Please read through the directions thoroughly and then move on to setting up your working environment.
In this workshop you will:
- Find a breast cancer cell line that is representative of your tumor sample from the NCI-60
- Analyze the NCI-60 screening dataset to select compounds that are potent against your cell line
- Build a model that uses chemical structure to predict protein target profiles
- Predict which protein targets your selected compounds bind
- Analyze ligand profiles of targets (known binders) that commonly appear in your predictions to identify important chemical substructures
- Repeat step 4 with a neural network to compare results across different ML methods.
thank you to @gtgask for putting this script together in 2017
moved to python 3.9 in 2021 by @wconnell
updated 2023 by @laurashub and @ghorbanimahdi73
The overall goal of this workship is (1) to gain experience with computational tools such as Python, conda, Jupyter notebooks, and (2) apply machine learning methods to the DTP-NCI60 cancer drug screening dataset (https://dtp.cancer.gov/discovery_development/nci-60) to predict which compounds are most relevant for treatment given the tumor data provided to you. The NCI60 dataset consists of approximately 27k compounds screened against 60 individual cancer cell lines, originating from 10 tissue subtypes.
cancer_compounds.sample.csv
- This file can be opened in Excel, or a simple text editor, and contains the SMILES and names or IDs of various cancer-related compounds.
- It demonstrates the file format for to be used for the query drugs as input to the "Predict" notebook.
candidate_compounds.sample.csv
This file can be opened in Excel, or a simple text editor, and likewise contains SMILES and IDs of several cancer compounds from the first data file.
- It demonstrates the file format for to be used for the query drugs as input to the "Compare" notebook.
chembl_21_binding_molecules.csv.gz
- This file contains: Chembl_ID, SMILE, Fingerprint
- The code requires this file to be gzipped, but you can explore this file by unzipping into a new file:
gunzip -c chembl_21_binding_molecules.csv.gz > chembl_21_binding_molecules.csv
chembl_21_binding_targets.csv.gz
- This file contains: Target ID, UniProt Name, Molecule IDs, Description
--The code requires this file to be gzipped, but you can explore this file by unzipping into a new file:
gunzip -c chembl_21_binding_targets.csv.gz > chembl_21_binding_targets.csv
This workshop will be conducted entirely in Python, a programming language with broad scientific use. We will be using a number of packages, including scikit-learn and Pytorch for model training, as visualization libraries such as matplotlib and seaborne.
Let's create a directory for us to work from. This directory is where we will store all the files we plan to use.
First clone this repository on your local computer by opening a terminal window and typing:
cd ~/Documents
git clone git@github.com:laurashub/PSPG-CDD.git
Conda (and miniconda) is an environment management system that will help us to install packages and resolve dependencies. If you do not have Conda installed, install the latest version of Miniconda for you OS from here: https://docs.conda.io/en/latest/miniconda.html
Once Miniconda has been installed, close your terminal window, and reopen another one. Check to see conda was installed correctly. Type the following command to see the list of packages that were installed.
conda list
Now that we have the infrastructure in place, we need to create a new conda environment, and install the additional packages, such as rdkit and jupyter, we will need to generate our predictions.
Create a new environment using the following code:
cd PSPG-CDD/ # change location to PSPG-CDD directory
conda env create -f environment.yml # create environment 'cdd_workshop'
Activate the environment and confirm:
conda activate cdd_workshop # activates our installed environment and gives us access to included packages
conda env list # list of installed environments - should see * beside cdd_workshop
For this workshop, we will be using an interactive Python session with Jupyter notebooks. This opens notebook files (.ipynb extensions) inside a web browser that can be run manually, as opposed to creating a single .py file that is run all at once. First, we need to make the conda environment we created available to Jupyter with the following command:
ipython kernel install --user --name=ccd_workshop
To test if everything is working, let's open up a jupyter lab session. We'll import the modules we need to run our scripts correctly.
cd WorkshopNotebooks
jupyter notebook
This should bring up a jupyter notebook session that is being hosted from your local machine in your browser. Open PSPG245B-NaiveBayes-Classification-STUDENTS.ipynb and try running the import cell by either hitting Shift+Enter or clicking "Run" at the top of the screen. If everything is set up correctly, that should throw no errors.
In this workshop, you will use machine learning tools from scikit-learn (aka sklearn, http://scikit-learn.org) along with chemical fingerprinting and visualization tools from RDKit (http://rdkit.org) to predict the protein targets of cancer-related compounds. This workshop uses two Python notebooks that are mostly complete. Your job is to fill in the missing sections and use these notebooks to explore the protein (target) profiles of these cancer-related compounds. Your goal is to decide which compound(s) are most applicable to your tumor for treatment and then predict which protein targets these compounds will bind.
In order to determine which compounds are optimal for treatment, you will need to identify which breast cancer cell line is most relevant to your tumor. Before the second portion of the workshop, your goal is to explore the datasets provided by the NCI60 and to make a decision on which cell line(s) to focus on. You will share how you made your decision.
The NCI's genomics and bioinformatics group has created a web portal consolidating all the data and metadata related to the screen that is publicly available and convenient to access. The sites homepage is located here: https://discover.nci.nih.gov/cellminer/home.do We encourage you to examine the upper blue tabs, especially the ones labeled "Data Set Metadata", "Cell Line Metadata", and "Download Data Sets"
Perform exploratory data analysis on your cell line to decide which compounds you would like to make target profile predictions for.
PSPG245B-NaiveBayes-Classification-STUDENTS.ipynb
This notebook explains common molecular formats such as SMILES and molecular fingerprint as well as molecular similarity calculations using the RDKit package. In addition, this notebook takes in a file containing compounds (SMILES), along with two reference files containing data on molecules and targets from ChEMBL. It loads the data, converts the SMILES into in-memory representations of molecules, calculates chemical fingerprints from them, and trains a machine-learning method called a Naïve Bayesian Classifier to predict the drug targets of each compound.
http://scikit-learn.org/stable/modules/naive_bayes.html#multinomial-naive-bayes
Some parts of the script are incomplete, and these are marked with question marks (?). It is your job to fill them in.
PSPG245B-NeuralNetwork-Classification-STUDENTS.ipynb
This notebook is similar to the later part of 2b, except instead of using a neural network this notebook demonstrates how to setup and train a deep neural network using the PyTorch library. It details setting up the Datasets and DataLoaders, defining the model architecture, and evaluating performance. It also shows how to apply the model to get predictions for novel compounds to compare predictions between the two different classification methods.
https://pytorch.org/docs/stable/index.html
The script is largely complete, but there are directions to follow and questions to answer.
PSPG245B-Compare-Ligands-STUDENTS.ipynb
This notebook is used to explore why a prediction from the first script was made. It takes a file of drugs along with a target ID (e.g., “CHEMBL...”), and the two ChEMBL reference files as mentioned before. It uses RDKit to calculate the molecular similarity of all the ligands for the target you specified as compared to the drugs you gave it, and shows the most similar ligands.
http://www.rdkit.org/docs/GettingStartedInPython.html
Some parts of the script are incomplete, and these are marked with question marks (?). It is your job to fill them in.
Do these ligands share common patterns, functional groups, “warheads”, etc with your compound? Which ones?