About

Phytebyte is an extensible software framework used to train machine learning models to identify bioactive compounds found in food. The use case captured in 'run.py' is training models using known drug compounds that target a specific gene, and identifying plant compounds in FooDB that may behave similarly.

Install

Download openbabel

 curl -L https://github.com/openbabel/openbabel/archive/refs/tags/openbabel-3-1-1.tar.gz --output openbabel-3-1-1.tar.gz
2. Install openbabel

tar zxf openbabel-3-1-1.tar.gz mkdir build cd build cmake ../openbabel-openbabel-3-1-1 -DPYTHON_BINDINGS=ON -DPYTHON_EXECUTABLE=/usr/local/bin/python3 -DRUN_SWIG=ON make -j4 make install

3. Create virtual environment & install dependencies

# In phytebyte root directory
/usr/local/bin/python3 -m venv env
# ^ Needs to be the same as PYTHON_EXECUTABLE flag in installation flag in prev step
source env/bin/activate
pip install -r requirements.txt

4. Install pybel (Python wrapper for openbabel)

pip install openbabel

5. Download the 'FooDB SQL file' from [http://foodb.ca/downloads](http://foodb.ca/downloads)

curl -L https://foodb.ca/public/system/downloads/foodb_2020_4_7_mysql.tar.gz --output foodb.tar.gz

Check the URL above, new version may be available!

6. Unzip the file

mkdir foodb tar zxf foodb.tar.gz -C foodb

6. Install mysql

brew install mysql brew tap homebrew/services brew services start mysql

Change local admin password to whatever you'd like.

mysqladmin -u root password 'phytebyte'

Verify that you can connect

mysql -uroot -pphytebyte

Create the foodb Database

CREATE DATABASE foodb;

7. Load foodb Database

cat foodb_*.sql | mysql -uroot -pphytebyte foo_db

Using wildcard to catch future versions of foodb database


8. Install postgresql

brew install postgres brew services start postgres

Verify you can connect

psql # Should open psql> prompt

Create the 'chembl' database

CREATE DATABASE chembl_30;

8. Download & decompress ChEMBL Database (for postgresql)

curl -O https://ftp.ebi.ac.uk/pub/databases/chembl/ChEMBLdb/latest/chembl_30_postgresql.tar.gz gunzip chembl_30_postgresql.tar.gz tar xopf chembl_30_postgresql.tar cd chembl_30/chembl_30_postgresql pg_restore -d chembl_30 chembl_30_postgresql.dmp

9. Set ENV variables in ~/.bash_profile

Open up ~/.bash_profile with text-editor, enter following at bottom

export FOODB_URL="mysql://root:phytebyte@localhost/foo_db" export CHEMBL_DB_URL="postgresql:///chembl_30" export PYTHONPATH=/usr/local/lib/python3.9/site-packages:$PYTHONPATH

Replace "python3.9" with the version with bindings linked in step 2/3

10. Re-load ~/.bash_profile

source ~/.bash_profile

11. Run Tests

From phytebyte root dir

source env/bin/activate # Active python virtual env pytest -vv tests


12. Run Phytebyte!

python run.py


# Customization

**BioactiveCompoundSource**
The `BioactiveCompoundSource` abstract method defines an interface for fetching compounds by gene target, by name and via exclusion (by checking SMILES for equality). It represents a data source of compounds that will be queried to train a `BinaryClassifierModel`.

**TargetInputs**
The `TargetInput` is a simple data structure that tells the `BioactiveCompoundSource` what method to use when fetching input data for the `BinaryClassifierModel`. One can leverage the `GeneTargetsInput` (as done in run.py) to train a model on all compounds that either `agonize` or `antagonize` the given gene, in this use case, `PPARG`. Alternatively, the `CompoundNamesTargetInput` can be used to give a list of compound names found in ChEMBL to manually select the compounds used to train the model.

**Fingerprinters**
The `FP_TYPE` variable can be set to either `daylight` or `spectrophore` in `run.py`. Alternative fingerprints can be created by implementing the Fingerprinter abstract base class.

**FoodCmpdSource**
The `FoodCmpdSource` abstract base class can be further implemented to support additional food database sources. Our case study uses FooDB, but we encourage extending to other sources!

**BinaryClassifierModel**
The `BinaryClassifierModel` abstract base class can be further implemented to support alternative binary classifiers (SVM, LogisticRegression, etc.). Our use case leverages sklearn's `Random Forest`, and we also provide support for the `TanimotoBinaryClassifier` that uses tanimoto index to classify compounds.

# Config

**Fingerprinters**
In `run.py`, `FP_TYPE` can be set to either `spectrophore` or `daylight`

**Database URLs**
In `run.py` we use environment variables storing `CHEMBL_DB_URL` and `FOODB_URL`. These get passed to each respective `Source` for querying compounds.

**Negative Sample Size Factor**
The `neg_sample_size_factor` indicates the multiple applied to the number of `positive` compound samples (i.e. `20` means "20 times the number of positive samples").

seanharr11/phytebyte