| evidence-kf1 | Human cell atlas | The observations are at cell and donor-sample levels in a small retrospective human lung series; cell counts are not participant counts. | Single-cell profiling showed PAH-associated heterogeneity across nine annotated lung cell types. | Retrospective descriptive human transcriptomics | Few donor samples; cells are not independent patients | PMC11513636: Materials and methods — scRNA-seq data collecting, processing and analysis; Results — scRNA-seq analysis revealed cell landscape and characteristics in PAH; Figure 2A; supplementary Figures S1–S2. | KF1 |
|---|
| evidence-kf2 | Endothelial trajectory | Slingshot and Monocle2 infer ordering from cross-sectional gene-expression data. They do not directly observe an individual endothelial cell converting in vivo or prove causality. | Endothelial trajectory analyses identified expression states consistent with an EndMT-related program, including an EC3 state. | Computational state and pseudotime inference | Associated expression program, not lineage-tracing proof | PMC11513636: Materials and methods — Pseudotime trajectory inference; Results — Trajectory analyses revealed EndMT in endothelial populations; Figure 4A–G; Figure 5A. | KF2 |
|---|
| evidence-kf3 | Candidate-gene prioritization | This is computational prioritization of disease-associated expression features. A secondary discrepancy in the EC3 marker count is disclosed rather than resolved by assumption. | Network and expression-based selection nominated six EndMT pattern genes from the EC3-related analysis. | Co-expression and marker-based feature selection | Candidate association; ancillary marker-count discrepancy retained | PMC11513636: Results — Identification of EndMT-related modules and genes; Figure 6B–F; main-text description of Table S6. Table S4 has 604 unique gene rows while the main text reports 602 EC3 markers; the discrepant marker count is not used as an accepted quantitative claim here. | KF3 |
|---|
| evidence-kf4 | PETS classifier | The testing subset came from the same GSE117261 dataset as discovery and was also used to compare the nine algorithms and select the preferred model. The reported performance may therefore be optimistic, rather than an untouched final validation estimate. The C-index is reported as such, not relabeled as ROC AUC, and the supplementary residual column is not relabeled as RMSR. | A two-gene glmNet score showed PAH-versus-control discrimination in an internal held-out test set. | Internal held-out retrospective classification | Same-dataset split; no independent external clinical validation | PMC11513636: Materials and methods — PETS generated from integrative machine-learning program; Results — Machine learning-based integrative program generates PETS; Figure 7A–D; Additional file 13, Table S7, glmNet row (worksheet row 3, columns B–G). | KF4 |
|---|
| evidence-kf5 | Expression experiments | Expression-level concordance does not establish that RARRES2 or TC2N causes EndMT or that inhibiting either gene improves PAH. No numerical fold changes or replicate counts are inferred from figures. | Targeted experiments supported expression changes in selected candidate genes under hypoxic conditions. | Targeted hypoxia-associated expression evidence | No causal perturbation or demonstrated therapeutic benefit | PMC11513636: Materials and methods — Cell lines and cultures; RNA extraction and real-time quantitative PCR; Results — Identification of EndMT-related modules and genes, and Machine learning-based integrative program generates PETS; Figure 7E and supplementary Figure S6 caption. Experimental n and numerical fold changes are not supplied by these text/caption locators. | KF5 |
|---|