Table 2.

Pros and cons of COBRA methods.

Category	Method/Tool	Pros	Cons
Reconstruction	AuReMe	- Support for eukaryotes model - Good traceability - Automatic integration of experimental data	- Some manual refinement assistance - Not FBA-ready
	CarveMe	- GEMs ready for FBA - Fast - Customizable for large number of genomes	- No manual refinement assistance - Some support for eukaryotes model
	MetaDraft	-Support for eukaryotes model - Fast	- No manual refinement assistance - Not FBA-ready
	CoReCo	- Support for eukaryotes model - GEMs nearly ready for FBA - Simultaneous reconstruction for multiple species (parallelizable)	- Requires KEGG license - No manual refinement assistance
FBA	FBA	- Does not require kinetic parameters	- Requires objective function - Requires reaction bounds (especially exchange flux)
Dynamic modeling	Dynamic FBA (SOA and DAE)	- Couples pseudo-steady states to dynamical systems - Does not require kinetic parameters	- SOA requires small steps and thus more computation
	DMPy	- Infers missing kinetic parameters using thermodynamics constraints	- Requires >80% of kinetic parameters for accuracy
Alternative optima	Geometric FBA	- Gives single representative solution – Reproducible typical solution (avoids randomly picking one solution from flux cone)	- Weak correlation with protein levels (without omics constraint)
	FVA/VFFVA	- Determines min and max flux for a reaction would achieve optimal objective state - (VFFVA) Increased speed and reduced memory usage	- Varies one reaction at a time
	Sampling	- Estimates probability distribution of feasible fluxes - Can be unbiased (not using an objective function)	- Computationally intensive
Omics constraints	E-flux	- Constraints reaction bounds only - No discretization of data	- May over-constrain model based on noisy data - Poor growth rate prediction
	GIMME	- LP problem (fast) - Ensures operability of required metabolic function - Predicts growth rate, uptake/secretion rates, essential genes, and oncogenes	- Discretizes data - Models have high fractions of blocked reactions, moderate resolution power, poor robustness to missing data/noise
	GIM3E	- Ensures operability of required metabolic function - Integrates metabolomics data	- Discretizes data - MILP problem (slow)
	(t)INIT	- Ensures operability of required metabolic functions - (INIT) predicts oncogenes and tumor suppressor genes, consistent model, good resolution power, robust to noise/missing data	- MILP problem (slow) - (INIT) Poor predictions of growth rate, uptake/secretion rates, and essential genes
	iMAT	- No objective required - Consistent model, good resolution power, robust to noise/missing data - Predicts oncogenes	- Discretizes data - MILP problem (slow) - Weak predictions of growth rate, uptake/secretion rates, and essential genes
	FASTCORE	- LP problem (fast) - Obtains minimal consistent model - Predicts oncogenes and loss of function mutations - Moderately consistent model, good resolution power, robust to noise	- Requires specification of core reactions - Poor predictions of growth rate, uptake/secretion rates, and essential genes
	CORDA	- LP problem (fast) - Non-parsimonious pruning - Predicts oncogenes and loss of function mutations	- Requires specification of core reactions - Weak predictions of growth rate and essential genes - Poor predictions of uptake/secretion rates
Regulatory constraints	rFBA	- Predicts flux over time intervals - Models transcriptional regulation	- Uses boolean TRN - Stepwise calculation of metabolic and regulatory states - Chooses only one solution per time interval
	SR-FBA	- Combined calculation using metabolic and regulatory constraints - Models transcriptional regulation	- Uses boolean TRN - Calculates flux for one time step (steady-state) - Does not account for metabolic transitions and feedback loops
	PROM	- Uses continuous TRN - Models transcriptional regulation	- Requires TF-target gene relationships
	GEM-PRO	- Models protein instability	- Requires protein structures
	arFBA	- Models allosteric regulation	- Requires regulation matrix defining effector-reaction relationship - Small-scale applications
Thermodynamics	ll-FBA	- Does not require metabolite concentrations or free energies	- MILP problem (slow)
	CycleFreeFlux	- Post-process using LP problem (fast) - Can be applied to any flux distribution including sampled solutions - Does not require metabolite concentrations or free energies	- Biased towards solutions with small total flux and those with same direction as their overlapping internal cycles
	TFA, TVA	- Explicitly models thermodynamics	- Requires metabolite concentrations and free energies - Over-approximates uncertainty
	PTA	- Explicitly models thermodynamics for optimization and sampling - Models uncertainty of free energies and metabolite concentrations	- Requires metabolite concentrations and free energies - Computationally intensive
Protein constraints	pFBA	- Predicts growth rate, uptake/secretion rates, and essential genes	- Assumes that flux distribution with smallest magnitude minimizes protein costs
	Enzymatic constraints (GECKO, sMOMENT, ECMpy)	- Model proteome limitation at enzyme resolution - (sMOMENT) Automates enzyme database query - (ECMpy) Automates enzyme parameters calibration - (ECMpy) Does not increase model size	- Requires experimentally measured enzyme turnover numbers - (GECKO) Increases model size - (sMOMENT) Moderately increases model size - (ECMpy) Manually obtains protein subunit composition data
ME-modeling	COBRAme	- Modeling proteome composition improves predictive accuracy - Framework for building ME-models for new organisms	- Large model size and complexity - No standardized SBML format for ME-models - Only applied to bacteria so far
Ensemble modeling	Medusa	- Compresses multiple models into compact ensemble objects - Reduces memory usage of storing ensembles - Interfaces with machine learning	- No standardized SBML format for ensemble objects
Single cell modeling	Compass	- Genome-scale modeling - Maximizes agreement with gene expression - Handles sparsity by sharing information across neighbors - Uses multiple objective functions	- Map gene expression to reaction expression using boolean relationships (GPR)
	scFEA	- Minimizes flux imbalance of all cells to simulate exchange of metabolites - Less stringent flux balance and steady-state assumption - Uses neural net to model nonlinear relationship between gene expression and reaction rates	- Not easily scalable due to large memory usage - Applied to small-scale models
Community modeling	MICOM	- Models exchanges and interactions between communities and environment - Automates building community models from a model database - Predicts replication rates in human gut microbiome	- Assumes trade-offs between individual and community growth rate (gut microbiome specific) - Metabolic models may not be accurate (labratory vs. gut conditions, species differences)
	Dynamic FBA (surfin_fba)	- Reduces optimizations problems (and parameter space) required for dynamic FBA for communities	- Non-biological approach to choosing between non-unique optima
Pathway Analysis	EFM	- Unbiased characterization of models (no objective function required) - (EFMlrs) Pre- and post-process models for EFM calculations	- (EFMlrs) EFM calculation performed by other tools not included in program - EFM calculations are memory intensive and not scalable

Some method comparisons extracted from literature for reconstruction (87, 88), dynamic modeling (89), omics constraints (90, 91), and regulatory constraints (92). Growth rate, uptake/secretion rates, and cancer essential gene prediction performances from Jamialahmadi et al. are based on human metabolic models and available only for GIMME, INIT, iMAT, FASTCORE, CORDA, and pFBA (91).