Spaces:
Sleeping
Sleeping
Upload 4 files
Browse filesFiltering and new features added version.
- About.py +84 -0
- pages/1_🔥_WarmMolGen.py +193 -0
- pages/__pycache__/util.cpython-37.pyc +0 -0
- pages/util.py +87 -0
About.py
ADDED
@@ -0,0 +1,84 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright 2018-2022 Streamlit Inc.
|
2 |
+
#
|
3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
4 |
+
# you may not use this file except in compliance with the License.
|
5 |
+
# You may obtain a copy of the License at
|
6 |
+
#
|
7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
8 |
+
#
|
9 |
+
# Unless required by applicable law or agreed to in writing, software
|
10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12 |
+
# See the License for the specific language governing permissions and
|
13 |
+
# limitations under the License.
|
14 |
+
|
15 |
+
import streamlit as st
|
16 |
+
from streamlit.logger import get_logger
|
17 |
+
LOGGER = get_logger(__name__)
|
18 |
+
|
19 |
+
|
20 |
+
def run():
|
21 |
+
st.set_page_config(
|
22 |
+
page_title="About WarmMolGen",
|
23 |
+
page_icon="🚀",
|
24 |
+
layout='wide'
|
25 |
+
)
|
26 |
+
|
27 |
+
st.write("## [Exploiting Pretrained Biochemical Language Models for Targeted Drug Design](https://arxiv.org/abs/2209.00981)")
|
28 |
+
#st.sidebar.title("Model Demos")
|
29 |
+
st.sidebar.success("Select a model demo above.")
|
30 |
+
|
31 |
+
st.markdown(
|
32 |
+
"""
|
33 |
+
This application demonstrates the generation capabilities of the models trained as part of the study below, which has been published in *Bioinformatics* Published by Oxford University Press. The available models are:
|
34 |
+
* WarmMolGen
|
35 |
+
- WarmMolGenOne (i.e. EncDecBase)
|
36 |
+
- WarmMolGenTwo (i.e. EncDecLM)
|
37 |
+
* ChemBERTaLM
|
38 |
+
👈 Select a model demo from the sidebar to generate molecules right away 🚀
|
39 |
+
### Abstract
|
40 |
+
**Motivation:** The development of novel compounds targeting proteins of interest is one of the most important tasks in
|
41 |
+
the pharmaceutical industry. Deep generative models have been applied to targeted molecular design and have shown
|
42 |
+
promising results. Recently, target-specific molecule generation has been viewed as a translation between the protein
|
43 |
+
language and the chemical language. However, such a model is limited by the availability of interacting protein–ligand
|
44 |
+
pairs. On the other hand, large amounts of unlabelled protein sequences and chemical compounds are available and
|
45 |
+
have been used to train language models that learn useful representations. In this study, we propose exploiting pretrained
|
46 |
+
biochemical language models to initialize (i.e. warm start) targeted molecule generation models. We investigate
|
47 |
+
two warm start strategies: (i) a one-stage strategy where the initialized model is trained on targeted molecule generation
|
48 |
+
and (ii) a two-stage strategy containing a pre-finetuning on molecular generation followed by target-specific training. We
|
49 |
+
also compare two decoding strategies to generate compounds: beam search and sampling.
|
50 |
+
**Results:** The results show that the warm-started models perform better than a baseline model trained from scratch.
|
51 |
+
The two proposed warm-start strategies achieve similar results to each other with respect to widely used metrics
|
52 |
+
from benchmarks. However, docking evaluation of the generated compounds for a number of novel proteins suggests
|
53 |
+
that the one-stage strategy generalizes better than the two-stage strategy. Additionally, we observe that
|
54 |
+
beam search outperforms sampling in both docking evaluation and benchmark metrics for assessing compound
|
55 |
+
quality.
|
56 |
+
**Availability and implementation:** The source code is available at https://github.com/boun-tabi/biochemical-lms-for-drug-design and the materials (i.e., data, models, and outputs) are archived in Zenodo at https://doi.org/10.5281/zenodo.6832145.
|
57 |
+
### Citation
|
58 |
+
```bibtex
|
59 |
+
@article{10.1093/bioinformatics/btac482,
|
60 |
+
author = {Uludoğan, Gökçe and Ozkirimli, Elif and Ulgen, Kutlu O and Karalı, Nilgün and Özgür, Arzucan},
|
61 |
+
title = "{Exploiting pretrained biochemical language models for targeted drug design}",
|
62 |
+
journal = {Bioinformatics},
|
63 |
+
volume = {38},
|
64 |
+
number = {Supplement_2},
|
65 |
+
pages = {ii155-ii161},
|
66 |
+
year = {2022},
|
67 |
+
doi = {10.1093/bioinformatics/btac482},
|
68 |
+
url = {https://doi.org/10.1093/bioinformatics/btac482},
|
69 |
+
}
|
70 |
+
```
|
71 |
+
"""
|
72 |
+
)
|
73 |
+
# page_names_to_funcs = {
|
74 |
+
# "—": intro,
|
75 |
+
# "Plotting Demo": plotting_demo,
|
76 |
+
# "Mapping Demo": mapping_demo,
|
77 |
+
# "DataFrame Demo": data_frame_demo
|
78 |
+
# }
|
79 |
+
|
80 |
+
# demo_name = st.sidebar.selectbox("Choose a demo", page_names_to_funcs.keys())
|
81 |
+
# page_names_to_funcs[demo_name]()
|
82 |
+
|
83 |
+
if __name__ == "__main__":
|
84 |
+
run()
|
pages/1_🔥_WarmMolGen.py
ADDED
@@ -0,0 +1,193 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import streamlit as st
|
2 |
+
import streamlit.components.v1 as components
|
3 |
+
import pandas as pd
|
4 |
+
import mols2grid
|
5 |
+
from ipywidgets import interact, widgets
|
6 |
+
import textwrap
|
7 |
+
import numpy as np
|
8 |
+
from transformers import EncoderDecoderModel, RobertaTokenizer
|
9 |
+
|
10 |
+
from moses.metrics.utils import QED, SA, logP, NP, weight, get_n_rings
|
11 |
+
from moses.utils import mapper, get_mol
|
12 |
+
|
13 |
+
# @st.cache(allow_output_mutation=False, hash_funcs={Tokenizer: str})
|
14 |
+
from typing import List
|
15 |
+
|
16 |
+
from util import filter_dataframe
|
17 |
+
|
18 |
+
|
19 |
+
@st.cache(suppress_st_warning=True)
|
20 |
+
def load_models():
|
21 |
+
# protein_tokenizer = RobertaTokenizer.from_pretrained("gokceuludogan/WarmMolGenTwo")
|
22 |
+
# mol_tokenizer = RobertaTokenizer.from_pretrained("seyonec/PubChem10M_SMILES_BPE_450k")
|
23 |
+
model1 = EncoderDecoderModel.from_pretrained("gokceuludogan/WarmMolGenOne")
|
24 |
+
model2 = EncoderDecoderModel.from_pretrained("gokceuludogan/WarmMolGenTwo")
|
25 |
+
return model1, model2 # , protein_tokenizer, mol_tokenizer
|
26 |
+
|
27 |
+
|
28 |
+
def count(smiles_list: List[str]):
|
29 |
+
counts = []
|
30 |
+
for smiles in smiles_list:
|
31 |
+
counts.append(len(smiles))
|
32 |
+
|
33 |
+
return counts
|
34 |
+
|
35 |
+
|
36 |
+
def remove_none_elements(mol_list, smiles_list):
|
37 |
+
filtered_mol_list = []
|
38 |
+
filtered_smiles_list = []
|
39 |
+
indices = []
|
40 |
+
for i, element in enumerate(mol_list):
|
41 |
+
if element is not None:
|
42 |
+
filtered_mol_list.append(element)
|
43 |
+
else:
|
44 |
+
indices.append(i)
|
45 |
+
removed_len = len(indices)
|
46 |
+
|
47 |
+
for i in range(len(smiles_list)):
|
48 |
+
if i not in indices:
|
49 |
+
filtered_smiles_list.append(smiles_list.__getitem__(i))
|
50 |
+
|
51 |
+
return filtered_mol_list, filtered_smiles_list, removed_len
|
52 |
+
|
53 |
+
|
54 |
+
def format_list_numbers(lst):
|
55 |
+
for i, value in enumerate(lst):
|
56 |
+
lst[i] = float("{:.3f}".format(value))
|
57 |
+
|
58 |
+
|
59 |
+
def generate_molecules(model_name, num_mols, max_new_tokens, do_sample, num_beams, target, pool):
|
60 |
+
protein_tokenizer = RobertaTokenizer.from_pretrained("gokceuludogan/WarmMolGenTwo")
|
61 |
+
mol_tokenizer = RobertaTokenizer.from_pretrained("seyonec/PubChem10M_SMILES_BPE_450k")
|
62 |
+
# model1, model2, protein_tokenizer, mol_tokenizer = load_models()
|
63 |
+
model1, model2 = load_models()
|
64 |
+
inputs = protein_tokenizer(target, return_tensors="pt")
|
65 |
+
|
66 |
+
model = model1 if model_name == 'WarmMolGenOne' else model2
|
67 |
+
outputs = model.generate(**inputs, decoder_start_token_id=mol_tokenizer.bos_token_id,
|
68 |
+
eos_token_id=mol_tokenizer.eos_token_id, pad_token_id=mol_tokenizer.eos_token_id,
|
69 |
+
max_length=int(max_new_tokens), num_return_sequences=int(num_mols),
|
70 |
+
do_sample=do_sample, num_beams=num_beams)
|
71 |
+
output_smiles = mol_tokenizer.batch_decode(outputs, skip_special_tokens=True)
|
72 |
+
st.write("### Generated Molecules")
|
73 |
+
# mol_list = list(map(MolFromSmiles, output_smiles))
|
74 |
+
# print(mol_list)
|
75 |
+
# QED_scores = list(map(QED.qed, mol_list))
|
76 |
+
# print(QED_scores)
|
77 |
+
# st.write(output_smiles)
|
78 |
+
mol_list = mapper(pool)(get_mol, output_smiles)
|
79 |
+
mol_list, output_smiles, removed_len = remove_none_elements(mol_list, output_smiles)
|
80 |
+
if removed_len != 0:
|
81 |
+
st.write(f"#### Note that: {removed_len} numbers of generated invalid molecules are discarded.")
|
82 |
+
|
83 |
+
QED_scores = mapper(pool)(QED, mol_list)
|
84 |
+
SA_scores = mapper(pool)(SA, mol_list)
|
85 |
+
logP_scores = mapper(pool)(logP, mol_list)
|
86 |
+
NP_scores = mapper(pool)(NP, mol_list)
|
87 |
+
weight_scores = mapper(pool)(weight, mol_list)
|
88 |
+
|
89 |
+
format_list_numbers(QED_scores)
|
90 |
+
format_list_numbers(SA_scores)
|
91 |
+
format_list_numbers(logP_scores)
|
92 |
+
format_list_numbers(NP_scores)
|
93 |
+
format_list_numbers(weight_scores)
|
94 |
+
|
95 |
+
df_smiles = pd.DataFrame(
|
96 |
+
{'SMILES': output_smiles, "QED": QED_scores, "SA": SA_scores, "logP": logP_scores, "NP": NP_scores,
|
97 |
+
"Weight": weight_scores})
|
98 |
+
|
99 |
+
return df_smiles
|
100 |
+
|
101 |
+
|
102 |
+
def warm_molgen_demo():
|
103 |
+
with st.form("my_form"):
|
104 |
+
with st.sidebar:
|
105 |
+
st.sidebar.subheader("Configurable parameters")
|
106 |
+
|
107 |
+
model_name = st.sidebar.selectbox(
|
108 |
+
"Model Selector",
|
109 |
+
options=[
|
110 |
+
"WarmMolGenOne",
|
111 |
+
"WarmMolGenTwo",
|
112 |
+
],
|
113 |
+
index=0,
|
114 |
+
)
|
115 |
+
|
116 |
+
num_mols = st.sidebar.number_input(
|
117 |
+
"Number of generated molecules",
|
118 |
+
min_value=0,
|
119 |
+
max_value=20,
|
120 |
+
value=10,
|
121 |
+
help="The number of molecules to be generated.",
|
122 |
+
)
|
123 |
+
|
124 |
+
max_new_tokens = st.sidebar.number_input(
|
125 |
+
"Maximum length",
|
126 |
+
min_value=0,
|
127 |
+
max_value=1024,
|
128 |
+
value=128,
|
129 |
+
help="The maximum length of the sequence to be generated.",
|
130 |
+
)
|
131 |
+
do_sample = st.sidebar.selectbox(
|
132 |
+
"Sampling?",
|
133 |
+
(True, False),
|
134 |
+
help="Whether or not to use sampling; use beam decoding otherwise.",
|
135 |
+
)
|
136 |
+
target = st.text_area(
|
137 |
+
"Target Sequence",
|
138 |
+
"MENTENSVDSKSIKNLEPKIIHGSESMDSGISLDNSYKMDYPEMGLCIIINNKNFHKSTG",
|
139 |
+
)
|
140 |
+
generate_new_molecules = st.form_submit_button("Generate Molecules")
|
141 |
+
|
142 |
+
num_beams = None if do_sample is True else int(num_mols)
|
143 |
+
|
144 |
+
pool = 1
|
145 |
+
|
146 |
+
if generate_new_molecules:
|
147 |
+
st.session_state.df = generate_molecules(model_name, num_mols, max_new_tokens, do_sample, num_beams,
|
148 |
+
target, pool)
|
149 |
+
if 'df' not in st.session_state:
|
150 |
+
st.session_state.df = generate_molecules(model_name, num_mols, max_new_tokens, do_sample, num_beams,
|
151 |
+
target, pool)
|
152 |
+
df = st.session_state.df
|
153 |
+
|
154 |
+
filtered_df = filter_dataframe(df)
|
155 |
+
if filtered_df.empty:
|
156 |
+
st.markdown(
|
157 |
+
"""
|
158 |
+
<span style='color: blue; font-size: 30px;'>No molecules were found with specified properties.</span>
|
159 |
+
""",
|
160 |
+
unsafe_allow_html=True
|
161 |
+
)
|
162 |
+
else:
|
163 |
+
raw_html = mols2grid.display(filtered_df, height=1000)._repr_html_()
|
164 |
+
components.html(raw_html, width=900, height=450, scrolling=True)
|
165 |
+
|
166 |
+
st.markdown("## How to Generate")
|
167 |
+
generation_code = f"""
|
168 |
+
from transformers import EncoderDecoderModel, RobertaTokenizer
|
169 |
+
protein_tokenizer = RobertaTokenizer.from_pretrained("gokceuludogan/{model_name}")
|
170 |
+
mol_tokenizer = RobertaTokenizer.from_pretrained("seyonec/PubChem10M_SMILES_BPE_450k")
|
171 |
+
model = EncoderDecoderModel.from_pretrained("gokceuludogan/{model_name}")
|
172 |
+
inputs = protein_tokenizer("{target}", return_tensors="pt")
|
173 |
+
outputs = model.generate(**inputs, decoder_start_token_id=mol_tokenizer.bos_token_id,
|
174 |
+
eos_token_id=mol_tokenizer.eos_token_id, pad_token_id=mol_tokenizer.eos_token_id,
|
175 |
+
max_length={max_new_tokens}, num_return_sequences={num_mols}, do_sample={do_sample}, num_beams={num_beams})
|
176 |
+
mol_tokenizer.batch_decode(outputs, skip_special_tokens=True)
|
177 |
+
"""
|
178 |
+
st.code(textwrap.dedent(generation_code)) # textwrap.dedent("".join("Halletcez")))
|
179 |
+
|
180 |
+
|
181 |
+
st.set_page_config(page_title="WarmMolGen Demo", page_icon="🔥", layout='wide')
|
182 |
+
st.markdown("# WarmMolGen Demo")
|
183 |
+
st.sidebar.header("WarmMolGen Demo")
|
184 |
+
st.markdown(
|
185 |
+
"""
|
186 |
+
This demo illustrates WarmMolGen models' generation capabilities.
|
187 |
+
Given a target sequence and a set of parameters, the models generate molecules targeting the given protein sequence.
|
188 |
+
Please enter an input sequence below 👇 and configure parameters from the sidebar 👈 to generate molecules!
|
189 |
+
See below for saving the output molecules and the code snippet generating them!
|
190 |
+
"""
|
191 |
+
)
|
192 |
+
|
193 |
+
warm_molgen_demo()
|
pages/__pycache__/util.cpython-37.pyc
ADDED
Binary file (2.04 kB). View file
|
|
pages/util.py
ADDED
@@ -0,0 +1,87 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import pandas as pd
|
2 |
+
import streamlit as st
|
3 |
+
import streamlit.components.v1 as components
|
4 |
+
from pandas.api.types import (
|
5 |
+
is_categorical_dtype,
|
6 |
+
is_datetime64_any_dtype,
|
7 |
+
is_numeric_dtype,
|
8 |
+
is_object_dtype,
|
9 |
+
)
|
10 |
+
|
11 |
+
def filter_dataframe(df: pd.DataFrame) -> pd.DataFrame:
|
12 |
+
"""
|
13 |
+
Adds a UI on top of a dataframe to let viewers filter columns
|
14 |
+
|
15 |
+
Args:
|
16 |
+
df (pd.DataFrame): Original dataframe
|
17 |
+
|
18 |
+
Returns:
|
19 |
+
pd.DataFrame: Filtered dataframe
|
20 |
+
"""
|
21 |
+
modify = st.checkbox("Add filters")
|
22 |
+
|
23 |
+
if not modify:
|
24 |
+
return df
|
25 |
+
|
26 |
+
df = df.copy()
|
27 |
+
|
28 |
+
# Try to convert datetimes into a standard format (datetime, no timezone)
|
29 |
+
for col in df.columns:
|
30 |
+
if is_object_dtype(df[col]):
|
31 |
+
try:
|
32 |
+
df[col] = pd.to_datetime(df[col])
|
33 |
+
except Exception:
|
34 |
+
pass
|
35 |
+
|
36 |
+
if is_datetime64_any_dtype(df[col]):
|
37 |
+
df[col] = df[col].dt.tz_localize(None)
|
38 |
+
|
39 |
+
modification_container = st.container()
|
40 |
+
|
41 |
+
with modification_container:
|
42 |
+
limit_non_unique = 1
|
43 |
+
to_filter_columns = st.multiselect("Filter dataframe on", df.columns)
|
44 |
+
for column in to_filter_columns:
|
45 |
+
if df[column].dtype == 'O': # Check if the column is of 'object' dtype (i.e., string)
|
46 |
+
df[column] = df[column].astype(pd.CategoricalDtype())
|
47 |
+
left, right = st.columns((1, 20))
|
48 |
+
# Treat columns with < 10 unique values as categorical
|
49 |
+
if is_categorical_dtype(df[column]) or df[column].nunique() < limit_non_unique:
|
50 |
+
user_cat_input = right.multiselect(
|
51 |
+
f"Values for {column}",
|
52 |
+
df[column].unique(),
|
53 |
+
default=list(df[column].unique()),
|
54 |
+
)
|
55 |
+
df = df[df[column].isin(user_cat_input)]
|
56 |
+
elif is_numeric_dtype(df[column]):
|
57 |
+
_min = float(df[column].min())
|
58 |
+
_max = float(df[column].max())
|
59 |
+
step = (_max - _min) / 100
|
60 |
+
user_num_input = right.slider(
|
61 |
+
f"Values for {column}",
|
62 |
+
min_value=_min,
|
63 |
+
max_value=_max,
|
64 |
+
value=(_min, _max),
|
65 |
+
step=step,
|
66 |
+
)
|
67 |
+
df = df[df[column].between(*user_num_input)]
|
68 |
+
elif is_datetime64_any_dtype(df[column]):
|
69 |
+
user_date_input = right.date_input(
|
70 |
+
f"Values for {column}",
|
71 |
+
value=(
|
72 |
+
df[column].min(),
|
73 |
+
df[column].max(),
|
74 |
+
),
|
75 |
+
)
|
76 |
+
if len(user_date_input) == 2:
|
77 |
+
user_date_input = tuple(map(pd.to_datetime, user_date_input))
|
78 |
+
start_date, end_date = user_date_input
|
79 |
+
df = df.loc[df[column].between(start_date, end_date)]
|
80 |
+
else:
|
81 |
+
user_text_input = right.text_input(
|
82 |
+
f"Substring or regex in {column}",
|
83 |
+
)
|
84 |
+
if user_text_input:
|
85 |
+
df = df[df[column].astype(str).str.contains(user_text_input)]
|
86 |
+
|
87 |
+
return df
|