phylim should accept likelihood functions

[GavinHuttley]

The current implementation of the phylim app only accepts a model_result object. This does not work if the model result has come from a split_codon=True model. In the latter case, there is no single likelihood function so the line

tree = model_result.lf.tree

fails because model_result.lf is an ordered dictionary instead of a likelihood function.

Here are several possible solutions.

solution 1

To this is to implement a separate method that uses singledispatchmethod decorator to split the inference based on whether the input data is a model_result or a likelihood function object. A user could then extract the individual likelihood functions from a split_codon model result and pass each separately to the app.

This allows a user to handle custom cases themselves, but puts the burden of that solution on them.

solution 2

We examine whether the model result has multiple values and if it does we iterate over those values producing a single PhyloLimitRec as the output.

This is the lowest barrier to the user and does not require a change on the current phylim.main() method.

solution 3

We do both of the above. I think this is what's going to be required.

the workaround

For current users, if you encounter this problem here's the work around. In brief, you will need to create a separate model_result from each likelihood function in your split_codon=True result.

from cogent3.app.result import model_result
from cogent3.app.composable import define_app
from phylim.apps import phylim, PhyloLimitRec

@define_app
def phylim_split_codon(result: model_result, check_one: phylim) -> PhyloLimitRec:
    """checks individual likelihood functions from a split_codon model_result"""
    for k in range (1, 4):
        value = result[k]
        one = model_result(name=result.name, source=result.source)
        one['value'] = value
        checked = check_one(one)
        if not checked.is_identifiable:
            return checked
    return checked

[GavinHuttley]

@functools.singledispatch
def _get_lf(result: model_result) -> AlignmentLikelihoodFunction:
    # same as current main method
    if len(result) != 1:
        raise ValueError("Model result must contain exactly one likelihood function.")
    return result.lf

@_get_lf.register
def _(result: AlignmentLikelihoodFunction) -> AlignmentLikelihoodFunction:
    # same as current main method
    return result

@define_app
class phylim:
    """record psubs classes, identifiability, boundary values etc of a model_result.
    Args:
        "strict" controls the sensitivity for Identity matrix (I); if false,
        treat I as DLC.
    Return:
        PhyloLimitRec object
    """

    def __init__(self, strict: bool = False) -> None:
        self.strict = strict

    
    def main(self, model_result: model_result | AlignmentLikelihoodFunction) -> PhyloLimitRec:
        tree = _get_lf(model_result).tree  # type: ignore

        check_bound_app = check_fit_boundary()
        classify_psubs_app = classify_model_psubs()

        boundary_values = check_bound_app(model_result).vio # pylint: disable=not-callable
        psubs_labelled = classify_psubs_app(model_result) # pylint: disable=not-callable
        result = eval_identifiability(psubs_labelled, tree, self.strict)

[YapengLang]

• apps phylim, check_fit_boundary, classify_model_psubs support likelihood function
• 100% tests
• app/pipeline in phylim to handle cases that multiple lf in a single model_result

[YapengLang]

Update to the workaround

for users who want to check identifiability of model_result contains more than 1 likelihood function:

from cogent3.app.result import model_result
from cogent3.app.composable import define_app
from phylim.apps import phylim

@define_app
def phylim_split_codon(result: model_result, check_one: phylim) -> bool:
    """checks individual likelihood functions from a split_codon model_result"""
    checker = check_one()
    for value in result.values():
        checked = checker(value)
        if not checked.is_identifiable:
            return False
        
    return  True