Description
Drug Molecular Toxicity Prediction
1. Project Specification
Almost all people are exposed to different chemicals during their lifetimes through different
sources including food, household cleaning products and medicines. However, in some cases,
these chemicals can be toxic and affect human health. As a matter of fact, over 30% of drugs have
failed in human clinical trials because they are determined to be toxic despite promising preclinical studies in animal models. Consider real-world clinical trials for assessing drugs are
extremely time-consuming, it is ideal if a computational drug molecular toxicity assessment
method can be developed to quickly test whether certain chemicals have the potential to disrupt
processes in the human body of great concern to human health.
Deep neural network has become a hot research topic in machine learning in recent years.
Compared to other methods, deep learning has shown its advantages in handling large amount of
data and achieving better performance. In this individual project, you will be given a dataset of
drug molecules with their SMILES expressions (which will be explained later) and the binary
labels indicating whether one drug molecule is toxic or not. You are going to develop a Deep
Neural Network which can learn useful patterns from the data provided and predict the toxicity of
a new list of molecules based on learned knowledge using TensorFlow package.
TensorFlow is an open-source software library designed for numerical computation using data
flow graph and is widely used in deep learning community. It has many convenient APIs for
implementing deep neural networks, more details would be introduced in our tutorial.
2. SMILES Expression
Simplified Molecular-Input Line-Entry System (SMILES) is a linear representation for
molecular structure using 1D ASCII strings. For example, aspirin, a commonly used drug in daily
life, its 2D structure is
and its SMILES is
CC(=O)OC1=CC=CC=C1C(=O)O
The one hot format of SMILES is a 2D {0,1} matrix, where each column represents a symbol in
the SMILES notation of the current molecule, and each row is one ASCII character appeared in the
dataset’s SMILES dictionary. The size of the 2D matrices is the size of the dataset’s SMILES
dictionary * the length of the longest molecule SMILES, which means we have zeros padded after
short molecule SMILES. For a SMILES notation, one at row i, col j means the jth symbol of that
SMILES is the ith character in the dictionary. The one-hot example for aspirin is:
CSCI3230, Fall 2020, Neural Network Project Page 2 of 4
C C ( = O ) O C 1 = C C = C C = C 1 C ( = O ) O
C
(
=
O
)
1
3. Dataset
The dataset provided is about the toxicity of some small molecules. We provide two folders for
you, one is the training data SR-ARE-train (7167 samples), and the other one is testing data SRARE-test (234 samples). There are three files in each folder:
File Type Description
names_smiles.txt String A text file, each line contains a drug molecule’s
name and its SMILES expression, separated by
comma (,)
names_labels.txt String A text file, each line contains a drug molecule’s
name and its toxicity label, where 0 means nontoxic and 1 means toxic, separated by comma (,)
names_onehots.pickle Numeric A pickle file which can be loaded by pickle
package, storing a python dictionary containing (1)
a list of names of the molecules, and (2) an numpy
array of the one-hot representations for the
SMILES expressions of drug molecules
The source data are names_smiles.txt and names_labels.txt files. Your neural network is
supposed to learn from (SMILES, label) records, and be able to predict label from SMILES in the
end. The names_onehots.pickle file is derived from names_smiles.txt, storing one-hot
representations of SMILES expressions of drug molecules. Providing names_onehots.pickle is to
ease the burden of you on data preprocessing and focus on neural network construction. You are
not constrained on how to use these data files as long as they are the only training data. You can
build Convolutional Neural Networks or other kinds of neural networks as you like.
The molecules in SR-ARE-train and SR-ARE-test do not overlap with each other. The data we
use to mark your model is in a folder named SR-ARE-score (hundreds of samples), and you have
no access to it. There are two files in the SR-ARE-score folder: names_smiles.txt and
names_onehots.pickle. The format of these two files are exactly the same with those for training
and testing, but the molecules are new.
4. Assignment Requirement
1) This is an individual project.
2) Data
a) SR-ARE-train
b) SR-ARE-test
c) SR-ARE-score (not accessible)
You can train your model on the SR-ARE-train and test its performance on SR-ARE-test,
or you can train your model with both.
3) Features
You can either use the one hot of SMILES as features for molecules, or directly use the
SMILES notation as you wish.
CSCI3230, Fall 2020, Neural Network Project Page 3 of 4
4) Model
You can use any deep neural network architecture in this assignment to achieve good
performance, e.g. convolutional neural network (CNN), recurrent neural network (RNN),
graph neural network (GNN), etc.
5) Prediction task
You are going to predict toxicity of the molecules based on their structures. The output of
your model is the {0,1} label indicating that the current molecule is toxic or not, where 0
means non-toxic and 1 means toxic.
6) Output and Marking
Your model will be tested on the SR-ARE-score dataset. Your submitted folder will be
extracted and put alongside the SR-ARE-score folder, and it means that you must use the
relative path “../SR-ARE-score/” in your submitted test.py file to access the marking data.
In your test.py file, you need to first restore your model parameters and then test your
model on the marking data (“../SR-ARE-score/”), and you should output your predictions
into a file named labels.txt, which should be in the same directory as test.py. Each line in
labels.txt file is the {0,1} toxicity label for the corresponding sample.
Your final score of this assignment will depend on the Balanced Accuracy among your
predictions and the true labels:
𝑏𝑎𝑙𝑎𝑛𝑐𝑒_𝑎𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =
1
2
× (
𝑇𝑃
𝑇𝑃 + 𝐹𝑁 +
𝑇𝑁
𝑇𝑁 + 𝐹𝑃)
7) Deep learning library
TensorFlow 1.15.4 and NumPy only. Python’s native packages are allowed. Please do not
use other libraries. Otherwise, you will get zero marks for this assignment.
8) Programming language
The only supported language for this assignment is python3. We do not accept python2
program.
5. Submission Requirement
1) Submission list
a) Source file for training. Name it as train.py
b) Source file for recovering your network model. Name it as test.py
c) TensorFlow generated files, which store your trained model.
d) Any other files that help your programs to work, such as preprocessing files, formatconverting files.
2) Submission packaging
Put everything in the submission list above into a folder, name the folder as your student id,
zip the folder WITHOUT encryption, also name the zip file as your_student_id.zip. Submit
the zip file to our submission system.
3) DO NOT add any of the data in training and test folder in your zip file, because we will
grade your model solely on the score folder, which is already put in the server.
6. Important Points
To make this project fair and meaningful, there are some other points you MUST follows:
1) The time limits to run your test.py is 60s
2) The size of the whole zip file should be less than 200 MB
3) Plagiarism will be SERIOUSLY punished (ZERO mark plus reporting to department)
CSCI3230, Fall 2020, Neural Network Project Page 4 of 4
