Review Papers Claim by Claim

Paper at-a-glance

• Goal: robustly compare trajectories across C conditions (topology, progression, fate-selection, gene dynamics).
• Core steps: topologyTest (permutations + weighted KS/classifier), progressionTest (KS/classifier), fateSelectionTest (classifier), conditionTest (NB‑GAM with condition-specific smoothers).
• Benchmarks: dyngen simulations (T1–T4) and 4 public datasets; code/data: GitHub + Bioconductor vignettes (reproducible).

Source: condiments paper (De Bézieux et al.)

Evidence-based strengths (paper claims & cited evidence)

1. Clear division of questions: topology → global abundance/progression → gene-level dynamics increases interpretability and reduces conflation of effects (authors implement topologyTest, progressionTest, fateSelectionTest, conditionTest) condiments workflow details
2. Reproducibility & software: authors distributed condiments (R/Bioconductor), Ecume package for distribution tests, and case-study vignettes — facilitating reuse and auditing code and data availability
3. Benchmarks: dyngen simulations (T1–T4) show higher TPR and F1 for progressionTest (downstream of slingshot/monocle3) compared to milo and DAseq for many scenarios, indicating improved sensitivity when leveraging trajectory structure simulation performance

Critical weaknesses, assumptions, and blindspots

1. Dependence on upstream TI estimates and "double-dipping": tests treat pseudotime/weights as observed but they are estimated from the full dataset — authors acknowledge this and frame p-values as exploratory numerical summaries rather than strict probabilistic statements dependence and interpretation caveat
2. TopologyTest mapping restrictions: requires TI methods that accept a prespecified skeleton (slingshot/TSCAN) — limits automation and generality when TI tools (e.g., monocle3) produce variable lineage counts or topologies topologyTest method constraints
3. Batch confounding risk: when conditions equal batches, normalization/integration may remove biology or leave residual batch signals that drive topologyTest rejections; authors discuss but empirical guidance is limited—strong experimental design remains essential batch confounding discussion
4. Classifier-test power/bias choices: classifier-based multivariate tests are flexible but sensitive to classifier choice, hyperparameters, and class-balance handling; authors default to k-NN/RandomForest and cross-validation, but downstream conclusions can vary with these choices (needs sensitivity checks) classifier-test implementation

Practical guidance for users (operational checklist)

• Pre-check 1 — batch layout: avoid perfect confounding of condition and batch; if unavoidable, interpret topologyTest rejections as potential batch artifacts.
• Pre-check 2 — TI stability: run multiple TI methods (slingshot, monocle3) and verify robustness of progression/conditionTest results; condiments authors show high correlation of Wald stats between slingshot/monocle3 in TGF-β case.
• Pre-check 3 — classifier sensitivity: run sensitivity analyses for classifier type (kNN vs RF) and training/test splits; report if conclusions hinge on classifier choice.
• Visualization: always show imbalance score map + pseudotime densities per condition (authors provide these plots in vignettes) to ground statistical tests in visuals.

All practical recommendations and caveats are drawn from the manuscript and its Methods/Discussion sections, which explicitly discuss the points above and provide code/vignettes for applied users recommendations and vignettes