConda packages and BioContainers, developed following nextflow and nf-core guidelines, and support resuming interrupted tasks and dynamic resource reallocation. quantms brings a flexible, efficient and user-friendly solution to the field of large-scale quantitative mass spectrometry data analysis.

Directional integration and pathway enrichment analysis for multi-omics data

High-throughput omics technologies can systematically describe genes, transcripts, proteins and epigenetic states in cells. Although methods for generating data have advanced rapidly, the corresponding integration methods remain a major challenge. Pathway enrichment analysis is a widely used technique for interpreting omics data through known gene functions and biological processes. The team of Jüri Reimand at the University of Toronto proposed a new method—directional P-value merging (DPM)—for integrating multiple omics datasets, including information at the gene, transcript and protein levels. Using user-defined directional constraints, the method prioritizes genes or pathways that change in concordant directions and penalizes those with discordant directions. Multiple experimental designs and analytical approaches—including simulated datasets, comparative analyses of real datasets and integrative analyses across several cancer types and biological samples—demonstrated the effectiveness and reliability of the method.

Down syndrome (DS) is a genetic disorder caused by trisomy of chromosome 21 (HSA21) and shows strong inter-individual variability in developmental phenotypes and in the diagnosis of comorbidities. The mechanisms behind this variability in developmental and clinical presentation remain to be clarified. By analyzing whole-blood transcriptomes from hundreds of DS patients, the team of Matthew D. Galbraith and Joaquin  M. Espinosa at the University of Colorado identified two distinct clusters of human HSA21 genes and, using clustering algorithms, classified DS into three unique molecular subtypes. Multi-omics comparative analysis revealed significant differences among the three subtypes in inflammation, immunity, cell growth and metabolism, and revealed distinct patterns of immune cell variation across subtypes. RNA sequencing was used to analyze the expression of 126 HSA21 genes in 356 DS patients, and the average expression of these genes was found to be higher than in normal controls. However, the expression of these genes varied widely among individuals with DS, indicating the existence of distinct HSA21 gene expression patterns. Cluster analysis then divided DS patients into three molecular subtypes (MS1, MS2 and MS3), each with a unique HSA21 gene expression pattern. Correlation analysis exploring the relationship between HSA21 genes and genome-wide expression profiles showed that the expression patterns of HSA21 gene cluster 1 and cluster 2 were associated with different genome-wide expression profiles. Principal component analysis comparing the three DS molecular subtypes with genome-wide expression profiles showed that each subtype has a distinct genome-wide expression profile. Finally, linear models and proteomic data analysis were used to compare the three DS molecular subtypes against the plasma proteome, revealing hundreds of unique plasma protein changes in each subtype. These findings reveal the diversity and complexity of HSA21 genes in DS patients and provide deeper insight into the molecular subtypes of DS, helping to better understand DS pathogenesis and laying the foundation for individualized clinical treatment.

Imputation of label-free quantitative mass spectrometry-based proteomics data using self-supervised deep learning

Proteomics is a technology for identifying and quantifying proteins. Because precursor fragment selection is semi-random, missing values are pervasive and must be replaced at least in some parts of the data analysis. At present, handling missing values in proteomics data usually assumes that protein abundance falls below the instrument's detection limit or that the protein is absent. In general, the research community distinguishes missing at random (MAR) from missing not at random (MNAR); the former affects all intensities across the entire dynamic range, while the latter means that missingness becomes more common as peptide intensity approaches the instrument's detection limit. However, not all missing values arise from this mechanism, and incorrectly assuming that the detection limit is the cause of missingness can lead to potentially erroneous inferences, producing biased statistical results and limiting the conclusions that can be drawn from the data. Simon Rasmussen and colleagues at the University of Copenhagen developed PIMMS (proteomics imputation modeling mass spectrometry), a method that uses self-supervised learning to handle large-scale datasets and can effectively predict disease progression and biomarkers. They compared different self-supervised learning models (such as autoencoders and variational autoencoders) on data at different levels (proteomics data, peptide data and precursor ion data), and evaluated the predictive ability of these models on real data with simulated missing values. The results showed that self-supervised learning models can effectively impute missing values without relying on prior knowledge, and perform well on data at different levels. The study also explored several properties of self-supervised learning models, such as their ability to preserve within-sample correlations and correlations among features. Overall, the work offers a new direction for the application of self-supervised learning in bioinformatics and provides inspiration for further research. The authors applied the method in a study of patients with alcohol-related liver disease and found that it identified more differentially expressed proteins, and that these protein sets can be used to predict disease progression. The authors also provide the corresponding Python code and workflow so that other researchers can reproduce and extend the work.

Host-derived protein profiles of human neonatal meconium across gestational ages

Meconium is a non-invasive biological material that reflects the accumulation of substances before birth and can provide valuable information about neonatal health. To investigate how the protein composition of neonatal meconium varies in relation to disease, the team of Eiichiro Watanabe at the University of Tokyo Graduate School of Medicine performed deep proteomic analysis of meconium samples stratified by gestational age, sex and disease, identifying 5,370 human proteins and finding differences in meconium protein composition across gestational ages and diseases. Sex-difference analysis revealed significant differences between female and male newborns: certain specific proteins were present at higher levels in females and the reverse in males. These results suggest that newborns of different sexes may have distinct physiological characteristics and disease risks. In addition, the team developed a machine learning model that can predict neonatal outcomes from meconium proteins. This study provides a new approach to assessing neonatal intestinal health and reveals the complex relationships between meconium protein composition and clinical parameters, opening new directions for further research on neonatal health.

Melanoma is one of the most common skin cancers, with a high rate of metastasis and poor prognosis. Understanding its molecular pathogenesis is essential to improving diagnosis and treatment. The team of Chen Ding at Fudan University and Yingyong Hou at Zhongshan Hospital, Fudan University, conducted an integrated multi-omics analysis of 207 untreated melanoma patients, revealing the biological characteristics of melanoma and approaches to treating it. The results showed that PRKDC amplification is one of the molecular markers of melanoma prognosis, and that its cis effect may promote tumor growth by activating DNA repair and folate metabolism pathways. Proteomics-based typing divided primary melanoma into three subtypes—the ECM subtype, the angiogenic subtype (with a high metastasis rate) and the cell proliferation subtype—providing an important framework for subtype-specific targeted therapy. Immune classification identified three immune subtypes, and further revealed that up-regulation of the MAPK7-NFKB signaling pathway can promote T-cell recruitment and increase patient sensitivity to immunotherapy, whereas PRKDC may reduce the sensitivity of melanoma patients to immunotherapy. These results underscore the value of multi-omics data analysis in clinical practice and hold promise for improving melanoma treatment.

ADP-ribosylation (ADPr) plays an important role in cellular processes such as the DNA damage response, and is mediated mainly by PARP1 and HPF1. BRCA1 and BRCA2 are key proteins in DNA damage repair, and their inactivation makes cells sensitive to PARP inhibitors. PARP inhibitors have been used to treat certain types of cancer, but the emergence of resistance is a major challenge for long-term treatment. The team of Michael Lund Nielsen at the University of Copenhagen investigated the role of DNA-damage-induced ADPr responses in wild-type and BRCA-mutant breast cancer cell lines, and analyzed the relationship between PARP inhibitor sensitivity and resistance. The researchers used liquid chromatography-tandem mass spectrometry (LC-MS/MS) to analyze ADPr responses in six breast cancer cell lines with different PARP inhibitor sensitivities. The results showed that in all cell lines, DNA-damage-induced ADPr occurred predominantly on serine residues, and that these target sites were highly conserved. In addition, the researchers observed differences between PARPi-sensitive BRCA mutants and the PARPi-resistant BRCA mutant HCC1937 cell line: the former showed higher USF1_S189  ADPr levels, whereas the latter had lower PARG levels and longer ADP-ribose chains. These findings provide new insight into the role of ADPr signaling in the DNA damage response and into the challenges of PARP inhibitor therapy.

Parkinson's disease (PD) is a complex neurodegenerative disorder for which effective curative therapies are still lacking. Further research into the molecular pathways of PD is essential to understanding its etiology and developing new treatments. Mingguang He of the University of Melbourne and Zhuoting Zhu and Xiaohong Yang of Southern Medical University used nuclear magnetic resonance (NMR) technology to analyze baseline metabolomics and genetic data from 109,790 participants, investigating the role of metabolic profiles in the onset and progression of PD. The results showed that 68 metabolites were associated with incident PD, among which polyunsaturated fatty acids (PUFAs) and omega-6 fatty acids were associated with both incident and prevalent PD; a further 14 metabolites were associated with prevalent PD, including amino acids, fatty acids, several lipoprotein subclasses and lipid ratios. Comparing the metabolomics data with conventional risk factor models showed that adding metabolomics data improved the predictive ability of the model, although the incremental contribution was not significant. Metabolic profiling can therefore provide additional information for understanding the different pathways involved in the development of PD and help reveal potential intervention targets.