from __future__ import absolute_import from __future__ import division from __future__ import print_function import collections import json import numpy as np import sys # import keras_genomics from keras import backend as K from keras.layers import ( Activation, AveragePooling1D, BatchNormalization, Convolution1D, Dense, Dropout, Flatten, Input, MaxPooling1D, Merge, Permute, Reshape, PReLU, Lambda, merge ) from tfdragonn import layers from keras.models import Model from tfdragonn import pwms def model_from_config(model_config_file_path): """Load a model from a json config file.""" thismodule = sys.modules[__name__] with open(model_config_file_path, 'r') as fp: config = json.load(fp) model_class_name = config['model_class'] model_class = getattr(thismodule, model_class_name) del config['model_class'] return model_class(**config) def model_from_config_and_queue(model_config_file_path, queue): """ Uses queue output shapes and json file with architecture params in to initialize a model """ thismodule = sys.modules[__name__] with open(model_config_file_path, 'r') as fp: config = json.load(fp) model_class_name = config['model_class'] model_class = getattr(thismodule, model_class_name) del config['model_class'] return model_class(queue.output_shapes, **config) def model_from_minimal_config(model_config_file_path, shapes, num_tasks): """ Uses queue output shapes and json file with architecture params in to initialize a model """ thismodule = sys.modules[__name__] with open(model_config_file_path, 'r') as fp: config = json.load(fp) model_class_name = config['model_class'] model_class = getattr(thismodule, model_class_name) del config['model_class'] return model_class(shapes, num_tasks, **config) def reshape_bigwig_input(x): """ Reshapes (interval_size) shaped dnase data from example queues into (interval_size, 1) """ interval_size = K.int_shape(x)[-1] return Reshape((interval_size, 1))(x) _input_reshape_func = { # "data/genome_data_dir": Permute((2, 1)), "data/HelT_data_dir": reshape_bigwig_input, "data/MGW_data_dir": reshape_bigwig_input, "data/OC2_data_dir": reshape_bigwig_input, "data/ProT_data_dir": reshape_bigwig_input, "data/H3K4me1_data_dir": reshape_bigwig_input, "data/H3K4me2_data_dir": reshape_bigwig_input, "data/H3K4me3_data_dir": reshape_bigwig_input, "data/H3K9ac_data_dir": reshape_bigwig_input, "data/H3K9me3_data_dir": reshape_bigwig_input, "data/H3K27ac_data_dir": reshape_bigwig_input, "data/H3K27me3_data_dir": reshape_bigwig_input, "data/H3K36me3_data_dir": reshape_bigwig_input, "data/H3K79me2_data_dir": reshape_bigwig_input, "data/H4K20me1_data_dir": reshape_bigwig_input, "data/MNAse_data_dir": reshape_bigwig_input, "data/dnase_data_dir": reshape_bigwig_input, "data/methylation_data_dir": reshape_bigwig_input, "data/conf_data_dir":reshape_bigwig_input } model_inputs = { "SequenceClassifier": [ "data/genome_data_dir"], "SequenceReverseComplementClassifier": [ "data/genome_data_dir"], "AmrSequenceClassifier": [ "data/genome_data_dir"], "SequenceBaselineClassifier": [ "data/genome_data_dir"], "SequenceAndDnaseClassifier": [ "data/genome_data_dir", "data/dnase_data_dir"], "SequenceAndMethylationClassifier":[ "data/genome_data_dir", "data/methylation_data_dir"], "SequenceAndMethylationClassifier_simple":[ "data/genome_data_dir", "data/methylation_data_dir"], "SequenceAndDnaseBaselineClassifier": [ "data/genome_data_dir", "data/dnase_data_dir"], "ShapeAndDnaseClassifier": [ "data/HelT_data_dir", "data/MGW_data_dir", "data/OC2_data_dir", "data/ProT_data_dir", "data/Roll_data_dir", "data/dnase_data_dir"], "SequenceDnaseTssDhsCountAndTssExpressionClassifier": [ "data/genome_data_dir", "data/dnase_data_dir", "data/dhs_counts", "data/tss_counts", "data/tss_mean_tpm", "data/tss_max_tpm"], "SequenceDnaseMethylationClassifier":[ "data/genome_data_dir", "data/methylation_data_dir", "data/dnase_data_dir"], "SequenceAndMethylationClassifierSeparate":[ "data/genome_data_dir", "data/methylation_data_dir"], "SequenceDnaseMethylationClassifier_SingleMode":[ "data/genome_data_dir", "data/methylation_data_dir", "data/dnase_data_dir"], "SequenceDnaseMethylationConfidenceClassifier":[ "data/genome_data_dir", "data/methylation_data_dir", "data/dnase_data_dir", "data/conf_data_dir"], "SequenceAndDNAseAnd11HistoneMarksClassifier_SingleMode":[ "data/genome_data_dir", "data/H3K4me1_data_dir", "data/H3K4me2_data_dir", "data/H3K4me3_data_dir", "data/H3K9ac_data_dir", "data/H3K9me3_data_dir", "data/H3K27ac_data_dir", "data/H3K27me3_data_dir", "data/H3K36me3_data_dir", "data/H3K79me2_data_dir", "data/H4K20me1_data_dir", "data/MNAse_data_dir", "data/dnase_data_dir"], "SequenceAndDNAseAnd11HistoneMarksClassifier_Separate":[ "data/genome_data_dir", "data/H3K4me1_data_dir", "data/H3K4me2_data_dir", "data/H3K4me3_data_dir", "data/H3K9ac_data_dir", "data/H3K9me3_data_dir", "data/H3K27ac_data_dir", "data/H3K27me3_data_dir", "data/H3K36me3_data_dir", "data/H3K79me2_data_dir", "data/H4K20me1_data_dir", "data/MNAse_data_dir", "data/dnase_data_dir"], "SequenceAndDNAseAnd11HistoneMarksClassifier_Separate3":[ "data/genome_data_dir", "data/H3K4me1_data_dir", "data/H3K4me2_data_dir", "data/H3K4me3_data_dir", "data/H3K9ac_data_dir", "data/H3K9me3_data_dir", "data/H3K27ac_data_dir", "data/H3K27me3_data_dir", "data/H3K36me3_data_dir", "data/H3K79me2_data_dir", "data/H4K20me1_data_dir", "data/MNAse_data_dir", "data/dnase_data_dir"], "SequenceAndDNAseAnd6HistoneMarksClassifier_SingleMode":[ "data/genome_data_dir", "data/H3K4me1_data_dir", "data/H3K4me3_data_dir", "data/H3K9me3_data_dir", "data/H3K27ac_data_dir", "data/H3K27me3_data_dir", "data/H3K36me3_data_dir", "data/dnase_data_dir"], "SequenceAndDNAseAnd6HistoneMarksClassifier_Separate":[ "data/genome_data_dir", "data/H3K4me1_data_dir", "data/H3K4me3_data_dir", "data/H3K9me3_data_dir", "data/H3K27ac_data_dir", "data/H3K27me3_data_dir", "data/H3K36me3_data_dir", "data/dnase_data_dir"], "SequenceAndDNAseAnd6HistoneMarksClassifier_Separate3":[ "data/genome_data_dir", "data/H3K4me1_data_dir", "data/H3K4me3_data_dir", "data/H3K9me3_data_dir", "data/H3K27ac_data_dir", "data/H3K27me3_data_dir", "data/H3K36me3_data_dir", "data/dnase_data_dir"], "SequenceAndDNAseAnd4HistoneMarksClassifier_SingleMode":[ "data/genome_data_dir", "data/H3K4me1_data_dir", "data/H3K9me3_data_dir", "data/H3K27ac_data_dir", "data/H3K27me3_data_dir", "data/dnase_data_dir"], "SequenceAndDNAseAnd4HistoneMarksClassifier_Separate":[ "data/genome_data_dir", "data/H3K4me1_data_dir", "data/H3K9me3_data_dir", "data/H3K27ac_data_dir", "data/H3K27me3_data_dir", "data/dnase_data_dir"], "SequenceClassifierRevCompConv": [ "data/genome_data_dir"], "SequenceAndDNAseAnd4HistoneMarksClassifier_Separate3":[ "data/genome_data_dir", "data/H3K4me1_data_dir", "data/H3K9me3_data_dir", "data/H3K27ac_data_dir", "data/H3K27me3_data_dir", "data/dnase_data_dir"] } def model_inputs_from_config(model_config_file_path): with open(model_config_file_path, 'r') as fp: config = json.load(fp) return model_inputs[config['model_class']] class Classifier(object): """ Classifier interface. Args: shapes (dict): a dict of input/output shapes. Example: `{data/genome_data_dir: (4, 1000)}` Attributes: get_inputs (list): a list of input names. Derived from model_inputs unless implemented. """ @property def get_inputs(self): return model_inputs[self.__class__.__name__] def __init__(self, shapes, **hyperparameters): pass def save(self, prefix): arch_fname = prefix + '.arch.json' weights_fname = prefix + '.weights.h5' open(arch_fname, 'w').write(self.model.to_json()) self.model.save_weights(weights_fname, overwrite=True) def save_to_yaml(self, prefix): arch_fname = prefix + '.arch.yaml' weights_fname = prefix + '.weights.h5' open(arch_fname, 'w').write(self.model.to_yaml()) self.model.save_weights(weights_fname, overwrite=True) def load_weights(self, filepath): self.model.load_weights(filepath) def get_keras_inputs(self, shapes): """Returns dictionary of named keras inputs""" return collections.OrderedDict( [(name, Input(shape=shapes[name], name=name)) for name in self.get_inputs]) @staticmethod def reshape_keras_inputs(keras_inputs): """reshapes keras inputs based on example queues""" inputs = collections.OrderedDict() for k, v in keras_inputs.items(): if k in _input_reshape_func: # reshape inputs[k] = _input_reshape_func[k](v) else: # keep as is inputs[k] = v return inputs class SequenceClassifier(Classifier): def __init__(self, shapes, num_tasks, num_filters=(15, 15, 15), conv_width=(15, 15, 15), pool_width=35, dropout=0, batch_norm=False): assert len(num_filters) == len(conv_width) # configure inputs keras_inputs = self.get_keras_inputs(shapes) inputs = self.reshape_keras_inputs(keras_inputs) # convolve sequence seq_preds = inputs["data/genome_data_dir"] for i, (nb_filter, nb_col) in enumerate(zip(num_filters, conv_width)): seq_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(seq_preds) if batch_norm: seq_preds = BatchNormalization()(seq_preds) seq_preds = Activation('relu')(seq_preds) if dropout > 0: seq_preds = Dropout(dropout)(seq_preds) # pool and fully connect seq_preds = AveragePooling1D((pool_width))(seq_preds) seq_preds = Flatten()(seq_preds) seq_preds = Dense(output_dim=num_tasks)(seq_preds) seq_preds = Activation('sigmoid')(seq_preds) self.model = Model(input=keras_inputs.values(), output=seq_preds) class SequenceBaselineClassifier(Classifier): def __init__(self, shapes, num_tasks, pwm_paths, num_filters=(15, 15, 15), conv_width=(15, 15, 15), pool_width=35, dropout=0, batch_norm=False): assert len(num_filters) == len(conv_width) # configure inputs keras_inputs = self.get_keras_inputs(shapes) inputs = self.reshape_keras_inputs(keras_inputs) # configure initialization weights # (nb_filter, filter_length, input_dim) conv_weights = pwms.pwms2conv_weights(pwm_paths) # (filter_length, input_dim, nb_filter) conv_weights = np.rollaxis(conv_weights, 0, 3) # (filter_length, 1, input_dim, nb_filter) conv_weights = np.expand_dims(conv_weights, 1) conv_biases = np.zeros((conv_weights.shape[3])) weights = [conv_weights, conv_biases] # convolve sequence with fixed known pwms seq_preds = inputs["data/genome_data_dir"] seq_preds = Convolution1D( conv_weights.shape[3], conv_weights.shape[0], weights=weights, trainable=False)(seq_preds) seq_preds = BatchNormalization()(seq_preds) # this is necessary seq_preds = Activation('relu')(seq_preds) if dropout > 0: seq_preds = Dropout(dropout)(seq_preds) # de novo convolutions for nb_filter, nb_col in zip(num_filters, conv_width): seq_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(seq_preds) if batch_norm: seq_preds = BatchNormalization()(seq_preds) seq_preds = Activation('relu')(seq_preds) if dropout > 0: seq_preds = Dropout(dropout)(seq_preds) # pool and fully connect seq_preds = AveragePooling1D((pool_width))(seq_preds) seq_preds = Flatten()(seq_preds) seq_preds = Dense(output_dim=num_tasks)(seq_preds) seq_preds = Activation('sigmoid')(seq_preds) self.model = Model(input=keras_inputs.values(), output=seq_preds) class SequenceAndDnaseClassifier(Classifier): def __init__(self, shapes, num_tasks, num_seq_filters=(25, 25, 25), seq_conv_width=(25, 25, 25), num_dnase_filters=(25, 25, 25), dnase_conv_width=(25, 25, 25), num_combined_filters=(55,), combined_conv_width=(25,), pool_width=25, fc_layer_widths=(100,), seq_conv_dropout=0.0, dnase_conv_dropout=0.0, combined_conv_dropout=0.0, fc_layer_dropout=0.0, batch_norm=False): assert len(num_seq_filters) == len(seq_conv_width) assert len(num_dnase_filters) == len(dnase_conv_width) assert len(num_combined_filters) == len(combined_conv_width) # configure inputs keras_inputs = self.get_keras_inputs(shapes) inputs = self.reshape_keras_inputs(keras_inputs) # convolve sequence seq_preds = inputs["data/genome_data_dir"] for nb_filter, nb_col in zip(num_seq_filters, seq_conv_width): seq_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(seq_preds) if batch_norm: seq_preds = BatchNormalization()(seq_preds) seq_preds = Activation('relu')(seq_preds) if seq_conv_dropout > 0: seq_preds = Dropout(seq_conv_dropout)(seq_preds) # convolve dnase dnase_preds = inputs["data/dnase_data_dir"] for nb_filter, nb_col in zip(num_dnase_filters, dnase_conv_width): dnase_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(dnase_preds) if batch_norm: dnase_preds = BatchNormalization()(dnase_preds) dnase_preds = Activation('relu')(dnase_preds) if dnase_conv_dropout > 0: dnase_preds = Dropout(dnase_conv_dropout)(dnase_preds) # stack and convolve logits = Merge(mode='concat', concat_axis=-1)([seq_preds, dnase_preds]) for nb_filter, nb_col in zip(num_combined_filters, combined_conv_width): logits = Convolution1D(nb_filter, nb_col, 'he_normal')(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if combined_conv_dropout > 0: logits = Dropout(combined_conv_dropout)(logits) # pool and fully connect logits = AveragePooling1D((pool_width))(logits) logits = Flatten()(logits) for fc_layer_width in fc_layer_widths: logits = Dense(fc_layer_width)(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if fc_layer_dropout > 0: logits = Dropout(fc_layer_dropout)(logits) logits = Dense(num_tasks)(logits) logits = Activation('sigmoid')(logits) self.model = Model(input=keras_inputs.values(), output=logits) class SequenceAndDnaseBaselineClassifier(Classifier): def __init__(self, shapes, num_tasks, pwm_paths, num_seq_filters=(25, 25, 25), seq_conv_width=(25, 25, 25), num_dnase_filters=(25, 25, 25), dnase_conv_width=(25, 25, 25), num_combined_filters=(55,), combined_conv_width=(25,), pool_width=25, fc_layer_widths=(100,), seq_conv_dropout=0.0, dnase_conv_dropout=0.0, combined_conv_dropout=0.0, fc_layer_dropout=0.0, batch_norm=False): assert len(num_seq_filters) == len(seq_conv_width) assert len(num_dnase_filters) == len(dnase_conv_width) assert len(num_combined_filters) == len(combined_conv_width) # configure inputs keras_inputs = self.get_keras_inputs(shapes) inputs = self.reshape_keras_inputs(keras_inputs) # configure initialization weights # (nb_filter, filter_length, input_dim) conv_weights = pwms.pwms2conv_weights(pwm_paths) # (filter_length, input_dim, nb_filter) conv_weights = np.rollaxis(conv_weights, 0, 3) # (filter_length, 1, input_dim, nb_filter) conv_weights = np.expand_dims(conv_weights, 1) conv_biases = np.zeros((conv_weights.shape[3])) weights = [conv_weights, conv_biases] # convolve sequence with fixed known pwms seq_preds = inputs["data/genome_data_dir"] seq_preds = Convolution1D( conv_weights.shape[3], conv_weights.shape[0], weights=weights, trainable=False)(seq_preds) seq_preds = BatchNormalization()(seq_preds) # this is necessary seq_preds = Activation('relu')(seq_preds) if seq_conv_dropout > 0: seq_preds = Dropout(dropout)(seq_preds) # convolve with de novo convolutions seq_preds = inputs["data/genome_data_dir"] for nb_filter, nb_col in zip(num_seq_filters, seq_conv_width): seq_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(seq_preds) if batch_norm: seq_preds = BatchNormalization()(seq_preds) seq_preds = Activation('relu')(seq_preds) if seq_conv_dropout > 0: seq_preds = Dropout(seq_conv_dropout)(seq_preds) # convolve dnase dnase_preds = inputs["data/dnase_data_dir"] for nb_filter, nb_col in zip(num_dnase_filters, dnase_conv_width): dnase_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(dnase_preds) if batch_norm: dnase_preds = BatchNormalization()(dnase_preds) dnase_preds = Activation('relu')(dnase_preds) if dnase_conv_dropout > 0: dnase_preds = Dropout(dnase_conv_dropout)(dnase_preds) # stack and convolve logits = Merge(mode='concat', concat_axis=-1)([seq_preds, dnase_preds]) for nb_filter, nb_col in zip(num_combined_filters, combined_conv_width): logits = Convolution1D(nb_filter, nb_col, 'he_normal')(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if combined_conv_dropout > 0: logits = Dropout(combined_conv_dropout)(logits) # pool and fully connect logits = AveragePooling1D((pool_width))(logits) logits = Flatten()(logits) for fc_layer_width in fc_layer_widths: logits = Dense(fc_layer_width)(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if fc_layer_dropout > 0: logits = Dropout(fc_layer_dropout)(logits) logits = Dense(num_tasks)(logits) logits = Activation('sigmoid')(logits) self.model = Model(input=keras_inputs.values(), output=logits) class ShapeAndDnaseClassifier(Classifier): def __init__(self, shapes, num_tasks, num_shape_filters=(25, 25, 25), shape_conv_width=(25, 25, 25), num_dnase_filters=(25, 25, 25), dnase_conv_width=(25, 25, 25), num_combined_filters=(55,), combined_conv_width=(25,), pool_width=25, fc_layer_widths=(100,), shape_conv_dropout=0.0, dnase_conv_dropout=0.0, combined_conv_dropout=0.0, fc_layer_dropout=0.0, batch_norm=False): assert len(num_shape_filters) == len(shape_conv_width) assert len(num_dnase_filters) == len(dnase_conv_width) assert len(num_combined_filters) == len(combined_conv_width) # configure inputs keras_inputs = self.get_keras_inputs(shapes) inputs = self.reshape_keras_inputs(keras_inputs) # convolve sequence shape_preds = Merge(mode='concat', concat_axis=-1)([ inputs[k] for k in ["data/HelT_data_dir", "data/MGW_data_dir", "data/OC2_data_dir", "data/ProT_data_dir", "data/Roll_data_dir"]]) for i, (nb_filter, nb_col) in enumerate(zip(num_shape_filters, shape_conv_width)): shape_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(shape_preds) if batch_norm: shape_preds = BatchNormalization()(shape_preds) shape_preds = Activation('relu')(shape_preds) if shape_conv_dropout > 0: shape_preds = Dropout(shape_conv_dropout)(shape_preds) # convolve dnase dnase_preds = inputs["data/dnase_data_dir"] for nb_filter, nb_col in zip(num_dnase_filters, dnase_conv_width): dnase_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(dnase_preds) if batch_norm: dnase_preds = BatchNormalization()(dnase_preds) dnase_preds = Activation('relu')(dnase_preds) if dnase_conv_dropout > 0: dnase_preds = Dropout(dnase_conv_dropout)(dnase_preds) # stack and convolve logits = Merge(mode='concat', concat_axis=- 1)([shape_preds, dnase_preds]) for nb_filter, nb_col in zip(num_combined_filters, combined_conv_width): logits = Convolution1D(nb_filter, nb_col, 'he_normal')(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if combined_conv_dropout > 0: logits = Dropout(combined_conv_dropout)(logits) # pool and fully connect logits = AveragePooling1D((pool_width))(logits) logits = Flatten()(logits) for fc_layer_width in fc_layer_widths: logits = Dense(fc_layer_width)(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if fc_layer_dropout > 0: logits = Dropout(fc_layer_dropout)(logits) logits = Dense(num_tasks)(logits) logits = Activation('sigmoid')(logits) self.model = Model(input=keras_inputs.values(), output=logits) class SequenceDnaseTssDhsCountAndTssExpressionClassifier(Classifier): def __init__(self, shapes, num_tasks, num_seq_filters=(25, 25, 25), seq_conv_width=(25, 25, 25), num_dnase_filters=(25, 25, 25), dnase_conv_width=(25, 25, 25), num_combined_filters=(55,), combined_conv_width=(25,), pool_width=25, seq_dnase_fc_layer_widths=(100,), final_fc_layer_widths=(100,), seq_conv_dropout=0, dnase_conv_dropout=0, combined_conv_dropout=0, seq_dnase_fc_layer_dropout=0, final_fc_dropout=0, batch_norm=False): assert len(num_seq_filters) == len(seq_conv_width) assert len(num_dnase_filters) == len(dnase_conv_width) assert len(num_combined_filters) == len(combined_conv_width) # configure inputs keras_inputs = self.get_keras_inputs(shapes) inputs = self.reshape_keras_inputs(keras_inputs) # convolve sequence seq_preds = inputs["data/genome_data_dir"] for nb_filter, nb_col in zip(num_seq_filters, seq_conv_width): seq_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(seq_preds) if batch_norm: seq_preds = BatchNormalization()(seq_preds) seq_preds = Activation('relu')(seq_preds) if seq_conv_dropout > 0: seq_preds = Dropout(seq_conv_dropout)(seq_preds) # convolve dnase dnase_preds = inputs["data/dnase_data_dir"] for nb_filter, nb_col in zip(num_dnase_filters, dnase_conv_width): dnase_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(dnase_preds) if batch_norm: dnase_preds = BatchNormalization()(dnase_preds) dnase_preds = Activation('relu')(dnase_preds) if dnase_conv_dropout > 0: dnase_preds = Dropout(dnase_conv_dropout)(dnase_preds) # stack sequence + dnase and convolve logits = Merge(mode='concat', concat_axis=-1)([seq_preds, dnase_preds]) for nb_filter, nb_col in zip(num_combined_filters, combined_conv_width): logits = Convolution1D(nb_filter, nb_col, 'he_normal')(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if combined_conv_dropout > 0: logits = Dropout(combined_conv_dropout)(logits) # pool and fully connect seq + dnase logits = AveragePooling1D((pool_width))(logits) logits = Flatten()(logits) for fc_layer_width in seq_dnase_fc_layer_widths: logits = Dense(fc_layer_width)(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if seq_dnase_fc_layer_dropout > 0: logits = Dropout(seq_dnase_fc_layer_dropout)(logits) # merge in tss+dhs counts, tss tpms and fully connected logits = Merge(mode='concat', concat_axis=-1)([ logits, inputs['data/dhs_counts'], inputs['data/tss_counts'], inputs['data/tss_mean_tpm'], inputs['data/tss_max_tpm']]) for fc_layer_width in final_fc_layer_widths: logits = Dense(fc_layer_width)(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if final_fc_dropout > 0: logits = Dropout(final_fc_dropout)(logits) logits = Dense(num_tasks)(logits) logits = Activation('sigmoid')(logits) self.model = Model(input=keras_inputs.values(), output=logits) class AmrSequenceClassifier(Classifier): def __init__(self, shapes, num_tasks, num_filters=(32, 32, 32), conv_width=(15, 14, 14), batch_norm=True, pool_width=40, pool_stride=20, fc_layer_sizes=(10,), dropout=(0.5,), final_dropout=0.5): assert len(num_filters) == len(conv_width) assert len(fc_layer_sizes) == len(dropout) # configure inputs keras_inputs = self.get_keras_inputs(shapes) inputs = self.reshape_keras_inputs(keras_inputs) # convolve sequence seq_preds = inputs["data/genome_data_dir"] for i, (nb_filter, nb_col) in enumerate(zip(num_filters, conv_width)): seq_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(seq_preds) if batch_norm: seq_preds = BatchNormalization()(seq_preds) seq_preds = Activation('relu')(seq_preds) # pool seq_preds = MaxPooling1D(pool_width, pool_stride)(seq_preds) seq_preds = Flatten()(seq_preds) # fully connect, drop before fc layers for drop_rate, fc_layer_size in zip(dropout, fc_layer_sizes): seq_preds = Dropout(dropout)(seq_preds) seq_preds = Dense(fc_layer_size)(seq_preds) seq_preds = Activation('relu')(seq_preds) seq_preds = Dropout(final_dropout)(seq_preds) seq_preds = Dense(output_dim=num_tasks)(seq_preds) seq_preds = Activation('sigmoid')(seq_preds) self.model = Model(input=keras_inputs.values(), output=seq_preds) class SequenceAndMethylationClassifier(Classifier): def __init__(self, shapes, num_tasks, num_filters=(55,50,50,50),conv_width=(25, 25, 25,25), pool_width=25, meth_pool_width=50, fc_layer_widths=(100,), conv_dropout=0.0, fc_layer_dropout=0.0, batch_norm=False): assert len(num_filters)==len(conv_width) #assert len(num_combined_filters)==len(combined_conv_width) #configure inputs keras_inputs = self.get_keras_inputs(shapes) inputs = self.reshape_keras_inputs(keras_inputs) seq_preds=inputs["data/genome_data_dir"] methylation_preds=inputs["data/methylation_data_dir"] logits = Merge(mode='concat', concat_axis=-1,name='Merged')([seq_preds, methylation_preds]) for num_filter,width in zip(num_filters,conv_width): logits=Convolution1D(num_filter,width,'he_normal')(logits) #Caught a bug here if batch_norm: logits=BatchNormalization()(logits) logits=Activation('relu')(logits) if conv_dropout>0: logits=Dropout(conv_dropout)(logits) ##Pool and flatten logits=AveragePooling1D(pool_width)(logits) logits=Flatten()(logits) methylation_preds=AveragePooling1D(meth_pool_width)(methylation_preds) #Avg pooling the methylation_preds final_dimension=K.int_shape(methylation_preds)[1] methylation_preds=Reshape((final_dimension,))(methylation_preds) ##Fully connected layers for fc_layer_width in fc_layer_widths: logits = Dense(fc_layer_width)(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if fc_layer_dropout > 0: logits = Dropout(fc_layer_dropout)(logits) # from IPython import embed # embed() logits = Merge(mode='concat', concat_axis=-1)([logits, methylation_preds]) logits = Dense(num_tasks)(logits) logits = Activation('sigmoid')(logits) self.model = Model(input=keras_inputs.values(), output=logits) self.model.save('/srv/scratch/manyu/train/dummy_model') class SequenceAndMethylationClassifier_simple(Classifier): def __init__(self, shapes, num_tasks, num_filters=(55,50,50,50),conv_width=(25, 25, 25,25), pool_width=25, fc_layer_widths=(100,), conv_dropout=0.0, fc_layer_dropout=0.0, batch_norm=False): assert len(num_filters)==len(conv_width) #assert len(num_combined_filters)==len(combined_conv_width) #configure inputs keras_inputs = self.get_keras_inputs(shapes) inputs = self.reshape_keras_inputs(keras_inputs) seq_preds=inputs["data/genome_data_dir"] methylation_preds=inputs["data/methylation_data_dir"] logits = Merge(mode='concat', concat_axis=2)([seq_preds, methylation_preds]) for num_filter,width in zip(num_filters,conv_width): logits=Convolution1D(num_filter,width,'he_normal')(logits) if batch_norm: logits=BatchNormalization()(logits) logits=Activation('relu')(logits) if conv_dropout>0: logits=Dropout(conv_dropout)(logits) logits=AveragePooling1D(pool_width)(logits) logits=Flatten()(logits) for fc_layer_width in fc_layer_widths: logits = Dense(fc_layer_width)(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if fc_layer_dropout > 0: logits = Dropout(fc_layer_dropout)(logits) # from IPython import embed # embed() #logits = Merge(mode='concat', concat_axis=-1)([logits, methylation_preds]) logits = Dense(num_tasks)(logits) logits = Activation('sigmoid')(logits) self.model = Model(input=keras_inputs.values(), output=logits) class SequenceDnaseMethylationClassifier(Classifier): def __init__(self, shapes, num_tasks, num_seq_filters=(50, 50, 50), seq_conv_width=(25, 25, 25), num_dnase_filters=(25, 25, 25), dnase_conv_width=(25, 25, 25), num_combined_filters=(55,), combined_conv_width=(25,), pool_width=25, fc_layer_widths=(100,), seq_conv_dropout=0.0, dnase_conv_dropout=0.0, combined_conv_dropout=0.0, fc_layer_dropout=0.0, batch_norm=False): assert len(num_seq_filters) == len(seq_conv_width) assert len(num_dnase_filters) == len(dnase_conv_width) assert len(num_combined_filters) == len(combined_conv_width) # configure inputs keras_inputs = self.get_keras_inputs(shapes) inputs = self.reshape_keras_inputs(keras_inputs) # convolve sequence after stacking with Methylation # from IPython import embed # embed() seq_preds = inputs["data/genome_data_dir"] methylation_preds=inputs["data/methylation_data_dir"] logits = Merge(mode='concat', concat_axis=2)([seq_preds, methylation_preds]) for nb_filter, nb_col in zip(num_seq_filters, seq_conv_width): logits = Convolution1D( nb_filter, nb_col, 'he_normal')(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if seq_conv_dropout > 0: logits = Dropout(seq_conv_dropout)(logits) # convolve dnase dnase_preds = inputs["data/dnase_data_dir"] for nb_filter, nb_col in zip(num_dnase_filters, dnase_conv_width): dnase_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(dnase_preds) if batch_norm: dnase_preds = BatchNormalization()(dnase_preds) dnase_preds = Activation('relu')(dnase_preds) if dnase_conv_dropout > 0: dnase_preds = Dropout(dnase_conv_dropout)(dnase_preds) # stack and convolve logits = Merge(mode='concat', concat_axis=2)([logits, dnase_preds]) for nb_filter, nb_col in zip(num_combined_filters, combined_conv_width): logits = Convolution1D(nb_filter, nb_col, 'he_normal')(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if combined_conv_dropout > 0: logits = Dropout(combined_conv_dropout)(logits) # pool and fully connect logits = AveragePooling1D((pool_width))(logits) logits = Flatten()(logits) for fc_layer_width in fc_layer_widths: logits = Dense(fc_layer_width)(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if fc_layer_dropout > 0: logits = Dropout(fc_layer_dropout)(logits) logits = Dense(num_tasks)(logits) logits = Activation('sigmoid')(logits) self.model = Model(input=keras_inputs.values(), output=logits) class SequenceDnaseMethylationConfidenceClassifier(Classifier): def __init__(self, shapes, num_tasks, num_seq_filters=(50, 50, 50), seq_conv_width=(25, 25, 25), num_dnase_filters=(25, 25, 25), dnase_conv_width=(25, 25, 25), num_combined_filters=(55,), combined_conv_width=(25,), pool_width=25, fc_layer_widths=(100,), seq_conv_dropout=0.0, dnase_conv_dropout=0.0, combined_conv_dropout=0.0, fc_layer_dropout=0.0, batch_norm=False): assert len(num_seq_filters) == len(seq_conv_width) assert len(num_dnase_filters) == len(dnase_conv_width) assert len(num_combined_filters) == len(combined_conv_width) # configure inputs keras_inputs = self.get_keras_inputs(shapes) inputs = self.reshape_keras_inputs(keras_inputs) # convolve sequence after stacking with Methylation # from IPython import embed # embed() seq_preds = inputs["data/genome_data_dir"] methylation_preds=inputs["data/methylation_data_dir"] conf_preds=inputs["data/conf_data_dir"] logits = Merge(mode='concat', concat_axis=2)([seq_preds, methylation_preds,conf_preds]) for nb_filter, nb_col in zip(num_seq_filters, seq_conv_width): logits = Convolution1D( nb_filter, nb_col, 'he_normal')(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if seq_conv_dropout > 0: logits = Dropout(seq_conv_dropout)(logits) # convolve dnase dnase_preds = inputs["data/dnase_data_dir"] for nb_filter, nb_col in zip(num_dnase_filters, dnase_conv_width): dnase_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(dnase_preds) if batch_norm: dnase_preds = BatchNormalization()(dnase_preds) dnase_preds = Activation('relu')(dnase_preds) if dnase_conv_dropout > 0: dnase_preds = Dropout(dnase_conv_dropout)(dnase_preds) # stack and convolve logits = Merge(mode='concat', concat_axis=2)([logits, dnase_preds]) for nb_filter, nb_col in zip(num_combined_filters, combined_conv_width): logits = Convolution1D(nb_filter, nb_col, 'he_normal')(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if combined_conv_dropout > 0: logits = Dropout(combined_conv_dropout)(logits) # pool and fully connect logits = AveragePooling1D((pool_width))(logits) logits = Flatten()(logits) for fc_layer_width in fc_layer_widths: logits = Dense(fc_layer_width)(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if fc_layer_dropout > 0: logits = Dropout(fc_layer_dropout)(logits) logits = Dense(num_tasks)(logits) logits = Activation('sigmoid')(logits) self.model = Model(input=keras_inputs.values(), output=logits) class SequenceDnaseMethylationClassifier_SingleMode(Classifier): def __init__(self, shapes, num_tasks, num_filters=(100, 100, 50,50), conv_width=(25, 25, 25,25), pool_width=25, fc_layer_widths=(100,), seq_conv_dropout=0.0, fc_layer_dropout=0.0, batch_norm=False): assert len(num_filters) == len(conv_width) # assert len(num_dnase_filters) == len(dnase_conv_width) # assert len(num_combined_filters) == len(combined_conv_width) # configure inputs keras_inputs = self.get_keras_inputs(shapes) inputs = self.reshape_keras_inputs(keras_inputs) # convolve sequence after stacking with Methylation # from IPython import embed # embed() seq_preds = inputs["data/genome_data_dir"] methylation_preds=inputs["data/methylation_data_dir"] dnase_preds=inputs["data/dnase_data_dir"] logits = Merge(mode='concat', concat_axis=2)([seq_preds, methylation_preds,dnase_preds]) for nb_filter, nb_col in zip(num_filters, conv_width): logits = Convolution1D( nb_filter, nb_col, 'he_normal')(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if seq_conv_dropout > 0: logits = Dropout(seq_conv_dropout)(logits) # pool and fully connect logits = AveragePooling1D((pool_width))(logits) logits = Flatten()(logits) for fc_layer_width in fc_layer_widths: logits = Dense(fc_layer_width)(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if fc_layer_dropout > 0: logits = Dropout(fc_layer_dropout)(logits) logits = Dense(num_tasks)(logits) logits = Activation('sigmoid')(logits) self.model = Model(input=keras_inputs.values(), output=logits) class SequenceAndMethylationClassifierSeparate(Classifier): def __init__(self, shapes, num_tasks, num_seq_filters=(25, 25, 25), seq_conv_width=(25, 25, 25), num_methylation_filters=(25, 25, 25), methylation_conv_width=(25, 25, 25), num_combined_filters=(55,), combined_conv_width=(25,), pool_width=25, fc_layer_widths=(100,), seq_conv_dropout=0.0, meth_conv_dropout=0.0, combined_conv_dropout=0.0, fc_layer_dropout=0.0, batch_norm=True): assert len(num_seq_filters) == len(seq_conv_width) assert len(num_methylation_filters) == len(methylation_conv_width) assert len(num_combined_filters) == len(combined_conv_width) # configure inputs keras_inputs = self.get_keras_inputs(shapes) inputs = self.reshape_keras_inputs(keras_inputs) # convolve sequence # from IPython import embed # embed() seq_preds = inputs["data/genome_data_dir"] for nb_filter, nb_col in zip(num_seq_filters, seq_conv_width): seq_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(seq_preds) if batch_norm: seq_preds = BatchNormalization()(seq_preds) seq_preds = Activation('relu')(seq_preds) if seq_conv_dropout > 0: seq_preds = Dropout(seq_conv_dropout)(seq_preds) # convolve methylation meth_preds = inputs["data/methylation_data_dir"] for nb_filter, nb_col in zip(num_methylation_filters, methylation_conv_width): meth_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(meth_preds) if batch_norm: meth_preds = BatchNormalization()(meth_preds) meth_preds = Activation('relu')(meth_preds) if meth_conv_dropout > 0: meth_preds = Dropout(meth_conv_dropout)(meth_preds) # stack and convolve logits = Merge(mode='concat', concat_axis=2)([seq_preds, meth_preds]) for nb_filter, nb_col in zip(num_combined_filters, combined_conv_width): logits = Convolution1D(nb_filter, nb_col, 'he_normal')(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if combined_conv_dropout > 0: logits = Dropout(combined_conv_dropout)(logits) # pool and fully connect logits = AveragePooling1D((pool_width))(logits) logits = Flatten()(logits) for fc_layer_width in fc_layer_widths: logits = Dense(fc_layer_width)(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if fc_layer_dropout > 0: logits = Dropout(fc_layer_dropout)(logits) logits = Dense(num_tasks)(logits) logits = Activation('sigmoid')(logits) self.model = Model(input=keras_inputs.values(), output=logits) class SequenceAndDNAseAnd11HistoneMarksClassifier_SingleMode(Classifier): def __init__(self, shapes, num_tasks, num_filters=(100, 100, 50,50), conv_width=(25, 25, 25,25), pool_width=25, fc_layer_widths=(100,), seq_conv_dropout=0.0, fc_layer_dropout=0.0, batch_norm=False): assert len(num_filters) == len(conv_width) # assert len(num_dnase_filters) == len(dnase_conv_width) # assert len(num_combined_filters) == len(combined_conv_width) # configure inputs keras_inputs = self.get_keras_inputs(shapes) inputs = self.reshape_keras_inputs(keras_inputs) # convolve sequence after stacking with histone marks # from IPython import embed # embed() seq_preds = inputs["data/genome_data_dir"] dnase_preds=inputs["data/dnase_data_dir"] H3K4me1_preds=inputs["data/H3K4me1_data_dir"] H3K4me2_preds=inputs["data/H3K4me2_data_dir"] H3K4me3_preds=inputs["data/H3K4me3_data_dir"] H3K9ac_preds=inputs["data/H3K9ac_data_dir"] H3K9me3_preds=inputs["data/H3K9me3_data_dir"] H3K27ac_preds=inputs["data/H3K27ac_data_dir"] H3K27me3_preds=inputs["data/H3K27me3_data_dir"] H3K36me3_preds=inputs["data/H3K36me3_data_dir"] H3K79me2_preds=inputs["data/H3K79me2_data_dir"] H4K20me1_preds=inputs["data/H4K20me1_data_dir"] MNAse_preds=inputs["data/MNAse_data_dir"] logits = Merge(mode='concat', concat_axis=2)([seq_preds, dnase_preds, MNAse_preds, H3K4me1_preds, H3K4me2_preds, H3K4me3_preds, H3K9ac_preds, H3K9me3_preds, H3K27ac_preds, H3K27me3_preds, H3K36me3_preds, H3K79me2_preds, H4K20me1_preds]) for nb_filter, nb_col in zip(num_filters, conv_width): logits = Convolution1D( nb_filter, nb_col, 'he_normal')(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if seq_conv_dropout > 0: logits = Dropout(seq_conv_dropout)(logits) # pool and fully connect logits = AveragePooling1D((pool_width))(logits) logits = Flatten()(logits) for fc_layer_width in fc_layer_widths: logits = Dense(fc_layer_width)(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if fc_layer_dropout > 0: logits = Dropout(fc_layer_dropout)(logits) logits = Dense(num_tasks)(logits) logits = Activation('sigmoid')(logits) self.model = Model(input=keras_inputs.values(), output=logits) class SequenceAndDNAseAnd11HistoneMarksClassifier_Separate(Classifier): def __init__(self, shapes, num_tasks, num_seq_filters=(25, 25, 25), seq_conv_width=(25, 25, 25), num_methylation_filters=(25, 25, 25), methylation_conv_width=(25, 25, 25), num_combined_filters=(55,), combined_conv_width=(25,), pool_width=25, fc_layer_widths=(100,), seq_conv_dropout=0.0, meth_conv_dropout=0.0, combined_conv_dropout=0.0, fc_layer_dropout=0.0, batch_norm=True): assert len(num_seq_filters) == len(seq_conv_width) assert len(num_methylation_filters) == len(methylation_conv_width) assert len(num_combined_filters) == len(combined_conv_width) # configure inputs keras_inputs = self.get_keras_inputs(shapes) inputs = self.reshape_keras_inputs(keras_inputs) # convolve sequence # from IPython import embed # embed() seq_preds = inputs["data/genome_data_dir"] for nb_filter, nb_col in zip(num_seq_filters, seq_conv_width): seq_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(seq_preds) if batch_norm: seq_preds = BatchNormalization()(seq_preds) seq_preds = Activation('relu')(seq_preds) if seq_conv_dropout > 0: seq_preds = Dropout(seq_conv_dropout)(seq_preds) # convolve histone marks dnase_preds=inputs["data/dnase_data_dir"] H3K4me1_preds=inputs["data/H3K4me1_data_dir"] H3K4me2_preds=inputs["data/H3K4me2_data_dir"] H3K4me3_preds=inputs["data/H3K4me3_data_dir"] H3K9ac_preds=inputs["data/H3K9ac_data_dir"] H3K9me3_preds=inputs["data/H3K9me3_data_dir"] H3K27ac_preds=inputs["data/H3K27ac_data_dir"] H3K27me3_preds=inputs["data/H3K27me3_data_dir"] H3K36me3_preds=inputs["data/H3K36me3_data_dir"] H3K79me2_preds=inputs["data/H3K79me2_data_dir"] H4K20me1_preds=inputs["data/H4K20me1_data_dir"] MNAse_preds=inputs["data/MNAse_data_dir"] meth_preds = Merge(mode='concat', concat_axis=2)([dnase_preds, MNAse_preds, H3K4me1_preds, H3K4me2_preds, H3K4me3_preds, H3K9ac_preds, H3K9me3_preds, H3K27ac_preds, H3K27me3_preds, H3K36me3_preds, H3K79me2_preds, H4K20me1_preds]) for nb_filter, nb_col in zip(num_methylation_filters, methylation_conv_width): meth_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(meth_preds) if batch_norm: meth_preds = BatchNormalization()(meth_preds) meth_preds = Activation('relu')(meth_preds) if meth_conv_dropout > 0: meth_preds = Dropout(meth_conv_dropout)(meth_preds) # stack and convolve logits = Merge(mode='concat', concat_axis=2)([seq_preds, meth_preds]) for nb_filter, nb_col in zip(num_combined_filters, combined_conv_width): logits = Convolution1D(nb_filter, nb_col, 'he_normal')(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if combined_conv_dropout > 0: logits = Dropout(combined_conv_dropout)(logits) # pool and fully connect logits = AveragePooling1D((pool_width))(logits) logits = Flatten()(logits) for fc_layer_width in fc_layer_widths: logits = Dense(fc_layer_width)(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if fc_layer_dropout > 0: logits = Dropout(fc_layer_dropout)(logits) logits = Dense(num_tasks)(logits) logits = Activation('sigmoid')(logits) self.model = Model(input=keras_inputs.values(), output=logits) class SequenceAndDNAseAnd11HistoneMarksClassifier_Separate3(Classifier): def __init__(self, shapes, num_tasks, num_seq_filters=(25, 25, 25), seq_conv_width=(25, 25, 25), num_methylation_filters=(25, 25, 25), methylation_conv_width=(25, 25, 25), num_combined_filters=(55,), combined_conv_width=(25,), pool_width=25, fc_layer_widths=(100,), seq_conv_dropout=0.0, meth_conv_dropout=0.0, combined_conv_dropout=0.0, fc_layer_dropout=0.0, batch_norm=True): assert len(num_seq_filters) == len(seq_conv_width) assert len(num_methylation_filters) == len(methylation_conv_width) assert len(num_combined_filters) == len(combined_conv_width) # configure inputs keras_inputs = self.get_keras_inputs(shapes) inputs = self.reshape_keras_inputs(keras_inputs) # convolve sequence # from IPython import embed # embed() seq_preds = inputs["data/genome_data_dir"] for nb_filter, nb_col in zip(num_seq_filters, seq_conv_width): seq_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(seq_preds) if batch_norm: seq_preds = BatchNormalization()(seq_preds) seq_preds = Activation('relu')(seq_preds) if seq_conv_dropout > 0: seq_preds = Dropout(seq_conv_dropout)(seq_preds) # convolve DNAse dnase_preds=inputs["data/dnase_data_dir"] for nb_filter, nb_col in zip(num_methylation_filters, methylation_conv_width): dnase_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(dnase_preds) if batch_norm: dnase_preds = BatchNormalization()(dnase_preds) dnase_preds = Activation('relu')(dnase_preds) if meth_conv_dropout > 0: dnase_preds = Dropout(meth_conv_dropout)(dnase_preds) # convolve histone marks H3K4me1_preds=inputs["data/H3K4me1_data_dir"] H3K4me2_preds=inputs["data/H3K4me2_data_dir"] H3K4me3_preds=inputs["data/H3K4me3_data_dir"] H3K9ac_preds=inputs["data/H3K9ac_data_dir"] H3K9me3_preds=inputs["data/H3K9me3_data_dir"] H3K27ac_preds=inputs["data/H3K27ac_data_dir"] H3K27me3_preds=inputs["data/H3K27me3_data_dir"] H3K36me3_preds=inputs["data/H3K36me3_data_dir"] H3K79me2_preds=inputs["data/H3K79me2_data_dir"] H4K20me1_preds=inputs["data/H4K20me1_data_dir"] MNAse_preds=inputs["data/MNAse_data_dir"] meth_preds = Merge(mode='concat', concat_axis=2)([MNAse_preds, H3K4me1_preds, H3K4me2_preds, H3K4me3_preds, H3K9ac_preds, H3K9me3_preds, H3K27ac_preds, H3K27me3_preds, H3K36me3_preds, H3K79me2_preds, H4K20me1_preds]) for nb_filter, nb_col in zip(num_methylation_filters, methylation_conv_width): meth_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(meth_preds) if batch_norm: meth_preds = BatchNormalization()(meth_preds) meth_preds = Activation('relu')(meth_preds) if meth_conv_dropout > 0: meth_preds = Dropout(meth_conv_dropout)(meth_preds) # stack and convolve logits = Merge(mode='concat', concat_axis=2)([seq_preds, dnase_preds, meth_preds]) for nb_filter, nb_col in zip(num_combined_filters, combined_conv_width): logits = Convolution1D(nb_filter, nb_col, 'he_normal')(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if combined_conv_dropout > 0: logits = Dropout(combined_conv_dropout)(logits) # pool and fully connect logits = AveragePooling1D((pool_width))(logits) logits = Flatten()(logits) for fc_layer_width in fc_layer_widths: logits = Dense(fc_layer_width)(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if fc_layer_dropout > 0: logits = Dropout(fc_layer_dropout)(logits) logits = Dense(num_tasks)(logits) logits = Activation('sigmoid')(logits) self.model = Model(input=keras_inputs.values(), output=logits) class SequenceAndDNAseAnd6HistoneMarksClassifier_SingleMode(Classifier): def __init__(self, shapes, num_tasks, num_filters=(100, 100, 50,50), conv_width=(25, 25, 25,25), pool_width=25, fc_layer_widths=(100,), seq_conv_dropout=0.0, fc_layer_dropout=0.0, batch_norm=False): assert len(num_filters) == len(conv_width) # assert len(num_dnase_filters) == len(dnase_conv_width) # assert len(num_combined_filters) == len(combined_conv_width) # configure inputs keras_inputs = self.get_keras_inputs(shapes) inputs = self.reshape_keras_inputs(keras_inputs) # convolve sequence after stacking with histone marks # from IPython import embed # embed() seq_preds = inputs["data/genome_data_dir"] dnase_preds=inputs["data/dnase_data_dir"] H3K4me1_preds=inputs["data/H3K4me1_data_dir"] H3K4me3_preds=inputs["data/H3K4me3_data_dir"] H3K9me3_preds=inputs["data/H3K9me3_data_dir"] H3K27ac_preds=inputs["data/H3K27ac_data_dir"] H3K27me3_preds=inputs["data/H3K27me3_data_dir"] H3K36me3_preds=inputs["data/H3K36me3_data_dir"] logits = Merge(mode='concat', concat_axis=2)([seq_preds, dnase_preds, H3K4me1_preds, H3K4me3_preds, H3K9me3_preds, H3K27ac_preds, H3K27me3_preds, H3K36me3_preds]) for nb_filter, nb_col in zip(num_filters, conv_width): logits = Convolution1D( nb_filter, nb_col, 'he_normal')(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if seq_conv_dropout > 0: logits = Dropout(seq_conv_dropout)(logits) # pool and fully connect logits = AveragePooling1D((pool_width))(logits) logits = Flatten()(logits) for fc_layer_width in fc_layer_widths: logits = Dense(fc_layer_width)(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if fc_layer_dropout > 0: logits = Dropout(fc_layer_dropout)(logits) logits = Dense(num_tasks)(logits) logits = Activation('sigmoid')(logits) self.model = Model(input=keras_inputs.values(), output=logits) class SequenceAndDNAseAnd6HistoneMarksClassifier_Separate(Classifier): def __init__(self, shapes, num_tasks, num_seq_filters=(25, 25, 25), seq_conv_width=(25, 25, 25), num_methylation_filters=(25, 25, 25), methylation_conv_width=(25, 25, 25), num_combined_filters=(55,), combined_conv_width=(25,), pool_width=25, fc_layer_widths=(100,), seq_conv_dropout=0.0, meth_conv_dropout=0.0, combined_conv_dropout=0.0, fc_layer_dropout=0.0, batch_norm=True): assert len(num_seq_filters) == len(seq_conv_width) assert len(num_methylation_filters) == len(methylation_conv_width) assert len(num_combined_filters) == len(combined_conv_width) # configure inputs keras_inputs = self.get_keras_inputs(shapes) inputs = self.reshape_keras_inputs(keras_inputs) # convolve sequence # from IPython import embed # embed() seq_preds = inputs["data/genome_data_dir"] for nb_filter, nb_col in zip(num_seq_filters, seq_conv_width): seq_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(seq_preds) if batch_norm: seq_preds = BatchNormalization()(seq_preds) seq_preds = Activation('relu')(seq_preds) if seq_conv_dropout > 0: seq_preds = Dropout(seq_conv_dropout)(seq_preds) # convolve histone marks dnase_preds=inputs["data/dnase_data_dir"] H3K4me1_preds=inputs["data/H3K4me1_data_dir"] H3K4me3_preds=inputs["data/H3K4me3_data_dir"] H3K9me3_preds=inputs["data/H3K9me3_data_dir"] H3K27ac_preds=inputs["data/H3K27ac_data_dir"] H3K27me3_preds=inputs["data/H3K27me3_data_dir"] H3K36me3_preds=inputs["data/H3K36me3_data_dir"] meth_preds = Merge(mode='concat', concat_axis=2)([dnase_preds, H3K4me1_preds, H3K4me3_preds, H3K9me3_preds, H3K27ac_preds, H3K27me3_preds, H3K36me3_preds]) for nb_filter, nb_col in zip(num_methylation_filters, methylation_conv_width): meth_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(meth_preds) if batch_norm: meth_preds = BatchNormalization()(meth_preds) meth_preds = Activation('relu')(meth_preds) if meth_conv_dropout > 0: meth_preds = Dropout(meth_conv_dropout)(meth_preds) # stack and convolve logits = Merge(mode='concat', concat_axis=2)([seq_preds, meth_preds]) for nb_filter, nb_col in zip(num_combined_filters, combined_conv_width): logits = Convolution1D(nb_filter, nb_col, 'he_normal')(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if combined_conv_dropout > 0: logits = Dropout(combined_conv_dropout)(logits) # pool and fully connect logits = AveragePooling1D((pool_width))(logits) logits = Flatten()(logits) for fc_layer_width in fc_layer_widths: logits = Dense(fc_layer_width)(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if fc_layer_dropout > 0: logits = Dropout(fc_layer_dropout)(logits) logits = Dense(num_tasks)(logits) logits = Activation('sigmoid')(logits) self.model = Model(input=keras_inputs.values(), output=logits) class SequenceAndDNAseAnd6HistoneMarksClassifier_Separate3(Classifier): def __init__(self, shapes, num_tasks, num_seq_filters=(25, 25, 25), seq_conv_width=(25, 25, 25), num_methylation_filters=(25, 25, 25), methylation_conv_width=(25, 25, 25), num_combined_filters=(55,), combined_conv_width=(25,), pool_width=25, fc_layer_widths=(100,), seq_conv_dropout=0.0, meth_conv_dropout=0.0, combined_conv_dropout=0.0, fc_layer_dropout=0.0, batch_norm=True): assert len(num_seq_filters) == len(seq_conv_width) assert len(num_methylation_filters) == len(methylation_conv_width) assert len(num_combined_filters) == len(combined_conv_width) # configure inputs keras_inputs = self.get_keras_inputs(shapes) inputs = self.reshape_keras_inputs(keras_inputs) # convolve sequence # from IPython import embed # embed() seq_preds = inputs["data/genome_data_dir"] for nb_filter, nb_col in zip(num_seq_filters, seq_conv_width): seq_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(seq_preds) if batch_norm: seq_preds = BatchNormalization()(seq_preds) seq_preds = Activation('relu')(seq_preds) if seq_conv_dropout > 0: seq_preds = Dropout(seq_conv_dropout)(seq_preds) # convolve DNAse dnase_preds=inputs["data/dnase_data_dir"] for nb_filter, nb_col in zip(num_methylation_filters, methylation_conv_width): dnase_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(dnase_preds) if batch_norm: dnase_preds = BatchNormalization()(dnase_preds) dnase_preds = Activation('relu')(dnase_preds) if meth_conv_dropout > 0: dnase_preds = Dropout(meth_conv_dropout)(dnase_preds) # convolve histone marks H3K4me1_preds=inputs["data/H3K4me1_data_dir"] H3K4me3_preds=inputs["data/H3K4me3_data_dir"] H3K9me3_preds=inputs["data/H3K9me3_data_dir"] H3K27ac_preds=inputs["data/H3K27ac_data_dir"] H3K27me3_preds=inputs["data/H3K27me3_data_dir"] H3K36me3_preds=inputs["data/H3K36me3_data_dir"] meth_preds = Merge(mode='concat', concat_axis=2)([H3K4me1_preds, H3K4me3_preds, H3K9me3_preds, H3K27ac_preds, H3K27me3_preds, H3K36me3_preds]) for nb_filter, nb_col in zip(num_methylation_filters, methylation_conv_width): meth_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(meth_preds) if batch_norm: meth_preds = BatchNormalization()(meth_preds) meth_preds = Activation('relu')(meth_preds) if meth_conv_dropout > 0: meth_preds = Dropout(meth_conv_dropout)(meth_preds) # stack and convolve logits = Merge(mode='concat', concat_axis=2)([seq_preds, dnase_preds, meth_preds]) for nb_filter, nb_col in zip(num_combined_filters, combined_conv_width): logits = Convolution1D(nb_filter, nb_col, 'he_normal')(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if combined_conv_dropout > 0: logits = Dropout(combined_conv_dropout)(logits) # pool and fully connect logits = AveragePooling1D((pool_width))(logits) logits = Flatten()(logits) for fc_layer_width in fc_layer_widths: logits = Dense(fc_layer_width)(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if fc_layer_dropout > 0: logits = Dropout(fc_layer_dropout)(logits) logits = Dense(num_tasks)(logits) logits = Activation('sigmoid')(logits) self.model = Model(input=keras_inputs.values(), output=logits) class SequenceAndDNAseAnd4HistoneMarksClassifier_SingleMode(Classifier): def __init__(self, shapes, num_tasks, num_filters=(100, 100, 50,50), conv_width=(25, 25, 25,25), pool_width=25, fc_layer_widths=(100,), seq_conv_dropout=0.0, fc_layer_dropout=0.0, batch_norm=False): assert len(num_filters) == len(conv_width) # assert len(num_dnase_filters) == len(dnase_conv_width) # assert len(num_combined_filters) == len(combined_conv_width) # configure inputs keras_inputs = self.get_keras_inputs(shapes) inputs = self.reshape_keras_inputs(keras_inputs) # convolve sequence after stacking with histone marks # from IPython import embed # embed() seq_preds = inputs["data/genome_data_dir"] dnase_preds=inputs["data/dnase_data_dir"] H3K4me1_preds=inputs["data/H3K4me1_data_dir"] H3K9me3_preds=inputs["data/H3K9me3_data_dir"] H3K27ac_preds=inputs["data/H3K27ac_data_dir"] H3K27me3_preds=inputs["data/H3K27me3_data_dir"] logits = Merge(mode='concat', concat_axis=2)([seq_preds, dnase_preds, H3K4me1_preds, H3K9me3_preds, H3K27ac_preds, H3K27me3_preds]) for nb_filter, nb_col in zip(num_filters, conv_width): logits = Convolution1D( nb_filter, nb_col, 'he_normal')(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if seq_conv_dropout > 0: logits = Dropout(seq_conv_dropout)(logits) # pool and fully connect logits = AveragePooling1D((pool_width))(logits) logits = Flatten()(logits) for fc_layer_width in fc_layer_widths: logits = Dense(fc_layer_width)(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if fc_layer_dropout > 0: logits = Dropout(fc_layer_dropout)(logits) logits = Dense(num_tasks)(logits) logits = Activation('sigmoid')(logits) self.model = Model(input=keras_inputs.values(), output=logits) class SequenceAndDNAseAnd4HistoneMarksClassifier_Separate(Classifier): def __init__(self, shapes, num_tasks, num_seq_filters=(25, 25, 25), seq_conv_width=(25, 25, 25), num_methylation_filters=(25, 25, 25), methylation_conv_width=(25, 25, 25), num_combined_filters=(55,), combined_conv_width=(25,), pool_width=25, fc_layer_widths=(100,), seq_conv_dropout=0.0, meth_conv_dropout=0.0, combined_conv_dropout=0.0, fc_layer_dropout=0.0, batch_norm=True): assert len(num_seq_filters) == len(seq_conv_width) assert len(num_methylation_filters) == len(methylation_conv_width) assert len(num_combined_filters) == len(combined_conv_width) # configure inputs keras_inputs = self.get_keras_inputs(shapes) inputs = self.reshape_keras_inputs(keras_inputs) # convolve sequence # from IPython import embed # embed() seq_preds = inputs["data/genome_data_dir"] for nb_filter, nb_col in zip(num_seq_filters, seq_conv_width): seq_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(seq_preds) if batch_norm: seq_preds = BatchNormalization()(seq_preds) seq_preds = Activation('relu')(seq_preds) if seq_conv_dropout > 0: seq_preds = Dropout(seq_conv_dropout)(seq_preds) # convolve histone marks dnase_preds=inputs["data/dnase_data_dir"] H3K4me1_preds=inputs["data/H3K4me1_data_dir"] H3K9me3_preds=inputs["data/H3K9me3_data_dir"] H3K27ac_preds=inputs["data/H3K27ac_data_dir"] H3K27me3_preds=inputs["data/H3K27me3_data_dir"] meth_preds = Merge(mode='concat', concat_axis=2)([dnase_preds, H3K4me1_preds, H3K9me3_preds, H3K27ac_preds, H3K27me3_preds]) for nb_filter, nb_col in zip(num_methylation_filters, methylation_conv_width): meth_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(meth_preds) if batch_norm: meth_preds = BatchNormalization()(meth_preds) meth_preds = Activation('relu')(meth_preds) if meth_conv_dropout > 0: meth_preds = Dropout(meth_conv_dropout)(meth_preds) # stack and convolve logits = Merge(mode='concat', concat_axis=2)([seq_preds, meth_preds]) for nb_filter, nb_col in zip(num_combined_filters, combined_conv_width): logits = Convolution1D(nb_filter, nb_col, 'he_normal')(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if combined_conv_dropout > 0: logits = Dropout(combined_conv_dropout)(logits) # pool and fully connect logits = AveragePooling1D((pool_width))(logits) logits = Flatten()(logits) for fc_layer_width in fc_layer_widths: logits = Dense(fc_layer_width)(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if fc_layer_dropout > 0: logits = Dropout(fc_layer_dropout)(logits) logits = Dense(num_tasks)(logits) logits = Activation('sigmoid')(logits) self.model = Model(input=keras_inputs.values(), output=logits) class SequenceAndDNAseAnd4HistoneMarksClassifier_Separate3(Classifier): def __init__(self, shapes, num_tasks, num_seq_filters=(25, 25, 25), seq_conv_width=(25, 25, 25), num_methylation_filters=(25, 25, 25), methylation_conv_width=(25, 25, 25), num_combined_filters=(55,), combined_conv_width=(25,), pool_width=25, fc_layer_widths=(100,), seq_conv_dropout=0.0, meth_conv_dropout=0.0, combined_conv_dropout=0.0, fc_layer_dropout=0.0, batch_norm=True): assert len(num_seq_filters) == len(seq_conv_width) assert len(num_methylation_filters) == len(methylation_conv_width) assert len(num_combined_filters) == len(combined_conv_width) # configure inputs keras_inputs = self.get_keras_inputs(shapes) inputs = self.reshape_keras_inputs(keras_inputs) # convolve sequence # from IPython import embed # embed() seq_preds = inputs["data/genome_data_dir"] for nb_filter, nb_col in zip(num_seq_filters, seq_conv_width): seq_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(seq_preds) if batch_norm: seq_preds = BatchNormalization()(seq_preds) seq_preds = Activation('relu')(seq_preds) if seq_conv_dropout > 0: seq_preds = Dropout(seq_conv_dropout)(seq_preds) # convolve DNAse dnase_preds=inputs["data/dnase_data_dir"] for nb_filter, nb_col in zip(num_methylation_filters, methylation_conv_width): dnase_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(dnase_preds) if batch_norm: dnase_preds = BatchNormalization()(dnase_preds) dnase_preds = Activation('relu')(dnase_preds) if meth_conv_dropout > 0: dnase_preds = Dropout(meth_conv_dropout)(dnase_preds) # convolve histone marks H3K4me1_preds=inputs["data/H3K4me1_data_dir"] H3K9me3_preds=inputs["data/H3K9me3_data_dir"] H3K27ac_preds=inputs["data/H3K27ac_data_dir"] H3K27me3_preds=inputs["data/H3K27me3_data_dir"] meth_preds = Merge(mode='concat', concat_axis=2)([H3K4me1_preds, H3K9me3_preds, H3K27ac_preds, H3K27me3_preds]) for nb_filter, nb_col in zip(num_methylation_filters, methylation_conv_width): meth_preds = Convolution1D( nb_filter, nb_col, 'he_normal')(meth_preds) if batch_norm: meth_preds = BatchNormalization()(meth_preds) meth_preds = Activation('relu')(meth_preds) if meth_conv_dropout > 0: meth_preds = Dropout(meth_conv_dropout)(meth_preds) # stack and convolve logits = Merge(mode='concat', concat_axis=2)([seq_preds, dnase_preds, meth_preds]) for nb_filter, nb_col in zip(num_combined_filters, combined_conv_width): logits = Convolution1D(nb_filter, nb_col, 'he_normal')(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if combined_conv_dropout > 0: logits = Dropout(combined_conv_dropout)(logits) # pool and fully connect logits = AveragePooling1D((pool_width))(logits) logits = Flatten()(logits) for fc_layer_width in fc_layer_widths: logits = Dense(fc_layer_width)(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) if fc_layer_dropout > 0: logits = Dropout(fc_layer_dropout)(logits) logits = Dense(num_tasks)(logits) logits = Activation('sigmoid')(logits) self.model = Model(input=keras_inputs.values(), output=logits) class SequenceReverseComplementClassifier(Classifier): def __init__(self, shapes, num_tasks, num_filters=(60,), conv_width=(35,), pool_width=75, dropout=0.2, batch_norm=False): assert len(num_filters) == len(conv_width) # configure inputs keras_inputs = self.get_keras_inputs(shapes) inputs = self.reshape_keras_inputs(keras_inputs) # convolve sequence seq_preds = inputs["data/genome_data_dir"] def reverse_comp(x): return K.reverse(K.reverse(x,axes=-1),axes=-2) def reverse_after_conv(x): #We need to use this after return K.reverse(x,axes=1) reverse_comp_seq = Lambda(reverse_comp,name = 'reverse_comp')(seq_preds) shared_convolution = Convolution1D(num_filters[0], conv_width[0], 'he_normal', input_shape=seq_preds.shape,name = 'conv_layer_1') seq_preds = shared_convolution(seq_preds) seq_preds = Activation('relu')(seq_preds) reverse_comp_seq = shared_convolution(reverse_comp_seq) reverse_comp_seq = Activation('relu')(reverse_comp_seq) #print("Reverse comp convolved shape %s"%str(reverse_comp_seq.shape)) reverse_comp_seq = Lambda(reverse_after_conv,name='reverse')(reverse_comp_seq) logits = Merge(mode='concat', concat_axis=2)([seq_preds, reverse_comp_seq]) # print("Concatenated Logits shape %s"%str(logits.shape)) #Note that the first convolution has already been done for i, (nb_filter, nb_col) in enumerate(zip(num_filters[1:], conv_width[1:])): logits = Convolution1D( nb_filter, nb_col, 'he_normal',name='conv_layer_%s'%(str(i+2)))(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) logits = Dropout(dropout)(logits) # pool and fully connect # print(logits.shape) logits = MaxPooling1D((pool_width))(logits) #print(logits.shape) logits = Flatten()(logits) #print(logits.shape) logits = Dense(output_dim=num_tasks)(logits) logits= Activation('sigmoid')(logits) self.model = Model(input=keras_inputs.values(), output=logits) class SequenceMethylationReverseComplementClassifier(Classifier): def __init__(self, shapes, num_tasks, num_filters=(60,),conv_width=(35,), pool_width=75, dropout=0.2, batch_norm=False): assert len(num_filters)==len(conv_width) #assert len(num_combined_filters)==len(combined_conv_width) #configure inputs keras_inputs = self.get_keras_inputs(shapes) inputs = self.reshape_keras_inputs(keras_inputs) #The shapes are (None,1000,4) and (None,1000,1) now seq_preds=inputs["data/genome_data_dir"] methylation_preds=inputs["data/methylation_data_dir"] def reverse_comp(x): return K.reverse(K.reverse(x,axes=-1),axes=-2) def reverse(x): return K.reverse(x,axes=-2) def get_C_locations(x): return K.expand_dims(x[:,:,1],dim=-1) def get_G_locations(x): return K.expand_dims(x[:,:,2],dim=-1) reverse_comp_seq = Lambda(reverse_comp,name="rev_comp_sequence")(seq_preds) C_locations_forward_strand = Lambda(get_C_locations,name="C_loc")(seq_preds) G_locations_forward_strand = Lambda(get_G_locations,name='G_loc')(seq_preds) forward_methylation_on_C_masked = merge([methylation_preds,C_locations_forward_strand],mode='mul', name = 'C_meth_plus_strand') forward_methylation_on_G_masked = merge([methylation_preds,G_locations_forward_strand],mode='mul', name = 'G_meth_minus_strand') forward_logit = Merge(mode='concat',concat_axis=-1,name = 'Seq_meth_forward_concat')([forward_methylation_on_C_masked, seq_preds, forward_methylation_on_G_masked]) #This ordering creates a 6 letter reverse comp alphabet reverse_comp_logit = Lambda(reverse_comp,name = 'Seq_meth_rev_concat')(forward_logit) shared_convolution = Convolution1D(num_filters[0], conv_width[0], 'he_normal', input_shape=seq_preds.shape,name='conv_layer_1') forward_logit = shared_convolution(forward_logit) forward_logit = Activation('relu')(forward_logit) reverse_comp_logit = shared_convolution(reverse_comp_logit) reverse_comp_logit = Activation('relu')(reverse_comp_logit) reverse_comp_logit = Lambda(reverse,name='flip')(reverse_comp_logit) #To line it up with original input logits = Merge(mode='concat', concat_axis=-1,name='merge_forward_rev')([forward_logit,reverse_comp_logit]) #Note that the first convolution has already been done for i, (nb_filter, nb_col) in enumerate(zip(num_filters[1:], conv_width[1:])): logits = Convolution1D( nb_filter, nb_col, 'he_normal',name='conv_layer_%s'%(str(i+2)))(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) logits = Dropout(dropout,name='dropout')(logits) logits = MaxPooling1D((pool_width),name='MaxPool')(logits) logits = Flatten(name='flatten')(logits) logits = Dense(output_dim=num_tasks,name='dense')(logits) logits = Activation('sigmoid',name='output')(logits) self.model = Model(input=keras_inputs.values(), output=logits) class SequenceA_MethylationReverseComplementClassifier(Classifier): def __init__(self, shapes, num_tasks, num_filters=(60,),conv_width=(35,), pool_width=75, dropout=0.2, batch_norm=False): assert len(num_filters)==len(conv_width) #assert len(num_combined_filters)==len(combined_conv_width) #configure inputs keras_inputs = self.get_keras_inputs(shapes) inputs = self.reshape_keras_inputs(keras_inputs) #The shapes are (None,1000,4) and (None,1000,1) now seq_preds=inputs["data/genome_data_dir"] methylation_preds=inputs["data/A_methylation_data_dir"] def reverse_comp(x): return K.reverse(K.reverse(x,axes=-1),axes=-2) def reverse(x): return K.reverse(x,axes=-2) def get_A_locations(x): return K.expand_dims(x[:,:,0],dim=-1) def get_T_locations(x): return K.expand_dims(x[:,:,-1],dim=-1) reverse_comp_seq = Lambda(reverse_comp,name="rev_comp_sequence")(seq_preds) A_locations_forward_strand = Lambda(get_A_locations,name="A_loc")(seq_preds) T_locations_forward_strand = Lambda(get_T_locations,name='T_loc')(seq_preds) forward_methylation_on_A_masked = merge([methylation_preds,A_locations_forward_strand],mode='mul', name = 'A_meth_plus_strand') forward_methylation_on_T_masked = merge([methylation_preds,T_locations_forward_strand],mode='mul', name = 'T_meth_minus_strand') forward_logit = Merge(mode='concat',concat_axis=-1,name = 'Seq_meth_forward_concat')([forward_methylation_on_A_masked, seq_preds, forward_methylation_on_T_masked]) #This ordering creates a 6 letter reverse comp alphabet reverse_comp_logit = Lambda(reverse_comp,name = 'Seq_meth_rev_concat')(forward_logit) shared_convolution = Convolution1D(num_filters[0], conv_width[0], 'he_normal', input_shape=seq_preds.shape,name='conv_layer_1') forward_logit = shared_convolution(forward_logit) forward_logit = Activation('relu')(forward_logit) reverse_comp_logit = shared_convolution(reverse_comp_logit) reverse_comp_logit = Activation('relu')(reverse_comp_logit) reverse_comp_logit = Lambda(reverse,name='flip')(reverse_comp_logit) #To line it up with original input logits = Merge(mode='concat', concat_axis=-1,name='merge_forward_rev')([forward_logit,reverse_comp_logit]) #Note that the first convolution has already been done for i, (nb_filter, nb_col) in enumerate(zip(num_filters[1:], conv_width[1:])): logits = Convolution1D( nb_filter, nb_col, 'he_normal',name='conv_layer_%s'%(str(i+2)))(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) logits = Dropout(dropout,name='dropout')(logits) logits = MaxPooling1D((pool_width),name='MaxPool')(logits) logits = Flatten(name='flatten')(logits) logits = Dense(output_dim=num_tasks,name='dense')(logits) logits = Activation('sigmoid',name='output')(logits) self.model = Model(input=keras_inputs.values(), output=logits) class SequenceACMethylationReverseComplementClassifier(Classifier): def __init__(self, shapes, num_tasks, num_filters=(60,),conv_width=(35,), pool_width=75, dropout=0.2, batch_norm=False): assert len(num_filters)==len(conv_width) #assert len(num_combined_filters)==len(combined_conv_width) #configure inputs keras_inputs = self.get_keras_inputs(shapes) inputs = self.reshape_keras_inputs(keras_inputs) #The shapes are (None,1000,4) and (None,1000,1) now seq_preds=inputs["data/genome_data_dir"] C_methylation_preds=inputs["data/methylation_data_dir"] #C methylation A_methylation_preds=inputs["data/A_methylation_data_dir"] #A methylation def reverse_comp(x): return K.reverse(K.reverse(x,axes=-1),axes=-2) def reverse(x): return K.reverse(x,axes=-2) def get_C_locations(x): return K.expand_dims(x[:,:,1],dim=-1) def get_G_locations(x): return K.expand_dims(x[:,:,2],dim=-1) def get_A_locations(x): return K.expand_dims(x[:,:,0],dim=-1) def get_T_locations(x): return K.expand_dims(x[:,:,3],dim=-1) reverse_comp_seq = Lambda(reverse_comp,name="rev_comp_sequence")(seq_preds) C_locations_forward_strand = Lambda(get_C_locations,name="C_loc")(seq_preds) G_locations_forward_strand = Lambda(get_G_locations,name='G_loc')(seq_preds) A_locations_forward_strand = Lambda(get_A_locations,name="A_loc")(seq_preds) T_locations_forward_strand = Lambda(get_T_locations,name='T_loc')(seq_preds) forward_methylation_on_C_masked = merge([C_methylation_preds,C_locations_forward_strand],mode='mul', name = 'C_meth_plus_strand') forward_methylation_on_G_masked = merge([C_methylation_preds,G_locations_forward_strand],mode='mul', name = 'G_meth_minus_strand') forward_methylation_on_A_masked = merge([A_methylation_preds,A_locations_forward_strand],mode='mul', name = 'A_meth_plus_strand') forward_methylation_on_T_masked = merge([A_methylation_preds,T_locations_forward_strand],mode='mul', name = 'T_meth_minus_strand') forward_logit = Merge(mode='concat',concat_axis=-1,name = 'Seq_meth_forward_concat')([forward_methylation_on_A_masked, forward_methylation_on_C_masked, seq_preds, forward_methylation_on_G_masked, forward_methylation_on_T_masked]) #This ordering creates a 6 letter reverse comp alphabet reverse_comp_logit = Lambda(reverse_comp,name = 'Seq_meth_rev_concat')(forward_logit) shared_convolution = Convolution1D(num_filters[0], conv_width[0], 'he_normal', input_shape=seq_preds.shape,name='conv_layer_1') forward_logit = shared_convolution(forward_logit) forward_logit = Activation('relu')(forward_logit) reverse_comp_logit = shared_convolution(reverse_comp_logit) reverse_comp_logit = Activation('relu')(reverse_comp_logit) reverse_comp_logit = Lambda(reverse,name='flip')(reverse_comp_logit) #To line it up with original input logits = Merge(mode='concat', concat_axis=-1,name='merge_forward_rev')([forward_logit,reverse_comp_logit]) #Note that the first convolution has already been done for i, (nb_filter, nb_col) in enumerate(zip(num_filters[1:], conv_width[1:])): logits = Convolution1D( nb_filter, nb_col, 'he_normal',name='conv_layer_%s'%(str(i+2)))(logits) if batch_norm: logits = BatchNormalization()(logits) logits = Activation('relu')(logits) logits = Dropout(dropout,name='dropout')(logits) logits = MaxPooling1D((pool_width),name='MaxPool')(logits) logits = Flatten(name='flatten')(logits) logits = Dense(output_dim=num_tasks,name='dense')(logits) logits = Activation('sigmoid',name='output')(logits) self.model = Model(input=keras_inputs.values(), output=logits) class SequenceClassifierRevCompConv(Classifier): def __init__(self, shapes, num_tasks, num_filters=(15, 15), conv_width=(15, 15), pool_size=20, stride=20, padding="same", dropout=0, batch_norm=True): assert len(num_filters) == len(conv_width) # configure inputs keras_inputs = self.get_keras_inputs(shapes) inputs = self.reshape_keras_inputs(keras_inputs) # convolve sequence seq_preds = inputs["data/genome_data_dir"] for i, (nb_filter, nb_col) in enumerate(zip(num_filters, conv_width)): seq_preds = RevCompConv1D( filters=nb_filter, kernel_size=11, input_shape=(input_length,4), activation="relu", padding="same")(seq_preds) if batch_norm: seq_preds = RevCompConv1DBatchNorm()(seq_preds) seq_preds = Activation('relu')(seq_preds) # pool and fully connect seq_preds = MaxPooling1D(pool_size, stride, padding)(seq_preds) seq_preds = DenseAfterRevcompConv1D(units=10, activation="relu")(seq_preds) seq_preds = Dense(units=1)(seq_preds) self.model = Model(input=keras_inputs.values(), output=seq_preds)