Shipped

D-MPNN drug property predictor

A directed message-passing neural network with ensemble learning for ADME property prediction, trained on 12,859 molecular SMILES entries with substructure features.

Role: Author
Date: May 2024
Kind: Academic

System sketch. Boxes are the real components; packets show the direction of flow.

R²: 0.9607
Dataset: 12,859 SMILES entries
Model: D-MPNN + ensemble

Python
PyTorch
Cheminformatics
Ensemble learning

Summary

Final-year work at VIT. Developed a D-MPNN model with ensemble learning for ADME property prediction, integrating substructure features to improve screening accuracy. Reached an R² of 0.9607 on the held-out set.

It is the project that taught me how much of machine learning is data plumbing, which turned out to be good preparation for backend work.