Skip to content
Snippets Groups Projects
Commit 31585737 authored by Tim O'Donnell's avatar Tim O'Donnell Committed by GitHub
Browse files

Merge pull request #107 from hammerlab/pypi0.9.2 

update setup.py
parents ac709bb7 0eda4dcb
No related branches found
No related tags found
No related merge requests found
Expand all Hide whitespace changes
Inline Side-by-side
Showing
with 13 additions and 1764 deletions
.travis.yml
+ 1
− 1
View file @ 31585737
......@@ -29,7 +29,7 @@ addons:
install:
- >
conda create -q -n test-environment python=$TRAVIS_PYTHON_VERSION
numpy scipy nose pandas matplotlib mkl-service tensorflow
numpy scipy nose pandas matplotlib mkl-service tensorflow theano
- source activate test-environment
- pip install pypandoc pylint
- pip install -r requirements.txt
......
README.md
+ 12
− 8
View file @ 31585737
......@@ -62,17 +62,21 @@ You may also needs to `pip install theano`.
```shell
$ mhcflurry-predict --alleles HLA-A0201 HLA-A0301 --peptides SIINFEKL SIINFEKD SIINFEKQ
allele,peptide,mhcflurry_prediction,mhcflurry_prediction_low,mhcflurry_prediction_high
HLA-A0201,SIINFEKL,6029.079749556217,4474.10333152741,7771.2922076773575
HLA-A0201,SIINFEKD,18950.310303704624,15317.127851792027,22490.05728778504
HLA-A0201,SIINFEKQ,18776.978315260818,14899.359763218705,22314.737180384865
HLA-A0301,SIINFEKL,25589.66470369661,22962.4956808368,29395.86949262485
HLA-A0301,SIINFEKD,25753.619337400796,22851.89399578629,29347.659901990868
HLA-A0301,SIINFEKQ,26870.51318688641,24198.39885651102,30364.15208364084
HLA-A0201,SIINFEKL,5326.541919062165,3757.86675352994,7461.37693353508
HLA-A0201,SIINFEKD,18763.70298522213,13140.82000240037,23269.82139560844
HLA-A0201,SIINFEKQ,18620.10057358322,13096.425874678192,23223.148184869413
HLA-A0301,SIINFEKL,24481.726678691946,21035.52779725433,27245.371837497867
HLA-A0301,SIINFEKD,24687.529360239587,21582.590014592537,27749.39869616437
HLA-A0301,SIINFEKQ,25923.062203902562,23522.5793450799,28079.456657427705
```
The predictions returned are affinities (KD) in nM. The `prediction_low` and
`prediction_high` fields give the 5-95 percentile predictions across the models
in the ensemble.
`prediction_high` fields give the 5-95 percentile predictions across the models
in the ensemble. The predictions above were generated with MHCflurry 0.9.2.
Your exact predictions may vary slightly from these (up to about 1 nM)
depending on the Keras backend in use and other numerical details even if you
match the MHCflurry version, whereas different versions of MHCflurry give
results that are considerably different from these.
You can also specify the input and output as CSV files.
Run `mhcflurry-predict -h` for details.
......
RELEASING.md deleted 100644 → 0
+ 0
− 21
View file @ ac709bb7
# Releasing MHCflurry
We currently don't have a PyPI release but that will change soon (and these instructions will be updated indicating how to push updates). For now the only "releasing" to deal with is changing the downloadable models and data.
## Changing the downloadable models or data
We publish our downloadable models and data as files associated with a GitHub release. Since we need to refer to the URLs for these files in [downloads.yml](mhcflurry/downloads.yml), which is checked into the repo, updating the models requires a few steps: we have to make a preliminary GitHub release containing the new models and data as GitHub attached files, update the code to point to these files, wait for travis to run, merge the PR, then modify the release's tag to now point to the new master. Here are these steps in more detail:
* Make a new release by going [here](https://github.com/hammerlab/mhcflurry/releases/new). The tag should be the version of MHCflurry you are releasing. Make sure you check "This is a pre-release." It actually doesn't matter what commit the tag is associated with as we will change it later, but you might as well make it point to HEAD of the branch you are working from. Attach your generated downloads to the release.
* Modify [downloads.yml](mhcflurry/downloads.yml) to point to the URLs of your files above. Commit and push your changes to your branch.
* When travis has suceeded and code review is complete, merge your PR to master.
* Now *change* the GitHub release's tag to point to the current latest master. Based on this stackoverflow [answer](http://stackoverflow.com/questions/24849362/change-connected-commit-on-release-github) you can run (from a checkout of the updated master branch):
```
git tag -f -a 0.0.1
git push -f --tags
```
mhcflurry/antigen_presentation/README.md deleted 100644 → 0
+ 0
− 5
View file @ ac709bb7
# Prediction of antigen presention
This submodule contains predictors for naturally presented MHC ligands. These predictors are typically trained on peptides eluted from cell surfaces and identified with mass-spec. The models combine MHC binding affinity with cleavage prediction and the level of expression of transcripts containing the given peptide.
This is a work in progress and not ready for production use.
mhcflurry/antigen_presentation/__init__.py deleted 100644 → 0
+ 0
− 11
View file @ ac709bb7
from .presentation_model import PresentationModel, build_presentation_models
from .percent_rank_transform import PercentRankTransform
from . import presentation_component_models, decoy_strategies
__all__ = [
"PresentationModel",
"build_presentation_models",
"PercentRankTransform",
"presentation_component_models",
"decoy_strategies",
]
mhcflurry/antigen_presentation/decoy_strategies/__init__.py deleted 100644 → 0
+ 0
− 9
View file @ ac709bb7
from .decoy_strategy import DecoyStrategy
from .same_transcripts_as_hits import SameTranscriptsAsHits
from .uniform_random import UniformRandom
__all__ = [
"DecoyStrategy",
"SameTranscriptsAsHits",
"UniformRandom",
]
mhcflurry/antigen_presentation/decoy_strategies/decoy_strategy.py deleted 100644 → 0
+ 0
− 57
View file @ ac709bb7
import pandas
class DecoyStrategy(object):
"""
A mechanism for selecting decoys (non-hit peptides) given hits (
peptides detected via mass-spec).
Subclasses should override either decoys() or decoys_for_experiment().
Whichever one is not overriden is implemented using the other.
"""
def __init__(self):
pass
def decoys(self, hits_df):
"""
Given a df of hits with columns 'experiment_name' and 'peptide',
return a df with the same structure giving decoys.
Subclasses should override either this or decoys_for_experiment()
"""
assert 'experiment_name' in hits_df.columns
assert 'peptide' in hits_df.columns
assert len(hits_df) > 0
grouped = hits_df.groupby("experiment_name")
dfs = []
for (experiment_name, sub_df) in grouped:
decoys = self.decoys_for_experiment(
experiment_name,
sub_df.peptide.values)
df = pandas.DataFrame({
'peptide': decoys,
})
df["experiment_name"] = experiment_name
dfs.append(df)
return pandas.concat(dfs, ignore_index=True)
def decoys_for_experiment(self, experiment_name, hit_list):
"""
Return decoys for a single experiment.
Parameters
------------
experiment_name : string
hit_list : list of string
List of hits
"""
# prevent infinite recursion:
assert self.decoys is not DecoyStrategy.decoys
hits_df = pandas.DataFrame({'peptide': hit_list})
hits_df["experiment_name"] = experiment_name
return self.decoys(hits_df)
mhcflurry/antigen_presentation/decoy_strategies/same_transcripts_as_hits.py deleted 100644 → 0
+ 0
− 58
View file @ ac709bb7
import numpy
from .decoy_strategy import DecoyStrategy
class SameTranscriptsAsHits(DecoyStrategy):
"""
Decoy strategy that selects decoys from the same transcripts the
hits come from. The transcript for each hit is taken to be the
transcript containing the hit with the the highest expression for
the given experiment.
Parameters
------------
experiment_to_expression_group : dict of string -> string
Maps experiment names to expression groups.
peptides_and_transcripts: pandas.DataFrame
Must have columns 'peptide' and 'transcript', index unimportant.
peptide_to_expression_group_to_transcript : pandas.DataFrame
Indexed by peptides, columns are expression groups. Values
give transcripts to use.
decoys_per_hit : int
"""
def __init__(
self,
experiment_to_expression_group,
peptides_and_transcripts,
peptide_to_expression_group_to_transcript,
decoys_per_hit=10):
DecoyStrategy.__init__(self)
assert decoys_per_hit > 0
self.experiment_to_expression_group = experiment_to_expression_group
self.peptides_and_transcripts = peptides_and_transcripts
self.peptide_to_expression_group_to_transcript = (
peptide_to_expression_group_to_transcript)
self.decoys_per_hit = decoys_per_hit
def decoys_for_experiment(self, experiment_name, hit_list):
assert len(hit_list) > 0, "No hits for %s" % experiment_name
expression_group = self.experiment_to_expression_group[experiment_name]
transcripts = self.peptide_to_expression_group_to_transcript.ix[
hit_list, expression_group
]
assert len(transcripts) > 0, experiment_name
universe = self.peptides_and_transcripts.ix[
self.peptides_and_transcripts.transcript.isin(transcripts) &
(~ self.peptides_and_transcripts.peptide.isin(hit_list))
].peptide.values
assert len(universe) > 0, experiment_name
return numpy.random.choice(
universe,
replace=True,
size=self.decoys_per_hit * len(hit_list))
mhcflurry/antigen_presentation/decoy_strategies/uniform_random.py deleted 100644 → 0
+ 0
− 21
View file @ ac709bb7
import numpy
from .decoy_strategy import DecoyStrategy
class UniformRandom(DecoyStrategy):
"""
Decoy strategy that selects decoys randomly from a provided universe
of peptides.
"""
def __init__(self, all_peptides, decoys_per_hit=999):
DecoyStrategy.__init__(self)
self.all_peptides = set(all_peptides)
self.decoys_per_hit = decoys_per_hit
def decoys_for_experiment(self, experiment_name, hit_list):
decoy_pool = self.all_peptides.difference(set(hit_list))
return numpy.random.choice(
list(decoy_pool),
replace=True,
size=self.decoys_per_hit * len(hit_list))
mhcflurry/antigen_presentation/percent_rank_transform.py deleted 100644 → 0
+ 0
− 39
View file @ ac709bb7
import numpy
class PercentRankTransform(object):
"""
Transform arbitrary values into percent ranks.
"""
def __init__(self, n_bins=1e5):
self.n_bins = int(n_bins)
self.cdf = None
self.bin_edges = None
def fit(self, values):
"""
Fit the transform using the given values, which are used to
establish percentiles.
"""
assert self.cdf is None
assert self.bin_edges is None
assert len(values) > 0
(hist, self.bin_edges) = numpy.histogram(values, bins=self.n_bins)
self.cdf = numpy.ones(len(hist) + 3) * numpy.nan
self.cdf[0] = 0.0
self.cdf[1] = 0.0
self.cdf[-1] = 100.0
numpy.cumsum(hist * 100.0 / numpy.sum(hist), out=self.cdf[2:-1])
assert not numpy.isnan(self.cdf).any()
def transform(self, values):
"""
Return percent ranks (range [0, 100]) for the given values.
"""
assert self.cdf is not None
assert self.bin_edges is not None
indices = numpy.searchsorted(self.bin_edges, values)
result = self.cdf[indices]
assert len(result) == len(values)
return result
mhcflurry/antigen_presentation/presentation_component_models/__init__.py deleted 100644 → 0
+ 0
− 18
View file @ ac709bb7
from .presentation_component_model import PresentationComponentModel
from .expression import Expression
from .mhcflurry_released import MHCflurryReleased
from .mhcflurry_trained_on_hits import MHCflurryTrainedOnHits
from .fixed_affinity_predictions import FixedAffinityPredictions
from .fixed_per_peptide_quantity import FixedPerPeptideQuantity
from .fixed_per_peptide_and_transcript_quantity import (
FixedPerPeptideAndTranscriptQuantity)
__all__ = [
"PresentationComponentModel",
"Expression",
"MHCflurryReleased",
"MHCflurryTrainedOnHits",
"FixedAffinityPredictions",
"FixedPerPeptideQuantity",
"FixedPerPeptideAndTranscriptQuantity",
]
mhcflurry/antigen_presentation/presentation_component_models/expression.py deleted 100644 → 0
+ 0
− 45
View file @ ac709bb7
from .presentation_component_model import PresentationComponentModel
from ...common import assert_no_null
class Expression(PresentationComponentModel):
"""
Model input for transcript expression.
Parameters
------------
experiment_to_expression_group : dict of string -> string
Maps experiment names to expression groups.
expression_values : pandas.DataFrame
Columns should be expression groups. Indices should be peptide.
"""
def __init__(
self, experiment_to_expression_group, expression_values, **kwargs):
PresentationComponentModel.__init__(self, **kwargs)
assert all(
group in expression_values.columns
for group in experiment_to_expression_group.values())
assert_no_null(expression_values)
self.experiment_to_expression_group = experiment_to_expression_group
self.expression_values = expression_values
def column_names(self):
return ["expression"]
def requires_fitting(self):
return False
def predict_for_experiment(self, experiment_name, peptides):
expression_group = self.experiment_to_expression_group[experiment_name]
return {
"expression": (
self.expression_values.ix[peptides, expression_group]
.values)
}
mhcflurry/antigen_presentation/presentation_component_models/fixed_affinity_predictions.py deleted 100644 → 0
+ 0
− 62
View file @ ac709bb7
from .presentation_component_model import PresentationComponentModel
from ...common import assert_no_null
class FixedAffinityPredictions(PresentationComponentModel):
"""
Parameters
------------
experiment_to_alleles : dict: string -> string list
Normalized allele names for each experiment.
panel : pandas.Panel
Dimensions should be:
- "value", "percentile_rank" (IC50 and percent rank)
- peptide (string)
- allele (string)
name : string
Used to name output columns and in debug messages
"""
def __init__(
self,
experiment_to_alleles,
panel,
name='precomputed',
**kwargs):
PresentationComponentModel.__init__(self, **kwargs)
self.experiment_to_alleles = experiment_to_alleles
for key in panel.items:
assert_no_null(panel[key])
self.panel = panel
self.name = name
def column_names(self):
return [
"%s_affinity" % self.name,
"%s_percentile_rank" % self.name
]
def requires_fitting(self):
return False
def predict_min_across_alleles(self, alleles, peptides):
return {
("%s_affinity" % self.name): (
self.panel
.value[alleles]
.min(axis=1)
.ix[peptides].values),
("%s_percentile_rank" % self.name): (
self.panel
.percentile_rank[alleles]
.min(axis=1)
.ix[peptides].values)
}
def predict_for_experiment(self, experiment_name, peptides):
alleles = self.experiment_to_alleles[experiment_name]
return self.predict_min_across_alleles(alleles, peptides)
mhcflurry/antigen_presentation/presentation_component_models/fixed_per_peptide_and_transcript_quantity.py deleted 100644 → 0
+ 0
− 93
View file @ ac709bb7
import logging
from .presentation_component_model import PresentationComponentModel
from ...common import assert_no_null
class FixedPerPeptideAndTranscriptQuantity(PresentationComponentModel):
"""
Model input for arbitrary fixed (i.e. not fitted) quantities that
depend only on the peptide and the transcript it comes from, which
is taken to be the most-expressed transcript in the experiment.
Motivating example: netChop cleavage predictions.
Parameters
------------
name : string
Name for this final model input. Used in debug messages.
experiment_to_expression_group : dict of string -> string
Maps experiment names to expression groups.
top_transcripts : pandas.DataFrame
Columns should be expression groups. Indices should be peptide. Values
should be transcript names.
df : pandas.DataFrame
Must have columns 'peptide' and 'transcript'. Remaining columns are
the values emitted by this model input.
"""
def __init__(
self,
name,
experiment_to_expression_group,
top_transcripts,
df,
**kwargs):
PresentationComponentModel.__init__(self, **kwargs)
self.name = name
self.experiment_to_expression_group = experiment_to_expression_group
self.top_transcripts = top_transcripts.copy()
self.df = df.drop_duplicates(['peptide', 'transcript'])
# This hack seems to be faster than using a multindex.
self.df.index = self.df.peptide.str.cat(self.df.transcript, sep=":")
del self.df["peptide"]
del self.df["transcript"]
assert_no_null(self.df)
df_set = set(self.df.index)
missing = set()
for expression_group in self.top_transcripts.columns:
self.top_transcripts[expression_group] = (
self.top_transcripts.index.str.cat(
self.top_transcripts[expression_group],
sep=":"))
missing.update(
set(self.top_transcripts[expression_group]).difference(df_set))
if missing:
logging.warn(
"%s: missing %d (peptide, transcript) pairs from df: %s" % (
self.name,
len(missing),
sorted(missing)[:1000]))
def column_names(self):
return list(self.df.columns)
def requires_fitting(self):
return False
def predict_for_experiment(self, experiment_name, peptides):
expression_group = self.experiment_to_expression_group[experiment_name]
indices = self.top_transcripts.ix[peptides, expression_group]
assert len(indices) == len(peptides)
sub_df = self.df.ix[indices]
assert len(sub_df) == len(peptides)
result = {}
for col in self.column_names():
result_series = sub_df[col]
num_nulls = result_series.isnull().sum()
if num_nulls > 0:
logging.warning("%s: mean-filling for %d nulls" % (
self.name, num_nulls))
result_series = result_series.fillna(self.df[col].mean())
result[col] = result_series.values
return result
mhcflurry/antigen_presentation/presentation_component_models/fixed_per_peptide_quantity.py deleted 100644 → 0
+ 0
− 40
View file @ ac709bb7
from .presentation_component_model import PresentationComponentModel
from ...common import assert_no_null
class FixedPerPeptideQuantity(PresentationComponentModel):
"""
Model input for arbitrary fixed (i.e. not fitted) quantities that
depend only on the peptide. Motivating example: Mike's cleavage
predictions.
Parameters
------------
name : string
Name for this final model input. Used in debug messages.
df : pandas.DataFrame
index must be named 'peptide'. The columns of the dataframe are
the columns emitted by this final modle input.
"""
def __init__(self, name, df, **kwargs):
PresentationComponentModel.__init__(self, **kwargs)
self.name = name
assert df.index.name == "peptide"
assert_no_null(df)
self.df = df
def column_names(self):
return list(self.df.columns)
def requires_fitting(self):
return False
def predict(self, peptides_df):
sub_df = self.df.ix[peptides_df.peptide]
return dict(
(col, sub_df[col].values)
for col in self.column_names())
mhcflurry/antigen_presentation/presentation_component_models/mhc_binding_component_model_base.py deleted 100644 → 0
+ 0
− 226
View file @ ac709bb7
import logging
import pandas
from numpy import array
from ...common import dataframe_cryptographic_hash
from .presentation_component_model import PresentationComponentModel
from ..decoy_strategies import SameTranscriptsAsHits
from ..percent_rank_transform import PercentRankTransform
class MHCBindingComponentModelBase(PresentationComponentModel):
"""
Base class for single-allele MHC binding predictors.
Parameters
------------
predictor_name : string
used on column name. Example: 'vanilla'
experiment_to_alleles : dict: string -> string list
Normalized allele names for each experiment.
experiment_to_expression_group : dict of string -> string
Maps experiment names to expression groups.
transcripts : pandas.DataFrame
Index is peptide, columns are expression groups, values are
which transcript to use for the given peptide.
Not required if decoy_strategy specified.
peptides_and_transcripts : pandas.DataFrame
Dataframe with columns 'peptide' and 'transcript'
Not required if decoy_strategy specified.
decoy_strategy : decoy_strategy.DecoyStrategy
how to pick decoys. If not specified peptides_and_transcripts and
transcripts must be specified.
fallback_predictor : function: (allele, peptides) -> predictions
Used when missing an allele.
iedb_dataset : mhcflurry.AffinityMeasurementDataset
IEDB data for this allele. If not specified no iedb data is used.
decoys_per_hit : int
random_peptides_for_percent_rank : list of string
If specified, then percentile rank will be calibrated and emitted
using the given peptides.
**kwargs : dict
passed to PresentationComponentModel()
"""
def __init__(
self,
predictor_name,
experiment_to_alleles,
experiment_to_expression_group=None,
transcripts=None,
peptides_and_transcripts=None,
decoy_strategy=None,
fallback_predictor=None,
iedb_dataset=None,
decoys_per_hit=10,
random_peptides_for_percent_rank=None,
**kwargs):
PresentationComponentModel.__init__(self, **kwargs)
self.predictor_name = predictor_name
self.experiment_to_alleles = experiment_to_alleles
self.fallback_predictor = fallback_predictor
self.iedb_dataset = iedb_dataset
self.fit_alleles = set()
if decoy_strategy is None:
assert peptides_and_transcripts is not None
assert transcripts is not None
self.decoy_strategy = SameTranscriptsAsHits(
experiment_to_expression_group=experiment_to_expression_group,
peptides_and_transcripts=peptides_and_transcripts,
peptide_to_expression_group_to_transcript=transcripts,
decoys_per_hit=decoys_per_hit)
else:
self.decoy_strategy = decoy_strategy
if random_peptides_for_percent_rank is None:
self.percent_rank_transforms = None
self.random_peptides_for_percent_rank = None
else:
self.percent_rank_transforms = {}
self.random_peptides_for_percent_rank = array(
random_peptides_for_percent_rank)
def stratification_groups(self, hits_df):
return [
self.experiment_to_alleles[e][0]
for e in hits_df.experiment_name
]
def column_name_value(self):
return "%s_value" % self.predictor_name
def column_name_percentile_rank(self):
return "%s_percentile_rank" % self.predictor_name
def column_names(self):
columns = [self.column_name_value()]
if self.percent_rank_transforms is not None:
columns.append(self.column_name_percentile_rank())
return columns
def requires_fitting(self):
return True
def fit_percentile_rank_if_needed(self, alleles):
for allele in alleles:
if allele not in self.percent_rank_transforms:
logging.info('fitting percent rank for allele: %s' % allele)
self.percent_rank_transforms[allele] = PercentRankTransform()
self.percent_rank_transforms[allele].fit(
self.predict_affinity_for_allele(
allele,
self.random_peptides_for_percent_rank))
def fit(self, hits_df):
assert 'experiment_name' in hits_df.columns
assert 'peptide' in hits_df.columns
if 'hit' in hits_df.columns:
assert (hits_df.hit == 1).all()
grouped = hits_df.groupby("experiment_name")
for (experiment_name, sub_df) in grouped:
self.fit_to_experiment(experiment_name, sub_df.peptide.values)
# No longer required after fitting.
self.decoy_strategy = None
self.iedb_dataset = None
def fit_allele(self, allele, hit_list, decoys_list):
raise NotImplementedError()
def predict_allele(self, allele, peptide_list):
raise NotImplementedError()
def supports_predicting_allele(self, allele):
raise NotImplementedError()
def fit_to_experiment(self, experiment_name, hit_list):
assert len(hit_list) > 0
alleles = self.experiment_to_alleles[experiment_name]
if len(alleles) != 1:
raise ValueError("Monoallelic data required")
(allele,) = alleles
decoys = self.decoy_strategy.decoys_for_experiment(
experiment_name, hit_list)
self.fit_allele(allele, hit_list, decoys)
self.fit_alleles.add(allele)
def predict_affinity_for_allele(self, allele, peptides):
if self.cached_predictions is None:
cache_key = None
cached_result = None
else:
cache_key = (
allele,
dataframe_cryptographic_hash(pandas.Series(peptides)))
cached_result = self.cached_predictions.get(cache_key)
if cached_result is not None:
print("Cache hit in predict_affinity_for_allele: %s %s %s" % (
allele, str(self), id(cached_result)))
return cached_result
else:
print("Cache miss in predict_affinity_for_allele: %s %s" % (
allele, str(self)))
if self.supports_predicting_allele(allele):
result = self.predict_allele(allele, peptides)
elif self.fallback_predictor:
print("Falling back on allee %s" % allele)
result = self.fallback_predictor(allele, peptides)
else:
raise ValueError("No model for allele: %s" % allele)
if self.cached_predictions is not None:
self.cached_predictions[cache_key] = result
return result
def predict_for_experiment(self, experiment_name, peptides):
peptides_deduped = pandas.unique(peptides)
print(len(peptides_deduped))
alleles = self.experiment_to_alleles[experiment_name]
predictions = pandas.DataFrame(index=peptides_deduped)
for allele in alleles:
predictions[allele] = self.predict_affinity_for_allele(
allele, peptides_deduped)
result = {
self.column_name_value(): (
predictions.min(axis=1).ix[peptides].values)
}
if self.percent_rank_transforms is not None:
self.fit_percentile_rank_if_needed(alleles)
percentile_ranks = pandas.DataFrame(index=peptides_deduped)
for allele in alleles:
percentile_ranks[allele] = (
self.percent_rank_transforms[allele]
.transform(predictions[allele].values))
result[self.column_name_percentile_rank()] = (
percentile_ranks.min(axis=1).ix[peptides].values)
assert all(len(x) == len(peptides) for x in result.values()), (
"Result lengths don't match peptide lengths. peptides=%d, "
"peptides_deduped=%d, %s" % (
len(peptides),
len(peptides_deduped),
", ".join(
"%s=%d" % (key, len(value))
for (key, value) in result.items())))
return result
mhcflurry/antigen_presentation/presentation_component_models/mhcflurry_released.py deleted 100644 → 0
+ 0
− 102
View file @ ac709bb7
import logging
import numpy
import pandas
from mhcnames import normalize_allele_name
from ..percent_rank_transform import PercentRankTransform
from ...encodable_sequences import EncodableSequences
from .presentation_component_model import PresentationComponentModel
from ...class1_affinity_prediction.class1_affinity_predictor import (
Class1AffinityPredictor)
class MHCflurryReleased(PresentationComponentModel):
"""
Final model input that uses the standard downloaded MHCflurry models.
Parameters
------------
experiment_to_alleles : dict: string -> string list
Normalized allele names for each experiment.
random_peptides_for_percent_rank : list of string
If specified, then percentile rank will be calibrated and emitted
using the given peptides.
predictor : Class1EnsembleMultiAllelePredictor-like object
Predictor to use.
"""
def __init__(
self,
experiment_to_alleles,
random_peptides_for_percent_rank=None,
predictor=None,
predictor_name="mhcflurry_released",
**kwargs):
PresentationComponentModel.__init__(self, **kwargs)
self.experiment_to_alleles = experiment_to_alleles
if predictor is None:
predictor = Class1AffinityPredictor.load()
self.predictor = predictor
self.predictor_name = predictor_name
if random_peptides_for_percent_rank is None:
self.percent_rank_transforms = None
self.random_peptides_for_percent_rank = None
else:
self.percent_rank_transforms = {}
self.random_peptides_for_percent_rank = numpy.array(
random_peptides_for_percent_rank)
def column_names(self):
columns = [self.predictor_name + '_affinity']
if self.percent_rank_transforms is not None:
columns.append(self.predictor_name + '_percentile_rank')
return columns
def requires_fitting(self):
return False
def fit_percentile_rank_if_needed(self, alleles):
for allele in alleles:
if allele not in self.percent_rank_transforms:
logging.info('fitting percent rank for allele: %s' % allele)
self.percent_rank_transforms[allele] = PercentRankTransform()
self.percent_rank_transforms[allele].fit(
self.predictor.predict(
allele=allele,
peptides=self.random_peptides_for_percent_rank))
def predict_min_across_alleles(self, alleles, peptides):
alleles = list(set([
normalize_allele_name(allele)
for allele in alleles
]))
peptides = EncodableSequences.create(peptides)
df = pandas.DataFrame()
df["peptide"] = peptides.sequences
for allele in alleles:
df[allele] = self.predictor.predict(peptides, allele=allele)
result = {
self.predictor_name + '_affinity': (
df[list(df.columns)[1:]].min(axis=1))
}
if self.percent_rank_transforms is not None:
self.fit_percentile_rank_if_needed(alleles)
percentile_ranks = pandas.DataFrame(index=df.index)
for allele in alleles:
percentile_ranks[allele] = (
self.percent_rank_transforms[allele]
.transform(df[allele].values))
result[self.predictor_name + '_percentile_rank'] = (
percentile_ranks.min(axis=1).values)
for (key, value) in result.items():
assert len(value) == len(peptides), (len(peptides), result)
return result
def predict_for_experiment(self, experiment_name, peptides):
alleles = self.experiment_to_alleles[experiment_name]
return self.predict_min_across_alleles(alleles, peptides)
mhcflurry/antigen_presentation/presentation_component_models/mhcflurry_trained_on_hits.py deleted 100644 → 0
+ 0
− 96
View file @ ac709bb7
from copy import copy
import pandas
from numpy import log, exp, nanmean
from ...class1_affinity_prediction import Class1AffinityPredictor
from mhcnames import normalize_allele_name
from .mhc_binding_component_model_base import MHCBindingComponentModelBase
class MHCflurryTrainedOnHits(MHCBindingComponentModelBase):
"""
Final model input that is a mhcflurry predictor trained on mass-spec
hits and, optionally, affinity measurements (for example from IEDB).
Parameters
------------
mhcflurry_hyperparameters : dict
hit_affinity : float
nM affinity to use for hits
decoy_affinity : float
nM affinity to use for decoys
**kwargs : dict
Passed to MHCBindingComponentModel()
"""
def __init__(
self,
mhcflurry_hyperparameters={},
hit_affinity=100,
decoy_affinity=20000,
ensemble_size=1,
**kwargs):
MHCBindingComponentModelBase.__init__(self, **kwargs)
self.mhcflurry_hyperparameters = mhcflurry_hyperparameters
self.hit_affinity = hit_affinity
self.decoy_affinity = decoy_affinity
self.ensemble_size = ensemble_size
self.predictor = Class1AffinityPredictor()
def combine_ensemble_predictions(self, column_name, values):
# Geometric mean
return exp(nanmean(log(values), axis=1))
def supports_predicting_allele(self, allele):
return allele in self.predictor.supported_alleles
def fit_allele(self, allele, hit_list, decoys_list):
allele = normalize_allele_name(allele)
hit_list = set(hit_list)
df = pandas.DataFrame({
"peptide": sorted(set(hit_list).union(decoys_list))
})
df["allele"] = allele
df["species"] = "human"
df["affinity"] = ((
~df.peptide.isin(hit_list))
.astype(float) * (
self.decoy_affinity - self.hit_affinity) + self.hit_affinity)
df["sample_weight"] = 1.0
df["peptide_length"] = 9
self.predictor.fit_allele_specific_predictors(
n_models=self.ensemble_size,
architecture_hyperparameters=self.mhcflurry_hyperparameters,
allele=allele,
peptides=df.peptide.values,
affinities=df.affinity.values,
)
def predict_allele(self, allele, peptides_list):
return self.predictor.predict(peptides=peptides_list, allele=allele)
def get_fit(self):
return {
'model': 'MHCflurryTrainedOnMassSpec',
'predictor': self.predictor,
}
def restore_fit(self, fit_info):
fit_info = dict(fit_info)
self.predictor = fit_info.pop('predictor')
model = fit_info.pop('model')
assert model == 'MHCflurryTrainedOnMassSpec', model
assert not fit_info, "Extra info in fit: %s" % str(fit_info)
def clone(self):
result = copy(self)
result.reset_cache()
result.predictor = copy(result.predictor)
return result
mhcflurry/antigen_presentation/presentation_component_models/presentation_component_model.py deleted 100644 → 0
+ 0
− 269
View file @ ac709bb7
import weakref
from copy import copy
import numpy
import pandas
from ...common import (
dataframe_cryptographic_hash, assert_no_null, freeze_object)
def cache_dict_for_policy(policy):
if policy == "weak":
return weakref.WeakValueDictionary()
elif policy == "strong":
return {}
elif policy == "none":
return None
else:
raise ValueError("Unsupported cache policy: %s" % policy)
class PresentationComponentModel(object):
'''
Base class for component models to a presentation model.
The component models are things like mhc binding affinity and cleavage,
and the presentation model is typically a logistic regression model
over these.
'''
def __init__(
self, fit_cache_policy="weak", predictions_cache_policy="weak"):
self.fit_cache_policy = fit_cache_policy
self.predictions_cache_policy = predictions_cache_policy
self.reset_cache()
def reset_cache(self):
self.cached_fits = cache_dict_for_policy(self.fit_cache_policy)
self.cached_predictions = cache_dict_for_policy(
self.predictions_cache_policy)
def __getstate__(self):
d = dict(self.__dict__)
d["cached_fits"] = None
d["cached_predictions"] = None
return d
def __setstate__(self, state):
self.__dict__.update(state)
self.reset_cache()
def combine_ensemble_predictions(self, column_name, values):
return numpy.nanmean(values, axis=1)
def stratification_groups(self, hits_df):
return hits_df.experiment_name
def column_names(self):
"""
Names for the values this final model input emits.
Some final model inputs emit multiple related quantities, such as
"binding affinity" and "binding percentile rank".
"""
raise NotImplementedError(str(self))
def requires_fitting(self):
"""
Does this model require fitting to mass-spec data?
For example, the 'expression' componenet models don't need to be
fit, but some cleavage predictors and binding predictors can be
trained on the ms data.
"""
raise NotImplementedError(str(self))
def clone_and_fit(self, hits_df):
"""
Clone the object and fit to given dataset with a weakref cache.
"""
if not self.requires_fitting():
return self
if self.cached_fits is None:
key = None
result = None
else:
key = dataframe_cryptographic_hash(
hits_df[["experiment_name", "peptide"]])
result = self.cached_fits.get(key)
if result is None:
print("Cache miss in clone_and_fit: %s" % str(self))
result = self.clone()
result.fit(hits_df)
if self.cached_fits is not None:
self.cached_fits[key] = result
else:
print("Cache hit in clone_and_fit: %s" % str(self))
return result
def clone_and_restore_fit(self, fit_info):
if not self.requires_fitting():
assert fit_info is None
return self
if self.cached_fits is None:
key = None
result = None
else:
key = freeze_object(fit_info)
result = self.cached_fits.get(key)
if result is None:
print("Cache miss in clone_and_restore_fit: %s" % str(self))
result = self.clone()
result.restore_fit(fit_info)
if self.cached_fits is not None:
self.cached_fits[key] = result
else:
print("Cache hit in clone_and_restore_fit: %s" % str(self))
return result
def fit(self, hits_df):
"""
Train the model.
Parameters
-----------
hits_df : pandas.DataFrame
dataframe of hits with columns 'experiment_name' and 'peptide'
"""
if self.requires_fitting():
raise NotImplementedError(str(self))
def predict_for_experiment(self, experiment_name, peptides):
"""
A more convenient prediction method to implement.
Subclasses should override this method or predict().
Returns
------------
A dict of column name -> list of predictions for each peptide
"""
assert self.predict != PresentationComponentModel.predict, (
"Must override predict_for_experiment() or predict()")
peptides_df = pandas.DataFrame({
'peptide': peptides,
})
peptides_df["experiment_name"] = experiment_name
return self.predict(peptides_df)
def predict(self, peptides_df):
"""
Subclasses can override either this or predict_for_experiment.
This is the high-level predict method that users should call.
This convenience method groups the peptides_df by experiment
and calls predict_for_experiment on each experiment.
"""
assert (
self.predict_for_experiment !=
PresentationComponentModel.predict_for_experiment)
assert 'experiment_name' in peptides_df.columns
assert 'peptide' in peptides_df.columns
if self.cached_predictions is None:
cache_key = None
cached_result = None
else:
cache_key = dataframe_cryptographic_hash(peptides_df)
cached_result = self.cached_predictions.get(cache_key)
if cached_result is not None:
print("Cache hit in predict: %s" % str(self))
return cached_result
else:
print("Cache miss in predict: %s" % str(self))
grouped = peptides_df.groupby("experiment_name")
if len(grouped) == 1:
print("%s : using single-experiment predict optimization" % (
str(self)))
return_value = pandas.DataFrame(
self.predict_for_experiment(
str(peptides_df.iloc[0].experiment_name),
peptides_df.peptide.values))
assert len(return_value) == len(peptides_df), (
"%d != %d" % (len(return_value), len(peptides_df)),
str(self),
peptides_df.peptide.nunique(),
return_value,
peptides_df)
assert_no_null(return_value, str(self))
else:
peptides_df = (
peptides_df[["experiment_name", "peptide"]]
.reset_index(drop=True))
columns = self.column_names()
result_df = peptides_df.copy()
for col in columns:
result_df[col] = numpy.nan
for (experiment_name, sub_df) in grouped:
assert (
result_df.loc[sub_df.index, "experiment_name"] ==
experiment_name).all()
unique_peptides = list(set(sub_df.peptide))
if len(unique_peptides) == 0:
continue
result_dict = self.predict_for_experiment(
experiment_name, unique_peptides)
for col in columns:
assert len(result_dict[col]) == len(unique_peptides), (
"Final model input %s: wrong number of predictions "
"%d (expected %d) for col %s:\n%s\n"
"Input was: experiment: %s, peptides:\n%s" % (
str(self),
len(result_dict[col]),
len(unique_peptides),
col,
result_dict[col],
experiment_name,
unique_peptides))
prediction_series = pandas.Series(
result_dict[col],
index=unique_peptides)
prediction_values = (
prediction_series.ix[sub_df.peptide.values]).values
result_df.loc[
sub_df.index, col
] = prediction_values
assert len(result_df) == len(peptides_df), "%s != %s" % (
len(result_df),
len(peptides_df))
return_value = result_df[columns]
if self.cached_predictions is not None:
self.cached_predictions[cache_key] = return_value
return dict(
(col, return_value[col].values) for col in self.column_names())
def clone(self):
"""
Copy this object so that the original and copy can be fit
independently.
"""
if self.requires_fitting():
# shallow copy won't work here, subclass must override.
raise NotImplementedError(str(self))
result = copy(self)
# We do not want to share a cache with the clone.
result.reset_cache()
return result
def get_fit(self):
if self.requires_fitting():
raise NotImplementedError(str(self))
return None
def restore_fit(self, fit_info):
if self.requires_fitting():
raise NotImplementedError(str(self))
assert fit_info is None, (str(self), str(fit_info))
mhcflurry/antigen_presentation/presentation_model.py deleted 100644 → 0
+ 0
− 583
View file @ ac709bb7
This diff is collapsed.
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment