Skip to content
Snippets Groups Projects
class1_affinity_predictor.py 57.80 KiB
import collections
import hashlib
import json
import logging
import time
import warnings
from os.path import join, exists, abspath
from os import mkdir, environ
from socket import gethostname
from getpass import getuser
from functools import partial
from six import string_types

import numpy
from numpy.testing import assert_equal
import pandas

import mhcnames

from .class1_neural_network import Class1NeuralNetwork
from .common import random_peptides, positional_frequency_matrix
from .downloads import get_default_class1_models_dir
from .encodable_sequences import EncodableSequences
from .percent_rank_transform import PercentRankTransform
from .regression_target import to_ic50
from .version import __version__
from .ensemble_centrality import CENTRALITY_MEASURES
from .allele_encoding import AlleleEncoding


# Default function for combining predictions across models in an ensemble.
# See ensemble_centrality.py for other options.
DEFAULT_CENTRALITY_MEASURE = "mean"

# Any value > 0 will result in attempting to optimize models after loading.
OPTIMIZATION_LEVEL = int(environ.get("MHCFLURRY_OPTIMIZATION_LEVEL", 1))


class Class1AffinityPredictor(object):
    """
    High-level interface for peptide/MHC I binding affinity prediction.

    This class manages low-level `Class1NeuralNetwork` instances, each of which
    wraps a single Keras network. The purpose of `Class1AffinityPredictor` is to
    implement ensembles, handling of multiple alleles, and predictor loading and
    saving. It also provides a place to keep track of metadata like prediction
    histograms for percentile rank calibration.
    """
    def __init__(
            self,
            allele_to_allele_specific_models=None,
            class1_pan_allele_models=None,
            allele_to_sequence=None,
            manifest_df=None,
            allele_to_percent_rank_transform=None,
            metadata_dataframes=None):
        """
        Parameters
        ----------
        allele_to_allele_specific_models : dict of string -> list of `Class1NeuralNetwork`
            Ensemble of single-allele models to use for each allele.

        class1_pan_allele_models : list of `Class1NeuralNetwork`
            Ensemble of pan-allele models.

        allele_to_sequence : dict of string -> string
            MHC allele name to fixed-length amino acid sequence (sometimes
            referred to as the pseudosequence). Required only if
            class1_pan_allele_models is specified.
        
        manifest_df : `pandas.DataFrame`, optional
            Must have columns: model_name, allele, config_json, model.
            Only required if you want to update an existing serialization of a
            Class1AffinityPredictor. Otherwise this dataframe will be generated
            automatically based on the supplied models.

        allele_to_percent_rank_transform : dict of string -> `PercentRankTransform`, optional
            `PercentRankTransform` instances to use for each allele

        metadata_dataframes : dict of string -> pandas.DataFrame, optional
            Optional additional dataframes to write to the models dir when
            save() is called. Useful for tracking provenance.
        """

        if allele_to_allele_specific_models is None:
            allele_to_allele_specific_models = {}
        if class1_pan_allele_models is None:
            class1_pan_allele_models = []

        self.allele_to_sequence = (
            dict(allele_to_sequence)
            if allele_to_sequence is not None else None)  # make a copy

        self._master_allele_encoding = None
        if class1_pan_allele_models:
            assert self.allele_to_sequence

        self.allele_to_allele_specific_models = allele_to_allele_specific_models
        self.class1_pan_allele_models = class1_pan_allele_models
        self._manifest_df = manifest_df

        if not allele_to_percent_rank_transform:
            allele_to_percent_rank_transform = {}
        self.allele_to_percent_rank_transform = allele_to_percent_rank_transform
        self.metadata_dataframes = (
            dict(metadata_dataframes) if metadata_dataframes else {})
        self._cache = {}
        self.optimization_info = {}

        assert isinstance(self.allele_to_allele_specific_models, dict)
        assert isinstance(self.class1_pan_allele_models, list)

    @property
    def manifest_df(self):
        """
        A pandas.DataFrame describing the models included in this predictor.

        Based on:
        - self.class1_pan_allele_models
        - self.allele_to_allele_specific_models

        Returns
        -------
        pandas.DataFrame
        """
        if self._manifest_df is None:
            rows = []
            for (i, model) in enumerate(self.class1_pan_allele_models):
                rows.append((
                    self.model_name("pan-class1", i),
                    "pan-class1",
                    json.dumps(model.get_config()),
                    model
                ))
            for (allele, models) in self.allele_to_allele_specific_models.items():
                for (i, model) in enumerate(models):
                    rows.append((
                        self.model_name(allele, i),
                        allele,
                        json.dumps(model.get_config()),
                        model
                    ))
            self._manifest_df = pandas.DataFrame(
                rows,
                columns=["model_name", "allele", "config_json", "model"])
        return self._manifest_df

    def clear_cache(self):
        """
        Clear values cached based on the neural networks in this predictor.

        Users should call this after mutating any of the following:
            - self.class1_pan_allele_models
            - self.allele_to_allele_specific_models
            - self.allele_to_sequence

        Methods that mutate these instance variables will call this method on
        their own if needed.
        """
        self._cache.clear()

    @property
    def neural_networks(self):
        """
        List of the neural networks in the ensemble.

        Returns
        -------
        list of `Class1NeuralNetwork`
        """
        result = []
        for models in self.allele_to_allele_specific_models.values():
            result.extend(models)
        result.extend(self.class1_pan_allele_models)
        return result

    @classmethod
    def merge(cls, predictors):
        """
        Merge the ensembles of two or more `Class1AffinityPredictor` instances.

        Note: the resulting merged predictor will NOT have calibrated percentile
        ranks. Call `calibrate_percentile_ranks` on it if these are needed.

        Parameters
        ----------
        predictors : sequence of `Class1AffinityPredictor`

        Returns
        -------
        `Class1AffinityPredictor` instance

        """
        assert len(predictors) > 0
        if len(predictors) == 1:
            return predictors[0]

        allele_to_allele_specific_models = collections.defaultdict(list)
        class1_pan_allele_models = []
        allele_to_sequence = predictors[0].allele_to_sequence

        for predictor in predictors:
            for (allele, networks) in (
                    predictor.allele_to_allele_specific_models.items()):
                allele_to_allele_specific_models[allele].extend(networks)
            class1_pan_allele_models.extend(
                predictor.class1_pan_allele_models)

        return Class1AffinityPredictor(
            allele_to_allele_specific_models=allele_to_allele_specific_models,
            class1_pan_allele_models=class1_pan_allele_models,
            allele_to_sequence=allele_to_sequence
        )

    def merge_in_place(self, others):
        """
        Add the models present other predictors into the current predictor.

        Parameters
        ----------
        others : list of Class1AffinityPredictor
            Other predictors to merge into the current predictor.

        Returns
        -------
        list of string : names of newly added models
        """
        new_model_names = []
        original_manifest = self.manifest_df
        new_manifest_rows = []
        for predictor in others:
            for model in predictor.class1_pan_allele_models:
                model_name = self.model_name(
                    "pan-class1",
                    len(self.class1_pan_allele_models))
                row = pandas.Series(collections.OrderedDict([
                    ("model_name", model_name),
                    ("allele", "pan-class1"),
                    ("config_json", json.dumps(model.get_config())),
                    ("model", model),
                ])).to_frame().T
                new_manifest_rows.append(row)
                self.class1_pan_allele_models.append(model)
                new_model_names.append(model_name)

            for allele in predictor.allele_to_allele_specific_models:
                if allele not in self.allele_to_allele_specific_models:
                    self.allele_to_allele_specific_models[allele] = []
                current_models = self.allele_to_allele_specific_models[allele]
                for model in predictor.allele_to_allele_specific_models[allele]:
                    model_name = self.model_name(allele, len(current_models))
                    row = pandas.Series(collections.OrderedDict([
                        ("model_name", model_name),
                        ("allele", allele),
                        ("config_json", json.dumps(model.get_config())),
                        ("model", model),
                    ])).to_frame().T
                    new_manifest_rows.append(row)
                    current_models.append(model)
                    new_model_names.append(model_name)

        self._manifest_df = pandas.concat(
            [original_manifest] + new_manifest_rows,
            ignore_index=True)

        self.clear_cache()
        self.check_consistency()
        return new_model_names

    @property
    def supported_alleles(self):
        """
        Alleles for which predictions can be made.
        
        Returns
        -------
        list of string
        """
        if 'supported_alleles' not in self._cache:
            result = set(self.allele_to_allele_specific_models)
            if self.allele_to_sequence:
                result = result.union(self.allele_to_sequence)
            self._cache["supported_alleles"] = sorted(result)
        return self._cache["supported_alleles"]

    @property
    def supported_peptide_lengths(self):
        """
        (minimum, maximum) lengths of peptides supported by *all models*,
        inclusive.

        Returns
        -------
        (int, int) tuple

        """
        if 'supported_peptide_lengths' not in self._cache:
            length_ranges = set(
                network.supported_peptide_lengths
                for network in self.neural_networks)
            result = (
                max(lower for (lower, upper) in length_ranges),
                min(upper for (lower, upper) in length_ranges))
            self._cache["supported_peptide_lengths"] = result
        return self._cache["supported_peptide_lengths"]

    def check_consistency(self):
        """
        Verify that self.manifest_df is consistent with:
        - self.class1_pan_allele_models
        - self.allele_to_allele_specific_models

        Currently only checks for agreement on the total number of models.

        Throws AssertionError if inconsistent.
        """
        num_models = len(self.class1_pan_allele_models) + sum(
            len(v) for v in self.allele_to_allele_specific_models.values())
        assert len(self.manifest_df) == num_models, (
            "Manifest seems out of sync with models: %d vs %d entries: "
            "\n%s\npan-allele: %s\nallele-specific: %s"% (
                len(self.manifest_df),
                num_models,
                str(self.manifest_df),
                str(self.class1_pan_allele_models),
                str(self.allele_to_allele_specific_models)))

    def save(self, models_dir, model_names_to_write=None, write_metadata=True):
        """
        Serialize the predictor to a directory on disk. If the directory does
        not exist it will be created.
        
        The serialization format consists of a file called "manifest.csv" with
        the configurations of each Class1NeuralNetwork, along with per-network
        files giving the model weights. If there are pan-allele predictors in
        the ensemble, the allele sequences are also stored in the
        directory. There is also a small file "index.txt" with basic metadata:
        when the models were trained, by whom, on what host.
        
        Parameters
        ----------
        models_dir : string
            Path to directory. It will be created if it doesn't exist.
            
        model_names_to_write : list of string, optional
            Only write the weights for the specified models. Useful for
            incremental updates during training.

        write_metadata : boolean, optional
            Whether to write optional metadata
        """
        self.check_consistency()

        if model_names_to_write is None:
            # Write all models
            model_names_to_write = self.manifest_df.model_name.values

        if not exists(models_dir):
            mkdir(models_dir)

        sub_manifest_df = self.manifest_df.loc[
            self.manifest_df.model_name.isin(model_names_to_write)
        ].copy()

        # Network JSON configs may have changed since the models were added,
        # for example due to changes to the allele representation layer.
        # So we update the JSON configs here also.
        updated_network_config_jsons = []
        for (_, row) in sub_manifest_df.iterrows():
            updated_network_config_jsons.append(
                json.dumps(row.model.get_config()))
            weights_path = self.weights_path(models_dir, row.model_name)
            Class1AffinityPredictor.save_weights(
                row.model.get_weights(), weights_path)
            logging.info("Wrote: %s", weights_path)
        sub_manifest_df["config_json"] = updated_network_config_jsons
        self.manifest_df.loc[
            sub_manifest_df.index,
            "config_json"
        ] = updated_network_config_jsons

        write_manifest_df = self.manifest_df[[
            c for c in self.manifest_df.columns if c != "model"
        ]]
        manifest_path = join(models_dir, "manifest.csv")
        write_manifest_df.to_csv(manifest_path, index=False)
        logging.info("Wrote: %s", manifest_path)

        if write_metadata:
            # Write "info.txt"
            info_path = join(models_dir, "info.txt")
            rows = [
                ("trained on", time.asctime()),
                ("package   ", "mhcflurry %s" % __version__),
                ("hostname  ", gethostname()),
                ("user      ", getuser()),
            ]
            pandas.DataFrame(rows).to_csv(
                info_path, sep="\t", header=False, index=False)

            if self.metadata_dataframes:
                for (name, df) in self.metadata_dataframes.items():
                    metadata_df_path = join(models_dir, "%s.csv.bz2" % name)
                    df.to_csv(metadata_df_path, index=False, compression="bz2")

        # Save allele sequences
        if self.allele_to_sequence is not None:
            allele_to_sequence_df = pandas.DataFrame(
                list(self.allele_to_sequence.items()),
                columns=['allele', 'sequence']
            )
            allele_to_sequence_df.to_csv(
                join(models_dir, "allele_sequences.csv"), index=False)
            logging.info("Wrote: %s", join(models_dir, "allele_sequences.csv"))

        if self.allele_to_percent_rank_transform:
            percent_ranks_df = None
            for (allele, transform) in self.allele_to_percent_rank_transform.items():
                series = transform.to_series()
                if percent_ranks_df is None:
                    percent_ranks_df = pandas.DataFrame(index=series.index)
                assert_equal(series.index.values, percent_ranks_df.index.values)
                percent_ranks_df[allele] = series
            percent_ranks_path = join(models_dir, "percent_ranks.csv")
            percent_ranks_df.to_csv(
                percent_ranks_path,
                index=True,
                index_label="bin")
            logging.info("Wrote: %s", percent_ranks_path)

    @staticmethod
    def load(models_dir=None, max_models=None, optimization_level=None):
        """
        Deserialize a predictor from a directory on disk.
        
        Parameters
        ----------
        models_dir : string
            Path to directory. If unspecified the default downloaded models are
            used.
            
        max_models : int, optional
            Maximum number of `Class1NeuralNetwork` instances to load

        optimization_level : int
            If >0, model optimization will be attempted. Defaults to value of
            environment variable MHCFLURRY_OPTIMIZATION_LEVEL.

        Returns
        -------
        `Class1AffinityPredictor` instance
        """
        if models_dir is None:
            models_dir = get_default_class1_models_dir()
        if optimization_level is None:
            optimization_level = OPTIMIZATION_LEVEL

        manifest_path = join(models_dir, "manifest.csv")
        manifest_df = pandas.read_csv(manifest_path, nrows=max_models)

        allele_to_allele_specific_models = collections.defaultdict(list)
        class1_pan_allele_models = []
        all_models = []
        for (_, row) in manifest_df.iterrows():
            weights_filename = Class1AffinityPredictor.weights_path(
                models_dir, row.model_name)
            config = json.loads(row.config_json)

            # We will lazy-load weights when the network is used.
            model = Class1NeuralNetwork.from_config(
                config,
                weights_loader=partial(
                    Class1AffinityPredictor.load_weights,
                    abspath(weights_filename)))
            if row.allele == "pan-class1":
                class1_pan_allele_models.append(model)
            else:
                allele_to_allele_specific_models[row.allele].append(model)
            all_models.append(model)

        manifest_df["model"] = all_models

        # Load allele sequences
        allele_to_sequence = None
        if exists(join(models_dir, "allele_sequences.csv")):
            allele_to_sequence = pandas.read_csv(
                join(models_dir, "allele_sequences.csv"),
                index_col=0).iloc[:, 0].to_dict()

        allele_to_percent_rank_transform = {}
        percent_ranks_path = join(models_dir, "percent_ranks.csv")
        if exists(percent_ranks_path):
            percent_ranks_df = pandas.read_csv(percent_ranks_path, index_col=0)
            for allele in percent_ranks_df.columns:
                allele_to_percent_rank_transform[allele] = (
                    PercentRankTransform.from_series(percent_ranks_df[allele]))

        logging.info(
            "Loaded %d class1 pan allele predictors, %d allele sequences, "
            "%d percent rank distributions, and %d allele specific models: %s",
            len(class1_pan_allele_models),
            len(allele_to_sequence) if allele_to_sequence else 0,
            len(allele_to_percent_rank_transform),
            sum(len(v) for v in allele_to_allele_specific_models.values()),
            ", ".join(
                "%s (%d)" % (allele, len(v))
                for (allele, v)
                in sorted(allele_to_allele_specific_models.items())))

        result = Class1AffinityPredictor(
            allele_to_allele_specific_models=allele_to_allele_specific_models,
            class1_pan_allele_models=class1_pan_allele_models,
            allele_to_sequence=allele_to_sequence,
            manifest_df=manifest_df,
            allele_to_percent_rank_transform=allele_to_percent_rank_transform,
        )
        if optimization_level >= 1:
            optimized = result.optimize()
            logging.info(
                "Model optimization %s",
                "succeeded" if optimized else "not supported for these models")
        return result

    def optimize(self, warn=True):
        """
        EXPERIMENTAL: Optimize the predictor for faster predictions.

        Currently the only optimization implemented is to merge multiple pan-
        allele predictors at the tensorflow level.

        The optimization is performed in-place, mutating the instance.

        Returns
        ----------
        bool
            Whether optimization was performed

        """
        num_class1_pan_allele_models = len(self.class1_pan_allele_models)
        if num_class1_pan_allele_models > 1:
            try:
                self.class1_pan_allele_models = [
                    Class1NeuralNetwork.merge(
                        self.class1_pan_allele_models,
                        merge_method="concatenate")
                ]
            except NotImplementedError as e:
                if warn:
                    logging.warning("Optimization failed: %s", str(e))
                return False
            self._manifest_df = None
            self.clear_cache()
            self.optimization_info["pan_models_merged"] = True
            self.optimization_info["num_pan_models_merged"] = (
                num_class1_pan_allele_models)
        else:
            return False
        return True

    @staticmethod
    def model_name(allele, num):
        """
        Generate a model name
        
        Parameters
        ----------
        allele : string
        num : int

        Returns
        -------
        string

        """
        random_string = hashlib.sha1(
            str(time.time()).encode()).hexdigest()[:16]
        return "%s-%d-%s" % (
            allele.upper().replace("*", "_").replace(":", "_"),
            num,
            random_string)

    @staticmethod
    def weights_path(models_dir, model_name):
        """
        Generate the path to the weights file for a model
        
        Parameters
        ----------
        models_dir : string
        model_name : string

        Returns
        -------
        string
        """
        return join(models_dir, "weights_%s.npz" % model_name)

    @property
    def master_allele_encoding(self):
        """
        An AlleleEncoding containing the universe of alleles specified by
        self.allele_to_sequence.

        Returns
        -------
        AlleleEncoding
        """
        if (self._master_allele_encoding is None or
                self._master_allele_encoding.allele_to_sequence !=
                self.allele_to_sequence):
            self._master_allele_encoding = AlleleEncoding(
                allele_to_sequence=self.allele_to_sequence)
        return self._master_allele_encoding

    def fit_allele_specific_predictors(
            self,
            n_models,
            architecture_hyperparameters_list,
            allele,
            peptides,
            affinities,
            inequalities=None,
            train_rounds=None,
            models_dir_for_save=None,
            verbose=0,
            progress_preamble="",
            progress_print_interval=5.0):
        """
        Fit one or more allele specific predictors for a single allele using one
        or more neural network architectures.
        
        The new predictors are saved in the Class1AffinityPredictor instance
        and will be used on subsequent calls to `predict`.
        
        Parameters
        ----------
        n_models : int
            Number of neural networks to fit
        
        architecture_hyperparameters_list : list of dict
            List of hyperparameter sets.
               
        allele : string
        
        peptides : `EncodableSequences` or list of string
        
        affinities : list of float
            nM affinities

        inequalities : list of string, each element one of ">", "<", or "="
            See `Class1NeuralNetwork.fit` for details.

        train_rounds : sequence of int
            Each training point i will be used on training rounds r for which
            train_rounds[i] > r, r >= 0.
        
        models_dir_for_save : string, optional
            If specified, the Class1AffinityPredictor is (incrementally) written
            to the given models dir after each neural network is fit.
        
        verbose : int
            Keras verbosity

        progress_preamble : string
            Optional string of information to include in each progress update

        progress_print_interval : float
            How often (in seconds) to print progress. Set to None to disable.

        Returns
        -------
        list of `Class1NeuralNetwork`
        """

        allele = mhcnames.normalize_allele_name(allele)
        if allele not in self.allele_to_allele_specific_models:
            self.allele_to_allele_specific_models[allele] = []

        encodable_peptides = EncodableSequences.create(peptides)
        peptides_affinities_inequalities_per_round = [
            (encodable_peptides, affinities, inequalities)
        ]

        if train_rounds is not None:
            for round in sorted(set(train_rounds)):
                round_mask = train_rounds > round
                if round_mask.any():
                    sub_encodable_peptides = EncodableSequences.create(
                        encodable_peptides.sequences[round_mask])
                    peptides_affinities_inequalities_per_round.append((
                        sub_encodable_peptides,
                        affinities[round_mask],
                        None if inequalities is None else inequalities[round_mask]))
        n_rounds = len(peptides_affinities_inequalities_per_round)

        n_architectures = len(architecture_hyperparameters_list)

        # Adjust progress info to indicate number of models and
        # architectures.
        pieces = []
        if n_models > 1:
            pieces.append("Model {model_num:2d} / {n_models:2d}")
        if n_architectures > 1:
            pieces.append(
                "Architecture {architecture_num:2d} / {n_architectures:2d}")
        if len(peptides_affinities_inequalities_per_round) > 1:
            pieces.append("Round {round:2d} / {n_rounds:2d}")
        pieces.append("{n_peptides:4d} peptides")
        progress_preamble_template = "[ %s ] {user_progress_preamble}" % (
            ", ".join(pieces))

        models = []
        for model_num in range(n_models):
            for (architecture_num, architecture_hyperparameters) in enumerate(
                    architecture_hyperparameters_list):
                model = Class1NeuralNetwork(**architecture_hyperparameters)
                for round_num in range(n_rounds):
                    (round_peptides, round_affinities, round_inequalities) = (
                        peptides_affinities_inequalities_per_round[round_num]
                    )
                    model.fit(
                        round_peptides,
                        round_affinities,
                        inequalities=round_inequalities,
                        verbose=verbose,
                        progress_preamble=progress_preamble_template.format(
                            n_peptides=len(round_peptides),
                            round=round_num,
                            n_rounds=n_rounds,
                            user_progress_preamble=progress_preamble,
                            model_num=model_num + 1,
                            n_models=n_models,
                            architecture_num=architecture_num + 1,
                            n_architectures=n_architectures),
                        progress_print_interval=progress_print_interval)

                model_name = self.model_name(allele, model_num)
                row = pandas.Series(collections.OrderedDict([
                    ("model_name", model_name),
                    ("allele", allele),
                    ("config_json", json.dumps(model.get_config())),
                    ("model", model),
                ])).to_frame().T
                self._manifest_df = pandas.concat(
                    [self.manifest_df, row], ignore_index=True)
                self.allele_to_allele_specific_models[allele].append(model)
                if models_dir_for_save:
                    self.save(
                        models_dir_for_save, model_names_to_write=[model_name])
                models.append(model)

        self.clear_cache()
        return models

    def fit_class1_pan_allele_models(
            self,
            n_models,
            architecture_hyperparameters,
            alleles,
            peptides,
            affinities,
            inequalities,
            models_dir_for_save=None,
            verbose=1,
            progress_preamble="",
            progress_print_interval=5.0):
        """
        Fit one or more pan-allele predictors using a single neural network
        architecture.
        
        The new predictors are saved in the Class1AffinityPredictor instance
        and will be used on subsequent calls to `predict`.
        
        Parameters
        ----------
        n_models : int
            Number of neural networks to fit
            
        architecture_hyperparameters : dict
        
        alleles : list of string
            Allele names (not sequences) corresponding to each peptide
        
        peptides : `EncodableSequences` or list of string
        
        affinities : list of float
            nM affinities

        inequalities : list of string, each element one of ">", "<", or "="
            See Class1NeuralNetwork.fit for details.
        
        models_dir_for_save : string, optional
            If specified, the Class1AffinityPredictor is (incrementally) written
            to the given models dir after each neural network is fit.
        
        verbose : int
            Keras verbosity

        progress_preamble : string
            Optional string of information to include in each progress update

        progress_print_interval : float
            How often (in seconds) to print progress. Set to None to disable.

        Returns
        -------
        list of `Class1NeuralNetwork`
        """

        alleles = pandas.Series(alleles).map(mhcnames.normalize_allele_name)
        allele_encoding = AlleleEncoding(
            alleles,
            borrow_from=self.master_allele_encoding)

        encodable_peptides = EncodableSequences.create(peptides)
        models = []
        for i in range(n_models):
            logging.info("Training model %d / %d", i + 1, n_models)
            model = Class1NeuralNetwork(**architecture_hyperparameters)
            model.fit(
                encodable_peptides,
                affinities,
                inequalities=inequalities,
                allele_encoding=allele_encoding,
                verbose=verbose,
                progress_preamble=progress_preamble,
                progress_print_interval=progress_print_interval)

            model_name = self.model_name("pan-class1", i)
            row = pandas.Series(collections.OrderedDict([
                ("model_name", model_name),
                ("allele", "pan-class1"),
                ("config_json", json.dumps(model.get_config())),
                ("model", model),
            ])).to_frame().T
            self._manifest_df = pandas.concat(
                [self.manifest_df, row], ignore_index=True)
            self.class1_pan_allele_models.append(model)
            if models_dir_for_save:
                self.save(
                    models_dir_for_save, model_names_to_write=[model_name])
            models.append(model)

        self.clear_cache()
        return models

    def add_pan_allele_model(self, model, models_dir_for_save=None):
        """
        Add a pan-allele model to the ensemble and optionally do an incremental
        save.

        Parameters
        ----------
        model : Class1NeuralNetwork
        models_dir_for_save : string
            Directory to save resulting ensemble to
        """
        model_name = self.model_name("pan-class1", 1)
        row = pandas.Series(collections.OrderedDict([
            ("model_name", model_name),
            ("allele", "pan-class1"),
            ("config_json", json.dumps(model.get_config())),
            ("model", model),
        ])).to_frame().T
        self._manifest_df = pandas.concat(
            [self.manifest_df, row], ignore_index=True)
        self.class1_pan_allele_models.append(model)
        self.clear_cache()
        self.check_consistency()
        if models_dir_for_save:
            self.save(
                models_dir_for_save, model_names_to_write=[model_name])

    def percentile_ranks(self, affinities, allele=None, alleles=None, throw=True):
        """
        Return percentile ranks for the given ic50 affinities and alleles.

        The 'allele' and 'alleles' argument are as in the `predict` method.
        Specify one of these.

        Parameters
        ----------
        affinities : sequence of float
            nM affinities
        allele : string
        alleles : sequence of string
        throw : boolean
            If True, a ValueError will be raised in the case of unsupported
            alleles. If False, a warning will be logged and NaN will be returned
            for those percentile ranks.

        Returns
        -------
        numpy.array of float
        """
        if allele is not None:
            try:
                transform = self.allele_to_percent_rank_transform[allele]
                return transform.transform(affinities)
            except KeyError:
                msg = "Allele %s has no percentile rank information" % allele
                if throw:
                    raise ValueError(msg)
                warnings.warn(msg)
                return numpy.ones(len(affinities)) * numpy.nan  # Return NaNs

        if alleles is None:
            raise ValueError("Specify allele or alleles")

        df = pandas.DataFrame({"affinity": affinities})
        df["allele"] = alleles
        df["result"] = numpy.nan
        for (allele, sub_df) in df.groupby("allele"):
            df.loc[sub_df.index, "result"] = self.percentile_ranks(
                sub_df.affinity, allele=allele, throw=throw)
        return df.result.values

    def predict(
            self,
            peptides,
            alleles=None,
            allele=None,
            throw=True,
            centrality_measure=DEFAULT_CENTRALITY_MEASURE,
            model_kwargs={}):
        """
        Predict nM binding affinities.
        
        If multiple predictors are available for an allele, the predictions are
        the geometric means of the individual model (nM) predictions.
        
        One of 'allele' or 'alleles' must be specified. If 'allele' is specified
        all predictions will be for the given allele. If 'alleles' is specified
        it must be the same length as 'peptides' and give the allele
        corresponding to each peptide.
        
        Parameters
        ----------
        peptides : `EncodableSequences` or list of string
        alleles : list of string
        allele : string
        throw : boolean
            If True, a ValueError will be raised in the case of unsupported
            alleles or peptide lengths. If False, a warning will be logged and
            the predictions for the unsupported alleles or peptides will be NaN.
        centrality_measure : string or callable
            Measure of central tendency to use to combine predictions in the
            ensemble. Options include: mean, median, robust_mean.
        model_kwargs : dict
            Additional keyword arguments to pass to Class1NeuralNetwork.predict

        Returns
        -------
        numpy.array of predictions
        """
        df = self.predict_to_dataframe(
            peptides=peptides,
            alleles=alleles,
            allele=allele,
            throw=throw,
            include_percentile_ranks=False,
            include_confidence_intervals=False,
            centrality_measure=centrality_measure,
            model_kwargs=model_kwargs
        )
        return df.prediction.values

    def predict_to_dataframe(
            self,
            peptides,
            alleles=None,
            allele=None,
            throw=True,
            include_individual_model_predictions=False,
            include_percentile_ranks=True,
            include_confidence_intervals=True,
            centrality_measure=DEFAULT_CENTRALITY_MEASURE,
            model_kwargs={}):
        """
        Predict nM binding affinities. Gives more detailed output than `predict`
        method, including 5-95% prediction intervals.
        
        If multiple predictors are available for an allele, the predictions are
        the geometric means of the individual model predictions.
        
        One of 'allele' or 'alleles' must be specified. If 'allele' is specified
        all predictions will be for the given allele. If 'alleles' is specified
        it must be the same length as 'peptides' and give the allele
        corresponding to each peptide. 
        
        Parameters
        ----------
        peptides : `EncodableSequences` or list of string
        alleles : list of string
        allele : string
        throw : boolean
            If True, a ValueError will be raised in the case of unsupported
            alleles or peptide lengths. If False, a warning will be logged and
            the predictions for the unsupported alleles or peptides will be NaN.
        include_individual_model_predictions : boolean
            If True, the predictions of each individual model are included as
            columns in the result DataFrame.
        include_percentile_ranks : boolean, default True
            If True, a "prediction_percentile" column will be included giving
            the percentile ranks. If no percentile rank info is available,
            this will be ignored with a warning.
        centrality_measure : string or callable
            Measure of central tendency to use to combine predictions in the
            ensemble. Options include: mean, median, robust_mean.
        model_kwargs : dict
            Additional keyword arguments to pass to Class1NeuralNetwork.predict

        Returns
        -------
        `pandas.DataFrame` of predictions
        """
        if isinstance(peptides, string_types):
            raise TypeError("peptides must be a list or array, not a string")
        if isinstance(alleles, string_types):
            raise TypeError("alleles must be a list or array, not a string")
        if allele is None and alleles is None:
            raise ValueError("Must specify 'allele' or 'alleles'.")

        peptides = EncodableSequences.create(peptides)
        df = pandas.DataFrame({
            'peptide': peptides.sequences
        }, copy=False)

        if allele is not None:
            if alleles is not None:
                raise ValueError("Specify exactly one of allele or alleles")
            df["allele"] = allele
            normalized_allele = mhcnames.normalize_allele_name(allele)
            df["normalized_allele"] = normalized_allele
            unique_alleles = [normalized_allele]
        else:
            df["allele"] = numpy.array(alleles)
            df["normalized_allele"] = df.allele.map(
                mhcnames.normalize_allele_name)
            unique_alleles = df.normalized_allele.unique()

        if len(df) == 0:
            # No predictions.
            logging.warning("Predicting for 0 peptides.")
            empty_result = pandas.DataFrame(
                columns=[
                    'peptide',
                    'allele',
                    'prediction',
                    'prediction_low',
                    'prediction_high'
                ])
            return empty_result

        (min_peptide_length, max_peptide_length) = (
            self.supported_peptide_lengths)

        if (peptides.min_length < min_peptide_length or
                peptides.max_length > max_peptide_length):
            # Only compute this if needed
            all_peptide_lengths_supported = False
            sequence_length = df.peptide.str.len()
            df["supported_peptide_length"] = (
                (sequence_length >= min_peptide_length) &
                (sequence_length <= max_peptide_length))
            if (~df.supported_peptide_length).any():
                msg = (
                    "%d peptides have lengths outside of supported range [%d, %d]: "
                    "%s" % (
                        (~df.supported_peptide_length).sum(),
                        min_peptide_length,
                        max_peptide_length,
                        str(df.loc[~df.supported_peptide_length].peptide.unique())))
                logging.warning(msg)
                if throw:
                    raise ValueError(msg)
        else:
            # Handle common case efficiently.
            df["supported_peptide_length"] = True
            all_peptide_lengths_supported = True

        num_pan_models = (
            len(self.class1_pan_allele_models)
            if not self.optimization_info.get("pan_models_merged", False)
            else self.optimization_info["num_pan_models_merged"])
        max_single_allele_models = max(
            len(self.allele_to_allele_specific_models.get(allele, []))
            for allele in unique_alleles
        )
        predictions_array = numpy.zeros(
            shape=(df.shape[0], num_pan_models + max_single_allele_models),
            dtype="float64")
        predictions_array[:] = numpy.nan

        if self.class1_pan_allele_models:
            master_allele_encoding = self.master_allele_encoding
            unsupported_alleles = [
                allele for allele in
                df.normalized_allele.unique()
                if allele not in self.allele_to_sequence
            ]
            if unsupported_alleles:
                truncate_at = 100
                allele_string = " ".join(
                    sorted(self.allele_to_sequence)[:truncate_at])
                if len(self.allele_to_sequence) > truncate_at:
                    allele_string += " + %d more alleles" % (
                        len(self.allele_to_sequence) - truncate_at)
                msg = (
                    "No sequences for allele(s): %s.\n"
                    "Supported alleles: %s" % (
                        " ".join(unsupported_alleles), allele_string))
                logging.warning(msg)
                if throw:
                    raise ValueError(msg)
            mask = df.supported_peptide_length & (
                ~df.normalized_allele.isin(unsupported_alleles))

            row_slice = None
            if mask is None or mask.all():
                row_slice = slice(None, None, None)  # all rows
                masked_allele_encoding = AlleleEncoding(
                    df.normalized_allele,
                    borrow_from=master_allele_encoding)
                masked_peptides = peptides
            elif mask.sum() > 0:
                row_slice = mask
                masked_allele_encoding = AlleleEncoding(
                    df.loc[mask].normalized_allele,
                    borrow_from=master_allele_encoding)
                masked_peptides = peptides.sequences[mask]

            masked_peptides = EncodableSequences.create(masked_peptides)

            if row_slice is not None:
                # The following line is a performance optimization that may be
                # revisited. It causes the neural network to set to include
                # only the alleles actually being predicted for. This makes
                # the network much smaller. However, subsequent calls to
                # predict will need to reset these weights, so there is a
                # tradeoff.
                masked_allele_encoding = masked_allele_encoding.compact()

                if self.optimization_info.get("pan_models_merged"):
                    # Multiple pan-allele models have been merged into one
                    # at the tensorflow level.
                    assert len(self.class1_pan_allele_models) == 1
                    predictions = self.class1_pan_allele_models[0].predict(
                        masked_peptides,
                        allele_encoding=masked_allele_encoding,
                        output_index=None,
                        **model_kwargs)
                    predictions_array[row_slice, :num_pan_models] = predictions
                else:
                    for (i, model) in enumerate(self.class1_pan_allele_models):
                        predictions_array[row_slice, i] = model.predict(
                            masked_peptides,
                            allele_encoding=masked_allele_encoding,
                            **model_kwargs)

        if self.allele_to_allele_specific_models:
            unsupported_alleles = [
                allele for allele in unique_alleles
                if not self.allele_to_allele_specific_models.get(allele)
            ]
            if unsupported_alleles:
                msg = (
                    "No single-allele models for allele(s): %s.\n"
                    "Supported alleles are: %s" % (
                        " ".join(unsupported_alleles),
                        " ".join(sorted(self.allele_to_allele_specific_models))))
                logging.warning(msg)
                if throw:
                    raise ValueError(msg)

            for allele in unique_alleles:
                models = self.allele_to_allele_specific_models.get(allele, [])
                if len(unique_alleles) == 1 and all_peptide_lengths_supported:
                    mask = None
                else:
                    mask = (
                        (df.normalized_allele == allele) &
                        df.supported_peptide_length).values

                row_slice = None
                if mask is None or mask.all():
                    peptides_for_allele = peptides
                    row_slice = slice(None, None, None)
                elif mask.sum() > 0:
                    peptides_for_allele = EncodableSequences.create(
                        df.loc[mask].peptide.values)
                    row_slice = mask

                if row_slice is not None:
                    for (i, model) in enumerate(models):
                        predictions_array[
                            row_slice,
                            num_pan_models + i,
                        ] = model.predict(peptides_for_allele, **model_kwargs)

        if callable(centrality_measure):
            centrality_function = centrality_measure
        else:
            centrality_function = CENTRALITY_MEASURES[centrality_measure]

        logs = numpy.log(predictions_array)
        log_centers = centrality_function(logs)
        df["prediction"] = numpy.exp(log_centers)

        if include_confidence_intervals:
            df["prediction_low"] = numpy.exp(
                numpy.nanpercentile(logs, 5.0, axis=1))
            df["prediction_high"] = numpy.exp(
                numpy.nanpercentile(logs, 95.0, axis=1))

        if include_individual_model_predictions:
            for i in range(num_pan_models):
                df["model_pan_%d" % i] = predictions_array[:, i]

            for i in range(max_single_allele_models):
                df["model_single_%d" % i] = predictions_array[
                    :, num_pan_models + i
                ]

        if include_percentile_ranks:
            if self.allele_to_percent_rank_transform:
                df["prediction_percentile"] = self.percentile_ranks(
                    df.prediction,
                    alleles=df.normalized_allele.values,
                    throw=throw)
            else:
                warnings.warn("No percentile rank information available.")

        del df["supported_peptide_length"]
        del df["normalized_allele"]
        return df

    @staticmethod
    def save_weights(weights_list, filename):
        """
        Save the model weights to the given filename using numpy's ".npz"
        format.
    
        Parameters
        ----------
        weights_list : list of array
        
        filename : string
            Should end in ".npz".
    
        """
        numpy.savez(
            filename,
            **dict((("array_%d" % i), w) for (i, w) in enumerate(weights_list)))

    @staticmethod
    def load_weights(filename):
        """
        Restore model weights from the given filename, which should have been
        created with `save_weights`.
    
        Parameters
        ----------
        filename : string
            Should end in ".npz".

        Returns
        ----------
        list of array
        """
        with numpy.load(filename) as loaded:
            weights = [
                loaded["array_%d" % i]
                for i in range(len(loaded.keys()))
            ]
        return weights

    def calibrate_percentile_ranks(
            self,
            peptides=None,
            num_peptides_per_length=int(1e5),
            alleles=None,
            bins=None,
            motif_summary=False,
            summary_top_peptide_fractions=[0.001],
            verbose=False,
            model_kwargs={}):
        """
        Compute the cumulative distribution of ic50 values for a set of alleles
        over a large universe of random peptides, to enable taking quantiles
        of this distribution later.

        Parameters
        ----------
        peptides : sequence of string or EncodableSequences, optional
            Peptides to use
        num_peptides_per_length : int, optional
            If peptides argument is not specified, then num_peptides_per_length
            peptides are randomly sampled from a uniform distribution for each
            supported length
        alleles : sequence of string, optional
            Alleles to perform calibration for. If not specified all supported
            alleles will be calibrated.
        bins : object
            Anything that can be passed to numpy.histogram's "bins" argument
            can be used here, i.e. either an integer or a sequence giving bin
            edges. This is in ic50 space.
        motif_summary : bool
            If True, the length distribution and per-position amino acid
            frequencies are also calculated for the top x fraction of tightest-
            binding peptides, where each value of x is given in the
            summary_top_peptide_fractions list.
        summary_top_peptide_fractions : list of float
            Only used if motif_summary is True
        verbose : boolean
            Whether to print status updates to stdout
        model_kwargs : dict
            Additional low-level Class1NeuralNetwork.predict() kwargs.

        Returns
        ----------
        dict of string -> pandas.DataFrame

        If motif_summary is True, this will have keys  "frequency_matrices" and
        "length_distributions". Otherwise it will be empty.

        """
        if bins is None:
            bins = to_ic50(numpy.linspace(1, 0, 1000))

        if alleles is None:
            alleles = self.supported_alleles

        if peptides is None:
            peptides = []
            lengths = range(
                self.supported_peptide_lengths[0],
                self.supported_peptide_lengths[1] + 1)
            for length in lengths:
                peptides.extend(
                    random_peptides(num_peptides_per_length, length))

        encoded_peptides = EncodableSequences.create(peptides)

        if motif_summary:
            frequency_matrices = []
            length_distributions = []
        else:
            frequency_matrices = None
            length_distributions = None
        for allele in alleles:
            start = time.time()
            predictions = self.predict(
                encoded_peptides, allele=allele, model_kwargs=model_kwargs)
            if verbose:
                elapsed = time.time() - start
                print(
                    "Generated %d predictions for allele %s in %0.2f sec: "
                    "%0.2f predictions / sec" % (
                        len(encoded_peptides.sequences),
                        allele,
                        elapsed,
                        len(encoded_peptides.sequences) / elapsed))
            transform = PercentRankTransform()
            transform.fit(predictions, bins=bins)
            self.allele_to_percent_rank_transform[allele] = transform

            if frequency_matrices is not None:
                predictions_df = pandas.DataFrame({
                    'peptide': encoded_peptides.sequences,
                    'prediction': predictions
                }).drop_duplicates('peptide').set_index("peptide")
                predictions_df["length"] = predictions_df.index.str.len()
                for (length, sub_df) in predictions_df.groupby("length"):
                    for cutoff_fraction in summary_top_peptide_fractions:
                        selected = sub_df.prediction.nsmallest(
                            max(
                                int(len(sub_df) * cutoff_fraction),
                                1)).index.values
                        matrix = positional_frequency_matrix(selected).reset_index()
                        original_columns = list(matrix.columns)
                        matrix["allele"] = allele
                        matrix["length"] = length
                        matrix["cutoff_fraction"] = cutoff_fraction
                        matrix["cutoff_count"] = len(selected)
                        matrix = matrix[
                            ["allele", "length", "cutoff_fraction", "cutoff_count"]
                            + original_columns
                        ]
                        frequency_matrices.append(matrix)

                # Length distribution
                for cutoff_fraction in summary_top_peptide_fractions:
                    cutoff_count = max(
                        int(len(predictions_df) * cutoff_fraction), 1)
                    length_distribution = predictions_df.prediction.nsmallest(
                        cutoff_count).index.str.len().value_counts()
                    length_distribution.index.name = "length"
                    length_distribution /= length_distribution.sum()
                    length_distribution = length_distribution.to_frame()
                    length_distribution.columns = ["fraction"]
                    length_distribution = length_distribution.reset_index()
                    length_distribution["allele"] = allele
                    length_distribution["cutoff_fraction"] = cutoff_fraction
                    length_distribution["cutoff_count"] = cutoff_count
                    length_distribution = length_distribution[[
                        "allele",
                        "cutoff_fraction",
                        "cutoff_count",
                        "length",
                        "fraction"
                    ]].sort_values(["cutoff_fraction", "length"])
                    length_distributions.append(length_distribution)

        if frequency_matrices is not None:
            frequency_matrices = pandas.concat(
                frequency_matrices, ignore_index=True)

        if length_distributions is not None:
            length_distributions = pandas.concat(
                length_distributions, ignore_index=True)

        if motif_summary:
            return {
                'frequency_matrices': frequency_matrices,
                'length_distributions': length_distributions,
            }
        return {}

    def model_select(
            self,
            score_function,
            alleles=None,
            min_models=1,
            max_models=10000):
        """
        Perform model selection using a user-specified scoring function.

        This works only with allele-specific models, not pan-allele models.

        Model selection is done using a "step up" variable selection procedure,
        in which models are repeatedly added to an ensemble until the score
        stops improving.

        Parameters
        ----------
        score_function : Class1AffinityPredictor -> float function
            Scoring function

        alleles : list of string, optional
            If not specified, model selection is performed for all alleles.

        min_models : int, optional
            Min models to select per allele

        max_models : int, optional
            Max models to select per allele

        Returns
        -------
        Class1AffinityPredictor : predictor containing the selected models
        """

        if alleles is None:
            alleles = self.supported_alleles

        dfs = []
        allele_to_allele_specific_models = {}
        for allele in alleles:
            df = pandas.DataFrame({
                'model': self.allele_to_allele_specific_models[allele]
            })
            df["model_num"] = df.index
            df["allele"] = allele
            df["selected"] = False

            round_num = 1

            while not df.selected.all() and sum(df.selected) < max_models:
                score_col = "score_%2d" % round_num
                prev_score_col = "score_%2d" % (round_num - 1)

                existing_selected = list(df[df.selected].model)
                df[score_col] = [
                    numpy.nan if row.selected else
                    score_function(
                        Class1AffinityPredictor(
                            allele_to_allele_specific_models={
                                allele: [row.model] + existing_selected
                            }
                        )
                    )
                    for (_, row) in df.iterrows()
                ]

                if round_num > min_models and (
                        df[score_col].max() < df[prev_score_col].max()):
                    break

                # In case of a tie, pick a model at random.
                (best_model_index,) = df.loc[
                    (df[score_col] == df[score_col].max())
                ].sample(1).index
                df.loc[best_model_index, "selected"] = True
                round_num += 1

            dfs.append(df)
            allele_to_allele_specific_models[allele] = list(
                df.loc[df.selected].model)

        df = pandas.concat(dfs, ignore_index=True)

        new_predictor = Class1AffinityPredictor(
            allele_to_allele_specific_models,
            metadata_dataframes={
                "model_selection": df,
            })
        return new_predictor