Proteomic machine-learning signatures associated with carotid plaque: A population-based hypothesis-generating study

Figures

• 

  Fig 1. Age distribution. Age distribution in negative (left side) and positive classes (right side).

• 

  Fig 2. Importance from models Mbase and Mext Importance of traditional risk factors (top) (binary if possible) and extended traditional risk factors (bottom) (continuos if possible) calculated by Boruta algorithm. Green boxes – important variables; yellow boxes – tentative variables; red boxes – unimportant variables. DM – diabetes, Hch – hypercholesterolemia, HA – hypertension, BP – systolic blood pressure, Ob – obesity, Sm – smoking, MS – male sex, HbA1c - glycated hemoglobin, RAS -Renin-Angiotensin System agents, LLA – lipid-lowering agents, APA – antiplatelet agents, LDL – Low-density lipoprotein cholesterol, eGFR - Estimated glomerular filtration rate, BMI- body mass index, ySM- years of smoking.

• 

  Fig 3. Calibration curves for all models. The red diagonal line represents the ideal calibration (where predicted risks equal observed proportions). The solid black/grey line indicates the flexible calibration curve estimated via LOESS smoothing (with the 95% confidence interval shaded in grey). The spikes along the bottom axis represent the distribution of predicted risks. Models Mbaseprot and Mextprot demonstrated the best calibration performance among all evaluated models, showing near-perfect agreement between predicted risks and observed outcomes across the entire probability spectrum.

• 

  Fig 4. Observed-versus-predicted risk across risk deciles for all models. The red dashed line represents perfect calibration (y = x). Black dots indicate the grouped patient percentiles with corresponding confidence intervals. Models Mbaseprot, Mextprot, and Mageprot demonstrate the highest calibration accuracy, closely aligning with the ideal diagonal line across the entire risk spectrum, while minor deviations are observed in models Mbase, Mext, Mprot, and Mbasemod.

• 

  Fig 5. Decision-curve analysis for all models. Decision-curve analysis showed that all models yielded a positive net benefit compared with the “treat all” and “treat none” strategies across most threshold probabilities. The decision curves largely overlapped, suggesting that the models provided similar clinical utility. Although models Mbaseprot, Mextprot, and Mbasemod showed a slightly higher net benefit at some higher threshold probabilities, these differences were generally small. Mbase model appears to be the weakest, particularly in high-risk areas.

• 

  Fig 6. AUC and DOR distribution for all models. Distribution of AUC (left panel) and Diagnostic Odd Ratio (right panel) from 10 repetition of 5-fold cross-validation: Mbase – traditional risk factors (age, male sex, diabetes, hypercholesterolemia, obesity, hypertension and smoking, binary values if possible),Mext – extended traditional risk factors (age, male sex, HbA1c, LDL cholesterol, BMI, systolic blood pressure, years of smoking, eGFR, the use of agents acting on the renin–angiotensin system, lipid-lowering drugs, and antiplatelet agents), Mprot – proteomic variables, Mbaseprot – proteomic plus traditional risk factors, Mextprot- proteomic plus Mext data, Mageprot – proteomic plus “age” and “male sex”, Mbasemod – traditional risk factors plus artificial variable based on proteins.

• 

  Fig 7. Graph depicting fourty-two marker proteins relevant for the age-restricted dataset. Graph was generated with the help of the web interface of the STRING database (version 12.0). The edges of the graph correspond to the interactions between proteins identified in the database. Colours correspond to the gene ontology and KEGG pathways that are significantly enriched within the set of these proteins and that are relevant to the molecular processes underlying carotid atherosclerosis. GO terms: GO:0034097 Response to cytokine (cyan), GO:0007155 Cell adhesion (lavender blue), GO:0006955 Immune response (lime green), GO:0033993 Response to lipid (peach yellow), GO:0019221 Cytokine-mediated signalling pathway (shadow green), GO:0030195 Negative regulation of blood coagulation (light mauve), GO:0019218 Regulation of steroid metabolic process (purple). KEGG pathways: hsa04060 Cytokine-cytokine receptor interaction (light salmon pink), hsa04151 PI3K-Akt signaling pathway (blush pink) and hsa04010 MAPK signaling pathway (yellow). Permanent link: https://version-12-0.string-db.org/cgi/network?networkId=bQZnso4pYIPo.

• 

  Fig 8. Functional grouping of the 25 most relevant markers. The node area is proportional to the Boruta importance score. Coloured sectors indicate functional classes manually assigned by the authors, with the darker sector denoting the primary class. The six vertices of the hexagonal layout correspond to the six categories occurring as primary assignments, including one category for proteins not assigned to any functional class. Lipid regulation occurred only as a secondary assignment and therefore has no separate vertex. Nodes are positioned according to their primary class, and edges connect proteins sharing at least one assigned functional class.