Review papers with raw data transparency

Rapid critical summary

The authors present hiddenCOMET, a population-level hidden Markov / continuous-time Markov chain (CTMC) framework that links latent EMT states (E,H,M) to observed cell cycle fractions (G1,S,G2M) via an estimated emission matrix B and uses it to recover EMT trajectories from time-series cell-cycle data when direct EMT calling fails. The method is implemented in an R package plus Shiny app and validated on four cancer cell lines and multiple inducers (TGFβ, EGF, TNFα) with simulations demonstrating identifiability under reasonable sample sizes hiddenCOMET preprint

Paper Review: hiddenCOMET — Inferring EMT dynamics from cell cycle profiles using a hidden Markov framework

What the paper does well

• Formalizes EMT inference as a latent-state CTMC/HMM connecting a robust observable (cell-cycle fractions) to a fragile latent process (EMT), providing a clear mathematical pipeline (LS and MLE estimators, multinomial likelihood) and public software to reproduce analyses Methods and code
• Uses simulations to characterize identifiability (varying B entries, sample size N, noise) and demonstrates that with realistic sampling (hundreds–thousands of cells per timepoint) recovery is possible—helpful diagnostics for practitioners Simulations and identifiability
• Open-source R package hiddenCOMET plus web app (shiny) increases accessibility and reproducibility; repository documents inputs/usage and demo data which is excellent for method uptake Repository and app

Main weaknesses and caveats

1. Fixed emission matrix across inducers: In the main analyses the emission matrix B is learned from TGFβ paired data and held fixed for EGF/TNFα fits. The authors justify this as a cell-line specific coupling but acknowledge that inducer-specific emissions are plausible and that the fixed-B assumption may bias inferred Q rates if inducer changes alter how EMT maps to cell cycle (authors note this as future work) Fixed B limitation
2. Time-homogeneous CTMC simplifies biology: EMT induction is often time-varying and stimulus-dependent (dose, duration). Using a single Q ignores transient regime shifts (e.g., immediate signaling spikes) and could blur short fast transitions—authors constrained Q to [1e-5,5/day], and note several fitted rates hit bounds, indicating sampling resolution cannot resolve very fast transitions Time homogeneous Q caveat
3. Cell-line only, limited in vivo generalizability: All input datasets are cell lines; stromal and immune cells can confound cell-cycle signals in patient or tissue samples. The authors explicitly note potential confounding by stromal/immune cell populations and that generalization to mixed samples requires further study Cell line limitation
4. Simplified 3-state EMT: The three-state coarse-grain (E,H,M) masks known heterogeneity of hybrid phenotypes; authors acknowledge lack of consensus on number of intermediate states and that richer multi-state models will need more data and likely better identifiability diagnostics EMT state simplification
5. Growth/death and fractional sampling not modeled: The model assumes constant total cell count and does not include state-specific proliferation or death rates; differential growth could change observed fractions independent of state transitions (authors mention growth/death rates as future work) No growth death modeling

Concrete suggestions to strengthen the work

• Estimate B separately for each inducer where paired data exist and compare fits (test whether inducer-specific B materially changes Q estimates). If inducer-specific B improves multinomial likelihood markedly, avoid holding B fixed across treatments Shared B vs inducer specific
• Incorporate state-specific net growth terms (birth-death) into expected fractions model: Theta_t = B x_t where x_t evolves under both transitions and differential growth to separate proliferation effects from transitions; test identifiability via simulations similar to those already performed.
• Collect higher temporal resolution (hours) for early timepoints to resolve fast transitions—authors report upper-bound hits consistent with dwell times <~3.3 hours for rates =5/day which are faster than their sampling intervals.
• Validate on mixed-population datasets or spatially resolved transcriptomics (which can separate stromal vs tumor compartments) to test robustness to non-cancer cell contamination.

Reproducibility and software

The authors provide an R package (hiddenCOMET) and a Shiny app with demo data and example usage, and the GitHub repo documents core functions (estimate_B_ls, estimate_B_mle, fit_G_given_p0) and input formats—this is excellent for reproducibility. The simulation scripts used to produce identifiability diagnostics are included in the Rmd notebooks so others can reproduce the sensitivity experiments hiddenCOMET GitHub and app

Practical guidance for users

• Minimum sampling: simulations show accuracy improves with N (cells per timepoint); practitioners should aim for at least several hundred to a few thousand cells per timepoint when using estimated B from similar contexts Sampling guidance from simulations
• If paired EMT/cell-cycle measurements exist for a given cell line and inducer, always estimate B from those paired data (MLE preferred with multinomial model when N is available) rather than borrowing B from another inducer.
• Beware of parameters hitting bounds; if many rates hit the upper bound consider acquiring earlier sampling points or relaxing time-homogeneous assumptions.

Quick technical audit of key numbers provided

• Example emission matrix used in simulations B0 = [[0.7,0.3,0.05],[0.2,0.5,0.25],[0.1,0.2,0.7]] (columns = E,H,M mapping to G1,S,G2M rows) — this matrix is coherent (columns sum to 1) and plausible biologically (epithelial high in S/G2M depending on cell line) Emission matrices example
• Generator matrix example Q = [[-0.2,0.2,0.0],[0.25,-0.47,0.22],[0.0,0.18,-0.18]] used in simulations; matrix exponentials yield valid P(t) used to generate cell-cycle fractions and multinomial sampling (methodologically sound) Generator matrix example

Visualizations (interactive) — emission matrix and generator sanity checks

Conclusions and confidence

Overall this is a well-constructed, open, and useful methodological contribution: framing EMT inference as a latent-state identification problem tied to a robust observable is conceptually clean and practical, and the software release enables adoption. However, the strongest caveats are biological generalizability (cell lines only), the fixed-B across inducers assumption in primary fits, and the time-homogeneous CTMC simplification which together limit claims about in vivo EMT kinetics. The authors themselves clearly acknowledge these limitations and propose appropriate future directions including richer paired datasets, inducer-specific B estimation, adding growth/death modeling, and higher temporal resolution sampling Authors limitations and future work

If you want I can: 1) run the hiddenCOMET R pipeline on the example data and replicate figures, 2) extend the model to include state-specific growth rates and test identifiability via simulations, or 3) compute inducer-specific B by re-analyzing paired datasets if you provide them. Click Run AI Biology Analysis to proceed.