Proteomics Reveal Clusters of Hypertension Cases Associated with Differing Prevalence of Cardiovascular and Renal Complications

By applying machine learning to proteomic data from over 7,000 hypertension patients, researchers identified distinct molecular subtypes characterized by specific protein expression patterns that correlate with varying risks of cardiovascular and renal complications, suggesting a path toward more personalized precision medicine.

The Gist

The Big Picture: Why "One Size Fits All" Doesn't Work for High Blood Pressure

Imagine Hypertension (High Blood Pressure) as a massive, noisy crowd of people. For a long time, doctors have treated this crowd as a single group: "You have high blood pressure? Here is a pill for everyone."

But the problem is, this crowd is actually made up of many different sub-groups. Some people in the crowd are at high risk of having a heart attack or kidney failure, while others are relatively safe. Currently, we don't have a good way to tell them apart until after something bad happens.

This paper is like a detective story where the researchers used a new kind of "molecular magnifying glass" (proteomics) and a smart computer brain (machine learning) to sort this noisy crowd into distinct groups based on what's happening inside their bodies right now.

---

The Detective's Toolkit: From Genes to Proteins

To understand how they did it, let's use an analogy:

• Genomics (DNA) is like reading the blueprint of a house. It tells you what the house could look like, but the blueprint never changes, even if the house is currently on fire or being renovated.
• Proteomics (Proteins) is like taking a live photo of the house. It shows you the smoke, the broken windows, and the firefighters currently working. It reflects your current health, your lifestyle, your diet, and the medicines you are taking.

The researchers looked at 2,911 different "live photos" (proteins) from over 7,000 patients with high blood pressure. They wanted to see if these photos could reveal hidden patterns that standard blood pressure readings miss.

The Method: Teaching a Computer to Sort the Crowd

Instead of just looking at the photos and guessing, they built a smart computer model (an XGBoost classifier). Think of this model as a very strict bouncer at a club who has to decide: "Is this person a 'Case' (has high blood pressure) or a 'Control' (healthy)?"

Here is the clever part:

1. The bouncer learns to spot the difference by looking at the proteins.
2. Once the bouncer is smart enough, the researchers asked: "Okay, you know who has high blood pressure. But how do you know? Which specific proteins are you looking at to make that decision?"
3. They took those specific "decision-making" clues and used them to sort the high-blood-pressure patients into 10 different groups (clusters).

It's like realizing that while everyone in the crowd is wearing a red shirt (high blood pressure), some are wearing red shirts with blue stripes, others with green polka dots, and others with yellow stars. These patterns tell you something different about each person.

The Discovery: 10 Different "Types" of High Blood Pressure

The computer found 10 distinct groups. The most exciting finding was that these groups had very different risks for future disasters:

• The "High-Risk" Groups (Clusters 4, 8, 10):
  These patients had a specific protein signature that looked like a storm cloud. They were much more likely to suffer from heart attacks, strokes, heart failure, and kidney failure.

  • The Clue: These groups had high levels of a protein called REN (Renin). Think of Renin as a pressure valve that is stuck open, causing too much pressure in the pipes. They also had high levels of HAVCR1, a protein that acts like a "distress signal" from the kidneys, saying, "We are being damaged!"

• The "Protected" Groups (Clusters 2, 6, 9):
  These patients had high blood pressure, but their internal "live photos" looked much calmer. They had a much lower risk of heart or kidney failure.

  • The Clue: Their "pressure valve" (Renin) wasn't screaming. Their kidneys weren't sending distress signals. They seemed to have high blood pressure for a different, less dangerous reason.

• The "Middle" Groups:
  Some groups had specific risks, like a higher chance of heart rhythm problems (Atrial Fibrillation) but not kidney failure.

Why This Matters: The "Precision Medicine" Revolution

Currently, if you walk into a doctor's office with high blood pressure, you get the same treatment as the person next to you. This paper suggests that is like giving a fire extinguisher to someone with a flat tire. It might help, but it's not the right tool for the specific problem.

The Future Vision:

1. Better Sorting: In the future, a doctor could take a blood sample, run this protein test, and say, "Ah, you are in the 'High-Risk Kidney' group."
2. Targeted Treatment: Instead of a generic pill, they could give a treatment specifically designed to fix the "stuck pressure valve" (Renin) or protect the kidneys.
3. Saving Money and Lives: People in the "Protected" groups might not need heavy medication, saving them from side effects. People in the "High-Risk" groups could get intensive care before they have a heart attack.

The Limitations (The Fine Print)

The researchers are honest about the hurdles:

• The Data is a Snapshot: They only took one photo of each person. We don't know how these groups change over time or if the medicines they are taking are messing up the "live photo."
• The Crowd is Mostly One Type: Most of the people in the study were of European descent. We need to make sure these groups exist in people of all backgrounds.
• The "Bouncer" is Still Learning: The computer model is very good, but it's not perfect. The researchers are working on making the sorting even more stable so that if you run the test twice, you get the same result.

The Bottom Line

This paper is a major step toward treating high blood pressure not as a single disease, but as many different diseases wearing the same mask. By using a smart computer to look at the body's "live photos" (proteins), they found hidden subgroups of patients who need very different care. It's the beginning of a future where your treatment is as unique as your fingerprint.

Technical Summary

1. Problem Statement

Hypertension is a leading global risk factor for cardiovascular disease (CVD) and premature mortality, affecting over 30% of adults. Despite effective therapies, the condition remains widely undiagnosed and undertreated due to its "silent" nature and clinical heterogeneity.

• Limitations of Current Methods: Traditional risk scores (e.g., Framingham) rely on linear regression of demographic and clinical features, failing to capture underlying disease biology or dynamic health changes. Genomic approaches (Polygenic Risk Scores) offer modest improvements but are limited by the static nature of the genome and poor portability across ancestries.
• The Gap: Existing proteomic studies often treat hypertension as a binary case/control phenotype, ignoring underlying mechanistic subtypes. This lack of stratification leads to conflicting associations (e.g., regarding renin) and prevents the identification of personalized treatment options for specific patient subgroups.

2. Methodology

The authors developed a novel machine learning-based patient stratification pipeline using proteomics data from the UK Biobank Pharma Proteomics Project (UKB-PPP).

Data Source and Preprocessing:

• Cohort: 7,086 hypertension cases (ICD-10 I10, diagnosed prior to sample collection) and 13,016 controls (normotensive, no BP medication, specific BP ranges).
• Proteomics: 2,911 plasma proteins measured via Olink technology.
• Exclusions: Patients taking ACE inhibitors were removed to prevent drug-induced protein expression artifacts (specifically affecting ACE and Renin).
• Covariates: Age and sex were included as input features to control for their strong influence on protein expression.

Machine Learning Pipeline:

1. Model Training: An XGBoost classifier (gradient boosting decision trees) was trained to predict case/control status. XGBoost was chosen for its ability to capture non-linear interactions and identify meaningful thresholds in feature space.
2. Feature Space Transformation: Instead of clustering raw protein expression values (which may be confounded by non-phenotypic factors), the authors mapped samples based on the contribution of each protein to the model's prediction output. This "model contribution space" isolates variance specifically relevant to the hypertension phenotype.
3. Dimensionality Reduction & Clustering:
   • Model contributions were reduced to 2 dimensions using UMAP (Uniform Manifold Approximation and Projection).
   • Spectral Clustering (with RBF kernel) was applied to the 2D space to identify patient subgroups.
   • A k-nearest neighbor (k-NN) classifier was fitted to the training clusters to create a reproducible "stratification map" applicable to external data.
4. Validation:
   • Cross-Validation: A 5-fold cross-validation loop was used.
   • Stability: The Adjusted Rand Index (ARI) and Rand Index (RI) were calculated between stratifications generated by different folds to measure reproducibility.
   • Enrichment Analysis: Clusters were tested for enrichment of CVD and renal complications (e.g., MI, stroke, heart failure, renal failure) using Fisher's exact tests and logistic regression (controlling for age/sex).

3. Key Contributions

• Novel Stratification Approach: The paper introduces a method that leverages the internal structure of a tree-based risk model to define patient subgroups, rather than clustering raw omics data. This approach naturally separates patients based on the specific combinations of protein values that drive disease risk.
• Identification of Mechanistic Subtypes: The study identifies 10 distinct hypertension clusters defined by differential expression of five key proteins: HAVCR1, PLAT, PTPRB, REN, and RTN4R.
• Clinical Relevance: The clusters are not arbitrary; they show statistically significant and distinct prevalences of cardiovascular and renal complications, suggesting they represent biologically distinct mechanistic subtypes of hypertension.

4. Key Results

Model Performance:

• The XGBoost model achieved a high diagnostic AUROC of 0.89 (± 0.0026) on the test set, indicating strong predictive power based on proteomics alone.

Cluster Characterization:
The analysis revealed 10 clusters with distinct risk profiles:

• High-Risk Clusters (4, 8, 10):
  • Profile: Significantly enriched for cardiovascular and renal complications (e.g., MI, heart failure, renal failure, stroke).
  • Biomarkers: Characterized by elevated expression of REN (Renin) and HAVCR1 (Kidney Injury Molecule-1).
  • Interpretation: The high REN levels suggest a "high-renin" hypertension subtype driven by RAAS overactivation. Elevated HAVCR1 indicates active renal injury.
  • Specifics: Cluster 8 showed the strongest enrichment for renal failure (OR = 3.06) and chronic ischemic heart disease.
• Low-Risk / "Protected" Clusters (2, 6, 9):
  • Profile: Significantly lower prevalence of complications compared to the general hypertension population.
  • Biomarkers: REN and HAVCR1 levels were similar to controls. These clusters showed variable expression of PLAT (Plasminogen Activator), PTPRB, and RTN4R.
  • Interpretation: The absence of RAAS overactivation and renal injury markers suggests a different, potentially more benign, mechanistic etiology.
• Intermediate/Specific Risk: Cluster 3 showed specific enrichment for cardiac arrest, while others showed mixed profiles.

Validation and Stability:

• Replication: Enrichment patterns observed in training data largely replicated in the held-out test data, though statistical power was lower due to smaller sample sizes.
• Stability: The Adjusted Rand Index (ARI) between different cross-validation folds ranged from 0.21 to 0.51. The authors note that while not perfect, an ARI of ~0.5 is considered satisfactory for unsupervised clustering without ground truth, comparable to an AUROC of 0.75.
• Covariate Control: Logistic regression confirmed that most complication enrichments remained significant even after adjusting for age and sex, indicating the clusters reflect disease biology rather than demographic artifacts.

5. Significance and Future Implications

• Precision Medicine: The findings suggest that hypertension is not a monolithic disease but a collection of mechanistic subtypes. Identifying a patient's cluster could guide treatment selection (e.g., targeting RAAS for high-renin clusters vs. lifestyle management for low-risk clusters).
• Clinical Trial Design: These clusters can be used to enrich clinical trials, recruiting patients with a shared disease etiology to increase the likelihood of detecting a drug's efficacy.
• Biomarker Development: The identified proteins (HAVCR1, REN, PLAT, PTPRB, RTN4R) serve as potential mechanism-based biomarkers for developing diagnostic tests.
• Limitations & Future Work:
  • The study relied on a single time-point proteomic profile; longitudinal data is needed to assess stability and treatment response.
  • Ancestry bias exists (86% European), necessitating validation in diverse populations.
  • Future work should integrate multi-omics (genomics, metabolomics) and causal inference methods (Mendelian Randomization) to distinguish causal drivers from downstream effects.

In conclusion, this study demonstrates that proteomic-based stratification using machine learning model contributions can successfully deconstruct the heterogeneity of hypertension, revealing distinct subgroups with varying risks of severe complications and offering a pathway toward more effective, personalized hypertension management.