EnKF with Knowledge Transfer for CHO Bioprocess Modelling

Reproducible research code accompanying the paper:

[Paper title] [Authors] [Journal, Year]

This repository implements an Ensemble Kalman Filter (EnKF) for dual state and parameter estimation in Chinese Hamster Ovary (CHO) cell culture bioprocesses, with cross-condition knowledge transfer.

---

Overview

The EnKF is applied to six CHO cell culture datasets covering two cell lines (T127 and GS46), multiple bioreactor scales, and different feeding strategies. The workflow:

1. Integrates the bioreactor volume ODE
2. Runs a nominal forward simulation with literature parameters (Kotidis et al. 2019)
3. Tunes ensemble size by sweeping RMSE across sizes [10, 25, 50, 75]
4. Performs long-term state prediction using the best ensemble size
5. Tests robustness with an irregular 48/72 h measurement schedule
6. Quantifies prior width sensitivity and ±20% mean parameter sensitivity

---

Repository Structure

BiotechBioeng/
│
├── cho_enkf/                   # Python package
│   ├── config.py               # All constants: RUN_NAME, paths, parameters, noise
│   ├── data_loader.py          # Load Excel datasets
│   ├── model.py                # Volume integration, kinetic model (ODE step)
│   ├── enkf.py                 # EnKF classes + all runner functions
│   ├── analysis.py             # R², convergence tables, correlation matrix
│   ├── plotting.py             # All publication-quality figure functions
│   └── io_utils.py             # Pickle save/load, path construction
│
├── scripts/                    # Numbered execution pipeline
│   ├── 01_ensemble_tuning.py   # Load data → run EnKF sweep → save pkl
│   ├── 02_longterm_pred.py     # Long-term forecasting with best ensemble size
│   ├── 03_irregular.py         # EnKF with 48/72 h irregular measurements
│   ├── 04_sensitivity.py       # Prior width + ±% parameter sensitivity
│   └── 05_comparisons.py       # EnKF vs reparametrised model comparison
│
├── data/                       # CHO experimental datasets (Excel)
│   ├── CHO_T127_flask_PMJ.xlsx
│   ├── CHO_T127_SNS_36.5.xlsx
│   ├── CHO_T127_SNS_32.xlsx
│   ├── CHO_GS46_F_C_Inv.xlsx
│   ├── CHO_GS46_F_all.xlsx
│   └── CHO_GS46_F_all_pl40.xlsx
│
├── results/                    # Generated outputs (gitignored)
│   └── {RUN_NAME}/
│       ├── run_notes.txt       # Auto-created on first run — fill in manually
│       ├── 01_ensemble_tuning/
│       │   ├── pkl/
│       │   └── figures/
│       ├── 02_longterm_pred/
│       │   ├── pkl/
│       │   └── figures/
│       ├── 03_irregular/
│       │   ├── pkl/
│       │   └── figures/
│       ├── 04_sensitivity/
│       │   ├── pkl/
│       │   └── figures/
│       └── 05_comparisons/
│           ├── pkl/
│           └── figures/
│
├── original.ipynb              # Original monolithic notebook (reference only)
├── pyproject.toml
├── poetry.lock
└── README.md

---

Quick Start

1. Install dependencies

# Install Poetry (if not already installed)
pip install poetry

# Create virtual environment and install all dependencies
poetry install

# Activate the environment
poetry shell
# OR: source .venv/Scripts/activate   (Windows)
#     source .venv/bin/activate        (Linux/macOS)

2. Set the run name (optional)

Edit cho_enkf/config.py:

RUN_NAME = "run_v1"   # all outputs go to results/run_v1/

Change this string to version a new experiment. Old results are preserved.

3. Run the pipeline

Each script saves results to its own subfolder. Set LOAD_FROM_PKL = True (default) to skip recomputation and regenerate figures only.

# Step 1 (~2–4 h): ensemble tuning
poetry run python scripts/01_ensemble_tuning.py

# Step 2 (~30–60 min): long-term prediction with best ensemble size
poetry run python scripts/02_longterm_pred.py

# Step 3 (~30–60 min): irregular measurement schedule
poetry run python scripts/03_irregular.py

# Step 4 (~2–4 h): sensitivity analyses (prior width + ±% params)
poetry run python scripts/04_sensitivity.py

# Step 5 (< 5 min): EnKF vs reparametrised model comparison
poetry run python scripts/05_comparisons.py

Tip: Each script has LOAD_FROM_PKL = True at the top. With this set, scripts load saved pkl files and skip all computation — useful for regenerating figures without re-running the EnKF.

4. Find outputs

results/{RUN_NAME}/
├── run_notes.txt               # Fill in to document what changed in this run
├── 01_ensemble_tuning/pkl|figures/
├── 02_longterm_pred/pkl|figures/
├── 03_irregular/pkl|figures/
├── 04_sensitivity/pkl|figures/
└── 05_comparisons/pkl|figures/

---

Datasets

Dataset	Cell Line	Condition
CHO_T127_flask_PMJ	T127 (Cell Line A)	Shake flask, 36.5°C
CHO_T127_SNS_36.5	T127 (Cell Line A)	Bioreactor, 36.5°C
CHO_T127_SNS_32	T127 (Cell Line A)	Bioreactor, 32°C
CHO_GS46_F_C_Inv	GS46 (Cell Line B)	Feed C
CHO_GS46_F_all	GS46 (Cell Line B)	Feed U
CHO_GS46_F_all_pl40	GS46 (Cell Line B)	Feed U +40%

Each Excel file contains three sheets: schedule (feed schedule), feed (feed concentrations), exp_meas (measured state variables ± std).

---

State Variables

Symbol	Description	Unit
Xv	Viable cell density	cell L⁻¹
mAb	Monoclonal antibody titre	mg L⁻¹
Glc	Glucose	mM
Amm	Ammonia	mM
Gln	Glutamine	mM
Lac	Lactate	mM
Glu	Glutamate	mM
Asn	Asparagine	mM

---

Key Configuration (cho_enkf/config.py)

Variable	Description
RUN_NAME	Experiment version tag; controls output folder
TUNING_ENSEMBLE_SIZES	Sizes swept in Step 1 (default [10, 25, 50, 75])
BEST_ENSEMBLE_SIZES	Best size per dataset (set after reviewing Step 1 results)
MEAN_PARAMETERS	Nominal model parameters from Kotidis et al. 2019
PARAMETERS_ENSEMBLE_COVARIANCE	Prior width for each parameter
DATASET_NOISE_VARIANCES	Process (Q) and observation (R) noise per dataset
KQ_DICT / KR_DICT	Scaling factors for Q and R matrices
PRIOR_WIDTH_SCALES	Scales tested in sensitivity analysis
PARAM_SENS_PERTURBATIONS	Fractions for ±% sensitivity (default [0.10, 0.20, 0.30])

---

Package API Summary

cho_enkf.enkf

• run_enkf_with_tuning(...) — sweep ensemble sizes
• enkf_long_pred_best_ensemble_size(...) — long-term prediction
• run_pipeline_irregular_48_72(...) — irregular measurement schedule
• run_enkf_with_mean_params(...) — parameter sensitivity runs

cho_enkf.analysis

• compute_r2_table(...) — R² for all datasets
• compute_overall_convergence_table(...) — parameter convergence %
• get_posterior_param_matrix(...) — posterior ensemble for correlation

cho_enkf.plotting

• plot_rmse_variance_and_computation_time_all(...) — tuning figure
• overlay_T127_subplots_with_errorbars(...) — T127 comparison
• overlay_gs46_subplots_with_errorbars(...) — GS46 comparison
• plot_longterm_pred_ensemble_simulation_errorbar(...) — long-term pred
• plot_parameter_comparison_across_datasets(...) — cross-dataset params
• plot_posterior_param_correlation(...) — correlation heatmap
• plot_prior_width_sensitivity_rmse(...) — prior width RMSE bar
• plot_param_sensitivity_comparison(...) — ±20% sensitivity

---

Dependencies

Package	Purpose
numpy, scipy	Numerical computation, ODE integration
pandas	Data loading and tabular analysis
matplotlib, seaborn	Plotting
openpyxl	Reading Excel files
tqdm	Progress bars
jupyter, ipykernel	Interactive exploration (optional)

---

Nominal Model

Parameters from Kotidis et al. (2019) for CHO-T127 shake flask cultures. The model describes growth, death, and metabolite dynamics via Monod-type kinetics, with ammonia and lactate inhibition, and a full yield-based metabolic network.

---

Citation

If you use this code, please cite:

@article{[key],
  title   = {[Title]},
  author  = {[Authors]},
  journal = {[Journal]},
  year    = {[Year]},
  doi     = {[DOI]}
}

---

License

[License]