This repository is dedicated to Learning to Smell challenge organized by AIcrowd. My best Jaccard score prediction is 0.346 (the 1st round).

Use machine learning (or any other related methods/techniques) to build a model and train it with the available dataset (4316 molecules) and predict smell character of molecules in a given test set (1079 molecules). There are 109 odors present in the dataset (data/vocabulary.txt). More details can be found on the website of the competition.

1. train.csv - (4316 molecules) : This csv file contains the attributes describing the molecules along with their "Sentence"
2. test.csv - (1079 molecules) (Round-1) : File that will be used for actual evaluation for the leaderboard score but does not have the "Sentence" for molecules.
3. vocabulary.txt : A file containing the list of all odors present in the dataset

Jaccard (Tanimoto) score - Read more

My neural network is designed with a feedforward neural network using Keras/TensorFlow library. The features (descriptors) that I used for training models are the Morgan fingerprint and structural properties of a molecule such as functional groups, number of benzene rings, van der Waals volume, etc.

In alphabetical order

• AdamW.py - Implementation of AdamW optimizer
• augmentation.py - Data augmentation
• descriptors.py - Extract structural properties from molecules and use them as features
• fingerprint.py - Calculate Morgan fingerprint
• func.py - Miscellaneous functions
• loss.py - Customized loss functions
• metric.py - Customized metric
• predict.py - Prediction
• README.md - This file
• run_24c_1g.sh SLURM input file
• train-predict.ipynb - Google Colab file
• train.py - Train model
• vdW_volume.py - Calculate van der Waals volume
• data - Dataset and files
• submission - Submission files

1. Download dataset and install necessary libraries.
2. Calculate Morgan fingerprints (1024 bits) of molecules using RDKit.
3. Build a model containing 3 hidden layers and 128 neurons each. DropOut and batch normalization are applied.
4. Compile model with Adam optimizer. Use Categorical entropy as a loss function and accuracy as a metric.
5. Train model for 300 epochs. Learning rate schedule and early stop techniques are also applied when a metric has stopped improving.
6. Predict smell.
7. Choose the top 15 smell predictions for each sample (molecule) and group by 3. Here is an example of a submission file.
8. Submit the prediction results (.csv) and get the score.