Explainable machine learning for the identification of proteins associated with cardiorespiratory fitness
E J Santana, N Cauwenberghs, J Claes, R Brouwers, E Ntalianis, V Cornelissen, T KuznetsovaAbstract
Background
Cardiorespiratory fitness (CRF) reflects the integrated functioning of the respiratory, cardiovascular, and skeletal muscle systems and, as such, is a strong predictor of mortality. Direct measurement of peak oxygen consumption (peak VO2) in cardiopulmonary exercise testing is the gold-standard to assess CRF. As molecular patterns influence how the body responds to physical exercise, integrating proteomic information with peak VO2 data using explainable machine learning (ML) could provide a powerful way to characterize molecular CRF signatures.
Purpose
To identify the most impactful proteins associated with CRF in cross-sectional studies using explainable ML methods.
Methods
Two Belgian cohorts were pooled for the analysis, one from a general population study and another from a randomized clinical trial including patients at high cardiovascular (CV) risk. We included individuals who were profiled with a CV-related targeted proteomics panel (92 proteins) and who performed cycle ergometry with a respiratory exchange ratio≥1. CRF was evaluated in terms of peak VO2/kg. ML analysis consisted of three main steps. First, dimensionality reduction was performed using Partial Least Squares (PLS). Next, proteins with PLS’ Variable Importance in the Projection (VIP) statistic >1.2 were modelled along with age and sex to predict CRF using Extreme Gradient Boosting (XGBoost), exploring non-linear relationships among the features and the target. For deriving the model, 80% of the data was reserved for training and 5-fold cross-validation was used for hyperparameter tuning. Finally, Shapley additive explanations (SHAP) were used to evaluate the most important proteins for the model’s prediction and their directionality.
Results
Among 379 individuals, the average age was 60.9 ± 9.9 years and 170 (44.9%) were women (Table 1). From the 92 profiled proteins, 15 were selected based on PLS’ VIP >1.2. The selected proteins included proteins related to energy homeostasis (e.g. leptin, agouti-related protein), inflammation (e.g. interleukein-6), cardiac function (e.g. brain natriuretic peptide), endothelial health (e.g. thrombomodulin), and angiogenesis (e.g. angiopoietin-1 receptor, pro-adrenomedullin). The XGBoost model for prediction of peak VO2/kg presented an R2=0.70 and 0.62 in the training and test sets, respectively (mean absolute error of 2.5 mL/kg/min and 3.7 mL/kg/min). For explainability, the SHAP plot showed the rank of the features (Figure 1). For instance, higher leptin level was associated with a lower CRF and ranked first among the features in both training and test sets.
Conclusion
Non-linear and explainable ML identified proteins related to energy homeostasis, cardiac function, inflammation and angiogenesis that were predictive of CRF. Potential future clinical applications could consider proteomics molecular levels for a deeper, more personalised understanding of low CRF.