Protein Compositional Ratio Representation (PCRR)Systematically Improves Human Disease Prediction

This study demonstrates that modeling plasma proteomics data as compositional systems using pairwise protein ratios significantly outperforms traditional abundance-based approaches in predicting human diseases, achieving substantial accuracy gains in Alzheimer's subtypes and across a broad spectrum of conditions in the UK Biobank.

The Gist

The Big Idea: It's Not About the Volume, It's About the Balance

Imagine you are trying to understand a complex orchestra.

• The Old Way (Raw Data): You measure the absolute loudness of every single instrument. You write down: "The violin is at 80 decibels, the trumpet is at 75, and the drum is at 90."
  • The Problem: What if the whole orchestra is playing in a tiny, echoey room versus a massive stadium? The numbers change completely, even if the music (the relationships between the instruments) stays exactly the same. If you just look at the raw numbers, you might think the music changed, when really, the room just got bigger or smaller.
• The New Way (This Paper's Method): Instead of measuring how loud each instrument is, you measure the ratio between them. You ask: "Is the violin twice as loud as the trumpet?" or "Is the drum twice as loud as the violin?"
  • The Result: It doesn't matter if the room is big or small. If the violin is always twice as loud as the trumpet, that relationship tells you the true "shape" of the music.

The Paper's Discovery:
The researchers found that when studying human blood proteins (proteomics), looking at the ratios between proteins is much better at predicting diseases than looking at the raw amounts of proteins.

---

The Problem: The "Noisy Room"

In the past, scientists treated every protein in your blood as an independent number. They thought, "If Protein A is high, that's bad."
But the authors realized that biological systems are like a compositional soup.

• If you add a drop of water to a soup, everything gets slightly less salty, but the ratio of salt to pepper stays the same.
• In your blood, technical glitches, how much you drank that morning, or how the sample was stored can make all protein levels look higher or lower (the "volume" changes).
• By focusing on raw numbers, machines were getting confused by this "noise." They were trying to learn from the volume of the room rather than the music being played.

The Solution: The "See-Saw" Approach

The authors created a new method called Protein Compositional Ratio Representation (PCRR).
Instead of asking "How much Protein A is there?", they ask, "How does Protein A compare to Protein B?"

Think of it like a see-saw:

• It doesn't matter if the see-saw is on a mountain or in a valley (the absolute height).
• What matters is who is heavier. Is the person on the left side heavier than the person on the right?
• In the body, diseases often happen when the balance shifts. Maybe a "good" protein goes down and a "bad" protein goes up. Even if both numbers change slightly, the ratio between them screams "Something is wrong!"

The Results: A Magic Trick for Disease Prediction

The team tested this on two massive groups of people:

1. Alzheimer's Patients (ROSMAP): They tried to predict different stages of Alzheimer's (from mild memory loss to full disease).
   • The Result: Their new "ratio" method was significantly better than the old "raw number" method. It was like upgrading from a blurry black-and-white photo to a crystal-clear 4K video. It was especially good at spotting the tricky, early stages of the disease that other methods missed.
2. The UK Biobank (53,000+ People): They tested this on 587 different diseases, from heart disease to diabetes to infections.
   • The Result: The ratio method won 95% of the time. It improved predictions for almost every single disease they looked at.

Why This Matters (The "Aha!" Moment)

The paper suggests that our bodies don't work by having a fixed amount of "Protein X." They work by maintaining a delicate balance between different proteins.

• Analogy: Think of a recipe for a cake. If you double the flour but also double the sugar and eggs, the cake tastes the same. The ratio of ingredients is what matters, not the total weight of the bowl.
• The Insight: When we get sick, the recipe gets thrown off. The ratio of "flour" to "sugar" changes. By measuring the ratios, the computer can spot the "bad recipe" (the disease) much faster and more accurately than by just weighing the ingredients.

The Bottom Line

This paper is a game-changer because it tells scientists: "Stop looking at the absolute numbers; look at the relationships."

By treating blood protein data like a balanced scale rather than a pile of independent numbers, we can build better AI models to predict diseases earlier, understand them better, and potentially find new ways to treat them. It's a simple shift in perspective that unlocks a huge amount of hidden information.

Technical Summary

1. Problem Statement

The Core Issue: Current machine learning models applied to plasma proteomics typically treat protein abundances as independent variables in Euclidean space. This approach ignores the fundamental nature of biological systems, which are compositional.

• Compositional Data: In proteomics, the absolute concentration of a protein is often less biologically meaningful than its relative balance to other proteins within a pathway (e.g., receptor-ligand stoichiometry, enzyme-substrate ratios).
• Limitations of Raw Data: Models trained on absolute abundances are susceptible to batch effects, normalization artifacts, and inter-individual variability (global scaling factors). These factors obscure the true biological signal, limiting predictive power and interpretability.
• Hypothesis: The fundamental unit of proteomic variation is not the absolute level of a single protein, but the relative balance between proteins. Therefore, modeling pairwise log-ratios ($\log(A) - \log(B)$) should provide a more stable, scale-invariant, and biologically coherent representation for disease prediction.

2. Methodology

The authors propose a Protein Compositional Ratio Representation (PCRR) framework, which transforms proteomic data into a log-ratio space before applying machine learning.

A. Mathematical Foundation

• Log-Ratio Transformation: Instead of using raw normalized abundances ($x$), the method constructs features as pairwise differences of log-transformed values: $r_{ij} = \log(x_i) - \log(x_j) = \log(x_i/x_j)$.
• Scale Invariance: This transformation ensures that the features are invariant to global multiplicative scaling (e.g., sample dilution or assay batch effects). If $x$ is scaled by a constant $c$, the ratio remains unchanged: $\log(cx_i/cx_j) = \log(x_i/x_j)$.
• Geometry: The method maps data from the Aitchison simplex (the natural space for compositional data) into a Euclidean subspace where pairwise differences represent valid geodesic directions, preserving the intrinsic geometry of relative proportions.

B. Feature Engineering Pipeline

The pipeline involves a three-stage process to generate predictive ratios:

1. Initial Feature Prioritization: A LightGBM model is trained on the full set of raw proteins (plus demographics) using cross-validation. Proteins exhibiting non-zero feature importance in at least 2 out of 5 splits are selected as "consistently predictive."
2. Ratio Generation: All unique, non-redundant pairwise log-ratios are generated only from this shortlist of predictive proteins. This reduces dimensionality while operating in the compositional space.
3. Model Training: The final classifier is trained on the engineered ratio features.

C. Datasets and Evaluation

The framework was validated across two distinct cohorts:

1. ROSMAP Cohort (Alzheimer's Disease):
   • Data: $n=871$ individuals, 953 visits, 7,298 plasma proteins (SomaScan).
   • Task: Multi-class classification of four subtypes: No Cognitive Impairment (NCI), Mild Cognitive Impairment (MCI), Alzheimer's Disease (AD), and AD+ (AD with concurrent causes).
   • Baselines: Compared against (1) Stratified Random, (2) Demographics-only, and (3) Raw Proteomics + Demographics.
2. UK Biobank (Generalizability):
   • Data: $n > 53,000$ individuals, 3,000 proteins (Olink platform).
   • Task: Prediction of 587 distinct disease outcomes (neurological, metabolic, immune, infectious, etc.).
   • Validation: 5-fold cross-validation for each outcome.

3. Key Contributions

1. Novel Representation: Introduction of PCRR, a systematic framework that treats proteomics as compositional data using pairwise log-ratios, directly encoding biological constraints into the learning space.
2. Scale Invariance: Formal proof and empirical demonstration that this representation removes irrelevant multiplicative noise (batch effects/dilution) that plagues raw abundance models.
3. Interpretability: The top-ranked ratios map directly to known biological pathways (e.g., microglial activation, proteostasis), offering mechanistic insights that raw protein lists often miss.
4. Scalability: Successful application of the method from a deep longitudinal cohort (ROSMAP) to a massive cross-sectional population dataset (UK Biobank) across hundreds of phenotypes.

4. Key Results

A. Alzheimer's Disease Subtype Classification (ROSMAP)

The PCRR model significantly outperformed all baselines, including the "gold standard" of raw proteomics + demographics.

• AUROC Gains: The ratio-based model achieved an average AUROC improvement of +0.1274 over the strongest baseline (Raw Proteomics + Demographics).
  • Specific gains: +0.1281 for MCI, +0.1246 for AD, and +0.1848 for AD+.
• Average Precision (AP): The gains were even more dramatic for imbalanced minority classes.
  • For the difficult AD+ class, the AP improved by +0.3898 (an 8-fold increase), transforming a near-useless classifier into a viable one.
• Feature Insights: Top ratios included biologically relevant pairs such as:
  • SEMA3C:TMEM70 (linked to microglial activation).
  • IDUA:NPTXR (linked to proteostasis and lipid clearance).
  • ACHE-based contrasts (linked to cholinergic signaling and amyloid-$\beta$).

B. Generalizability (UK Biobank)

The approach proved robust across 587 disease outcomes.

• Win Rate: Ratio-based models outperformed raw protein models in 95.1% of all outcomes.
• Statistical Significance: Significant gains (FDR < 0.05) were observed in 56.7% of outcomes.
• Magnitude: The average AUROC improvement was 7.93%, with a maximum observed improvement of 46.6%.
• Broad Applicability: Improvements were seen across neurological (Parkinson's, Vascular Dementia), cardiometabolic (Heart Failure, Stroke), and infectious diseases. Notably, for acute infections, ratios appeared to capture "host vulnerability" (immune competence/inflammation) rather than pathogen-specific signals.

5. Significance and Implications

• Paradigm Shift: The paper argues that proteomic data should fundamentally be viewed and modeled as compositional systems. The information distinguishing disease states lies in the relative balance of proteins, not their absolute abundances.
• Biological Coherence: By focusing on ratios, the model recovers "biologically coherent axes of disease" that are often obscured by noise in raw data. This aligns with the biological reality that cellular processes depend on stoichiometry and feedback loops.
• Clinical Utility: The method offers a general-purpose, interpretable strategy for biomarker discovery. It is particularly effective for:
  • Heterogeneous diseases (like AD subtypes) where distinct molecular mechanisms drive different phenotypes.
  • Imbalanced classes (rare diseases or specific subtypes) where raw models fail to detect minority signals.
  • Noisy data environments where batch effects or sample dilution are concerns.
• Future Directions: The authors suggest this compositional principle may extend to other omics layers (transcriptomics, metabolomics, lipidomics) and could be integrated into multi-omic frameworks to reveal higher-order biological structures.

In conclusion, the study provides rigorous evidence that Protein Compositional Ratio Representation (PCRR) is a superior, mathematically grounded, and biologically interpretable approach for disease prediction, consistently outperforming traditional methods that rely on absolute protein abundances.