Modelsearch#
The Modelsearch tool is a general tool to decide the best structural model given a base model and a search space of model features. The tool supports different algorithms and selection criteria.
Running#
The modelsearch tool is available both in Pharmpy/pharmr and from the command line.
To initiate modelsearch in Python/R:
from pharmpy.modeling import read_model
from pharmpy.tools import read_modelfit_results, run_modelsearch
start_model = read_model('path/to/model')
start_model_results = read_modelfit_results('path/to/model')
res = run_modelsearch(search_space='ABSORPTION([FO,ZO]);PERIPHERALS([0,1]);LAGTIME(ON)',
algorithm='reduced_stepwise',
model=start_model,
results=start_model_results,
iiv_strategy='absorption_delay',
rank_type='bic',
cutoff=None)
start_model <- read_model('path/to/model')
start_model_results <- read_modelfit_results('path/to/model')
res <- run_modelsearch(search_space='ABSORPTION([FO,ZO]);PERIPHERALS([0,1]);LAGTIME(ON)',
algorithm='reduced_stepwise',
model=start_model,
results=start_model_results,
iiv_strategy='absorption_delay',
rank_type='bic',
cutoff=NULL)
This will take an input model model
with search_space
as the search space, meaning adding one peripheral
compartment and lagtime will be tried. The tool will use the ‘reduced_stepwise’ search algorithm
. IIVs on
structural parameters (such as mean absorption time) will not be added to candidates since iiv_strategy
is
set to be ‘absorption_delay’. The candidate models will have BIC as the rank_type
with default cutoff
,
which for BIC is None/NULL.
To run modelsearch from the command line, the example code is redefined accordingly:
pharmpy run modelsearch path/to/model 'PERIPHERALS(1);LAGTIME(ON)' 'reduced_stepwise' --iiv_strategy 'absorption_delay' --rank_type 'bic'
Arguments#
For a more detailed description of each argument, see their respective chapter on this page.
Mandatory#
Argument |
Description |
---|---|
|
Search space to test |
|
Algorithm
to use (e.g. |
|
Start model |
|
ModelfitResults of the start model |
Optional#
Argument |
Description |
---|---|
|
Which selection criteria to rank models on, e.g. OFV (default is BIC) |
|
Cutoff for the ranking function, exclude models that are below cutoff (default is None/NULL) |
|
If/how IIV should be added to candidate models (default is to add to absorption delay parameters). See Adding IIV to the candidate models during search |
|
Strictness criteria for model selection. Default is “minimization_successful or (rounding_errors and sigdigs>= 0.1)” |
The search space#
The model feature search space is a set of all possible combinations of model features that is allowed for the final model. The supported features cover absorption, absorption delay, elimination, and distribution. The search space is given as a string with a specific grammar, according to the Model Feature Language (MFL) (see detailed description). If an attribute is not given, the default value for that attribute will be used as seen below. If the input model is not part of the given search space, a base model will be created. This is done by performing the least amount of transformations to the input model in order to make the base model a part of the given search space.
Category |
DEFAULT |
---|---|
ABSORPTION |
|
ELIMINATION |
|
PERIPHERALS |
|
TRANSITS |
|
LAGTIME |
|
The logical flow for the creation of the base model can be seen below. In summary, given an input model and a search space, the first step is to examine if the input model is a part of the search space. If so, the model features for the input model is filtered from the search space. As these are already present in the input model, they are not needed in the search space. After filtration, all transformations that are left will be examined. However, if the input model is not part of the search space, the base model is created by which will be part of the search space. Following this, the model features from the base model is filtered from the search space which leaves the transformations left to be examined.
Some examples of this workflow :
Search space |
Input model |
Base model |
Transformations to apply |
---|---|---|---|
ABSORPTION([FO,ZO]) ELIMINATION(FO) PERIPHERALS([1,2]) |
ABSORPTION(FO) ELIMINATION(ZO) TRANSITS(0) PERIPHERALS(0) LAGTIME(ON) |
ABSORPTION(FO) ELIMINATION(FO) TRANSITS(0) PERIPHERALS(1) LAGTIME(OFF) |
ABSORPTION(ZO) PERIPHERALS(2) |
ABSORPTION(FO) ELIMINATION(FO) TRANSITS([1,2]) |
ABSORPTION(FO) ELIMINATION(ZO) TRANSITS(0) PERIPHERALS(2) LAGTIME(OFF) |
ABSORPTION(FO) ELIMINATION(FO) TRANSITS(1) PERIPHERALS(0) LAGTIME(OFF) |
TRANSITS(2) |
ABSORPTION([FO,ZO]) ELIMINATION([FO,ZO,MM]) PERIPHERALS([0,1,2]) LAGTIME([OFF,ON]) |
ABSORPTION(FO) ELIMINATION(FO) TRANSITS(0) PERIPHERALS(0) LAGTIME(OFF) |
Not needed since input model is part of search space |
ABSORPTION(ZO) ELIMINATION([ZO,MM]) PERIPHERALS([1,2] LAGTIME(ON) |
Algorithms#
The tool can conduct the model search using different algorithms. The available algorithms can be seen in the table below.
Algorithm |
Description |
---|---|
|
All possible combinations of the search space are tested |
|
Add one feature in each step in all possible orders |
|
Add one feature in each step in all possible orders. After each feature layer, choose best model between models with same features |
Common behaviours between algorithms#
Feature combination exclusions#
Some combinations of features are excluded in algorithms that are performed stepwise, the following combinations are never run:
Feature A |
Feature B |
---|---|
ABSORPTION(ZO) |
TRANSITS |
ABSORPTION(SEQ-ZO-FO) |
TRANSITS |
ABSORPTION(SEQ-ZO-FO) |
LAGTIME(ON) |
ABSORPTION(INST) |
LAGTIME(ON) |
ABSORPTION(INST) |
TRANSITS |
LAGTIME(ON) |
TRANSITS |
Additionally, peripheral compartments are always run sequentially, i.e. the algorithm will never add more than one compartment at a given step. This is done in order to allow for better initial estimates from previous peripherals.
Exhaustive search#
An exhaustive
search will test all possible combinations of features in the search space. All candidate models will be
created simultaneously from the input model.
ABSORPTION(ZO)
ELIMINATION(MM)
PERIPHERALS(1)
Exhaustive stepwise search#
The exhaustive_stepwise
search applies features in a stepwise manner such that only one feature is changed at a time.
Between each step, the initial estimates from the new candidate model will be updated from the final estimates from the
previous step.
Reduced stepwise search#
The reduced_stepwise
search is similar to the exhaustive stepwise search, but after each layer it compares models with
the same features, where the compared models were obtained by adding the features in a different order. Next, the
algorithm uses the best model from each comparison as the basis for the next layer, where the subsequent feature is
added.
Adding IIV to the candidate models during search#
The iiv_strategy
option determines whether or not IIV on the PK parameters should be added to the candidate models.
The different strategies can be seen here:
Strategy |
Description |
---|---|
|
No IIVs are added during the search |
|
IIV is added to all structural parameters as diagonal |
|
IIV is added to all structural parameters, and all IIVs will be in a full block |
|
IIV is added only to the absorption delay parameter (default) |
For more information regarding which parameters are counted as structural parameters, see
pharmpy.modeling.add_pk_iiv()
.
Comparing and ranking candidates#
The supplied rank_type
will be used to compare a set of candidate models and rank them. Each candidate model
will be compared to the input model. A cutoff may also be provided if the user does not want to use the default.
The following rank functions are available:
Rank type |
Description |
---|---|
|
ΔOFV. Default is to not rank candidates with ΔOFV < cutoff (default 3.84) |
|
ΔAIC. Default is to rank all candidates if no cutoff is provided. |
|
ΔBIC (mixed). Default is to rank all candidates if no cutoff is provided. |
Information about how BIC is calculated can be found in pharmpy.modeling.calculate_bic()
.
The Modelsearch results#
The results object contains various summary tables which can be accessed in the results object, as well as files in .csv/.json format. The name of the selected best model (based on the input selection criteria) is also included.
Consider a modelsearch run with the search space of zero and first order absorption, adding zero or one peripheral compartment and lagtime:
res = run_modelsearch(search_space='ABSORPTION([FO,ZO]);PERIPHERALS([0,1]);LAGTIME(ON)',
algorithm='reduced_stepwise',
model=start_model,
results=start_model_results,
iiv_strategy='absorption_delay',
rank_type='bic',
cutoff=None)
res <- run_modelsearch(search_space='ABSORPTION([FO,ZO]);PERIPHERALS([0,1]);LAGTIME(ON)',
algorithm='reduced_stepwise',
model=start_model,
results=start_model_results,
iiv_strategy='absorption_delay',
rank_type='bic',
cutoff=NULL)
The summary_tool
table contains information such as which feature each model candidate has, the difference to the
start model (in this case comparing BIC), and final ranking:
description | n_params | d_params | dbic | bic | rank | parent_model | |
---|---|---|---|---|---|---|---|
model | |||||||
base | LAGTIME(ON) | 8 | 0 | 0.000000 | -1273.792080 | 1.0 | base |
modelsearch_run2 | PERIPHERALS(1) | 11 | 3 | -3.305500 | -1270.486580 | 2.0 | base |
modelsearch_run1 | ABSORPTION(ZO) | 9 | 1 | -9.452187 | -1264.339892 | 3.0 | base |
modelsearch_run3 | ABSORPTION(ZO);PERIPHERALS(1) | 11 | 3 | NaN | NaN | NaN | modelsearch_run1 |
modelsearch_run4 | PERIPHERALS(1);ABSORPTION(ZO) | 11 | 3 | NaN | NaN | NaN | modelsearch_run2 |
To see information about the actual model runs, such as minimization status, estimation time, and parameter estimates,
you can look at the summary_models
table. The table is generated with
pharmpy.tools.summarize_modelfit_results()
.
description | minimization_successful | errors_found | warnings_found | ofv | runtime_total | estimation_runtime | POP_CL_estimate | POP_CL_SE | POP_CL_RSE | ... | POP_MDT_RSE | IIV_MDT_estimate | IIV_MDT_SE | IIV_MDT_RSE | POP_QP1_estimate | POP_QP1_SE | POP_QP1_RSE | POP_VP1_estimate | POP_VP1_SE | POP_VP1_RSE | ||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
step | model | |||||||||||||||||||||
0 | input | True | 0 | 0 | -1292.186761 | 4.0 | 0.10 | 24.5328 | NaN | NaN | ... | NaN | NaN | NaN | NaN | NaN | NaN | NaN | NaN | NaN | NaN | |
1 | base | LAGTIME(ON) | True | 0 | 0 | -1313.362311 | 6.0 | 0.20 | 24.2927 | NaN | NaN | ... | NaN | NaN | NaN | NaN | NaN | NaN | NaN | NaN | NaN | NaN |
modelsearch_run1 | ABSORPTION(ZO) | True | 0 | 1 | -1305.577305 | 8.0 | 0.31 | 24.1880 | NaN | NaN | ... | NaN | 0.000001 | NaN | NaN | NaN | NaN | NaN | NaN | NaN | NaN | |
modelsearch_run2 | PERIPHERALS(1) | True | 0 | 1 | -1325.551467 | 8.0 | 0.67 | 24.7779 | NaN | NaN | ... | NaN | 0.000001 | NaN | NaN | 96.4011 | NaN | NaN | 47.4555 | NaN | NaN | |
modelsearch_run3 | ABSORPTION(ZO);PERIPHERALS(1) | False | 2 | 1 | -1308.965498 | 9.0 | 1.08 | 24.2319 | NaN | NaN | ... | NaN | 0.000001 | NaN | NaN | 1098.1900 | NaN | NaN | 77.8857 | NaN | NaN | |
modelsearch_run4 | PERIPHERALS(1);ABSORPTION(ZO) | False | 2 | 0 | -1308.040354 | 7.0 | 0.96 | 24.1760 | NaN | NaN | ... | NaN | 0.011456 | NaN | NaN | 766.4720 | NaN | NaN | 56.9327 | NaN | NaN |
6 rows × 40 columns
Finally, you can see a summary of different errors and warnings in summary_errors
.
See pharmpy.tools.summarize_errors()
for information on the content of this table.
time | message | |||
---|---|---|---|---|
model | category | error_no | ||
modelsearch_run1 | WARNING | 0 | 2024-10-22 13:09:42.519 | PARAMETER ESTIMATE IS NEAR ITS BOUNDARY |
modelsearch_run2 | WARNING | 0 | 2024-10-22 13:09:43.936 | PARAMETER ESTIMATE IS NEAR ITS BOUNDARY |
modelsearch_run3 | ERROR | 1 | 2024-10-22 13:09:52.316 | MINIMIZATION TERMINATED\nDUE TO MAX. NO. OF FUNCTION EVALUATIONS EXCEEDED |
2 | 2024-10-22 13:09:52.316 | NO. OF SIG. DIGITS UNREPORTABLE | ||
WARNING | 0 | 2024-10-22 13:09:52.316 | PARAMETER ESTIMATE IS NEAR ITS BOUNDARY | |
modelsearch_run4 | ERROR | 0 | 2024-10-22 13:09:53.750 | MINIMIZATION TERMINATED\nDUE TO MAX. NO. OF FUNCTION EVALUATIONS EXCEEDED |
1 | 2024-10-22 13:09:53.750 | NO. OF SIG. DIGITS UNREPORTABLE |