def train_tars(mode="prod"):
# Interdict the Transformers cache
# TODO: Make this cleaner somehow
import os
from keter.stage import FileSystemStage, get_path
with FileSystemStage():
os.environ["TRANSFORMERS_CACHE"] = str(
get_path("external") / "transformers")
from keter.actors.flair import ChemicalUnderstandingTARS
ChemicalUnderstandingTARS()
def drug_discovery_on_moses(mode="prod"):
from tqdm.auto import tqdm
import pandas as pd
from keter.stage import FileSystemStage, get_path
from keter.actors.sklearn import Analyzer
from keter.datasets.raw import Moses
from keter.interfaces.chemistry import create_jamstack
with FileSystemStage():
if mode == "prod":
analyzer = Analyzer()
elif mode == "doc2vec":
analyzer = Analyzer("doc2vec")
elif mode == "lda":
analyzer = Analyzer("lda")
else:
raise ValueError(f"Invalid mode: {mode}")
moses = Moses().to_df()["SMILES"].tolist()
last = 0
block_size = 107609
blocks = []
get_path("output").mkdir(parents=True, exist_ok=True)
for i in tqdm(
range(0, len(moses), block_size),
total=len(moses) // block_size,
unit="block",
):
blocks.append(analyzer.analyze(moses[i:i + block_size]))
pd.concat(blocks).reset_index(drop=True).to_parquet(
get_path("output") / "moses_drugs.parquet")
create_jamstack()
def train(self):
tox_corpus = FlairTox21().to_corpus()
self.model = TARSClassifier(
task_name="Toxicity",
label_dictionary=tox_corpus.make_label_dictionary(),
document_embeddings="distilbert-base-uncased",
)
trainer = ModelTrainer(self.model, tox_corpus)
trainer.train(
base_path=get_path("model") / self.filename,
learning_rate=0.02,
mini_batch_size=1,
max_epochs=10,
)
def to_csv(self) -> Sequence[str]:
constructed_data_root = get_path("constructed")
csv_file = (constructed_data_root /
self.filename).with_suffix(".csv.xz")
if csv_file.exists():
with lzma.open(csv_file, "rt") as fd:
for line in fd:
yield line.rstrip()
return
corona_deaths = CoronaDeathsUSA().to_df()
corona_deaths = corona_deaths.rename(
columns={
column: int(parse(column).timestamp())
for column in corona_deaths.columns if "/" in column
})
timestamp_columns = [
column for column in corona_deaths.columns
if isinstance(column, int)
]
corona_deaths[timestamp_columns] = corona_deaths[
timestamp_columns].diff(axis=1)
corona_deaths = corona_deaths.dropna(axis=1)
constructed_data_root.mkdir(parents=True, exist_ok=True)
fd = lzma.open(csv_file, "wt")
for row in corona_deaths.iterrows():
_, series = row
for column, val in series.items():
if isinstance(column, int):
for record in construct_infection_records(
column, val, series.Lat, series.Long_):
fd.write(record + "\n")
yield record
fd.close()
def download(self, assay: str):
if assay not in self.tox21_assays:
raise ValueError(f"Not a valid Tox21 assay: {assay}")
raw_dir = get_path("raw") / "tox21"
raw_url = f"https://tripod.nih.gov/tox21/assays/download/{assay}.zip"
return urlopen(raw_url).read()
def create_jamstack():
app.config["FREEZER_IGNORE_MIMETYPE_WARNINGS"] = True
app.config["FREEZER_DESTINATION"] = get_path("output") / "static_html"
freezer = Freezer(app)
freezer.freeze()
def make_drug_db():
db.make_drug_db(get_path("output"))