Inter-individual variability in lipoprotein proteomics reveals distinct patient clusters informative for disease pathogenesis and severity

This study analyzes lipoprotein proteomics in over 1,000 sepsis patients to identify three distinct sub-phenotypes reflecting disease severity and mortality, ultimately developing a validated quantitative score and machine learning models to enable precision medicine approaches for lipoprotein-based therapeutics.

The Gist

The Big Picture: Sepsis is a Chaotic Traffic Jam

Imagine your bloodstream as a busy highway. On this highway, there are special delivery trucks called lipoproteins (specifically HDL, or "good cholesterol" trucks). In a healthy person, these trucks are well-maintained, carrying the right cargo and keeping the road smooth. They act like the body's "cleanup crew," fighting inflammation and protecting organs.

Sepsis is a massive, chaotic traffic jam caused by an infection. The paper suggests that when this jam happens, the delivery trucks get hijacked, damaged, or stripped of their parts. The researchers wanted to know: Do all patients' trucks get damaged in the same way, or are there different "types" of damage?

The Discovery: Three Types of "Truck Damage"

The researchers looked at blood samples from over 1,100 sepsis patients and 149 healthy people. They analyzed the proteins (the "parts") on these trucks. Instead of finding one big mess, they found three distinct groups of patients, like three different stages of a car breaking down:

1. The "Minor Fender Bender" (Cluster LP3):

   • What happened: The trucks are still mostly intact, but they've picked up some "emergency stickers" (proteins called SAA1 and SAA2) because of the inflammation.
   • Analogy: It's like your car is still driving fine, but someone stuck a "Road Work Ahead" sign on the bumper. It's a reaction to the stress, but the engine is still running.
   • Outcome: These patients are generally less sick.

2. The "Stripped Parts" (Cluster LP2):

   • What happened: The emergency stickers are still there, but now the trucks are starting to lose their essential engine parts (healthy proteins like APOA1).
   • Analogy: The car is still moving, but the mechanic is taking out the spark plugs and the alternator to use elsewhere. The car is running on fumes.
   • Outcome: These patients are getting sicker.

3. The "Totaled Wreck" (Cluster LP1):

   • What happened: The trucks have lost almost all their essential healthy parts. The "cleanup crew" is gone.
   • Analogy: The car has been stripped for parts. The engine is dead, the wheels are off, and it's just a shell. The body has lost its ability to fight back.
   • Outcome: This group had the highest organ failure and the highest risk of death.

The Key Insight: It's a Step-by-Step Process

The most exciting finding is that this isn't random. It happens in a sequence.

• Step 1: The body reacts to the infection by adding "emergency stickers" (SAA proteins).
• Step 2: If the infection is too strong, the body starts losing the "essential engine parts" (healthy proteins).

The researchers realized that the loss of the essential parts is what actually kills the patient, not just the presence of the emergency stickers. This explains why some patients get better (they stop at Step 1) and others get worse (they slide into Step 2).

The Solution: A "Car Diagnostic Tool"

Because sepsis is so different for every person, giving the same medicine to everyone often fails. You wouldn't give a new engine to a car that just needs a sticker, and you wouldn't give a sticker to a car that needs a new engine.

The team built two tools to help doctors:

1. A Score (LPq): A number that tells you exactly how "damaged" a patient's delivery trucks are, from 0 (healthy) to 1 (totaled).
2. A Prediction Model: A computer program that looks at just 5 specific proteins (like checking 5 specific gauges on a dashboard) to instantly tell a doctor which of the three groups a patient belongs to.

Why This Matters

Think of sepsis treatment like a mechanic trying to fix a car.

• Before: Doctors treated every broken car the same way, hoping it would work.
• Now: With this new tool, a doctor can look at a patient's "dashboard" and say, "Ah, this patient is in the 'Stripped Parts' phase. We need to give them a treatment that replaces those missing engine parts (like specific proteins)."

This moves medicine from "one size fits all" to precision medicine. It opens the door for future treatments that specifically target the missing parts of the delivery trucks, potentially saving lives by fixing the specific type of damage each patient has.

Summary

• The Problem: Sepsis damages the body's protective delivery trucks (lipoproteins), but we didn't know how.
• The Discovery: There are three stages of damage, and the final stage (losing essential parts) is what causes death.
• The Innovation: The team created a simple test to identify which stage a patient is in.
• The Future: This allows doctors to tailor treatments to the specific "broken part" of the patient's immune system, rather than guessing.

Technical Summary

1. Problem Statement

Sepsis is a life-threatening organ dysfunction caused by a dysregulated host response to infection, characterized by high morbidity and mortality. A major barrier to effective treatment is the extreme clinical and biological heterogeneity of the syndrome, which has led to numerous failed clinical trials. While lipoprotein composition is known to be altered during sepsis (e.g., decreased cholesterol, increased triglycerides, and changes in protein cargo), there has been no large-scale exploration of the inter-individual variability within the lipoprotein proteome. Understanding this variability is crucial for:

• Identifying specific sub-phenotypes of sepsis.
• Linking lipoprotein alterations to immune dysfunction, organ failure, and mortality.
• Developing precision medicine approaches and risk stratification tools for future therapeutic trials.

2. Methodology

The study employed a multi-omics approach combining proteomics, transcriptomics, and machine learning across two distinct cohorts.

• Datasets:
  • Discovery Cohort (GAinS): 1,134 patients with sepsis (1,781 samples collected on Day 1, 3, and 5 of ICU admission) and 149 healthy volunteers. Proteomic data was generated via LC-MS/MS, focusing on 18 specific proteins known to be components of lipoproteins.
  • External Validation Cohort (COMBAT): 265 patients (353 samples) comprising COVID-19 patients, non-COVID sepsis patients, and healthy controls.
• Analytical Pipeline:
  1. Proteomic Profiling: Analysis of 18 lipoprotein-associated proteins, categorized into "healthy HDL" (h-HDL) proteins (e.g., APOA1, APOA2, PON1) and "inflammatory HDL" (i-HDL) proteins (e.g., SAA1, SAA2).
  2. Unsupervised Clustering: Principal Component Analysis (PCA) followed by hierarchical clustering, K-means, and consensus clustering to identify patient sub-phenotypes based on the 18-protein profile.
  3. Clinical & Molecular Correlation: Association of clusters with clinical severity (SOFA scores), mortality, Sepsis Response Signature (SRS) scores, and differential gene expression (RNA-seq).
  4. Score Development: Canonical Correlation Analysis (CCA) was used to derive a continuous quantitative score (LPq) reflecting the degree of lipoprotein alteration.
  5. Machine Learning: Random Forest classifiers and regression models were trained to predict cluster membership and LPq scores using a reduced set of proteins.
  6. Validation: Models were externally validated on the COMBAT dataset and tested for longitudinal stability across Days 1, 3, and 5.

3. Key Contributions

• Discovery of Three Distinct Sub-phenotypes: The study defined three step-wise lipoprotein proteome clusters (LP1, LP2, LP3) representing a continuum of disease severity.
• Elucidation of a Two-Step Pathogenic Mechanism: The authors identified that lipoprotein alteration in sepsis occurs in two distinct phases rather than simultaneously:
  1. Phase 1 (Inflammatory Shift): Increased abundance of i-HDL proteins (SAA1, SAA2) integrating into HDL.
  2. Phase 2 (Depletion/Dysfunction): Decreased abundance of essential h-HDL proteins (APOA1, APOA2, APOL1, etc.), which is associated with the most severe outcomes.
• Development of Predictive Tools: Creation of a continuous severity score (LPq) and machine learning classifiers capable of assigning patients to sub-phenotypes using a minimal panel of proteins (as few as 5).
• External Validity: Successful validation of these models in a mixed cohort of sepsis and COVID-19 patients, suggesting broad applicability.

4. Key Results

• Cluster Characterization:
  • LP3 (Lowest Severity): Closest to the healthy profile; characterized primarily by increased i-HDL proteins (SAA1/2).
  • LP2 (Intermediate): Further increase in i-HDL proteins and beginning of h-HDL depletion.
  • LP1 (Highest Severity): Marked by significant depletion of h-HDL proteins (APOA1, APOA2, APOL1, APOM, PON1) and lower i-HDL levels compared to LP2.
• Clinical Correlations:
  • LP1 was strongly associated with the highest SOFA scores (specifically cardiovascular and renal failure), the highest SRS1 scores (indicating immune suppression/myeloid dysfunction), and the highest 30-day mortality rates.
  • LP2 showed upregulation of neutrophil/platelet pathways, while LP1 showed downregulation of T-cell and cytokine pathways.
• Temporal Dynamics:
  • The increase in i-HDL proteins (Phase 1) appeared to regress over time (Days 1–5).
  • The depletion of h-HDL proteins (Phase 2) persisted and was the primary driver of the transition to the most severe cluster (LP1).
• Model Performance:
  • A Random Forest model using 10 proteins achieved an overall accuracy of 84% and AUCs of 0.91–0.95 for distinguishing the three clusters.
  • A reduced 5-protein model (APOA1, APOL1, APOA2, FGB, IGK3D.15) maintained robust performance (77% accuracy) and was successfully applied to the external COMBAT dataset.
  • The LPq score was independently associated with 30-day mortality and correlated with SRSq scores, confirming it captures a distinct but complementary aspect of sepsis severity.

5. Significance and Implications

• Precision Medicine in Sepsis: This study provides a molecular framework for stratifying sepsis patients based on lipoprotein proteomics. This allows for the identification of patients who may benefit from specific interventions (e.g., those in LP1 with depleted h-HDL proteins might benefit from ApoA1 mimetics or HDL supplementation).
• Therapeutic Targeting: The findings suggest that the loss of specific h-HDL proteins (like APOA1 and APOM) is a critical, persistent event driving organ failure and mortality, highlighting these proteins as promising therapeutic targets.
• Dynamic Disease Progression: The identification of a "two-step" alteration process suggests that the timing of therapeutic intervention is critical; early interventions might target the inflammatory shift, while later interventions must address the depletion of protective HDL components.
• Scalability: The development of classifiers based on a small panel of proteins (5–10) makes this approach feasible for clinical implementation and future clinical trials, moving beyond complex multi-omics requirements.

In conclusion, the paper demonstrates that the lipoprotein proteome is a highly heterogeneous and dynamic marker in sepsis. By defining specific sub-phenotypes and developing predictive models, the authors pave the way for lipoprotein-based personalized therapeutics to improve outcomes in sepsis and related critical illnesses.