Non-Deterministic results PreProcess.reduce(extend=True) due to internal MCMC sampling

[PalashMendhe]

Describe the bug
When running PreProcess.reduce(extend=True) on the same model multiple times, the function returns different sets of removed reactions. This non-determinism prevents reproducibility for downstream analysis.

To Reproduce
Steps to reproduce the behavior:

1. Go to your root directory.
2. Create a test file.
3. here is a minimal script using e_coli_core.json to recreate the error

import numpy as np
import cobra.io
from dingo.preprocess import PreProcess

def test_extend_nondeterminism():
    ecoli = cobra.io.load_json_model("ext_data/e_coli_core.json")
    np.random.seed(42) 
    preprocessor_A = PreProcess(ecoli.copy(), tol=1e-6, verbose=False)
    removed_A, _ = preprocessor_A.reduce(extend=True)
    
    np.random.seed(42) 
    preprocessor_B = PreProcess(ecoli.copy(), tol=1e-6, verbose=False)
    removed_B, _ = preprocessor_B.reduce(extend=True)

    set_A, set_B = frozenset(removed_A), frozenset(removed_B)
    if set_A != set_B:
        print(f"Reactions removed ONLY in Run A: {sorted(set_A - set_B)}")
        print(f"Reactions removed ONLY in Run B: {sorted(set_B - set_A)}")
if __name__ == "__main__":
    test_extend_nondeterminism()

Expected behavior
Given the same metabolic model and the same algorithm parameters, reduce(extend=True) should always return the same set of removed reactions.

Screenshots

Proposed fix
Add an optional steady_states parameter to reduce().
When steady_states is provided by the caller, the internal sampling step is skipped entirely and the provided matrix is used directly for correlation estimation. When steady_states=None (the default), the current internal sampling behaviour is preserved as a fallback for convenience, but a UserWarning is emitted so the user is informed that results will not be reproducible.

Desktop

• OS: WSL / Ubuntu
• Browser : chrome

[mahicodes28]

Hi @PalashMendhe @TolisChal , I would like to work on this issue.

My plan is to add an optional 'steady states' parameter to
PreProcess.reduce(). If provided, the function will skip the
internal sampling step and use the given matrix for correlation
estimation. Otherwise, the current behavior will remain with a
UserWarning.

Please let me know if this approach aligns with the intended design.