Skip to content

Limitations

Experimental Design Assumptions

  1. Time points are assigned a priori — The pipeline assumes that the replicate numbering (1, 2, 3) corresponds to specific time points (e.g., 0, 15, 60 min). This must be verified against the actual experimental design before interpreting results.

  2. Triplicates as pseudo-time — The dataset (PXD000901) was not designed as a true dense time course. Using replicate indices as time proxies is an approximation. Ideal usage requires explicitly measured time points.

  3. Cell lines treated independently — Each cell line is modeled independently. No cross-cell-line information sharing is implemented.

Statistical Assumptions

  1. GP stationarity — The Matérn 5/2 kernel assumes a stationary covariance structure. Non-stationary dynamics (e.g., rapid early activation followed by slow decay) may be poorly captured.

  2. Gaussian noise — Both the GP likelihood and the observation model assume Gaussian measurement noise. Heteroscedastic noise is partially handled via the localization-probability-based sigma, but systematic biases are not modeled.

  3. Velocity as GP derivative — Velocity is computed as the derivative of the fitted GP mean, not from a mechanistic ODE model. This is a phenomenological, not mechanistic, definition of phosphosite velocity.

Data Quality Limitations

  1. Missing values — MaxQuant reports 0 for missing values; these are converted to NaN. Sites with many missing values will have poorly constrained GP fits.

  2. Localization probability filtering — Sites with localization probability below 0.75 (default) are filtered. This threshold is empirical and may need adjustment.

  3. No peptide-level deconvolution — Phosphosite intensities are taken as-is from MaxQuant's site-level table. Multiple peptides mapping to the same site are not explicitly deconvolved.

Computational Limitations

  1. Bayesian mode is slow — PyMC NUTS sampling can be slow for large numbers of sites. For exploratory analysis, use the phospho-velocity fit command (scikit-learn GP derivative).

  2. No multi-site modeling — Each phosphosite is modeled independently. Shared kinase regulation is not jointly modeled (network priors are used only as scalar weights, not as structural constraints).