Spironolactone and Fibrosis in Heart Failure Risk: Machine Learning Analysis of HOMAGE Trial Plasma Proteomics

FIGURE 1. Flowchart illustrating the machine learning and statistical analyses workflow. PICP means procollagen Type I C‐terminal propeptide; ML, machine learning; RMSE, root mean square of residuals; ANN, artificial neural network; cRF, conditional random forest; XGB, extreme gradient boosting; ENR, elastic net regression.

FIGURE 2. Selection of proteins associated with PICP changes after treatment. Venn diagrams highlighting baseline levels (black) and fold‐changes (green) of proteins selected if present within the 10% most important at least in two of the four machine learning models after 1 (A) and 9 (B) months of treatment, and the common proteins at both visits (C).

FIGURE 3. Functional enrichment analyses. Cnetplots showing the connections among the five most significantly enriched GO terms using the human genome as the background (A), and among the top enriched GO terms (fold enrichment > 5, unadjusted p < 0.05) using the list of all Olink proteins determined as the background (B), for the MLA‐selected proteins associated with PICP fold‐changes at 1 and 9 months. Proteins are shown in blue, GO terms in red, and dot size reflects the number of genes per term. GO, Gene Ontology; MLA, machine learning algorithms.

FIGURE 4. Mediation analyses in HOMAGE clinical trial. Effect decomposition plots for the combination of the selected candidates in mediator signature scores derived from linear regression models using PICP fold‐change as the continuous target and including all selected protein fold‐changes as independent variables at 1 (A) and 9 months (C). These scores rendered mediated proportions of 0.70 and 0.58, suggesting that the combined variation in these proteins mediated 70 and 58% of the total effect of spironolactone on PICP at 1 (A) and 9 (C) months, respectively. Graphs depicting mediation loss after removal of each protein from the mediator scores estimated at 1 and 9 months of spironolactone are shown in panels B and D, respectively. Estimates are accompanied by 95% percentile bootstrap confidence intervals based on 1000 bootstrap resamples. Variables in red are those with higher proportion mediation loss values, indicating greater importance in the mediation. ACME, average causal mediation effect; ADE, average direct effect.

FIGURE 5. Association of changes in plasma THBS2 and GAL9 with the decrease in serum PICP. 3D scatterplots with regression planes depicting THBS2 and GAL9 fold‐changes associated with PICP fold‐changes in patients from the HOMAGE trial after 1 month (A) and 9 months (B) of spironolactone treatment, and in patients from the Aldo‐DHF trial after 12 months of treatment (C). Multiple linear regressions were adjusted by the baseline values of PICP, THBS2, and GAL9.

FIGURE 6. Associations between plasma THBS2 and GAL9 slopes and the slopes of LAVI and E:A ratio in HOMAGE. 3D scatterplots with regression planes depict the predicted subject‐specific slopes of LAVI (A) and the E:A ratio (B) as a function of subject‐specific slopes of THBS2 and GAL9. Subject‐specific slopes were estimated as random slopes from linear mixed‐effects models using all three visits (basal, 1, and 9 months). Multivariable linear models assessing these associations were adjusted for age, sex, study arm, and eGFR. The color scale indicates the magnitude of the predicted slopes. Omnibus p values refer to the joint statistical significance of THBS2 and GAL9 slopes in the multivariable models.