Proto-Moleculo

Introduction

• This is the repository containing the end-to-end training and inference notebook for predicting the binding affinity of small molecules to specific protein targets.

Pre-Processing

• Pre-processing involves encoding of molecule smiles using a set vocabulary map.
• The entire process is parallelized using joblib library and the encoded smiles are stored for later usage.

Modelling

• The model used is a stack of 1D CNN layers over an the encoded smiles passed through an encoder layer.
• The output of the 1D CNN stack is passed through a stack of dense layers with the final output being the three classes of protiens required.

Training

Following are the training settings used:

• Optimizer : Adam
• Metric : AUC Curve for average precision (AP)
• Stratified K-Fold
• Scheduler : Reduce LR on Plateau
• Accelerator : TPU Backend (used 8 TPU Cores)