Bayesian_Hybrid_OU_Branching_Precision_Medicine: Bayesian Hybrid OU-Branching Precision Medicine v2.0.0

This release provides the revised and fully reproducible analysis workflow supporting the PLOS ONE manuscript:

“Stage-specific evolutionary dynamics in pediatric leukemia inferred using Bayesian stochastic modeling.”

Major updates

• Added a complete end-to-end workflow for preprocessing sparse and irregular longitudinal KMT2A-rearranged leukemia data.
• Added simulation-based benchmarking of Ornstein–Uhlenbeck parameter recovery under empirical patient sampling schedules.
• Added comparison of Bayesian, maximum-likelihood, and method-of-moments estimators.
• Added scenario-stratified robustness analyses examining the effects of observation number, temporal spacing, and generating parameter combinations.
• Added hierarchical Bayesian OU inference using the exact transition density for irregularly spaced observations.
• Added posterior summaries for group-level mean-reversion and diffusion parameters.
• Added patient-specific equilibrium-state estimates.
• Added posterior predictive trajectory generation and patient-level predictive summaries.
• Added patient-level and group-level model diagnostic analyses.
• Added all scripts required to reproduce the main and supplementary figures and tables.

New main outputs

• Figure 1: Study and modeling workflow schematic
• Figure 2: Simulation-based OU calibration benchmark
• Figure 3: Group-level posterior OU parameter estimates
• Figure 4: Representative patient-level posterior predictive dynamics
• Figure 5: Cohort-level model diagnostics

New supplementary outputs

• Figure S1: Patient sampling characteristics
• Figure S2: Posterior predictive dynamics for all patients
• Tables S1–S12: Sampling summaries, simulation results, posterior summaries, pairwise posterior probabilities, patient-level predictive summaries, and model diagnostics
• Supporting Data S1–S14: Processed data, simulation outputs, posterior trace, posterior summaries, convergence diagnostics, and run metadata

Reproducibility improvements

• Reorganized scripts in numerical execution order.
• Added standardized input, output, and manuscript-file directories.
• Added analysis-ready processed datasets and supporting-data files.
• Added the full posterior trace in NetCDF format.
• Added MCMC convergence summaries and computational run metadata.
• Added documentation linking each script to its inputs, outputs, figures, tables, and supporting datasets.
• Added environment and dependency specifications for Python 3.12, PyMC, ArviZ, NumPy, SciPy, pandas, Matplotlib, NetCDF4, and related packages.

Bayesian model configuration

The final hierarchical Bayesian analysis uses:

• 4 MCMC chains
• 2,500 tuning iterations per chain
• 2,000 retained posterior draws per chain
• NUTS sampling
• target_accept = 0.97

Data-processing updates

• Converted sampling time from days to years using Day / 365.25.
• Collapsed duplicate patient-day measurements by arithmetic mean.
• Preserved irregular observation intervals.
• Avoided temporal interpolation and imputation in the final hierarchical Bayesian analysis.

Archived release

A permanent archived version of the complete computational workflow is available through Zenodo:

10.5281/zenodo.21447281

Citation

Please cite the associated PLOS ONE article when using this repository, its code, or its derived outputs.