`ThinLayerPysces`: only fit a subset of the parameters

[jmrohwer]

Currently, the Pysces thinlayer fits all the parameters that are defined in the EnzymeML document to the measurement data:

PyEnzyme/pyenzyme/thinlayers/psyces.py

Lines 271 to 304 in 5f6cdac

	def _initialize_parameters(self):
	"""
	Initializes lmfit Parameters instance with model parameters.
	Returns:
	lmfit.Parameters: Parameters object ready for optimization.
	Raises:
	ValueError: If a parameter has neither an initial value nor a value attribute.
	"""
	# Initialize lmfit parameters
	parameters = lmfit.Parameters()
	# Add global parameters
	for param in self.enzmldoc.parameters:
	# Build kwargs dictionary with conditional assignments
	kwargs = {
	**({"min": param.lower_bound} if param.lower_bound is not None else {}),
	**({"max": param.upper_bound} if param.upper_bound is not None else {}),
	}
	# Determine parameter value
	if param.value:
	kwargs["value"] = param.value
	elif param.initial_value:
	kwargs["value"] = param.initial_value
	else:
	raise ValueError(
	f"Neither initial_value nor value given for parameter {param.name} in global parameters"
	)
	parameters.add(param.symbol, **kwargs)
	return parameters

I have a use case, however, where I only want to fit a subset of the parameters, because some parameters have been previously estimated by another method (plate assay) and the NMR data does not have sufficient information to uniquely fit all of the parameters.

The lmfit parameters dictionary as an option vary which can be assigned to a parameter, setting this to False causes this parameter not to be optimised but kept at its original value. I propose to add another keyword to an enzmldoc.parameter, e.g. vary or fit, set to True by default, but which can be overridden so that that the parameter does not get fitted by lmfit. The code quoted above can then be modified to check for this keyword.

I can work on this once #82 is fixed and the Pysces ThinLayer optimisation works in general.

[JR-1991]

We could utilize the existing constant attribute of Parameter to indicate that this parameter should not be fitted. Alternatively, as a more flexible solution, we could provide a list of parameters that should not be varied. What are your thoughts on these options?

[jmrohwer]

I am not in favour of using the constant attribute of Parameter. If you read the SBML specification (p. 53-55, specifically §4.7.4), specifically to quote:

The Parameter object has a mandatory boolean attribute named constant that indicates whether the parameter’s value can vary during a simulation

In the context of EnzymeML, all parameters are fixed in terms of the underlying SBML model; the variables are the species. We do "vary" the parameters but in the context of an optimisation to fit the best value. Once this value has been found in the context of the underlying data, the parameter is actually fixed. Also, during a single simulation run of the optimisation process, the parameter is fixed, thus constant. The optimisation consists of repeated runs, each time "fixing" the parameter at a different value.

Why would you want to provide a list of parameters that should not be varied, if this could be handled with a simple attribute as suggested in my original message? Seems unnecessarily complex. param already has attributes value, initial_value, lower_bound, etc. So why not add one more attribute param.fit (default True) to indicate if the parameter should be fitted to the data.

[JR-1991]

Fair, that makes sense. I have opened a PR at the EnzymeML specification repository to distribute the change to all of our libraries:

• Introduce 'fit' parameter for optimization context enzymeml-specifications#27

[jmrohwer]

Implemented in #85