def main():
args = parser.parse_args()
graph = get_graph_from_hyperparameters(args.pred)
batch_size = 32
##Load graph
epoch_range = map(int, args.epoch.split(','))
#graph.set_weights(initial_weights)
for epoch in range(epoch_range[0], epoch_range[1]) :
graph.load_weights(HOME_PATH + '/weights' + str(max_ic50) + '/' + args.pred + '/weights' + str(batch_size) + '_' + str(epoch) )
allele_sequence_data, max_allele_length = load_allele_sequence_data(HOME_PATH + '/pan_allele/files/pseudo/pseudo_sequences.fasta')
predictions = read_tcell_predictions(HOME_PATH + '/pan_allele/files/iedb-tcell-2009-negative.csv',HOME_PATH + '/pan_allele/files/iedb-tcell-2009-positive.csv')
allele_list = sorted(predictions.keys())
allele_list[:] = [x for x in allele_list if not x.startswith('C')]
Y_true = []
Y_pred = []
for allele in allele_list:
peptides = predictions[allele].keys()
for peptide in peptides:
if(len(peptide)>7 and len(peptide)<12):
#print allele, peptide, predictions[allele][peptide], 20000**(1-make_prediction(peptide, allele_sequence_data[allele], graph))
Y_true.append( predictions[allele][peptide])
Y_pred.append(make_prediction(peptide, allele_sequence_data[allele], graph))
#print "=====", allele, sum(Y_true), len(Y_true), "===="
score = scores(Y_true, Y_pred)
print epoch, ','.join(map(str,score[1:]))
def main():
args = parser.parse_args()
graph = get_graph_from_hyperparameters(args.pred)
batch_sizes = [32]
learning_rates = [0.001]
for lr in learning_rates:
for batch_size in batch_sizes:
save_model(graph, args.pred, args.batch_size, args.epochs, max_ic50=args.max_ic50)
def main():
args = parser.parse_args()
max_ic50 = args.max_ic50
#IEDB data
allele_groups, df = load_binding_data(BINDING_DATA_PATH, max_ic50=max_ic50, peptide_length=9)
#graph initialized here so that pseudo sequences are made accordingly
graph = get_graph_from_hyperparameters(args.pred)
#allele sequence data
allele_sequence_data, max_allele_length = load_allele_sequence_data(SEQUENCE_DATA_PATH)
allele_list = sorted(create_allele_list(allele_groups, allele_sequence_data))
#reading blind data from txt file that contains aggregated data for all alleles
blind_allele_groups, blind_df = load_binding_data('blind_data.txt', max_ic50=max_ic50, peptide_length=None)
blind_allele_list = sorted(create_allele_list(blind_allele_groups, allele_sequence_data))
nb_iter = 50 #number of networks to include in the ensemble
preds_allele = defaultdict(list)
actual_allele = defaultdict(list)
for i in range(0,nb_iter):
#get_model_data shuffles the data so theres no need for further shuffling
peptides, mhc, Y = get_model_data( allele_list,
allele_sequence_data,
allele_groups,
peptide_length = 9,
mhc_length=max_allele_length
)
#splitting peptides, mhcs and binding into training and test
peptides_train, peptides_test = split_train_test(peptides,5)
mhc_train, mhc_test = split_train_test(mhc,5)
Y_train, Y_test = split_train_test(Y,5)
#fit graph model
graph = get_graph_from_hyperparameters(args.pred)
graph.fit({'peptide':peptides_train, 'mhc':mhc_train, 'output': Y_train},
batch_size=32,
nb_epoch=12,
verbose = 0,
)
#calculate metrics for each allele
for allele in blind_allele_list:
print i, allele
predictions = read_blind_predictions(HOME_PATH + '/combined-test-data/'+ allele + '.csv')
peptides = predictions.keys()
preds = []
meas = []
for peptide in peptides:
preds.append(make_prediction(peptide, allele_sequence_data[allele],graph))
meas.append(predictions[peptide]['meas'])
preds = np.array(preds)
meas = np.array(meas)
try:
preds_allele[allele]+=preds/nb_iter
except:
preds_allele[allele]=preds/nb_iter
actual_allele[allele] = meas
#calculate average for all the alleles
calculated_metrics = np.zeros(6)
for allele in blind_allele_list:
Y_pred_allele = max_ic50**(1-preds_allele[allele])
Y_true_allele = actual_allele[allele]
score_allele = scores(Y_true_allele, Y_pred_allele)
calculated_metrics += score_allele ##sum metrics for all alleles
print calculated_metrics/len(blind_allele_list) #divide sum by number of alleles
def main():
#prediction input either "conv", "ffn_concat", "ffn_mult"
args = parser.parse_args()
graph = get_graph_from_hyperparameters(args.pred)
allele_sequence_data, max_allele_length = load_allele_sequence_data(SEQUENCE_DATA_PATH)
predictors = ['mhcflurry','netmhcpan','netmhc','smmpmbec_cpp']
#allele_list
allele_list = ['A0101', 'A0201', 'A0202', 'A0203', 'A0206', 'A0301',
'A1101', 'A2301', 'A2402', 'A2501', 'A2601', 'A2602',
'A2603', 'A2902', 'A3001', 'A3002', 'A3101', 'A3201',
'A3301', 'A6801', 'A6802', 'A6901', 'A8001', 'B0702',
'B0801', 'B0802', 'B0803', 'B1501', 'B1503', 'B1509',
'B1517', 'B1801', 'B2703', 'B2705', 'B3501', 'B3801',
'B3901', 'B4001', 'B4002', 'B4402', 'B4403', 'B4501',
'B4601', 'B5101', 'B5301', 'B5401', 'B5701', 'B5801' ]
#Load graph
batch_size = 32
epoch_range = map(int, args.epoch.split(','))
for epoch in range(epoch_range[0],epoch_range[1]):
graph.load_weights(HOME_PATH + '/weights' + str(args.max_ic50) + '/' + args.pred + '/weights' + str(batch_size) + '_' + str(epoch) )
#Initializing
data_len = sum(len(read_blind_predictions(HOME_PATH + '/combined-test-data/'+ allele + '.csv').keys()) for allele in allele_list)
Y_true_all = np.zeros(data_len)
total_metrics = collections.defaultdict(list)
for val in predictors:
total_metrics[val] = np.zeros(data_len)
pos = 0
calculated_metrics =collections.defaultdict(tuple)
for val in predictors:
calculated_metrics[val] = np.zeros(6)
#calculating metrics per allele
for allele in allele_list:
filename = HOME_PATH + '/combined-test-data/'+ allele + '.csv'
predictions = read_blind_predictions(filename)
peptides = predictions.keys()
for peptide in peptides:
predictions[peptide]['mhcflurry'] = max_ic50**(1-make_prediction(peptide, allele_sequence_data[allele], graph))
df_pred = pd.DataFrame(predictions)
Y_true_allele = np.array(df_pred.loc['meas'])
Y_true_all[pos:pos+len(peptides)] = Y_true_allele
if (args.allele_info == True):
print "\n=====", allele, sum(Y_true_allele <= ic50_cutoff), len(Y_true_allele), "===="
for val in predictors:
Y_pred_allele = np.array(df_pred.loc[val])
calculated_metrics[val] += scores(Y_true_allele, Y_pred_allele)
if (args.allele_info == True):
print val, scores(Y_true_allele, Y_pred_allele)
pos +=len(peptides)
print '\n',epoch,
for val in predictors:
calculated_metrics[val] = calculated_metrics[val]/len(allele_list)
print val,',',','.join(map(str,calculated_metrics[val][1:]))
peptides.append(np.random.randint(20, size=9))
allele_sequence_data, max_allele_length = load_allele_sequence_data(SEQUENCE_DATA_PATH)
allele_list = ['A0101', 'A0201', 'A0202', 'A0203', 'A0206', 'A0301',
'A1101', 'A2301', 'A2402', 'A2501', 'A2601', 'A2602',
'A2603', 'A2902', 'A3001', 'A3002', 'A3101', 'A3201',
'A3301', 'A6801', 'A6802', 'A6901', 'A8001', 'B0702',
'B0801', 'B0802', 'B0803', 'B1501', 'B1503', 'B1509',
'B1517', 'B1801', 'B2703', 'B2705', 'B3501', 'B3801',
'B3901', 'B4001', 'B4002', 'B4402', 'B4403', 'B4501',
'B4601', 'B5101', 'B5301', 'B5401', 'B5701', 'B5801' ]
graph = get_graph_from_hyperparameters('ffn_mult')
for epoch in range(1,99):
graph.load_weights(HOME_PATH + 'weights/ffn_mult/weights32_' + str(epoch))
predictions = np.empty(len(allele_list)*nb_peptides)
counter = 0
for allele in allele_list:
allele_sequence = mhc_seq = padded_indices([allele_sequence_data[allele]],
add_start_symbol=False,
add_end_symbol=False,
index_dict=amino_acid_letter_indices)
for peptide in peptides:
predictions[counter] = 20000**(1-graph.predict({'peptide':[[peptide]],'mhc':mhc_seq})['output'])
counter = counter + 1
print epoch, np.sum(predictions<500)
