Skip to content

Pablo-R0/MolProPreNet

Repository files navigation

MolProPreNet - GNN-based Polymer Property Prediction

This is a playground repository. This README.md was partly generated by Claude Sonnet 4.6.

Predicting glass transition temperature (Tg) of polymers from SMILES using Graph Neural Networks.


Motivation

Measuring polymer properties experimentally is costly and time-consuming. ML-based prediction allows screening thousands of candidates in silico before any synthesis — drastically accelerating materials discovery.

GNNs are particularly suited for polymers: unlike fingerprint-based models (Random Forest, SVM), GNNs operate directly on the molecular graph and learn structural patterns automatically, without manual feature engineering.

Predicting Tg before synthesis has direct industrial impact: it guides the design of high-performance polymers for applications in aerospace, electronics, and packaging.


Architecture

SMILES → RDKit graph → PyTorch Geometric Data object
             ↓
        GCNConv (9 → 128) + ReLU
             ↓
        GCNConv (128 → 64) + ReLU
             ↓
        global_mean_pool
             ↓
        Linear (64 → 1)  →  Tg (K)

Node features (9): atomic number, degree, formal charge, is in ring, is aromatic, hybridization, total Hs, is in ring of size 5, is in ring of size 6

Dataset: PolyMetriX CuratedGlassTempDataset — 7 367 labeled polymers Split: 80% train / 10% val / 10% test


Results

Metric Value
RMSE (test) ~64 K
MAE (test) ~50 K

Baseline model — not yet optimized. Further improvements planned (deeper architecture, edge features, PolyBERT comparison).


Installation

git clone https://github.com/Pablo-R0/MolProPreNet
cd MolProPreNet
conda create -n molproprenet python=3.10
conda activate molproprenet
pip install -r requirements.txt

Usage

# To be updated

References

  • PolyMetriX : github.com/lamalab-org/PolyMetriX
  • PyTorch Geometric : pytorch-geometric.readthedocs.io

About

This is a playground repository.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors