class1_affinity_predictor.py

                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)
        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.get_master_allele_encoding()
            unsupported_alleles = [
                allele for allele in
                df.normalized_allele.unique()
                if allele not in self.allele_to_sequence
            ]
            if unsupported_alleles:
                msg = (
                    "No sequences for allele(s): %s.\n"
                    "Supported alleles: %s" % (
                        " ".join(unsupported_alleles),
                        " ".join(sorted(self.allele_to_sequence))))
                logging.warning(msg)
                if throw:
                    raise ValueError(msg)
            mask = df.supported_peptide_length & (
                ~df.normalized_allele.isin(unsupported_alleles))
            if mask is None or mask.all():
                # Common case optimization
                allele_encoding = AlleleEncoding(
                    df.normalized_allele,
                    borrow_from=master_allele_encoding)
                for (i, model) in enumerate(self.class1_pan_allele_models):
                    predictions_array[:, i] = (
                        model.predict(
                            peptides,
                            allele_encoding=allele_encoding,
                            **model_kwargs))
            elif mask.sum() > 0:
                masked_allele_encoding = AlleleEncoding(
                    df.loc[mask].normalized_allele,
                    borrow_from=master_allele_encoding)
                masked_peptides = peptides.sequences[mask]
                for (i, model) in enumerate(self.class1_pan_allele_models):
                    predictions_array[mask, 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
                if mask is None or mask.all():
                    # Common case optimization
                    for (i, model) in enumerate(models):
                        predictions_array[:, num_pan_models + i] = (
                            model.predict(peptides, **model_kwargs))
                elif mask.sum() > 0:
                    peptides_for_allele = EncodableSequences.create(
                        df.ix[mask].peptide.values)
                    for (i, model) in enumerate(models):
                        predictions_array[
                            mask,
                            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):
        """
        Compute the cumulative distribution of ic50 values for a set of alleles
        over a large universe of random peptides, to enable computing quantiles in
        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.

        Returns
        ----------
        EncodableSequences : peptides used for calibration
        """
        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)

        for (i, allele) in enumerate(alleles):
            predictions = self.predict(encoded_peptides, allele=allele)
            transform = PercentRankTransform()
            transform.fit(predictions, bins=bins)
            self.allele_to_percent_rank_transform[allele] = transform

        return encoded_peptides

    def filter_networks(self, predicate):
        """
        Return a new Class1AffinityPredictor containing a subset of this
        predictor's neural networks.

        Parameters
        ----------
        predicate : Class1NeuralNetwork -> boolean
            Function specifying which neural networks to include

        Returns
        -------
        Class1AffinityPredictor
        """
        allele_to_allele_specific_models = {}
        for (allele, models) in self.allele_to_allele_specific_models.items():
            allele_to_allele_specific_models[allele] = [
                m for m in models if predicate(m)
            ]
        class1_pan_allele_models = [
            m for m in self.class1_pan_allele_models if predicate(m)
        ]

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

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

        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