Sex and race shape serum protein biomarkers and bias diagnostic reference intervals

Technical Summary: Sex and Race Shape Serum Protein Biomarkers and Bias Diagnostic Reference Intervals

Problem Statement
Clinical protein biomarkers, whether FDA-cleared in-vitro diagnostics (IVDs) or laboratory-developed tests (LDTs), are routinely interpreted against fixed, population-agnostic reference intervals. A critical, rarely examined assumption underpins this practice: that a reference interval derived from a general "healthy" population applies uniformly to all individuals, regardless of sex or self-reported race. While physiological variation (e.g., hormonal milieu, body composition) suggests that sex significantly alters protein baselines, and genetic literature indicates ancestry-associated differences in specific proteins, these factors are often ignored in diagnostic thresholds. Furthermore, the uncritical use of race as a biological variable in clinical algorithms has been increasingly scrutinized. The study addresses the lack of broad, standardized testing across a clinical panel to determine if baseline serum concentrations of these proteins differ systematically by sex and race in healthy individuals, and whether such differences are sufficient to cause diagnostic misclassification.

Methodology
The authors employed a targeted mass spectrometry approach to ensure high specificity, reproducibility, and portability across laboratories.

• Cohort: The study analyzed 60 commercially sourced, single-donor human serum samples from healthy young adults (ages 20–30). The cohort was strictly balanced into six groups ($n=10$ per group) based on three self-reported races (White, Hispanic, Black) and both sexes. Age distributions were matched across groups to eliminate age as a confounder.
• Assay: Scheduled multiple-reaction-monitoring mass spectrometry (MRM-MS) with stable-isotope-labelled internal standards (SIS) was used. This method quantified 93 peptides mapping to 87 proteins, comprising 39 FDA-cleared IVD analytes and 48 LDT/research markers.
• Quantitation: The primary readout was the peak-area ratio (PAR) of endogenous (light) to SIS (heavy) peptides. Protein-level values were aggregated as the median of constituent peptides.
• Statistical Analysis:
  • Global Structure: Principal-component analysis (PCA) and permutational multivariate ANOVA (PERMANOVA) were used to assess global separation by race and sex.
  • Differential Abundance: Two-way ANOVA models ($\text{log}_2\text{PAR} \sim \text{Race} + \text{Sex} + \text{Race} \times \text{Sex}$) were applied to each protein. P-values were adjusted for multiple testing using the Benjamini–Hochberg false-discovery-rate (FDR) procedure.
  • Misclassification Simulation: For FDA-target proteins, a single pooled 95% reference interval was calculated from all 60 donors. The authors then analytically computed the proportion of individuals in each race–sex subgroup falling outside these limits to quantify systematic misclassification rates.

Key Results

1. Sex as a Pervasive Axis of Variation: Unsupervised multivariate analysis revealed that donors separated significantly by sex (PERMANOVA $p = 0.020$) but not by race ($p = 0.14$). Sex acted as a distributed signal across the proteome.
2. Specific Sex-Associated Proteins: At an FDR of $q < 0.05$, three proteins showed significant sex differences: Sex hormone-binding globulin (SHBG), Haemoglobin-α (HBA), and Albumin (ALBU).
   • SHBG was higher in females; HBA and ALBU were higher in males.
   • These directions were consistent across all three racial groups, indicating a stable, cross-population biological signature.
3. Protein-Specific Race Differences: While race did not restructure the global proteome, specific proteins showed significant pairwise race differences (e.g., HBA, Transthyretin, Leucine-rich $\alpha$-2-glycoprotein, Haemopexin, and Soluble CD14). Notably, Soluble CD14 levels were higher in White donors compared to Black donors, consistent with known allele-frequency differences at the CD14 locus. These differences were localized rather than global.
4. Systematic Diagnostic Misclassification: Applying a single, pooled reference interval to all subgroups resulted in significant over-flagging of healthy individuals.
   • While a calibrated interval should flag 5% of a healthy subgroup as abnormal, specific subgroups were flagged at rates up to ~25% for certain proteins.
   • Hispanic females carried the heaviest cumulative burden, exceeding a 10% misclassification rate for eight FDA-target proteins.
   • This bias is systematic and direction-consistent, leading to differential false-positive rates across demographic groups.

Significance and Claims
The paper argues that demographic-agnostic reference intervals introduce a predictable and quantifiable bias into the interpretation of clinically used protein biomarkers. The authors claim that:

• Sex-specific partitioning is essential: Given the pervasive and direction-consistent nature of sex effects (e.g., SHBG, HBA), incorporating sex into reference intervals is a low-risk refinement that aligns with established physiology.
• Ancestry requires nuance: Race differences are real but protein-specific and often traceable to genetic ancestry (e.g., CD14 alleles) rather than the social construct of race itself. The authors suggest that where mechanistic links exist, reference intervals should reflect genetic ancestry or specific protein biology rather than broad racial categories.
• Equity in Diagnostics: Ignoring these demographic contexts results in the systematic misclassification of healthy individuals, eroding both diagnostic accuracy and equity. The study concludes that demographic context must be incorporated into reference intervals and decision thresholds to ensure that proteomic diagnostics are accurate and equitable, rather than relying on a single pooled interval that fails diverse patient populations.

The authors note limitations, including the modest sample size ($n=10$ per group), the narrow age range (20–30 years), and the use of self-reported race rather than measured genetic ancestry. They position their findings as a demonstration of the direction and magnitude of the bias, calling for larger, more granular cohorts to derive deployable, partitioned reference intervals.