import pickle import numpy as np from keras.models import Sequential from keras.models import model_from_json from keras.layers import Dense, Dropout, Flatten from keras.layers import Conv3D def loop_one_hot_encode(loop_seq, loop_struct): ''' Encodes a ribozyme loop into a 3-dimensional array meant to represent its 3-domensional physical form. The dimensions are as follows: width - 2, to represent the nucleotides as the directly emerge from a paired stem height - 15, to encapsulate long loops. Most positions will be blank for shorter loops channels - 8, to represent the 4 nucleotides, each in a bonded or unbonded state :param loop_seq: String of nucleotides representing the sequence of the loop :param loop_struct: String of dot-bracket notation representing the structure of the loop :return: numpy array of size (2, 1, 15, 8) ''' encoded_loop = np.zeros((2, 1, 15, 8)) loop_codes = ['A', 'A', 'U', 'U', 'C', 'C', 'G', 'G'] if len(loop_seq) <= 30: # Iterates forward from the 5' end of the loop and backwards from the 3' end simultaneously, marking at each # position the correct channel. for i in range(len(loop_seq) // 2): encoded_loop[0, 0, i, loop_codes.index(loop_seq[i]) + (loop_struct[i] == '(' or loop_struct[i] == ')')] = 1 encoded_loop[1, 0, i, loop_codes.index(loop_seq[-(i + 1)]) + (loop_struct[-(i + 1)] == '(' or loop_struct[-(i + 1)] == ')')] = 1 # If an odd number of nucleotides, put the on at the apex of the loop into the 1st column. if len(loop_seq) % 2 == 1: i = len(loop_seq) // 2 encoded_loop[0, 0, i, loop_codes.index(loop_seq[i]) + (loop_struct[i] == '(' or loop_struct[i] == ')')] = 1 else: # If the loop is too large to fit, just does the first 15 on each side. for i in range(15): encoded_loop[0, 0, i, loop_codes.index(loop_seq[i]) + (loop_struct[i] == '(' or loop_struct[i] == ')')] = 1 encoded_loop[1, 0, i, loop_codes.index(loop_seq[-(i + 1)]) + (loop_struct[-(i + 1)] == '(' or loop_struct[-(i + 1)] == ')')] = 1 return encoded_loop def struct_dict_to_array(in_dict): ''' Turns a dictionary of ribozyme sequences and structures into a 5-dimensional numpy array, containing 4-dimensional arrays representing paired ribozyme loops for each sequence, along with an array containing the basal gene-regulatory activity for each sequence. :param in_dict: Dictionary of ribozyme sequences where the keys are tuples like so: (sequence string, dot-bracket structure string) and the value is the basal gene-regulatory activity for that sequence. :return: Tuple containing 3 arrays: 5-dimensional numpy array representing paired ribozyme loops for each sequence 1-dimensional numpy array containing the basal gene-regulatory activity for each sequence 1-dimensional array of tuples, containing the sequences of the 2 loops for each sequence ''' out_X = [] out_y = [] out_loops = [] first, mid, last = 'GCUGUC', 'CUGAUGA', 'GAAACAGC' # Iterates through each sequence in the dictionary for seq in in_dict.keys(): test_seq = seq[0] test_struct = seq[1] # Gets the sequence and array of the first loop loop_seq = test_seq[test_seq.find(first) + len(first): test_seq.find(mid)] loop_struct = test_struct[test_seq.find(first) + len(first): test_seq.find(mid)] l1 = loop_one_hot_encode(loop_seq, loop_struct) l1_seq = loop_seq # Gets the sequence and array of the second loop loop_seq = test_seq[test_seq.find(mid) + len(mid): test_seq.find(last)] loop_struct = test_struct[test_seq.find(mid) + len(mid): test_seq.find(last)] l2 = loop_one_hot_encode(loop_seq, loop_struct) # Pairs the 2 loops up and stores paired array and value out_X.append(np.concatenate((l1, l2), axis=1)) out_y.append(in_dict[seq][0]) out_loops.append((l1_seq, loop_seq)) out_X = np.array(out_X).astype('float32') out_y = np.expand_dims(out_y, axis=1) return out_X, out_y, out_loops # Reads in training data rs_list = [] rs_list.extend(pickle.load(open('NGS_data/CMS_NGS1_5_RNAs_min_structures_all_activities.pkl', 'rb'), encoding = 'latin1')) rs_list.extend(pickle.load(open('NGS_data/CMS_NGS2_1_RNAs_min_structures_all_activities_normed.pkl', 'rb'), encoding = 'latin1')) rs_list.extend(pickle.load(open('NGS_data/CMS_NGS3_1_RNAs_min_structures_all_activities.pkl', 'rb'), encoding = 'latin1')) rest_dict = {} # Pulls training data into dictionary for later conversion to array for seq in rs_list: target_dict = rest_dict target_dict[seq[0][0][0]] = seq train_seq_dict = {} train_dict = rest_dict for i in train_dict.keys(): train_seq_dict[(tuple(train_dict[i][0][0][1][0]), tuple(train_dict[i][0][0][1][1]), train_dict[i][0][0][0], train_dict[i][0][0][2])] = \ [train_dict[i][0][1], train_dict[i][1]] # Segments training data by structure, creating a dictionaries where each sequence in that dictionary has the same # structure train_segmented_dict = {} for loop in train_seq_dict: if (len(loop[0][0]) / 2, len(loop[0][1]) / 2, loop[1][0], loop[1][1]) not in train_segmented_dict: train_segmented_dict[(len(loop[0][0]) / 2, len(loop[0][1]) / 2, loop[1][0], loop[1][1])] = {} train_segmented_dict[(len(loop[0][0]) / 2, len(loop[0][1]) / 2, loop[1][0], loop[1][1])][(loop[2], loop[3])] = \ train_seq_dict[loop] # List of structures in the segmented dictionary that have over 50 entries train_populous_loops = [i for i in train_segmented_dict.keys() if len(train_segmented_dict[i]) > 50] # List detailing how to expand the structure search parameter if not enough sequences in training set to train model diff_list = [(1, 1, 1, 1), (2, 1, 1, 1), (2, 2, 1, 1), (3, 2, 1, 1), (3, 3, 1, 1), (3, 3, 2, 1), (3, 3, 2, 2), (4, 4, 2, 2)] # Get the loop size of model to save te_seg = [0, 0, 0, 0] te_seg[0] = int(input("Loop 1 size: ")) te_seg[1] = int(input("Loop 2 size: ")) te_seg[2] = int(input("Stem 1 length: ")) te_seg[3] = int(input("Stem 2 length: ")) training_dict = {} # Iterates through the diff_list, relaxing the requirements for structural similarity until 1000 sequences are in the # training set for i in diff_list: # Updates the dictionary with sequences that have the prescribed difference in structure for tr_seg in train_populous_loops: if abs(te_seg[0] - tr_seg[0]) < i[0] and abs(te_seg[1] - tr_seg[1]) < i[1] and \ abs(te_seg[2] - tr_seg[2]) < i[2] and abs(te_seg[3] - tr_seg[3]) < i[3]: training_dict.update(train_segmented_dict[tr_seg]) # If at least 1000 sequences are in the training set, moves forward trX, trY, trloops = struct_dict_to_array(training_dict) if len(trY) > 1000: break # Defines the model model = Sequential() layer = Conv3D(32, (2, 2, 2), activation='relu', input_shape=(2, 2, 15, 8)) model.add(layer) model.add(Dropout(0.25)) model.add(Flatten()) model.add(Dense(64, activation='relu')) model.add(Dropout(0.5)) model.add(Dense(output_dim=1)) # Optimize with SGD model.compile(loss='mean_squared_error', optimizer='adam') # Fit model in batches model.fit(trX, trY[:, 0], nb_epoch=100, batch_size=1000, verbose=1) # Dumps model to json for later predictions model_json = model.to_json() with open('Models/' + str(te_seg) + ".json", "w") as json_file: json_file.write(model_json) model.save_weights('Models/' + str(te_seg) + "model.h5") print("Saved model to disk")