def main(in_file, out_file, go_file, model_file, terms_file, annotations_file,
chunk_size, diamond_file, threshold, batch_size, alpha):
# Load GO and read list of all terms
go = Ontology(go_file, with_rels=True)
terms_df = pd.read_pickle(terms_file)
terms = terms_df['terms'].values.flatten()
# Read known experimental annotations
annotations = {}
df = pd.read_pickle(annotations_file)
for row in df.itertuples():
annotations[row.proteins] = set(row.prop_annotations)
go.calculate_ic(annotations.values())
diamond_preds = {}
mapping = {}
with gzip.open(diamond_file, 'rt') as f:
for line in f:
it = line.strip().split()
if it[0] not in mapping:
mapping[it[0]] = {}
mapping[it[0]][it[1]] = float(it[2])
for prot_id, sim_prots in mapping.items():
annots = {}
allgos = set()
total_score = 0.0
for p_id, score in sim_prots.items():
allgos |= annotations[p_id]
total_score += score
allgos = list(sorted(allgos))
sim = np.zeros(len(allgos), dtype=np.float32)
for j, go_id in enumerate(allgos):
s = 0.0
for p_id, score in sim_prots.items():
if go_id in annotations[p_id]:
s += score
sim[j] = s / total_score
for go_id, score in zip(allgos, sim):
annots[go_id] = score
diamond_preds[prot_id] = annots
# Load CNN model
model = load_model(model_file)
# Alphas for the latest model
alphas = {NAMESPACES['mf']: 0.55, NAMESPACES['bp']: 0.59, NAMESPACES['cc']: 0.46}
# Alphas for the cafa2 model
# alphas = {NAMESPACES['mf']: 0.63, NAMESPACES['bp']: 0.68, NAMESPACES['cc']: 0.48}
start_time = time.time()
total_seq = 0
w = gzip.open(out_file, 'wt')
for prot_ids, sequences in read_fasta(in_file, chunk_size):
total_seq += len(prot_ids)
deep_preds = {}
ids, data = get_data(sequences)
preds = model.predict(data, batch_size=batch_size)
assert preds.shape[1] == len(terms)
for i, j in enumerate(ids):
prot_id = prot_ids[j]
if prot_id not in deep_preds:
deep_preds[prot_id] = {}
for l in range(len(terms)):
if preds[i, l] >= 0.01: # Filter out very low scores
if terms[l] not in deep_preds[prot_id]:
deep_preds[prot_id][terms[l]] = preds[i, l]
else:
deep_preds[prot_id][terms[l]] = max(
deep_preds[prot_id][terms[l]], preds[i, l])
# Combine diamond preds and deepgo
for prot_id in prot_ids:
annots = {}
if prot_id in diamond_preds:
for go_id, score in diamond_preds[prot_id].items():
annots[go_id] = score * alphas[go.get_namespace(go_id)]
for go_id, score in deep_preds[prot_id].items():
if go_id in annots:
annots[go_id] += (1 - alphas[go.get_namespace(go_id)]) * score
else:
annots[go_id] = (1 - alphas[go.get_namespace(go_id)]) * score
# Propagate scores with ontology structure
gos = list(annots.keys())
for go_id in gos:
for g_id in go.get_anchestors(go_id):
if g_id in annots:
annots[g_id] = max(annots[g_id], annots[go_id])
else:
annots[g_id] = annots[go_id]
sannots = sorted(annots.items(), key=lambda x: x[1], reverse=True)
for go_id, score in sannots:
if score >= threshold:
w.write(prot_id + '\t' + go_id + '\t' + go.get_term(go_id)['name'] + '\t%.2f' % go.get_ic(go_id) + '\t%.3f\n' % score)
w.write('\n')
w.close()
total_time = time.time() - start_time
print('Total prediction time for %d sequences is %d' % (total_seq, total_time))
def main(train_data_file, test_data_file, terms_file, diamond_scores_file, ont,
alpha):
alpha /= 100.0
go_rels = Ontology('data-cafa/go.obo', with_rels=True)
terms_df = pd.read_pickle(terms_file)
terms = terms_df['terms'].values.flatten()
terms_dict = {v: i for i, v in enumerate(terms)}
train_df = pd.read_pickle(train_data_file)
test_df = pd.read_pickle(test_data_file)
annotations = train_df['annotations'].values
annotations = list(map(lambda x: set(x), annotations))
test_annotations = test_df['annotations'].values
test_annotations = list(map(lambda x: set(x), test_annotations))
go_rels.calculate_ic(annotations + test_annotations)
# Print IC values of terms
ics = {}
for term in terms:
ics[term] = go_rels.get_ic(term)
prot_index = {}
for i, row in enumerate(train_df.itertuples()):
prot_index[row.proteins] = i
# BLAST Similarity (Diamond)
diamond_scores = {}
with open(diamond_scores_file) as f:
for line in f:
it = line.strip().split()
if it[0] not in diamond_scores:
diamond_scores[it[0]] = {}
diamond_scores[it[0]][it[1]] = float(it[2])
blast_preds = []
for i, row in enumerate(test_df.itertuples()):
annots = {}
prot_id = row.proteins
# BlastKNN
if prot_id in diamond_scores:
sim_prots = diamond_scores[prot_id]
allgos = set()
total_score = 0.0
for p_id, score in sim_prots.items():
allgos |= annotations[prot_index[p_id]]
total_score += score
allgos = list(sorted(allgos))
sim = np.zeros(len(allgos), dtype=np.float32)
for j, go_id in enumerate(allgos):
s = 0.0
for p_id, score in sim_prots.items():
if go_id in annotations[prot_index[p_id]]:
s += score
sim[j] = s / total_score
ind = np.argsort(-sim)
for go_id, score in zip(allgos, sim):
annots[go_id] = score
blast_preds.append(annots)
# DeepGOPlus
go_set = go_rels.get_namespace_terms(NAMESPACES[ont])
go_set.remove(FUNC_DICT[ont])
labels = test_df['annotations'].values
labels = list(map(lambda x: set(filter(lambda y: y in go_set, x)), labels))
# print(len(go_set))
deep_preds = []
alphas = {
NAMESPACES['mf']: 0.55,
NAMESPACES['bp']: 0.59,
NAMESPACES['cc']: 0.46
}
for i, row in enumerate(test_df.itertuples()):
annots_dict = blast_preds[i].copy()
for go_id in annots_dict:
annots_dict[go_id] *= alphas[go_rels.get_namespace(go_id)]
for j, score in enumerate(row.preds):
go_id = terms[j]
score *= 1 - alphas[go_rels.get_namespace(go_id)]
if go_id in annots_dict:
annots_dict[go_id] += score
else:
annots_dict[go_id] = score
deep_preds.append(annots_dict)
print('AUTHOR DeepGOPlus')
print('MODEL 1')
print('KEYWORDS sequence alignment.')
for i, row in enumerate(test_df.itertuples()):
prot_id = row.proteins
for go_id, score in deep_preds[i].items():
print(f'{prot_id}\t{go_id}\t{score:.2f}')
print('END')
return
# Propagate scores
# deepgo_preds = []
# for annots_dict in deep_preds:
# annots = {}
# for go_id, score in annots_dict.items():
# for a_id in go_rels.get_anchestors(go_id):
# if a_id in annots:
# annots[a_id] = max(annots[a_id], score)
# else:
# annots[a_id] = score
# deepgo_preds.append(annots)
fmax = 0.0
tmax = 0.0
precisions = []
recalls = []
smin = 1000000.0
rus = []
mis = []
for t in range(0, 101):
threshold = t / 100.0
preds = []
for i, row in enumerate(test_df.itertuples()):
annots = set()
for go_id, score in deep_preds[i].items():
if score >= threshold:
annots.add(go_id)
new_annots = set()
for go_id in annots:
new_annots |= go_rels.get_anchestors(go_id)
preds.append(new_annots)
# Filter classes
preds = list(
map(lambda x: set(filter(lambda y: y in go_set, x)), preds))
fscore, prec, rec, s, ru, mi, fps, fns = evaluate_annotations(
go_rels, labels, preds)
avg_fp = sum(map(lambda x: len(x), fps)) / len(fps)
avg_ic = sum(
map(lambda x: sum(map(lambda go_id: go_rels.get_ic(go_id), x)),
fps)) / len(fps)
print(f'{avg_fp} {avg_ic}')
precisions.append(prec)
recalls.append(rec)
print(
f'Fscore: {fscore}, Precision: {prec}, Recall: {rec} S: {s}, RU: {ru}, MI: {mi} threshold: {threshold}'
)
if fmax < fscore:
fmax = fscore
tmax = threshold
if smin > s:
smin = s
print(f'Fmax: {fmax:0.3f}, Smin: {smin:0.3f}, threshold: {tmax}')
precisions = np.array(precisions)
recalls = np.array(recalls)
sorted_index = np.argsort(recalls)
recalls = recalls[sorted_index]
precisions = precisions[sorted_index]
aupr = np.trapz(precisions, recalls)
print(f'AUPR: {aupr:0.3f}')
plt.figure()
lw = 2
plt.plot(recalls,
precisions,
color='darkorange',
lw=lw,
label=f'AUPR curve (area = {aupr:0.2f})')
plt.xlim([0.0, 1.0])
plt.ylim([0.0, 1.05])
plt.xlabel('Recall')
plt.ylabel('Precision')
plt.title('Area Under the Precision-Recall curve')
plt.legend(loc="lower right")
plt.savefig(f'aupr_{ont}_{alpha:0.2f}.pdf')
df = pd.DataFrame({'precisions': precisions, 'recalls': recalls})
df.to_pickle(f'PR_{ont}_{alpha:0.2f}.pkl')
def main(train_data_file, test_data_file, terms_file, diamond_scores_file,
ont):
go_rels = Ontology('data/go.obo', with_rels=True)
terms_df = pd.read_pickle(terms_file)
terms = terms_df['terms'].values.flatten()
terms_dict = {v: i for i, v in enumerate(terms)}
train_df = pd.read_pickle(train_data_file)
test_df = pd.read_pickle(test_data_file)
print("Length of test set: " + str(len(test_df)))
annotations = train_df['prop_annotations'].values
annotations = list(map(lambda x: set(x), annotations))
test_annotations = test_df['prop_annotations'].values
test_annotations = list(map(lambda x: set(x), test_annotations))
go_rels.calculate_ic(annotations + test_annotations)
# Print IC values of terms
ics = {}
for term in terms:
ics[term] = go_rels.get_ic(term)
prot_index = {}
for i, row in enumerate(train_df.itertuples()):
prot_index[row.proteins] = i
# BLAST Similarity (Diamond)
diamond_scores = {}
with open(diamond_scores_file) as f:
for line in f:
it = line.strip().split()
if it[0] not in diamond_scores:
diamond_scores[it[0]] = {}
diamond_scores[it[0]][it[1]] = float(it[2])
blast_preds = []
#print('Diamond preds')
for i, row in enumerate(test_df.itertuples()):
annots = {}
prot_id = row.proteins
# BlastKNN
if prot_id in diamond_scores:
sim_prots = diamond_scores[prot_id]
allgos = set()
total_score = 0.0
for p_id, score in sim_prots.items():
allgos |= annotations[prot_index[p_id]]
total_score += score
allgos = list(sorted(allgos))
sim = np.zeros(len(allgos), dtype=np.float32)
for j, go_id in enumerate(allgos):
s = 0.0
for p_id, score in sim_prots.items():
if go_id in annotations[prot_index[p_id]]:
s += score
sim[j] = s / total_score
ind = np.argsort(-sim)
for go_id, score in zip(allgos, sim):
annots[go_id] = score
blast_preds.append(annots)
last_release_metadata = 'metadata/last_release.json'
with open(last_release_metadata, 'r') as f:
last_release_data = json.load(f)
last_release_data['alphas'][ont] = find_alpha(ont, test_df, blast_preds,
go_rels, terms)
with open(last_release_metadata, 'w') as f:
json.dump(last_release_data, f)
def main(train_data_file, valid_data_file, terms_file, diamond_scores_file,
ont):
go_rels = Ontology('data-cafa/go.obo', with_rels=True)
terms_df = pd.read_pickle(terms_file)
terms = terms_df['terms'].values.flatten()
terms_dict = {v: i for i, v in enumerate(terms)}
train_df = pd.read_pickle(train_data_file)
valid_df = pd.read_pickle(valid_data_file)
annotations = train_df['annotations'].values
annotations = list(map(lambda x: set(x), annotations))
valid_annotations = valid_df['annotations'].values
valid_annotations = list(map(lambda x: set(x), valid_annotations))
go_rels.calculate_ic(annotations + valid_annotations)
# Print IC values of terms
ics = {}
for term in terms:
ics[term] = go_rels.get_ic(term)
prot_index = {}
for i, row in enumerate(train_df.itertuples()):
prot_index[row.proteins] = i
# BLAST Similarity (Diamond)
diamond_scores = {}
with open(diamond_scores_file) as f:
for line in f:
it = line.strip().split()
if it[0] not in diamond_scores:
diamond_scores[it[0]] = {}
diamond_scores[it[0]][it[1]] = float(it[2])
blast_preds = []
for i, row in enumerate(valid_df.itertuples()):
annots = {}
prot_id = row.proteins
# BlastKNN
if prot_id in diamond_scores:
sim_prots = diamond_scores[prot_id]
allgos = set()
total_score = 0.0
for p_id, score in sim_prots.items():
allgos |= annotations[prot_index[p_id]]
total_score += score
allgos = list(sorted(allgos))
sim = np.zeros(len(allgos), dtype=np.float32)
for j, go_id in enumerate(allgos):
s = 0.0
for p_id, score in sim_prots.items():
if go_id in annotations[prot_index[p_id]]:
s += score
sim[j] = s / total_score
ind = np.argsort(-sim)
for go_id, score in zip(allgos, sim):
annots[go_id] = score
blast_preds.append(annots)
# DeepGOPlus
go_set = go_rels.get_namespace_terms(NAMESPACES[ont])
go_set.remove(FUNC_DICT[ont])
labels = valid_df['annotations'].values
labels = list(map(lambda x: set(filter(lambda y: y in go_set, x)), labels))
print(len(go_set))
best_fmax = 0.0
best_alpha = 0.0
for alpha in range(44, 70):
alpha /= 100.0
deep_preds = []
for i, row in enumerate(valid_df.itertuples()):
annots_dict = blast_preds[i].copy()
for go_id in annots_dict:
annots_dict[go_id] *= alpha
for j, score in enumerate(row.preds):
go_id = terms[j]
score *= 1 - alpha
if go_id in annots_dict:
annots_dict[go_id] += score
else:
annots_dict[go_id] = score
deep_preds.append(annots_dict)
fmax = 0.0
tmax = 0.0
precisions = []
recalls = []
smin = 1000000.0
rus = []
mis = []
for t in range(14, 20):
threshold = t / 100.0
preds = []
for i, row in enumerate(valid_df.itertuples()):
annots = set()
for go_id, score in deep_preds[i].items():
if score >= threshold:
annots.add(go_id)
new_annots = set()
for go_id in annots:
new_annots |= go_rels.get_anchestors(go_id)
preds.append(new_annots)
# Filter classes
preds = list(
map(lambda x: set(filter(lambda y: y in go_set, x)), preds))
fscore, prec, rec, s, ru, mi, fps, fns = evaluate_annotations(
go_rels, labels, preds)
avg_fp = sum(map(lambda x: len(x), fps)) / len(fps)
avg_ic = sum(
map(lambda x: sum(map(lambda go_id: go_rels.get_ic(go_id), x)),
fps)) / len(fps)
print(
f'Fscore: {fscore}, Precision: {prec}, Recall: {rec} S: {s}, RU: {ru}, MI: {mi} threshold: {threshold}'
)
if fmax < fscore:
fmax = fscore
tmax = threshold
if smin > s:
smin = s
if best_fmax < fmax:
best_fmax = fmax
best_alpha = alpha
print(
f'Alpha: {alpha} Fmax: {fmax:0.3f}, Smin: {smin:0.3f}, threshold: {tmax}'
)
print(f'{best_alpha} {best_fmax}')
def main(train_data_file, test_data_file, terms_file, diamond_scores_file, ont,
alpha):
alpha /= 100.0
go_rels = Ontology('data-cafa/go.obo', with_rels=True)
terms_df = pd.read_pickle(terms_file)
terms = terms_df['terms'].values.flatten()
# terms_dict = {v: i for i, v in enumerate(terms)}
train_df = pd.read_pickle(train_data_file)
test_df = pd.read_pickle(test_data_file)
annotations = train_df['annotations'].values
annotations = list(map(lambda x: set(x), annotations))
test_annotations = test_df['annotations'].values
test_annotations = list(map(lambda x: set(x), test_annotations))
#### ? notice that @annotations and @test_annotations are used to get IC scores, so we are not allowed to do pre-filtering
go_rels.calculate_ic(annotations + test_annotations)
go_set = go_rels.get_namespace_terms(
NAMESPACES[ont]) #? consider all the MF or CC or BP
#### ? filter terms to have only mf ?
# terms = [t for t in terms if t in go_set]
# print ('number of terms kept from terms_file {}'.format(len(terms)))
# Print IC values of terms
ics = {}
for term in terms:
ics[term] = go_rels.get_ic(term)
##!! let's save this
pickle.dump(ics, open("data-cafa/ICsValueTable.pickle", "wb"))
prot_index = {}
for i, row in enumerate(train_df.itertuples()):
prot_index[row.proteins] = i
####
# BLAST Similarity (Diamond) #! we can use same call, we have their output
diamond_scores = {}
with open(diamond_scores_file) as f:
for line in f:
it = line.strip().split()
if it[0] not in diamond_scores:
diamond_scores[it[0]] = {}
diamond_scores[it[0]][it[1]] = float(it[2])
blast_preds = []
for i, row in enumerate(test_df.itertuples()):
annots = {}
prot_id = row.proteins
# BlastKNN
if prot_id in diamond_scores:
sim_prots = diamond_scores[prot_id]
allgos = set()
total_score = 0.0
for p_id, score in sim_prots.items():
allgos |= annotations[prot_index[p_id]]
total_score += score
allgos = list(sorted(allgos))
sim = np.zeros(len(allgos), dtype=np.float32)
for j, go_id in enumerate(allgos):
s = 0.0
for p_id, score in sim_prots.items():
if go_id in annotations[prot_index[p_id]]:
s += score
sim[j] = s / total_score
ind = np.argsort(-sim)
for go_id, score in zip(allgos, sim):
annots[go_id] = score
blast_preds.append(annots)
####
# DeepGOPlus
# go_set = go_rels.get_namespace_terms(NAMESPACES[ont]) #? consider all the MF or CC or BP
go_set.remove(FUNC_DICT[ont])
labels = test_df['annotations'].values
labels = list(map(lambda x: set(filter(lambda y: y in go_set, x)),
labels)) ##! filter true labels by @go_set
print("total labels {}".format(len(go_set)))
deep_preds = []
# alphas = {NAMESPACES['mf']: 0.55, NAMESPACES['bp']: 0.59, NAMESPACES['cc']: 0.46}
for i, row in enumerate(
test_df.itertuples()): #! read in prediction of neural net
annots_dict = {}
# annots_dict = blast_preds[i].copy() #! copy blast score
# for go_id in annots_dict: # * set 0 for all @blast_prediction
# annots_dict[go_id] = 0 # *= alphas[go_rels.get_namespace(go_id)] #! scale down blast score.
for j, score in enumerate(row.preds): #! prediction of @test_df
go_id = terms[j]
# if go_id not in go_set: #? faster filter of labels because we don't add ancestor anyway
# continue
# score *= 1 - alphas[go_rels.get_namespace(go_id)] # x *= 1-0.5 --> x = x * (1-0.5)
# if go_id in annots_dict: #? should not need this line??
# annots_dict[go_id] += score #! add into blast score
# else: #! are we going to see error??
annots_dict[go_id] = score #! replace blast score
deep_preds.append(annots_dict) #! later on, we use only @deep_preds
# print('AUTHOR DeepGOPlus')
# print('MODEL 1')
# print('KEYWORDS sequence alignment.')
# for i, row in enumerate(test_df.itertuples()):
# prot_id = row.proteins
# for go_id, score in deep_preds[i].items():
# print(f'{prot_id}\t{go_id}\t{score:.2f}')
# print('END')
# return
# Propagate scores
# deepgo_preds = []
# for annots_dict in deep_preds:
# annots = {}
# for go_id, score in annots_dict.items():
# for a_id in go_rels.get_anchestors(go_id):
# if a_id in annots:
# annots[a_id] = max(annots[a_id], score)
# else:
# annots[a_id] = score
# deepgo_preds.append(annots)
fmax = 0.0
tmax = 0.0
precisions = []
recalls = []
smin = 1000000.0
rus = []
mis = []
print('\nontology {}\n'.format(ont))
####
for threshold in np.arange(0.005, .4, .01): # np.arange(0.005,1,.01)
# threshold = t / 100.0
print('\n')
preds = []
for i, row in enumerate(test_df.itertuples()):
annots = set()
for go_id, score in deep_preds[i].items():
if go_id not in go_set: #? faster filter of labels because we don't add ancestor anyway
continue
if score >= threshold:
annots.add(go_id)
preds.append(annots)
##!! append parent terms or something ??
# new_annots = set()
# for go_id in annots:
# new_annots |= go_rels.get_anchestors(go_id)
# preds.append(new_annots)
# Filter classes
preds = list(
map(lambda x: set(filter(lambda y: y in go_set, x)), preds))
# print ('see 1 prediction')
# print (preds[10])
# print ('see 1 label')
# print (labels[10])
fscore, prec, rec, s, ru, mi, fps, fns = evaluate_annotations(
go_rels, labels, preds)
avg_fp = sum(map(lambda x: len(x), fps)) / len(fps)
avg_ic = sum(
map(lambda x: sum(map(lambda go_id: go_rels.get_ic(go_id), x)),
fps)) / len(fps)
print(f'{avg_fp} {avg_ic}')
precisions.append(prec)
recalls.append(rec)
print(
f'Fscore: {fscore}, Precision: {prec}, Recall: {rec} S: {s}, RU: {ru}, MI: {mi} threshold: {threshold}'
)
if fmax < fscore:
fmax = fscore
tmax = threshold
if smin > s:
smin = s
print(f'\nFmax: {fmax:0.3f}, Smin: {smin:0.3f}, threshold: {tmax}')
precisions = np.array(precisions)
recalls = np.array(recalls)
sorted_index = np.argsort(recalls)
recalls = recalls[sorted_index]
precisions = precisions[sorted_index]
aupr = np.trapz(precisions, recalls)
print(f'AUPR: {aupr:0.3f}')
plt.figure()
lw = 2
plt.plot(recalls,
precisions,
color='darkorange',
lw=lw,
label=f'AUPR curve (area = {aupr:0.2f})')
plt.xlim([0.0, 1.0])
plt.ylim([0.0, 1.05])
plt.xlabel('Recall')
plt.ylabel('Precision')
plt.title('Area Under the Precision-Recall curve')
plt.legend(loc="lower right")
plt.savefig(f'aupr_{ont}_{alpha:0.2f}.pdf')
df = pd.DataFrame({'precisions': precisions, 'recalls': recalls})
df.to_pickle(f'PR_{ont}_{alpha:0.2f}.pkl')