TL;DR
This study introduces a novel deep learning-based method (DeepPRM) for efficiently designing mass spectrometry assays to identify circulating protein biomarkers for intracranial aneurysms (IAs). They successfully developed two panels of proteins that can accurately distinguish IA patients from healthy controls and classify IA rupture risk, offering a promising avenue for non-invasive diagnosis and risk stratification.
What Problem Does This Paper Address?
Intracranial aneurysms (IAs) are a significant health concern, affecting 2-3% of the population, with a high risk of rupture leading to severe consequences, including death. Current diagnostic methods like DSA are invasive, and there's a critical need for non-invasive, sensitive biomarkers to detect IAs early and predict rupture risk. This would enable better treatment decisions, balancing the risks of preventive procedures against the danger of rupture. Existing research has identified numerous proteins associated with IA, but a comprehensive, clinically applicable serum biomarker panel has been elusive.
Methods
The researchers employed a multi-stage approach:
- Biomarker Bank Construction: They compiled an extensive "Serum Protein Candidate Biomarker Bank" (SPCBB) of 717 proteins. This bank integrated proteins identified from their own IA tissue and serum proteomic analyses with previously reported IA-associated proteins from the literature.
- DeepPRM for Assay Development: A key innovation is their DeepPRM algorithm, a deep learning-based method designed to predict the detectability and retention time of peptides for mass spectrometry (MS) assays. This significantly streamlines the development of targeted MS assays for a large number of candidate proteins, a process that is typically tedious and costly. They demonstrated DeepPRM's superiority over existing tools like Picky in generating targeted peptides.
- Targeted Proteomics (PRM): Using DeepPRM, they developed MS assays for the SPCBB proteins and quantified 113 potential markers in two serum cohorts (Cohort I: n=212, Cohort II: n=32).
- Machine Learning for Classification: A machine-learning pipeline was used to build predictive models. This involved differential feature reservation (DFR), candidate feature selection, and final model construction (CFS & FMC) using logistic regression and recursive feature elimination (RFE) with cross-validation.
Datasets
- IA Tissue and Serum Proteomics: Data generated from IA tissues and matched superficial temporal artery (STA) tissues, as well as serum samples from healthy controls (NC), unruptured IA (UR), and ruptured IA (R) patients.
- Literature Data: Previously reported IA-associated proteins were integrated to build the SPCBB.
- Cohort I (n=212): Serum samples used for initial PRM quantification and machine learning model training and internal validation. This cohort was divided into training (75%) and internal validation (25%) sets.
- Cohort II (n=32): Serum samples used for external validation of the developed biomarker panels.
- HeLa Dataset: Used to validate the performance of the DeepPRM algorithm.
Key Insights and Ideas
- Deep Learning for Assay Design: The development of DeepPRM is a significant contribution, addressing a major bottleneck in large-scale proteomic biomarker discovery by automating and improving the efficiency of targeted MS assay design.
- Comprehensive Biomarker Bank: The creation of a 717-protein SPCBB, integrating multiple data sources, provides a valuable resource for IA research.
- Serum Biomarkers Reflect Disease State: The study strongly suggests that circulating protein profiles in serum can effectively differentiate between healthy individuals and IA patients, and even stratify risk for IA rupture.
- Inflammation as a Key Player: The findings reinforce the role of inflammation and immune system activation in IA, particularly in the context of rupture.
Key Results
- Biomarker Panels:
- P6 Panel: A six-protein panel accurately distinguished IA patients from healthy controls with an accuracy of 87.50% in the external validation set.
- P8 Panel: An eight-protein panel achieved an accuracy of 91.67% in classifying IA rupture risk in the external validation set.
- DeepPRM Performance: DeepPRM demonstrated a significantly higher detection rate of targeted peptides (85.23%) compared to the Picky method (30.70%) when validated against HeLa cell digests and serum samples.
- Protein Quantification: 113 peptides (corresponding to 100 proteins) were quantified with high confidence in serum samples.
- Pathway Enrichment: Analysis revealed that ruptured IA patients showed distinct enrichment in pathways related to the innate immune system and neutrophil degranulation.
Limitations and Future Work
- Limited Tissue Samples: The study acknowledges a limitation in the number of tissue samples available for discovery due to technical complexity and treatment strategies. However, they argue that the focus on serum validation mitigates this.
- Therapeutic Strategy Variability: Differences in therapeutic approaches among patients could potentially influence results, although the consistency across cohorts suggests this impact was managed.
- Detailed Protein Roles: The precise roles of some identified biomarker proteins (e.g., PDLIM1, CTSG) in IA pathogenesis require further investigation and experimental validation.
- Clinical Translation: While promising, the developed panels require further validation in larger, diverse clinical trials to establish their full clinical utility for routine diagnostics and risk assessment. The authors envision the widespread application of their serum biomarker strategy for other proteomic signature discovery studies.
Summary updated