diff --git a/modelpruner/core/model_pruner.py b/modelpruner/core/model_pruner.py
index 806ac249b937de7ef2666b2153cd6d53efa2fcec..fabf46491f0c88ffda6e3e825b93a9b4ab9bc525 100644
--- a/modelpruner/core/model_pruner.py
+++ b/modelpruner/core/model_pruner.py
@@ -4,52 +4,110 @@ import sys
import os
import re
import time
+import math
import copy
import numpy as np
import multiprocessing as mp
-import pickle
+import json
from modelpruner.models.fba_model import FbaModel
from modelpruner.core.protected_parts import ProtectedParts
global _fba_model
+global _pf
+
+
+def init_worker(fba_model, pf):
+ global _fba_model
+ global _pf
+
+ _fba_model = fba_model.copy()
+ _pf = pf
+
+
+def worker(rid):
+ global _fba_model
+ global _pf
+
+ return _fba_model.fva_pf_flux_ranges(_pf, [rid])
class ModelPruner:
- def __init__(self, sbml_fname, protected_parts, reduced_fname=None, resume=False):
+ def __init__(self, sbml_fname, protected_parts, resume=False):
- self.tolerance = 1e-8 # later accessible by set_params() ?
- self.cpus = int(os.cpu_count()*.8)
+ self.tolerance = 1e-7 # later accessible by set_params() ?
+ self.cpus = int(os.cpu_count() * 0.8)
- if reduced_fname is None:
- self.reduced_fname = re.sub(r'.xml$', '_red.xml', sbml_fname)
self._snapshot_sbml = re.sub(r'.xml$', '_snapshot.xml', sbml_fname)
- self._snapshot_pkl = re.sub(r'.xml$', '_snapshot.pkl', sbml_fname)
+ self._snapshot_json = re.sub(r'.xml$', '_snapshot.json', sbml_fname)
- if resume is True:
+ if resume is True and os.path.exists(self._snapshot_sbml):
sbml_fname = self._snapshot_sbml
self.fba_model = FbaModel(sbml_fname)
self.nrp = ProtectedParts(protected_parts)
- self.protected_sids = self.nrp.initial_protected_sids
-
- if resume is True and os.path.exists(self._snapshot_pkl):
- with open(self._snapshot_pkl, 'rb') as f:
- self.protected_rids = pickle.load(f)
- print(f'snapshot restored at {len(self.fba_model.rids):4d} remaining reactions '
- f'from: {self._snapshot_pkl}')
+ self.protected_sids = self.nrp.protected_sids
+
+ if resume is True and os.path.exists(self._snapshot_json):
+ with open(self._snapshot_json, 'r') as f:
+ snapshot_data = json.load(f)
+ self.protected_rids = set(snapshot_data['protected_rids'])
+ print(f'snapshot restored at {len(self.fba_model.rids):4d} remaining reactions '
+ f'from: {self._snapshot_json}')
else:
- self.protected_rids = self.get_total_protected_rids(self.nrp.initial_protected_rids)
-
- self.fba_model.update_flux_bounds(self.nrp.overwrite_bounds)
+ self.protected_rids = self.nrp.protected_rids
+ # self.protected_rids = self.get_total_protected_rids()
- self.base_obj_dir = self.fba_model.obj_dir
- self.base_flux_bounds = self.fba_model.flux_bounds.copy()
- self.base_coefs = self.fba_model.obj_coefs.copy()
+ # overwrite the fba model bounds by general bound of protected parts (priot to LpProblem instantiation)
+ self.fba_model.update_model_flux_bounds(self.nrp.overwrite_bounds)
+ self.fba_model.update_model_reversible(self.get_pp_reversible_reactions())
+ print('determine optimal values for protected functions')
self.set_function_optimum()
+ print('Check consistency of protected parts')
self.check_protected_parts()
+ def get_total_protected_rids(self):
+ # identify reactions that need protection due to protected metabolites
+ # over and above the protected rids provided in protected parts
+ sid_mask = np.array([True if sid in self.protected_sids else False
+ for sid in self.fba_model.sids])
+ psf_mask = np.any(self.fba_model.s_mat[sid_mask] != 0, axis=0)
+ protected_sid_rids = set(self.fba_model.rids[psf_mask])
+
+ protected_rids = self.nrp.protected_rids.union(protected_sid_rids)
+ return protected_rids
+
+ def get_pp_reversible_reactions(self):
+ """Identify reversible reactions from protected parts
+
+ :return: reaction ids from protected parts with negative fluxes
+ :rtype: set of str
+ """
+ rev_pp_rids = set()
+ for rid, bounds in self.nrp.overwrite_bounds.items():
+ for idx in [0, 1]:
+ if math.isfinite(bounds[idx]):
+ rev_pp_rids.add(rid)
+ for pf in self.nrp.protected_functions.values():
+ for rid, bounds in pf.overwrite_bounds.items():
+ for idx in [0, 1]:
+ if math.isfinite(bounds[idx]) and bounds[idx] < 0.0:
+ rev_pp_rids.add(rid)
+ return rev_pp_rids
+
+ def set_function_optimum(self):
+ """using FBA to set optimal value for protected functions"""
+ for pf in self.nrp.protected_functions.values():
+ res = self.fba_model.fba_pf_optimize(pf)
+ pf.optimal = res['fun']
+ pf.fba_success = res['success']
+ if hasattr(pf, 'target_frac'):
+ if pf.obj_dir == 'maximize':
+ pf.set_target_lb(pf.target_frac * pf.optimal)
+ else:
+ pf.set_target_ub(pf.target_frac * pf.optimal)
+
def check_protected_parts(self):
extra_p_rids = self.protected_rids.difference(set(self.fba_model.rids))
@@ -74,123 +132,108 @@ class ModelPruner:
print('reactions in protected functions that are not protected:', pf_unprotected_rids)
blocked_rids = self.reaction_types_fva(self.fba_model.rids)['blocked_rids']
- blocked_p_rids = self.nrp.initial_protected_sids.intersection(blocked_rids)
+ blocked_p_rids = self.nrp.protected_sids.intersection(blocked_rids)
blocked_pf_rids = pf_rids.intersection(blocked_rids)
if len(blocked_p_rids) > 0:
print('protected reactions that are blocked (no flux):', blocked_p_rids)
if len(blocked_pf_rids) > 0:
print('reactions in protected functions that are blocked (no flux):', blocked_pf_rids)
- def prune(self, snapshot_freq=None):
- print(f'initial S-matrix: {self.fba_model.s_mat.shape}; '
- f' protected reactions: {len(self.protected_rids)}:', self.protected_rids)
- self.remove_parallel_reactions()
- print('protected functions feasibility:', self.check_pf_feasibility())
-
- if type(snapshot_freq) is int:
- next_snapshot = max(0, len(self.fba_model.rids) - snapshot_freq)
- else:
- next_snapshot = None
-
- while len(self.protected_rids) < self.fba_model.s_mat.shape[1]:
- free_rids = list(set(self.fba_model.rids).difference(self.protected_rids))
- reaction_types = self.reaction_types_fva(free_rids)
- if len(reaction_types['blocked_rids']) > 0:
- self.fba_model.remove_reactions(reaction_types['blocked_rids'], self.protected_rids)
- print(f"{len(reaction_types['blocked_rids'])} blocked reaction(s) removed:",
- reaction_types['blocked_rids'])
- if len(reaction_types['essential_rids']) > 0:
- self.protected_rids |= reaction_types['essential_rids']
- print(f"{len(reaction_types['essential_rids'])} reaction(s) added to protected reactions:",
- reaction_types['essential_rids'])
-
- print(f"{len(reaction_types['candidate_rids_sorted'])} candidates:",
- reaction_types['candidate_rids_sorted'][:4], '...')
- # try to remove one reaction while still satisfying all protected functions
- feasible = False
- essential_rids = set()
- for rid in reaction_types['candidate_rids_sorted']:
- orig_flux_bounds = self.fba_model.get_flux_bounds(rid)
- self.fba_model.update_flux_bounds({rid: [0.0, 0.0]})
- feasible = self.check_pf_feasibility()
- self.fba_model.update_flux_bounds({rid: orig_flux_bounds})
- if feasible:
- self.fba_model.remove_reactions({rid}, self.protected_rids)
- print(f'{rid} removed from the model')
- break
- else:
- essential_rids.add(rid)
- if len(essential_rids) > 0:
- self.protected_rids |= essential_rids
- print(f"{len(essential_rids)} reaction(s) added to protected reactions:", essential_rids)
-
- if feasible is False:
- print('no more reactions to remove')
-
- # store intermediate results (snapshots)
- if (next_snapshot is not None) and (len(self.fba_model.rids) < next_snapshot):
- self.export_pruned_model(self._snapshot_sbml)
- with open(self._snapshot_pkl, 'wb') as f:
- pickle.dump(self.protected_rids, f)
- next_snapshot = max(0, len(self.fba_model.rids) - snapshot_freq)
- print('snapshot data saved.')
-
- print(f'S-matrix: {self.fba_model.s_mat.shape}; '
- f'protected reactions: {len(self.protected_rids)}')
-
- # check if blocked reactions exist in the finally pruned model
- blocked_rids = self.reaction_types_fva(self.fba_model.rids)['blocked_rids']
- if len(blocked_rids) > 0:
- print(f'{len(blocked_rids)} blocked reaction(s) in pruned model', blocked_rids)
-
- # remove any snapshot files
- for fname in [self._snapshot_sbml, self._snapshot_pkl]:
- if os.path.exists(fname):
- os.remove(fname)
+ def identify_parallel_reactions(self):
+ """Identify parallel reactions having same stoichiometry (considering reverse).
- def init_worker(self):
- global _fba_model
- _fba_model = copy.deepcopy(self.fba_model)
+ protected reaction are preserved.
+ reversible reactions have higher priority
- @staticmethod
- def fva_single_rids(rid):
- return _fba_model.fva_single_rid(rid)
+ :return: reaction ids of parallel reactions
+ :rtype: set of str
+ """
+ drop_rids = set()
+ for group_rids in self.fba_model.get_parallel_rids():
+ protected = set()
+ rid_rev = None
+ for rid in group_rids:
+ if rid in self.protected_rids:
+ protected.add(rid)
+ if rid in self.fba_model.rids_rev:
+ rid_rev = rid
+ if len(protected) > 0:
+ drop_group_rids = set(group_rids) - protected
+ elif rid_rev is not None:
+ drop_group_rids = set(group_rids) - {rid_rev}
+ else:
+ drop_group_rids = set(group_rids[1:])
+ drop_rids |= drop_group_rids
+ return drop_rids
def reaction_types_fva(self, free_rids):
+ """Idenify reaction types using FVA applied to all protected functions
+
+ first part: run FVA across all conditions
+ :param free_rids: reaction ids used for FVA
+ :type free_rids: list of strings
+ :return: rid, min and max flux values
+ :rtype: dict
+ """
print(time.strftime("%H:%M:%S", time.localtime()))
+
n_pf = len(self.nrp.protected_functions)
flux_min_pfs = np.zeros((n_pf, len(free_rids)))
flux_max_pfs = np.zeros((n_pf, len(free_rids)))
+ res = {}
processes = min(self.cpus, len(free_rids))
for idx, pf in enumerate(self.nrp.protected_functions.values()):
- self.set_temp_func_conditions(pf)
- if processes > 20:
+ if processes > 100:
+ frid2idx = {rid: idx for idx, rid in enumerate(free_rids)}
# place to split fva for free_rids over several
- rid2idx = {rid: idx for idx, rid in enumerate(free_rids)}
- pool = mp.Pool(processes, initializer=self.init_worker)
- for rid, res in pool.imap_unordered(self.fva_single_rids, free_rids):
- flux_min_pfs[idx, rid2idx[rid]] = res['min']
- flux_max_pfs[idx, rid2idx[rid]] = res['max']
- pool.close()
- pool.join()
+ with mp.Pool(processes, initializer=init_worker, initargs=(self.fba_model, pf)) as pool:
+ for res in pool.imap_unordered(worker, free_rids):
+ rid = res['rids'][0]
+ flux_min_pfs[idx, frid2idx[rid]] = res['min'][0]
+ flux_max_pfs[idx, frid2idx[rid]] = res['max'][0]
else:
- res = self.fba_model.fva_optimize(free_rids, fract_of_optimum=0.0)
- flux_min_pfs[idx] = res['min']
- flux_max_pfs[idx] = res['max']
-
- self.restore_base_conditions()
-
- # if res['success'] is False:
- # print('FVA unsuccessful')
- # return {'blocked_rids': [], 'essential_rids': [], 'candidate_rids_sorted': []}
-
+ # print(f'\nCheck FVA for pf {pf.obj_id} with {len(free_rids)} free variables')
+ res = self.fba_model.fva_pf_flux_ranges(pf, free_rids)
+ if res['success'] is True:
+ flux_min_pfs[idx] = res['min']
+ flux_max_pfs[idx] = res['max']
+ else:
+ print(f'FBA for condition {pf.obj_id} is not feasible')
+ # TODO: can this happen? how best to continue?
+ flux_min_pfs[idx] = np.nan
+ flux_max_pfs[idx] = np.nan
+ break
sys.stdout.write('\rFVA[{}{}] {:3.0f}%'.format(
'=' * int(idx+1), ' ' * int(n_pf - idx-1), (idx+1)/n_pf*100))
sys.stdout.flush()
- print()
+ if res['success'] is False:
+ reaction_types = {'blocked_rids': set(), 'essential_rids': set(free_rids),
+ 'candidate_rids_sorted': []}
+ else:
+ reaction_types = self.classify_reactions(free_rids, flux_min_pfs, flux_max_pfs)
+ return reaction_types
+
+ def classify_reactions(self, rids, flux_min_pfs, flux_max_pfs):
+ """Classify reaction types based on FVA results.
+
+ determine flux span (min/max flux) for each reaction across all conditions
+ Identify blocked, essential and deletion candidate reactions based on flux span
+ blocked reactions: min and max of flux span is zero
+ essential reactions: min and max either > 0 or < 0
+ deletion candidates: flux span support zero (0) flux
+
+ :param rids: reaction ids used in FVA
+ :type rids: list of strings
+ :param flux_min_pfs: minimum fluxes per condition
+ :type flux_min_pfs: 2D ndarray of floats (reactions x conditions)
+ :param flux_max_pfs: maximum fluxes per condition
+ :type flux_max_pfs: 2D ndarray of floats (reactions x conditions)
+ :return: rid, min and max flux values
+ :rtype: dict
+ """
# analyze flux ranges per reaction, aggregate fluxes
flux_min_pfs[np.abs(flux_min_pfs) < self.tolerance] = 0.0
flux_max_pfs[np.abs(flux_max_pfs) < self.tolerance] = 0.0
@@ -199,116 +242,104 @@ class ModelPruner:
# blocked reactions do not carry any flux in all conditions
mask_b = np.all(np.vstack((flux_min == 0, flux_max == 0)), axis=0)
- blocked_rids = set(np.array(free_rids)[mask_b])
+ blocked_rids = set(np.array(rids)[mask_b])
# essential reactions carry either only forward or reverse flux, exluding zero flux
mask_e1 = np.all(np.vstack((flux_min < 0, flux_max < 0)), axis=0)
mask_e2 = np.all(np.vstack((flux_min > 0, flux_max > 0)), axis=0)
mask_e = np.any(np.vstack((mask_e1, mask_e2)), axis=0)
- essential_rids = set(np.array(free_rids)[mask_e])
+ essential_rids = set(np.array(rids)[mask_e])
# reactions that support zero flux are considered candiates for removal
# here such reactions are sorted from smallest to largest flux span
mask_c = np.all(np.vstack((np.logical_not(mask_b), np.logical_not(mask_e))), axis=0)
flux_span = flux_max - flux_min
ind = np.argsort(flux_span)
- candidate_rids_sorted = list(np.array(free_rids)[ind][np.array(mask_c)[ind]])
+ candidate_rids_sorted = list(np.array(rids)[ind][np.array(mask_c)[ind]])
- return {'blocked_rids': blocked_rids, 'essential_rids': essential_rids,
- 'candidate_rids_sorted': candidate_rids_sorted}
+ reaction_types = {'blocked_rids': blocked_rids, 'essential_rids': essential_rids,
+ 'candidate_rids_sorted': candidate_rids_sorted}
- def remove_parallel_reactions(self):
+ return reaction_types
- parallel_rid_groups = self.fba_model.get_identical_columns()
+ def init_worker(self):
+ global _fba_model
+ _fba_model = copy.deepcopy(self.fba_model)
- drop_rids = set()
- for group_rids in parallel_rid_groups:
- protected = set()
- reversible = None
- for rid in group_rids:
- if rid in self.protected_rids:
- protected.add(rid)
- if self.fba_model.reversible[self.fba_model.rid2idx[rid]]:
- reversible = rid
- if len(protected) > 0:
- drop_group_rids = set(group_rids) - protected
- elif reversible is not None:
- drop_group_rids = set(group_rids) - {reversible}
- else:
- drop_group_rids = set(group_rids[1:])
- drop_rids |= drop_group_rids
+ @staticmethod
+ def fva_single_rids(rid):
+ return _fba_model.fva_single_rid(rid)
- self.fba_model.remove_reactions(drop_rids, self.protected_sids)
- print(f'{len(drop_rids)} parallel reaction(s) dropped:', drop_rids)
+ def prune(self, snapshot_freq=None):
+ print(f'initial S-matrix: {self.fba_model.s_mat.shape}; '
+ f' protected reactions: {len(self.protected_rids)}:', self.protected_rids)
+ drop_rids = self.identify_parallel_reactions().difference(self.protected_rids)
+ if len(drop_rids) > 0:
+ self.fba_model.remove_reactions(drop_rids)
+ print(f'{len(drop_rids)} parallel reaction(s) dropped:', drop_rids)
- def set_function_optimum(self):
- """using FBA to set optimal value for protected functions"""
- for obj_id, pf in self.nrp.protected_functions.items():
- self.set_temp_func_conditions(pf)
- res = self.fba_model.fba_optimize(short_result=True)
- self.restore_base_conditions()
+ feasible = np.array([self.fba_model.check_pf_feasibility(pf) for pf in self.nrp.protected_functions.values()])
+ print('protected functions feasibility:', np.all(feasible))
- pf.optimal = res['fun']
- pf.fba_success = res['success']
- if hasattr(pf, 'target_frac'):
- if pf.obj_dir == 'minimize':
- pf.set_target_ub(pf.target_frac * pf.optimal)
- else:
- pf.set_target_lb(pf.target_frac * pf.optimal)
+ next_snapshot = 0
+ if type(snapshot_freq) is int:
+ next_snapshot = max(0, len(self.fba_model.rids) - snapshot_freq)
- def get_total_protected_rids(self, protected_rids):
- # identify reactions that need protection due to protected metabolites
- # over and above the protected rids provided in protected parts
- sid_mask = np.array([True if sid in self.protected_sids else False
- for sid in self.fba_model.sids])
+ while len(self.protected_rids) < self.fba_model.s_mat.shape[1]:
+ free_rids = list(set(self.fba_model.rids).difference(self.protected_rids))
+ reaction_types = self.reaction_types_fva(free_rids)
+ print()
+ blocked_rids = reaction_types['blocked_rids']
+ essential_rids = reaction_types['essential_rids']
+ candidate_rids = reaction_types['candidate_rids_sorted']
+ if len(blocked_rids) > 0:
+ self.fba_model.remove_reactions(blocked_rids)
+ print(f"{len(blocked_rids)} blocked reaction(s) removed:", blocked_rids)
+ print(f"{len(candidate_rids)} candidates:", candidate_rids[:4], '...')
- psf_mask = np.any(self.fba_model.s_mat[sid_mask] != 0, axis=0)
- protected_sid_rids = set(self.fba_model.rids[psf_mask])
+ # try to remove one reaction while still satisfying all protected functions
+ for rid in candidate_rids:
+ # print('check feasibility of', rid)
+ feasible = True
+ for pf in self.nrp.protected_functions.values():
+ feasible = self.fba_model.check_pf_feasibility(pf, zero_flux_rid=rid)
+ # print(f'feasibility of {rid} is {feasible}')
+ if feasible is False:
+ essential_rids.add(rid)
+ break
+ if feasible is True:
+ self.fba_model.remove_reactions({rid})
+ print(f'{rid} removed from the model')
+ break
- # identify rids in protected functions and ensure they are protected
- pf_rids = set()
- for pf in self.nrp.protected_functions.values():
- pf_rids |= {rid for rid in pf.objective}
- pf_rids |= {rid for rid in pf.overwrite_bounds}
+ if len(essential_rids) > 0:
+ self.protected_rids |= essential_rids
+ print(f"{len(essential_rids)} reaction(s) added to protected reactions:", essential_rids)
- protected_rids |= protected_sid_rids
- protected_rids |= pf_rids
- return protected_rids
+ if feasible is False:
+ print('no more reactions to remove')
- def check_pf_feasibility(self):
- """check that current model still satisfies the protected functions
- """
- feasible = True
- for pf in self.nrp.protected_functions.values():
- self.set_temp_func_conditions(pf)
- res = self.fba_model.fba_optimize(short_result=True)
- self.restore_base_conditions()
- if res['success'] is False:
- feasible = False
- break
- if pf.obj_dir == 'maximize':
- if res['fun'] < pf.target_lb:
- feasible = False
- break
- else:
- if res['fun'] > pf.target_ub:
- feasible = False
- break
- return feasible
+ # store intermediate results (snapshots)
+ if len(self.fba_model.rids) < next_snapshot:
+ self.export_pruned_model(self._snapshot_sbml)
+ snapshot_data = {'protected_rids': list(self.protected_rids)}
+ with open(self._snapshot_json, 'w') as f:
+ json.dump(snapshot_data, f)
+ next_snapshot = max(0, len(self.fba_model.rids) - snapshot_freq)
+ print('snapshot data saved.')
- def set_temp_func_conditions(self, pf):
- self.base_obj_dir = self.fba_model.obj_dir
- self.base_flux_bounds = self.fba_model.flux_bounds.copy()
- self.base_coefs = self.fba_model.obj_coefs.copy()
+ print(f'S-matrix: {self.fba_model.s_mat.shape}; '
+ f'protected reactions: {len(self.protected_rids)}')
- self.fba_model.obj_dir = pf.obj_dir
- self.fba_model.update_flux_bounds(pf.overwrite_bounds)
- self.fba_model.obj_coefs = self.fba_model.set_objective_coefficients(pf.objective)
+ # check if blocked reactions exist in the finally pruned model
+ # blocked_rids = self.reaction_types_fva(self.fba_model.rids)['blocked_rids']
+ # if len(blocked_rids) > 0:
+ # print(f'{len(blocked_rids)} blocked reaction(s) in pruned model', blocked_rids)
- def restore_base_conditions(self):
- self.fba_model.obj_dir = self.base_obj_dir
- self.fba_model.flux_bounds = self.base_flux_bounds.copy()
- self.fba_model.obj_coefs = self.base_coefs.copy()
+ # remove any snapshot files
+ # for fname in [self._snapshot_sbml, self._snapshot_json]:
+ # if os.path.exists(fname):
+ # os.remove(fname)
def get_reduction_status(self):
""" Return status in model reduction process
@@ -327,7 +358,7 @@ class ModelPruner:
status = self.get_reduction_status()
print("reactions removed/remaining/protected: "
- f"{status['n_full_rids' ] - status['n_rids']:4d}/{status['n_rids']:4d}/"
+ f"{status['n_full_rids' ] - status['n_rids']:4d}/{status['n_rids']:4d}"
f"/{len(self.protected_rids):4d}; dof: {status['dof']:3d}")
def export_pruned_model(self, pruned_sbml):
@@ -336,3 +367,6 @@ class ModelPruner:
print('pruned model exported to', pruned_sbml)
else:
print('pruned model could not be exported')
+
+ def __del__(self):
+ del self.fba_model
diff --git a/modelpruner/core/protected_parts.py b/modelpruner/core/protected_parts.py
index b35343a52aac00ab0ba6a884bc3d32a90ae7b415..4f2180d733bc3fc555e347e2f15c587af65a909d 100644
--- a/modelpruner/core/protected_parts.py
+++ b/modelpruner/core/protected_parts.py
@@ -23,8 +23,8 @@ class ProtectedParts:
assert type(protected_parts) == dict, 'expected filename or a dict'
nrp = protected_parts
- self.initial_protected_rids = set(nrp.get('reactions', []))
- self.initial_protected_sids = set(nrp.get('metabolites', []))
+ self.protected_rids = set(nrp.get('reactions', []))
+ self.protected_sids = set(nrp.get('metabolites', []))
if 'bounds' in nrp:
num_cols = ['fbcLb', 'fbcUb']
nrp['bounds'][num_cols] = nrp['bounds'][num_cols].applymap(
@@ -82,7 +82,7 @@ class ProtectedFunction:
self.obj_dir = row['type']
for record in sbmlxdf.misc.record_generator(row['fluxObjectives']):
params = sbmlxdf.misc.extract_params(record)
- self.objective[params['reac']] = params['coef']
+ self.objective[params['reac']] = float(params['coef'])
if math.isfinite(row['fraction']):
self.target_frac = row['fraction']
if math.isfinite(row['fbcLb']):
diff --git a/modelpruner/models/fba_model.py b/modelpruner/models/fba_model.py
index 1613be9ee1d4611b857a9e716e276987d0abd044..90d43b038870496b4d5184335a90bd56eb1d8d8f 100644
--- a/modelpruner/models/fba_model.py
+++ b/modelpruner/models/fba_model.py
@@ -26,268 +26,72 @@ class FbaModel:
model_dict = sbml_model.to_df()
df_reactions = model_dict['reactions']
df_species = model_dict['species']
- df_fbc_objectives = model_dict['fbcObjectives']
- self.full_model_shape = {'n_full_rids': len(df_reactions),
- 'n_full_sids': len(df_species)}
self.rids = df_reactions.index.to_numpy()
self.rid2idx = {rid: idx for idx, rid in enumerate(self.rids)}
self.sids = df_species.index.to_numpy()
self.sid2idx = {sid: idx for idx, sid in enumerate(self.sids)}
- self.reversible = df_reactions['reversible'].to_numpy().copy()
+ self.full_model_shape = {'n_full_rids': len(self.rids), 'n_full_sids': len(self.sids)}
self.flux_bounds = np.vstack((df_reactions['fbcLb'].to_numpy(),
df_reactions['fbcUb'].to_numpy()))
+ self.reversible = df_reactions['reversible'].to_numpy().copy()
+ for idx in range(len(self.reversible)):
+ if self.flux_bounds[0, idx] < 0 and np.logical_not(self.reversible[idx]):
+ self.reversible[idx] = True
+ print(f'reaction {idx} set to reversible')
+ self.rids_rev = self.rids[self.reversible]
+ self.rid2idx_rev = {rid: idx + len(self.rids) for idx, rid in enumerate(self.rids_rev)}
self.s_mat = sbml_model.get_s_matrix().to_numpy()
# TODO: create LP Problem only once (but what is impact on deepcopy()
# function to assign new LP problem
- # TODO: delete rows / cols in LP problem on removal of rids, sids
- # TODO: prior to solve, add objective, thereafter immediatly remove objective
-
- self.objective = {}
- for oid, row in df_fbc_objectives.iterrows():
- if row['active'] is True:
- self.obj_id = row.name
- self.obj_dir = row['type']
- for record in sbmlxdf.misc.record_generator(row['fluxObjectives']):
- params = sbmlxdf.misc.extract_params(record)
- self.objective[params['reac']] = params['coef']
-
- self.check_consistency()
- self.obj_coefs = self.set_objective_coefficients(self.objective)
-
- def set_objective_coefficients(self, objective):
- obj_coefs = np.zeros(len(self.rids))
- for rid, coef in objective.items():
- obj_coefs[self.rid2idx[rid]] = coef
- return obj_coefs
- def check_consistency(self):
+ # lp problem will only be created once required. This gives chances to update e.g. flux bounds
+ self.lp = None
- # check that irreversible reaction have a flux lower bound >= 0
+ def update_reversible_rids(self):
+ self.rids_rev = self.rids[self.reversible]
+ self.rid2idx_rev = {rid: idx + len(self.rids) for idx, rid in enumerate(self.rids_rev)}
- for idx in range(len(self.rids)):
- if np.logical_not(self.reversible[idx]):
- if self.flux_bounds[0, idx] < 0:
- self.reversible[idx] = True
- print(f'reaction {self.rids[idx]} set to reversible')
+ def update_model_reversible(self, reactions):
+ """Update reversible flag of reactions
- if len(self.objective) == 0:
- print('Active FBA objective function missing in the model!')
-
- for rid in self.objective:
- if rid not in self.rids:
- print('objective reaction not part of reactions')
- raise AttributeError
-
- def combine_fwd_bwd(self, vector):
- """combine forward and backward flux information.
-
- :param vector: contains data on all reactions + backward information for reversible reactions
- :type vector: 1D ndarray of shape (len(mr_rids) + sum(reversible))
- :return: combined flux information
- :rtype: 1D ndarray of shape (len(mr_rids))
+ :param reactions: reactions ids that have negative flux bounds
+ :type reactions: set of str
"""
- n_cols = len(self.rids)
- combined = vector[:n_cols]
- combined[self.reversible] -= vector[n_cols:]
- return combined
+ upd_rids = set(reactions).difference(set(self.rids_rev))
+ for rid in upd_rids:
+ self.reversible[self.rid2idx[rid]] = True
+ if len(upd_rids) > 0:
+ print('reactions in protected functions made reversible:', upd_rids)
+ self.update_reversible_rids()
- def create_fba_lp(self):
- """create Linear Programming Problem for FBA with split reactions
+ def update_model_flux_bounds(self, flux_bounds):
+ """Overwrite model flux bounds with protected parts flux bounds.
- Splitting reactions in forward/backward directions reduces high values during FBA
- Splitting required for pFBA (minimize absolute flux values)
+ lp problem might need to be instantiated
- :return: LpProblem configured for FBA
- :rtype: LpProblem
+ :param flux_bounds: lower and upper flux bounds for selected reactions
+ :type flux_bounds: dict: key: reaction id, value: list-like of two float values
"""
- lp = LpProblem()
- fwd_bwd_s_mat = np.hstack((self.s_mat, -self.s_mat[:, self.reversible]))
- fwd_bwd_obj_coefs = np.hstack((self.obj_coefs, -self.obj_coefs[self.reversible]))
- fwd_bwd_bnds = np.hstack((np.fmax(self.flux_bounds, 0),
- np.vstack((np.fmax(-self.flux_bounds[1, self.reversible], 0),
- np.fmax(-self.flux_bounds[0, self.reversible], 0)))))
- row_bnds = np.zeros((fwd_bwd_s_mat.shape[0], 2))
- lp.configure(fwd_bwd_obj_coefs, fwd_bwd_s_mat, row_bnds, fwd_bwd_bnds, self.obj_dir)
- return lp
-
- def fba_optimize(self, short_result=False):
- """FBA optimization of current problem
+ for rid, bound_upd in flux_bounds.items():
+ for i in [0, 1]:
+ if math.isfinite(bound_upd[i]):
+ self.flux_bounds[i, self.rid2idx[rid]] = bound_upd[i]
- :return: results information from the optimization
- :rtype: dict
- """
- lp = self.create_fba_lp()
- if lp is None:
- return {'success': False, 'message': 'not an FBA model'}
- res = lp.solve(short_result)
- del lp
- if short_result is False:
- res['x'] = self.combine_fwd_bwd(res['x'])
- res['reduced_costs'] = self.combine_fwd_bwd(res['reduced_costs'])
- return res
+ def get_parallel_rids(self):
+ """Identify identical reactions (forward/reverse).
- def fva_optimize(self, rids=None, fract_of_optimum=1.0):
- """FVA - flux variability analysis for all or selected reactions.
+ Linerar scaling is not considered.
- :param rids: selected reaction ids to run FVA (alternatively, all)
- :type rids: list of strings (default: None - all reactions)
- :param fract_of_optimum: value between 0.0 and 1.0.
- :type fract_of_optimum: float (default 1.0)
- :return: rid, min and max flux values
- :rtype: dict
+ :return: lists of reaction ids that are parallel
+ :rtype: list with lists of str
"""
- lp = self.create_fba_lp()
- if lp is None:
- return {'success': False, 'message': 'not an FBA model'}
- res = lp.solve(short_result=True)
-
- if res['success'] is False:
- del lp
- return {'success': False, 'message': res['generic_status']}
-
- if rids is None:
- rids = self.rids
- flux_min = np.zeros(len(rids))
- flux_max = np.zeros(len(rids))
-
- # fix FBA objective as constraint
- fwd_bwd_obj_coefs = np.hstack((self.obj_coefs, -self.obj_coefs[self.reversible]))
- if self.obj_dir == 'maximize':
- row_bds = np.array([res['fun'] * fract_of_optimum, np.nan])
- else:
- if fract_of_optimum > 0.0:
- row_bds = np.array([np.nan, res['fun'] / fract_of_optimum])
- else:
- row_bds = np.array([np.nan, 1000.0])
- lp.add_constraint(fwd_bwd_obj_coefs, row_bds)
-
- n_cols = len(self.rids)
- for idx, rid in enumerate(rids):
- # select reaction maximize/minimize flux
- new_obj_coefs = np.zeros(n_cols)
- new_obj_coefs[self.rid2idx[rid]] = 1
- new_fwd_bwd_obj_coefs = np.hstack((new_obj_coefs, -new_obj_coefs[self.reversible]))
- lp.set_obj_coefs(new_fwd_bwd_obj_coefs)
-
- lp.set_obj_dir('minimize')
- res = lp.solve(short_result=True)
- flux_min[idx] = res['fun'] if res['success'] else np.nan
-
- lp.set_obj_dir('maximize')
- res = lp.solve(short_result=True)
- flux_max[idx] = res['fun'] if res['success'] else np.nan
- del lp
- return {'success': True, 'rids': np.array(rids), 'min': flux_min, 'max': flux_max}
-
- def fva_single_rid(self, rids):
- """FVA - flux variability analysis for all or selected reactions.
-
- :param rids: selected reaction ids to run FVA (alternatively, all)
- :type rids: list of strings (default: None - all reactions)
- :return: rid, min and max flux values
- :rtype: dict
- """
- lp = self.create_fba_lp()
- if lp is None:
- return {'success': False, 'message': 'not an FBA model'}
- res = lp.solve(short_result=True)
-
- if res['success'] is False:
- del lp
- return {'success': False, 'message': res['generic_status']}
-
- if rids is None:
- rids = self.rids
- flux_min = np.zeros(len(rids))
- flux_max = np.zeros(len(rids))
- # fix FBA objective as constraint
- fwd_bwd_obj_coefs = np.hstack((self.obj_coefs, -self.obj_coefs[self.reversible]))
- if self.obj_dir == 'maximize':
- row_bds = np.array([0.0, np.nan])
- else:
- row_bds = np.array([np.nan, 1000.0])
- lp.add_constraint(fwd_bwd_obj_coefs, row_bds)
-
- n_cols = len(self.rids)
- for idx, rid in enumerate(rids):
- # select reaction maximize/minimize flux
- new_obj_coefs = np.zeros(n_cols)
- new_obj_coefs[self.rid2idx[rid]] = 1
- new_fwd_bwd_obj_coefs = np.hstack((new_obj_coefs, -new_obj_coefs[self.reversible]))
- lp.set_obj_coefs(new_fwd_bwd_obj_coefs)
-
- lp.set_obj_dir('minimize')
- res = lp.solve(short_result=True)
- flux_min[idx] = res['fun'] if res['success'] else np.nan
-
- lp.set_obj_dir('maximize')
- res = lp.solve(short_result=True)
- flux_max[idx] = res['fun'] if res['success'] else np.nan
- del lp
- return {'success': True, 'rids': np.array(rids), 'min': flux_min, 'max': flux_max}
-
- def update_flux_bounds(self, flux_bounds):
- """selectivly set lower/upper flux bounds for given reactions.
-
- Flux bounds with value of 'np.nan' will not be set
-
- :param flux_bounds: new lower and upper flux bounds selected reactions
- :type flux_bounds: dict (key: reaction id, value: list with two float values for lower/upper bound)
- """
- for rid, (lb, ub) in flux_bounds.items():
- assert rid in self.rids
- if math.isfinite(lb):
- self.flux_bounds[0, self.rid2idx[rid]] = lb
- if lb < 0:
- self.reversible[self.rid2idx[rid]] = True
- if math.isfinite(ub):
- self.flux_bounds[1, self.rid2idx[rid]] = ub
-
- def get_flux_bounds(self, rid):
- assert rid in self.rids
- return [self.flux_bounds[0, self.rid2idx[rid]], self.flux_bounds[1, self.rid2idx[rid]]]
-
- def remove_reactions(self, drop_rids, protected_sids):
- # remove_reactions
- # also remove corresponding metabolite columns
- # update, rids, rid2idx, sids, sid2idx, reversible, flux_bounds, s_mat
-
- keep_rids_mask = np.array([False if rid in drop_rids else True for rid in self.rids])
-
- self.rids = self.rids[keep_rids_mask]
- self.rid2idx = {rid: idx for idx, rid in enumerate(self.rids)}
- self.reversible = self.reversible[keep_rids_mask]
- self.obj_coefs = self.obj_coefs[keep_rids_mask]
- self.flux_bounds = self.flux_bounds[:, keep_rids_mask]
- self.s_mat = self.s_mat[:, keep_rids_mask]
-
- keep_sids_mask = np.any(self.s_mat != 0, axis=1)
- drop_sids = set(self.sids[np.logical_not(keep_sids_mask)])
-
- if len(drop_sids) > 0:
- drop_protected_sids = drop_sids.intersection(protected_sids)
- if len(drop_protected_sids) > 0:
- print('Something went worong, protected metabolites dropped due to reaction removal.',
- drop_protected_sids)
- self.sids = self.sids[keep_sids_mask]
- self.sid2idx = {sid: idx for idx, sid in enumerate(self.sids)}
- self.s_mat = self.s_mat[keep_sids_mask, :]
-
- def get_identical_columns(self):
- # Identify identical columns:
- # sort the columns
- # check parallel columns, if they are identical
- # report on identical columns
-
- # we also consider reversibility
- # we do not consider linearly scaled reactions
- fwd_bwd_s_mat = np.hstack((self.s_mat, -self.s_mat[:, self.reversible]))
- fwd_bwd_rids = np.hstack((self.rids, self.rids[self.reversible]))
+ fwd_rev_s_mat = np.hstack((self.s_mat, -self.s_mat[:, self.reversible]))
+ fwd_rev_rids = np.hstack((self.rids, self.rids_rev))
# identify groups of parallel reactions
- sorted_idx = np.lexsort(fwd_bwd_s_mat)
+ sorted_idx = np.lexsort(fwd_rev_s_mat)
groups = []
consumed_idxs = set()
for pos in range(len(sorted_idx)):
@@ -296,17 +100,30 @@ class FbaModel:
group = [idx]
for adjacent_pos in range(pos + 1, len(sorted_idx)):
candiate_idx = sorted_idx[adjacent_pos]
- if np.all(fwd_bwd_s_mat[:, idx] == fwd_bwd_s_mat[:, candiate_idx]):
- group.append(candiate_idx)
+ if np.all(fwd_rev_s_mat[:, idx] == fwd_rev_s_mat[:, candiate_idx]):
consumed_idxs.add(candiate_idx)
+ group.append(candiate_idx)
else:
break
if len(group) > 1:
groups.append(group)
- parallel_rid_groups = [list(fwd_bwd_rids[group]) for group in groups]
+ parallel_rid_groups = [list(fwd_rev_rids[group]) for group in groups]
return parallel_rid_groups
+ def combine_fwd_rev(self, vector):
+ """combine forward and backward flux information.
+
+ :param vector: contains data on all reactions + backward information for reversible reactions
+ :type vector: 1D ndarray of shape (len(mr_rids) + sum(reversible))
+ :return: combined flux information
+ :rtype: 1D ndarray of shape (len(mr_rids))
+ """
+ n_cols = len(self.rids)
+ combined = vector[:n_cols]
+ combined[self.reversible] -= vector[n_cols:]
+ return combined
+
def get_model_status(self):
""" Return status on model
@@ -320,17 +137,6 @@ class FbaModel:
status.update(self.full_model_shape)
return status
- @property
- def obj_dir(self):
- return self.__obj_dir
-
- @obj_dir.setter
- def obj_dir(self, dir_string):
- if 'min' in dir_string:
- self.__obj_dir = 'minimize'
- else:
- self.__obj_dir = 'maximize'
-
def export_pruned_model(self, pruned_sbml):
sbml_model = sbmlxdf.Model(self.sbml_fname)
@@ -383,3 +189,250 @@ class FbaModel:
if len(err) > 0:
print('SBML validation result: %s', ', '.join([k + ': ' + str(v) for k, v in err.items()]))
return pruned_model.export_sbml(pruned_sbml)
+
+ def create_core_fba_lp(self):
+ """create a core Linear Programming Problem for FBA with split reactions
+
+ Core Problem contains coefficient matrix, variable and row bounds
+ Objective function yet needs to be added prior to any optimization
+
+ Splitting reactions in forward/backward directions reduces high values during FBA
+ configured bounds are >= 0
+ Splitting required for pFBA (minimize absolute flux values)
+
+ :return: LpProblem configured for FBA
+ :rtype: LpProblem
+ """
+ fwd_rev_s_mat = np.hstack((self.s_mat[:], -self.s_mat[:, self.reversible]))
+ fwd_rev_col_bnds = np.hstack((np.fmax(self.flux_bounds[:], 0),
+ np.fmax(np.flip(-self.flux_bounds[:, self.reversible], axis=0), 0)))
+ row_bnds = np.zeros((fwd_rev_s_mat.shape[0], 2))
+ lp = LpProblem(fwd_rev_s_mat, row_bnds, fwd_rev_col_bnds)
+ return lp
+
+ def set_lp_objective(self, objective, direction='maximize'):
+ """Set objective coefficients and direction in linear problem
+
+ :param objective: objective coefficients
+ :type objective: dict with key: reaction id, value: stoichiometric coefficient
+ :param direction: direction of optimization
+ :type direction: str (optional, default: 'maximize')
+ """
+ assert self.lp is not None
+ self.lp.set_obj_coefs(self.get_coefficients(objective))
+ self.lp.set_obj_dir(direction)
+
+ def remove_lp_objective(self, objective):
+ """remove objective coefficients
+
+ :param objective: objective coefficients
+ :type objective: dict with key: reaction id, value: stoichiometric coefficient
+ """
+ self.lp.remove_obj_coefs(self.get_coefficients(objective))
+
+ def update_lp_flux_bounds(self, flux_bounds):
+ """update flux bounds for selected reactions in the linear problem.
+
+ lp problem might need to be instantiated
+ overwrite bounds of protected functions will be cleaned from already removed reactions
+
+ :param flux_bounds: lower and upper flux bounds for selected reactions
+ :type flux_bounds: dict: key: reaction id, value: list-like of two float values
+ """
+ if self.lp is None:
+ self.lp = self.create_core_fba_lp()
+ print('Core LP Problem created')
+ for rid, bound_upd in flux_bounds.items():
+ if rid in self.rids:
+ temp_bounds = self.flux_bounds[:, self.rid2idx[rid]].copy()
+ for i in [0, 1]:
+ if math.isfinite(bound_upd[i]):
+ temp_bounds[i] = bound_upd[i]
+ self.lp.set_col_bnd(self.rid2idx[rid], np.fmax(temp_bounds, 0))
+ if rid in self.rids_rev:
+ self.lp.set_col_bnd(self.rid2idx_rev[rid], np.fmax(np.flip(-temp_bounds), 0))
+
+ def reset_lp_flux_bounds(self, rids):
+ """restore original flux bounds for selected reactions in linear problem.
+
+ overwrite bounds of protected functions will be cleaned from already removed reactions
+
+ :param rids: reaction ids for which to restore the model orignal flux bounds
+ :type rids: set of str
+ """
+ assert self.lp is not None
+
+ for rid in rids:
+ if rid in self.rids:
+ assert rid in self.rids
+ bounds = self.flux_bounds[:, self.rid2idx[rid]]
+ self.lp.set_col_bnd(self.rid2idx[rid], np.fmax(bounds, 0))
+ if rid in self.rids_rev:
+ self.lp.set_col_bnd(self.rid2idx_rev[rid], np.fmax(np.flip(-bounds), 0))
+
+ def get_coefficients(self, objective):
+ """get coefficients for objective or constraint function
+
+ :param objective: objective/constraint coefficients
+ :type objective: dict with key: reaction id, value: stoichiometric coefficient
+ :return: coefficients aligned with the linear problem
+ :rtype: 1D ndarray with float
+ """
+ coefs = np.zeros(len(self.rids) + len(self.rids_rev))
+ for rid, coef in objective.items():
+ coefs[self.rid2idx[rid]] = coef
+ if rid in self.rids_rev:
+ coefs[self.rid2idx_rev[rid]] = -coef
+ return coefs
+
+ def remove_reactions(self, drop_rids):
+ """remove reactions and blocked metabolites from model and lp problem
+
+ :param drop_rids: reaction ids to be removed from the model
+ :type drop_rids: set of str
+ """
+ assert self.lp is not None
+
+ # remove variables (columns) from linear problem
+ col_idxs0_fwd = [self.rid2idx[rid] for rid in drop_rids]
+ col_idxs0_rev = [self.rid2idx_rev[rid] for rid in drop_rids if rid in self.rids_rev]
+ self.lp.remove_cols(col_idxs0_fwd + col_idxs0_rev)
+
+ keep_rids_mask = np.array([False if rid in drop_rids else True for rid in self.rids])
+ self.rids = self.rids[keep_rids_mask]
+ self.rid2idx = {rid: idx for idx, rid in enumerate(self.rids)}
+ self.reversible = self.reversible[keep_rids_mask]
+ self.update_reversible_rids()
+ self.flux_bounds = self.flux_bounds[:, keep_rids_mask]
+ self.s_mat = self.s_mat[:, keep_rids_mask]
+
+ # drop blocked metabolites due to deleted reactions
+ keep_sids_mask = np.any(self.s_mat != 0, axis=1)
+ row_idxs0 = [idx for idx, val in enumerate(keep_sids_mask) if val == np.bool_(False)]
+ if len(row_idxs0) > 0:
+ self.lp.remove_rows(row_idxs0)
+ self.sids = self.sids[keep_sids_mask]
+ self.sid2idx = {sid: idx for idx, sid in enumerate(self.sids)}
+ self.s_mat = self.s_mat[keep_sids_mask, :]
+
+ def fba_pf_optimize(self, pf, zero_flux_rid=None):
+ """FBA optimization of current problem
+
+ zero_flux_rids overwrites settings in protected functions
+
+ :param pf: protected function (objective)
+ :type pf: class ProtectedFunction
+ :param zero_flux_rid: reaction id of candiated checked for removal
+ :type zero_flux_rid: str, optional (default: None)
+ :return: results information from the optimization
+ :rtype: dict
+ """
+ if self.lp is None:
+ self.lp = self.create_core_fba_lp()
+ print('Core LP Problem created')
+ assert self.lp is not None
+
+ if zero_flux_rid is not None:
+ overwrite_bounds = pf.overwrite_bounds.copy()
+ overwrite_bounds[zero_flux_rid] = np.array([0.0, 0.0])
+ else:
+ overwrite_bounds = pf.overwrite_bounds
+
+ self.set_lp_objective(pf.objective, pf.obj_dir)
+ self.update_lp_flux_bounds(overwrite_bounds)
+ res = self.lp.solve()
+ # print(res['success'], res['fun'])
+ self.reset_lp_flux_bounds(overwrite_bounds.keys())
+ self.remove_lp_objective(pf.objective)
+ return res
+
+ def check_pf_feasibility(self, pf, zero_flux_rid=None):
+ """check that current model satisfies the protected functions
+
+ :param pf: protected function (objective)
+ :type pf: class ProtectedFunction
+ :param zero_flux_rid: reaction id of candiated checked for removal
+ :type zero_flux_rid: str, optional (default: None)
+ :return: status if protected function satisfied
+ :rtype: bool
+ """
+ feasible = True
+ res = self.fba_pf_optimize(pf, zero_flux_rid)
+ if res['success'] is False:
+ feasible = False
+ elif pf.obj_dir == 'maximize':
+ if res['fun'] < pf.target_lb:
+ feasible = False
+ else:
+ if res['fun'] > pf.target_ub:
+ feasible = False
+ return feasible
+
+ def fva_pf_flux_ranges(self, pf, rids):
+ """FVA for a specific protected function for all or selected reactions
+
+ :param pf: protected function (objective)
+ :type pf: class ProtectedFunction
+ :param rids: list of reaction ids to get flux range
+ :type rids: list of str
+ :return: rid, min and max flux values
+ :rtype: dict
+ """
+ # check that problem is feasible
+ if self.check_pf_feasibility(pf) is False:
+ return {'success': False}
+
+ # implement objective as a constraint and overwrite bounds
+ row_bnds = (np.array([pf.target_lb, np.nan]) if pf.obj_dir == 'maximize'
+ else np.array([np.nan, pf.target_ub]))
+ row_idx0 = self.lp.add_constraint(self.get_coefficients(pf.objective), row_bnds)
+ self.update_lp_flux_bounds(pf.overwrite_bounds)
+
+ flux_min = np.zeros(len(rids))
+ flux_max = np.zeros(len(rids))
+ for idx, rid in enumerate(rids):
+ self.set_lp_objective({rid: 1}, 'minimize')
+ res = self.lp.solve()
+ flux_min[idx] = res['fun'] if res['success'] else np.nan
+ self.lp.set_obj_dir('maximize')
+ res = self.lp.solve()
+ flux_max[idx] = res['fun'] if res['success'] else np.nan
+ self.remove_lp_objective({rid: 1})
+
+ self.reset_lp_flux_bounds(pf.overwrite_bounds.keys())
+ self.lp.del_constraint(row_idx0)
+
+ res = {'success': True, 'rids': np.array(rids), 'min': flux_min, 'max': flux_max}
+ # print(res)
+ return res
+
+ def fva_single_rid(self, rid):
+ """FVA - flux variability analysis for all or selected reactions.
+
+ :param rid: selected reaction ids to run FVA (alternatively, all)
+ :type rid: list of strings (default: None - all reactions)
+ :return: rid, min and max flux values
+ :rtype: dict
+ """
+ len(self.rids)
+ print('fva_single_rid not implemented', rid)
+ return {'success': False}
+
+ def copy(self):
+ """Copy the FbaModel instance (while removing linear problem).
+
+ A shallow copy, to support multiprocessing of FVA,
+ which do not impact FBA model attributes, except for linear problem
+ """
+ cls = self.__class__
+ result = cls.__new__(cls)
+ result.__dict__.update(self.__dict__)
+ result.__dict__['lp'] = None
+ return result
+
+ def __reduce__(self):
+ return self.__class__, (self.sbml_fname, )
+
+ def __del__(self):
+ print('deletion of fba_model')
+ del self.lp
diff --git a/modelpruner/problems/lp_problem.py b/modelpruner/problems/lp_problem.py
index e346a33011f0aaf3a223a29d2efed89d1b214495..0d094a52a0d1a5cc132b6f3a6ace6387b869f530 100644
--- a/modelpruner/problems/lp_problem.py
+++ b/modelpruner/problems/lp_problem.py
@@ -61,18 +61,58 @@ scaling_options = {'GM': glpk.GLP_SF_GM,
class LpProblem:
- def __init__(self):
- """create a glpk problem.
+ def __init__(self, constr_coefs, row_bnds, col_bnds):
+ """create core linear problem (constraints, variable and row bounds)
Problem needs to be properly deleted usin del(LpProblem)
+ Note: glpk related indices start at 1 (not zero)
+ :param constr_coefs: matrix with constraint coefficients
+ :type constr_coefs: 2D nparray with shape(nrows, ncols)
+ :param row_bnds: lower and upper bounds of auxhiliary (row) variables
+ :type row_bnds: 2D ndarray of shape(nrows, 2) (1st col: lower bounds, 2nd col: upper bounds)
+ :param col_bnds: lower and upper bounds of structural (col) variables
+ :type col_bnds: 2D ndarray of shape(2, ncols) (1st row: lower bounds, 2nd row: upper bounds)
"""
- self.lp = glpk.glp_create_prob()
- self.ncols = 0
- self.nrows = 0
- self.msg_lev = glpk.GLP_MSG_ERR
- self.tm_lim = 10 * 1000
+ glp_msg_lev = glpk.GLP_MSG_ERR
+ glp_tm_lim = 10 * 1000
+ glp_presolve = glpk.GLP_OFF # or glpk.GLP_OFF (default) or GLP_ON
self.res = {}
+ self.nrows, self.ncols = constr_coefs.shape
+ assert col_bnds.shape == (2, self.ncols)
+ assert row_bnds.shape == (self.nrows, 2)
+
+ self.lp = glpk.glp_create_prob()
+
+ # configure variables/columns and related bounds
+ glpk.glp_add_cols(self.lp, self.ncols)
+ for i, lb_ub in enumerate(col_bnds.T):
+ var_type, lb, ub = self._get_variable_type(lb_ub)
+ glpk.glp_set_col_bnds(self.lp, 1 + i, var_type, lb, ub)
+
+ # configure rows/constraints
+ glpk.glp_add_rows(self.lp, self.nrows)
+ for i, lb_ub in enumerate(row_bnds):
+ var_type, lb, ub = self._get_variable_type(lb_ub)
+ glpk.glp_set_row_bnds(self.lp, 1 + i, var_type, lb, ub)
+ constr_coefs_sparse = scipy.sparse.coo_matrix(constr_coefs)
+ n_coefs = len(constr_coefs_sparse.row)
+ ia = glpk.intArray(1 + n_coefs)
+ ja = glpk.intArray(1 + n_coefs)
+ ar = glpk.doubleArray(1 + n_coefs)
+ for i in range(n_coefs):
+ ia[1 + i] = int(1 + constr_coefs_sparse.row[i])
+ ja[1 + i] = int(1 + constr_coefs_sparse.col[i])
+ ar[1 + i] = constr_coefs_sparse.data[i]
+ glpk.glp_load_matrix(self.lp, n_coefs, ia, ja, ar)
+
+ # configure options
+ self._smcp = glpk.glp_smcp()
+ glpk.glp_init_smcp(self._smcp)
+ self._smcp.msg_lev = glp_msg_lev
+ self._smcp.tm_lim = glp_tm_lim
+ self._smcp.presolve = glp_presolve
+
@staticmethod
def _get_variable_type(bounds):
"""Determine variable type for upper/lower bounds.
@@ -105,76 +145,42 @@ class LpProblem:
var_type = glpk.GLP_LO
return var_type, lb, ub
- def configure(self, obj_coefs, constr_coefs, row_bnds, col_bnds, direction='maximize'):
- """Configure the LP Problem.
+ def set_obj_coefs(self, obj_coefs):
+ """set coefficients for objective function
+
+ configure coefficients that are not zero
Note: glpk related indices start at 1 (not zero)
:param obj_coefs: coefficients of objective function
:type obj_coefs: 1D ndarray of shape (1, ncols)
- :param constr_coefs: matrix with constraint coefficients
- :type constr_coefs: 2D nparray with shape(nrows, ncols)
- :param row_bnds: lower and upper bounds of auxhiliary (row) variables
- :type row_bnds: 2D ndarray of shape(nrows, 2) (1st col: lower bounds, 2nd col: upper bounds)
- :param col_bnds: lower and upper bounds of structural (col) variables
- :type col_bnds: 2D ndarray of shape(2, ncols) (1st row: lower bounds, 2nd row: upper bounds)
- :param direction: direction for the optimization ('maximize' or 'minimize')
- :type direction: string
"""
- self.nrows, self.ncols = constr_coefs.shape
-
assert len(obj_coefs) == self.ncols
- assert col_bnds.shape == (2, self.ncols)
- assert row_bnds.shape == (self.nrows, 2)
+ for i in range(self.ncols):
+ if obj_coefs[i] != 0.0:
+ glpk.glp_set_obj_coef(self.lp, int(1 + i), obj_coefs[i])
- # configure objective function
- glpk.glp_add_cols(self.lp, self.ncols)
- self.set_obj_coefs(obj_coefs)
- self.set_obj_dir(direction)
+ def remove_obj_coefs(self, obj_coefs):
+ """remove coefficients from objective function
- # configure bounds on structural (columns) variables
- for i, lb_ub in enumerate(col_bnds.T):
- var_type, lb, ub = self._get_variable_type(lb_ub)
- glpk.glp_set_col_bnds(self.lp, 1 + i, var_type, lb, ub)
+ set coefficients that are not zero to zero in the liner problem
- # configure constraints
- glpk.glp_add_rows(self.lp, self.nrows)
- for i, lb_ub in enumerate(row_bnds):
- var_type, lb, ub = self._get_variable_type(lb_ub)
- glpk.glp_set_row_bnds(self.lp, 1 + i, var_type, lb, ub)
- constr_coefs_sparse = scipy.sparse.coo_matrix(constr_coefs)
- n_coefs = len(constr_coefs_sparse.row)
- ia = glpk.intArray(1 + n_coefs)
- ja = glpk.intArray(1 + n_coefs)
- ar = glpk.doubleArray(1 + n_coefs)
- for i in range(n_coefs):
- ia[1 + i] = int(1 + constr_coefs_sparse.row[i])
- ja[1 + i] = int(1 + constr_coefs_sparse.col[i])
- ar[1 + i] = constr_coefs_sparse.data[i]
- glpk.glp_load_matrix(self.lp, n_coefs, ia, ja, ar)
-
- def replace_constraint(self, row, constr_coefs, row_bnds):
- # set/replace one bounded constraint
- #
- # Parameters:
- # row: row id to be replaced
- # coeffs: 1D np.array of floats of size number design variables
- # coefficients != 0.0 will be configured
- # bounds: 1D numpy.ndarray with float of size two
- # lists and tuples will be converted to numpy.ndarray
- # lower_bound, upper_bound, np.nan and np.inf supported
+ Note: glpk related indices start at 1 (not zero)
+ :param obj_coefs: coefficients of objective function
+ :type obj_coefs: list-like of floats
+ """
+ assert len(obj_coefs) == self.ncols
+ for i in range(self.ncols):
+ if obj_coefs[i] != 0.0:
+ glpk.glp_set_obj_coef(self.lp, int(1 + i), 0.0)
- assert row <= glpk.glp_get_num_rows(self.lp)
- glpk.glp_set_row_bnds(self.lp, row, *self._get_variable_type(row_bnds))
+ def set_obj_dir(self, direction):
+ """Configure the optimization direction
- n_coef = len(np.nonzero(constr_coefs)[0])
- ind = glpk.intArray(1 + n_coef)
- val = glpk.doubleArray(1 + n_coef)
- idx = 1
- for i in np.nonzero(constr_coefs)[0]:
- ind[idx] = int(1 + i)
- val[idx] = constr_coefs[i]
- idx += 1
- glpk.glp_set_mat_row(self.lp, row, n_coef, ind, val)
+ :param direction: direction for the optimization ('maximize' or 'minimize')
+ :type direction: string
+ """
+ obj_dir = glpk.GLP_MAX if 'max' in direction else glpk.GLP_MIN
+ glpk.glp_set_obj_dir(self.lp, obj_dir)
def add_constraint(self, constr_coefs, row_bnds):
"""add a single constraint to the lp problem
@@ -183,108 +189,93 @@ class LpProblem:
:type constr_coefs: 1d ndarray of length ncols
:param row_bnds: upper and lower bounds of the auxiliary (row) variable
:type row_bnds: 1D ndarray of length 2
- :return:
+ :return: row index (start from 0)
+ :rtype: int
"""
- row = glpk.glp_add_rows(self.lp, 1)
- self.replace_constraint(row, constr_coefs, row_bnds)
- return row
-
- def del_constraint(self, row):
- # delete one constraint
- #
- # Parameters:
- # row: row id to be replaced
- assert row <= glpk.glp_get_num_rows(self.lp)
+ row_idx1 = glpk.glp_add_rows(self.lp, 1)
+ glpk.glp_set_row_bnds(self.lp, row_idx1, *self._get_variable_type(row_bnds))
+
+ n_coef = len(np.nonzero(constr_coefs)[0])
+ ind = glpk.intArray(n_coef + 1)
+ val = glpk.doubleArray(n_coef + 1)
+ for num_idx0, col_idx0 in enumerate(np.nonzero(constr_coefs)[0]):
+ ind[num_idx0 + 1] = int(col_idx0 + 1)
+ val[num_idx0 + 1] = constr_coefs[col_idx0]
+ glpk.glp_set_mat_row(self.lp, row_idx1, n_coef, ind, val)
+ self.nrows += 1
+ row_idx0 = row_idx1 - 1
+ return row_idx0
+
+ def del_constraint(self, row_idx0):
+ """remove single constraint from the lp problem
+
+ :param row_idx0: row index (start from 0) to be removed
+ :type row_idx0: int
+ """
+ assert row_idx0 < self.nrows
num = glpk.intArray(1 + 1)
- num[1] = row
+ num[1] = row_idx0 + 1
glpk.glp_del_rows(self.lp, 1, num)
-
- def set_single_bound(self, react_idx, bounds):
- assert react_idx <= self.ncols
- if type(bounds) is not np.ndarray:
- bounds = np.array(bounds)
- glpk.glp_set_col_bnds(self.lp, 1 + react_idx, *self._get_variable_type(bounds))
-
- def set_obj_dir(self, direction):
- """Configure the optimization direction
-
- :param direction: direction for the optimization ('maximize' or 'minimize')
- :type direction: string
+ self.nrows -= 1
+
+ def set_col_bnd(self, col_idx0, bounds):
+ """Set a single column (variable) bound
+ :param col_idx0: column index (start at 0) of variable
+ :type col_idx0: int
+ :param bounds: lower and upper bound to set
+ :type bounds: list with two float values
"""
- obj_dir = glpk.GLP_MAX if 'max' in direction else glpk.GLP_MIN
- glpk.glp_set_obj_dir(self.lp, obj_dir)
-
- def reset_cost(self, cost_coef=None):
- # reset coefficients of objective function (cols) to 0.0
- #
- # Parameters:
- # cost_coef: None or 1d numpy.ndarray with floats
- # None: reset all cost coefficiencs
- # ndarray: reset only coest coefficiencs != 0
- if cost_coef is None:
- for i in range(glpk.glp_get_num_cols(self.lp)):
- glpk.glp_set_obj_coef(self.lp, int(1 + i), 0.0)
- else:
- for i in np.nonzero(cost_coef)[0]:
- glpk.glp_set_obj_coef(self.lp, int(1 + i), 0.0)
+ assert col_idx0 < self.ncols
+ glpk.glp_set_col_bnds(self.lp, col_idx0 + 1, *self._get_variable_type(bounds))
- def set_obj_coefs(self, obj_coefs):
- """set coefficients for objective function
+ def remove_cols(self, col_idxs0):
+ """ remove variables (columns) from the linear problem
- :param obj_coefs: coefficients of objective function
- :type obj_coefs: 1D ndarray of shape (1, ncols)
+ :param col_idxs0: col indices (start at 0) to be removed
+ :type col_idxs0: list or 1D ndarray of col indices
"""
- for i in range(self.ncols):
- glpk.glp_set_obj_coef(self.lp, int(1 + i), obj_coefs[i])
-
- def solve(self, short_result=False, scaling=None):
+ num = glpk.intArray(len(col_idxs0) + 1)
+ for num_idx0, col_idx0 in enumerate(col_idxs0):
+ assert col_idx0 < self.ncols
+ num[num_idx0 + 1] = col_idx0 + 1
+ glpk.glp_del_cols(self.lp, len(col_idxs0), num)
+ self.ncols -= len(col_idxs0)
+
+ def remove_rows(self, row_idxs0):
+ """ remove constraints (rows) from the linear problem.
+
+ :param row_idxs0: row indices (start at 0) to be removed
+ :type row_idxs0: list or 1D ndarray of row indices
+ """
+ num = glpk.intArray(len(row_idxs0) + 1)
+ for num_idx0, row_idx0 in enumerate(row_idxs0):
+ assert row_idx0 < self.nrows
+ num[num_idx0 + 1] = row_idx0 + 1
+ glpk.glp_del_rows(self.lp, len(row_idxs0), num)
+ self.nrows -= len(row_idxs0)
+
+ def solve(self, short_result=True, scaling=None):
"""Solve the Linear Programming Problem.
:param short_result: short result (only objective value and status) or long result
- :type short_result: bool (default:False)
+ :type short_result: bool (optional, default: True)
:param scaling: problem scaling options ('GM', 'EQ', '2N', 'SKIP', 'AUTO'
- :type scaling: string (default, None)
- :return: result information from the optimization stored in a dictionary
+ :type scaling: string (optional, default: None)
+ :return: result from linear optimization
:rtype: dict
"""
- # configure options
- smcp = glpk.glp_smcp()
- glpk.glp_init_smcp(smcp)
- smcp.msg_lev = self.msg_lev
- smcp.tm_lim = self.tm_lim
-
if scaling is not None:
- opt = scaling_options.get(scaling, glpk.GLP_SF_AUTO)
- glpk.glp_scale_prob(self.lp, opt)
- # sjj = [glpk.glp_get_sjj(self.lp, 1+i) for i in range(glpk.glp_get_num_cols(self.lp))]
- # rii = [glpk.glp_get_rii(self.lp, 1+i) for i in range(glpk.glp_get_num_rows(self.lp))]
- # TODO: glpk.glp_adv_basis(self.lp, 0)
- simplex_result = glpk.glp_simplex(self.lp, smcp)
- return self.results(simplex_result, short_result)
-
- def get_row_value(self, row):
- """Retrieve right-hand side of a constraint (auxiliary variable)
-
- :param row: index into constraints
- :type row: int
- :return: value of auxiliary varibale (value of rhs)
- :rtype: float
- """
- assert row <= glpk.glp_get_num_rows(self.lp)
- return glpk.glp_get_row_prim(self.lp, row)
+ glpk.glp_scale_prob(self.lp, scaling_options.get(scaling, glpk.GLP_SF_AUTO))
- def results(self, simplex_result, short_result=False):
- """Collect results for the optimization
+ simplex_result = glpk.glp_simplex(self.lp, self._smcp)
- Alternatively, reduced results, e.g. when doing pFBA
+ # in case of unsuccessful result, set basis and redo optimization
+ if glpk.glp_get_status(self.lp) != glpk.GLP_OPT:
+ glpk.glp_adv_basis(self.lp, 0)
+ # print('GLP ADV BASIS')
+ # glpk.glp_cpx_basis(self.lp)
+ simplex_result = glpk.glp_simplex(self.lp, self._smcp)
- :param simplex_result: result from the LP solver
- :type simplex_result: int
- :param short_result: short result (only objective value and status) or long result
- :type short_result: bool (default:False)
- :return: result from the FBA optimization
- :rtype: dict
- """
self.res = {
'fun': glpk.glp_get_obj_val(self.lp),
'success': glpk.glp_get_status(self.lp) == glpk.GLP_OPT,
@@ -292,16 +283,31 @@ class LpProblem:
}
if short_result is False:
self.res['x'] = np.array([glpk.glp_get_col_prim(self.lp, 1 + i) for i in range(self.ncols)])
- self.res['reduced_costs'] = np.array([glpk.glp_get_col_dual(self.lp, 1 + i) for i in range(self.ncols)])
- self.res['shadow_prices'] = np.array([glpk.glp_get_row_dual(self.lp, 1 + i) for i in range(self.nrows)])
+ self.res['reduced_costs'] = np.array([glpk.glp_get_col_dual(self.lp, 1 + i)
+ for i in range(self.ncols)])
+ self.res['shadow_prices'] = np.array([glpk.glp_get_row_dual(self.lp, 1 + i)
+ for i in range(self.nrows)])
self.res['simplex_result'] = simplex_result
self.res['simplex_message'] = simplex_status[simplex_result],
self.res['primal_status'] = prim_status[glpk.glp_get_prim_stat(self.lp)]
self.res['dual_status'] = dual_status[glpk.glp_get_dual_stat(self.lp)]
return self.res
+ def get_row_value(self, row):
+ """Retrieve right-hand side of a constraint (auxiliary variable)
+
+ :param row: index into constraints
+ :type row: int
+ :return: value of auxiliary varibale (value of rhs)
+ :rtype: float
+ """
+ assert row <= glpk.glp_get_num_rows(self.lp)
+ return glpk.glp_get_row_prim(self.lp, row)
+
def __del__(self):
"""Explicitly delete the LpProblem. (can we do without?)
"""
+ print('deletion of LpProblem')
+ assert self.lp is not None
if getattr(self, 'lp'):
glpk.glp_delete_prob(self.lp)