Blood Protein Selection Methods Show Modest Accuracy for Predicting Incident Depression

A comparative evaluation of machine learning feature selection methods has demonstrated that selecting plasma proteins yields modest predictive performance for identifying individuals at risk of incident major depressive disorder. In a study analyzing extensive proteomic data from the UK Biobank, researchers assessed four distinct feature selection techniques to address challenges common to high-dimensional biomedical datasets, including feature redundancy, model overfitting, and limited interpretability. While machine learning can capture complex, non-linear relationships across broad biological profiles, the resulting models achieved modest discrimination overall, revealing critical practical trade-offs among parsimony, stability, concordance, and computational demand.

The investigation utilized blood plasma data spanning 2,920 proteins collected from 48,378 UK Biobank participants, among whom 2,384 were documented incident cases of major depressive disorder. To benchmark feature discovery rigorously, the investigators compared the conventional univariate F-test against three multivariate algorithms: LASSO regression, PowerSHAP, and LLpowerSHAP. Each methodology was executed through repeated five-fold outer cross-validation consisting of 20 repeats and 100 total evaluations, providing an empirical basis to inspect whether different algorithmic approaches could reliably isolate robust circulating proteomic predictors across thousands of biological measurements.

Following feature selection, the selected protein subsets were modeled using four downstream classification algorithms: logistic regression, XGBoost, LightGBM, and CatBoost. Across all evaluated pipelines, predictive performance remained modest and largely comparable, showing only minor differences tied to the choice of downstream classifier. The models demonstrated mean receiver operating characteristic area under the curve values ranging between 0.614 and 0.629, while the precision-recall area under the curve ranged from 0.071 to 0.081, indicating that no single feature selection framework substantially outpaced the others in classification accuracy.

Despite this parity in predictive accuracy, the selection strategies produced starkly contrasting feature set sizes. The univariate F-test retained the broadest collections of biomarkers, selecting between 357 and 579 proteins per evaluation. In sharp contrast, PowerSHAP yielded the most parsimonious profiles, reducing candidate signatures to between 16 and 147 proteins. These findings underscore that while complex selection routines can compress proteomic panels into much smaller, compact subsets, this aggressive reduction neither significantly compromised nor enhanced the modest discriminative capacity of the downstream predictive classifiers.

The evaluation further highlighted measurable differences in algorithmic stability and cross-method concordance across repeated evaluations. The F-test and LLpowerSHAP demonstrated the highest internal selection stability across cross-validation runs, achieving adjusted stability measures of 0.75 and 0.73, respectively. When comparing how consistently methods identified the same biomarkers, LLpowerSHAP and PowerSHAP exhibited the highest degree of cross-method agreement, recording a Jaccard concordance index of 0.52, whereas other comparisons demonstrated greater divergence in their preferred candidate proteins.

Computational demands diverged by multiple orders of magnitude across the four evaluated selection procedures. The univariate F-test completed in an average of 1.25 seconds per evaluation, representing minimal computational overhead. In contrast, LLpowerSHAP required approximately 3.46 hours per evaluation, imposing substantially greater resource and time constraints without translating into superior diagnostic or predictive accuracy over faster, less resource-intensive analytical workflows.

The researchers also examined the influence of data processing choices, specifically correlation filtering and covariate adjustments. Applying moderate correlation filtering alongside covariate inclusion preserved predictive performance for the models. However, imposing stringent correlation filtering reduced predictive accuracy and altered which proteins were retained, while covariate inclusion primarily influenced the selection patterns of LASSO regression rather than changing overall predictive capacity across other pipelines.

Across all 100 evaluations and all four methods, a consensus set of 36 proteins was repeatedly chosen in at least 50 percent of the runs. Among these recurrently identified biomarkers, 30 proteins displayed nominal statistical associations with incident major depressive disorder, characterized by odds ratios between 0.88 and 0.95 for inverse associations and between 1.06 and 1.23 for positive associations. When this fixed 36-protein profile was evaluated directly, it maintained modest discrimination, achieving receiver operating characteristic area under the curve values of 0.622 to 0.641 and precision-recall area under the curve values between 0.078 and 0.083.

Subject of Research: Comparative evaluation of machine learning feature selection methods for proteomic predictors of incident major depressive disorder

Article Title: Comparative evaluation of feature selection methods for identifying proteomic predictors of incident major depressive disorder using UK Biobank proteomic data

Article References: Abebe, E. C., Madakkattel, I., Mulugeta, A., Stacey, D., & Hyppönen, E. (2026). Comparative evaluation of feature selection methods for identifying proteomic predictors of incident major depressive disorder using UK Biobank proteomic data. Journal of Big Data. https://doi.org/10.1186/s40537-026-01606-2

Image Credits: AI Generated

DOI: 10.1186/s40537-026-01606-2

Keywords: major depressive disorder, feature selection, proteomics, plasma proteins, UK Biobank, machine learning, LASSO, PowerSHAP, LLpowerSHAP, F-test, XGBoost, LightGBM, CatBoost, predictive performance, biomarkers, incident depression, computational cost, selection stability