Highlights
In this release we have made major changes to pretty much all corners
of cobrapy but we hope that any negative impact on existing workflows
will still be limited. The major change is the ongoing move away from
cobrapy's internal solver interfaces to those provided by
[optlang](https://github.com/biosustain/optlang) which provides a
single unified interface to glpk, cplex and gurobi enhanced by the
ability to deal with symbolic expressions. This means formulating
complex constraints no longer implies defining the necessary matrix
algebra, but instead simply writing the expression and assigning that
as on objective to the model.
We feel that the this, and the clarified scope and focus attained by
separating the topic of linear programming (optlang) and metabolic
flux analysis (cobrapy) to two packages is natural and makes both of
these tasks less confusing and more fun. We hope that you after
trying, feel the same and that in any case you let us know what you
think by
[raising an issue](https://github.com/opencobra/cobrapy/issues) or
talking directly to us on [gitter](gitter.im/opencobra/cobrapy) or
[google groups](https://groups.google.com/forum/#!forum/cobra-pie).
New features
The optlang solver interface
The main change is the addition of `model.solver` which is the optlang
interface to the chosen solver (cplex and glpk are currently well
supported, gurobi interface is at the time of writing mostly working
but improvements are still expected). The solver interface manages
variables, constraints and the objective of the model and the task of
turning these into a model formulation to be optimized by the
solver. From cobrapy's point-of-view, this means that all aspects
concerning generating problems, configuring solvers are handed over to
optlang and consequently the whole `cobra.solver` has been deprecated,
slated for removal in the next major release of cobrapy.
Importantly, configuring the solver by passing `**solver_args` or
`solver='solver'` is now instead done by assigning solver to
`model.solver` and then configuring via `model.solver.configuration`.
Creating new algorithms has been greatly facilitated as it no longer
requires formulating objectives and constraints by matrix algebra but
instead directly by expressions, e.g. see the implementation of
`cobra.flux_analysis.moma.add_moma` and
`cobra.flux_analysis.loopless.loopless_solution`.
Instead of having only reactions as variables and metabolites as
constraints, with optlang, cobrapy now supports arbitrary constraints
and variables and these can be added/removed by `model.add_cons_vars`
and `model.remove_cons_vars` which take care of adding these to
`model.problem` which is the optlang's mathematical model associated
with the cobra model.
Reactions are now modeled by two variables, forward and reverse, and
these can be seen by accessing `reaction.{forward,reverse}_variable`
and the combined `reaction.flux_expression`.
Objectives can now easily be made quite advanced by simply crafting
the right expression and assigning this as usual to `model.objective`,
see the
[contraints and objectives notebook](http://cobrapy.readthedocs.io/en/latest/constraints_objective.html).
Temporary changes to a model
Models are large complex objects and copying such objects is
inevitably slow. To avoid that, cobrapy has drawn on the experience
from the development of cameo to introduce the `HistoryManager` class
and the concept of models as contexts. Now, most changes that can be
made to a model such as changing the objective, setting reaction
bounds, adding and removing reactions, is reversed upon exit when done
inside a context, see the updated
[getting started notebook](http://cobrapy.readthedocs.io/en/latest/getting_started.html).
Improved solution handling
Previously, cobra models lugged around their latest solution to enable
providing `reaction.{flux,reduced_cost}` (formerly
`reaction.{x,y}`). This was problematic because if the model had
changed since last optimization, then this would effectively give the
wrong result. On top of that, it was not easy to make a change,
optimize and get values, and then undo that change to the model
without having to copy the whole model object. To solve this, and many
similar problem, we have completely refactored `cobra.Solution` so
that `model.optimize()` now returns a solution and it is the user's
responsibility to manage this object. `reaction.flux` gets its values
directly from the `model.problem`. To sugar the new solution class,
fluxes, reduced costs, and shadow prices are now pandas series!
Sampling
Cobrapy now has flux sampling supported by
`cobra.flux_analysis.sampling` see
[the sampling notebook](http://cobrapy.readthedocs.io/en/latest/sampling.html).
Loopless models and solutions
Added implementations of
[CycleFreeFlux](http://dx.doi.org/10.1093/bioinformatics/btv096) and
the loopless model of
[Schellenberger et al.](http://dx.doi.org/10.1016/j.bpj.2010.12.3707). See
the
[notebook on loopless](http://cobrapy.readthedocs.io/en/latest/loopless.html)
and [simulating](http://cobrapy.readthedocs.io/en/latest/simulating.html)
DataFrames as return values
`flux_variability_analysis`, `single_{gene,rection}_deletion`,
`cobra.flux_analysis.sampling` and
`cobra.util.create_stoichiometric_array` now return a pandas data frame
instead of nested dicts as these are more convenient and fun to work
with. Pandas (and numpy) are therefore now hard requirements for
cobrapy, which should not be a problem for neither linux, windows or
mac users as there are reliable wheels for these packages now.
Model medium
`model.medium` is now a dict and setter for getting boundary feeding
reactions and their bounds
Fixes
- Handle multiple IDs in Matlab models
[345](https://github.com/opencobra/cobrapy/issues)
- `DictList.query` behavior changed so that attribute is `None` if the
search parameter is not a regex or string, to enable
`reactions.query(lambda x: x.boundary)`
- Set charge from notes if not defined elsewhere
[352](https://github.com/opencobra/cobrapy/issues)
- Warnings are no longer issued on package import if soft requirement
scipy, python-libsbml is not available.
Deprecated features
These features are now deprecated and slated for complete removal in
the next major cobrapy release.
- The whole `cobra.solver` module is now deprecated, see New features.
- `ArrayBasedModel` / `Model.to_array_based_model` are
deprecated. This formulation makes little sense when handing over
the matrix algebra to optlang, for the stoichiometry matrix (aka S),
see `cobra.util.array.create_stoichiometric_array`.
- `Metabolite.y` in favor of `Metabolite.shadow_price`
- `Model.add_reaction` in favor of `Model.add_reactions`
- `Reaction.x` in favor of `Reaction.flux`
- `Reaction.y` in favor of `Reaction.reduced_cost`
- `Solution.{x, y, x_dict, y_dict, f}` in favor of `Solution.{fluxes,
reduced_costs}`. The setters are also deprecated.
- `phenotype_phase_plane` in favor of `production_envelope`. The
plotting capabilities are deprecated, to be re-implemented somewhere
else.
- `convert_to_irreverible`, `revert_to_irreversible`, `canonical_form`
deprecated without replacement.
- `check_reaction_bounds` deprecated without replacement.
Backwards incompatible changes
- optknock was completely removed, users are advised to use cameo for
this functionality
- dual_problem was removed
- gapfilling algorithms were removed without replacement, to be
re-introduced in a different package
- `cobra.topology` was removed, possibly to be reintroduced in a
different package
- flux_variability_analysis results must be transformed to have them
work as the previous nested dict,
i.e. `flux_variability_analysis(model).T` should give behavior as
previously.
- In a major linting effort we renamed capitalized modules to lower-case,
e.g. `cobra.core.Model` to `cobra.core.model`. Imports from `cobra`
are unchanged though.
- objective coefficients of reactions can now only be set once the
reaction is attached to a model.
- `Reaction.{x,y}`, `Metabolite.y` are defunct for legacy solvers.