Predicting Molecular Lipophilicity (LogP) Using a GNN and SVR

Description

This repository contains our team’s solution for the Element 119 competition organized by SIBUR, focusing on the application of artificial intelligence in chemistry. The goal of the project is to predict the lipophilicity coefficient (LogP) of organic molecules using two models:

• Graph Neural Network (GNN) implemented with PyTorch for analyzing molecular structures
  
• Support Vector Regression (SVR) with engineered physicochemical features

Key features

• Data Preprocessing: Conversion of SMILES to molecular graphs (for GNN) and RDKit descriptors (for SVR)
• Validation: Nested cross-validation and testing on blind datasets

Results

• Top 20 in the final leaderboard
• Public dataset RMSE: 0.64311
• Private dataset RMSE: 0.72190