POLY-X: Multi-Tier Polymer Property Prediction Platform

POLY-X is a computational platform for polymer thermophysical property prediction that integrates three complementary prediction paradigms:

Tier	Method	Description
Tier 1	Van Krevelen Group Contribution	25 SMARTS-defined functional groups, structural corrections, multi-level Boyer-Beaman Tm estimation
Tier 2	ML Ensemble (RF + GB)	ECFP4 fingerprints (2048-bit), trained on 7,365 polymers from PolyMetriX, scaffold-split evaluation
Tier 3	polyBERT Embeddings	Pre-trained DeBERTa transformer (600-D), lightweight RF+GB prediction heads

Key Results

• Tier 1 (GC): MAE = 3.9 K for Tg across 8 canonical homopolymers
• Tier 2 (ML): Test R² = 0.553, MAE = 43.6 K (scaffold split)
• Tier 3 (polyBERT): Test R² = 0.517, MAE = 47.0 K (scaffold split)

ECFP4 fingerprints outperform frozen polyBERT embeddings under scaffold splitting, demonstrating that scaffold-agnostic local representations generalize more effectively to structurally novel polymers.

Installation

pip install -r requirements.txt

Requirements: Python 3.10+, RDKit, scikit-learn, numpy, scipy, pandas, matplotlib

For Tier 3 (polyBERT), additionally install:

pip install transformers torch

Quick Start

Tier 1: Group Contribution Prediction

python -m poly_x.group_contribution "[*]CC(c1ccccc1)[*]"

Example Script

python examples/predict_example.py

Training (Tier 2)

python -m poly_x.train_tg --data-path <path_to_tg_data.csv> --output-dir results/

The training data (PolyMetriX, 7,365 polymers) is available from Zenodo: DOI: 10.5281/zenodo.14980914

Repository Structure

poly-x/
├── poly_x/                     # Core computational modules
│   ├── group_contribution.py   # Tier 1: Van Krevelen GC calculator
│   ├── fingerprints.py         # ECFP4 fingerprint generation
│   ├── scaffold_split.py       # Murcko scaffold splitting
│   └── train_tg.py             # Tier 2: ML training pipeline
├── figures/                    # Figure generation scripts
│   ├── fig1_architecture.py    # System architecture diagram
│   ├── fig2_gc_validation.py   # GC validation bar chart
│   ├── fig3_ml_parity.py       # ML parity plot (real predictions)
│   ├── fig4_tier_comparison.py # Multi-tier comparison
│   ├── fig5_enhanced_features.py # Enhanced features panel
│   └── data/                   # Pre-computed test predictions (.npy)
├── results/                    # Model performance metadata
│   ├── tier2_metrics.json      # ECFP4 ML metrics
│   └── tier3_metrics.json      # polyBERT metrics
├── examples/
│   └── predict_example.py      # Quick demo script
├── requirements.txt
└── LICENSE

Data Sources

• Training data: PolyMetriX dataset (7,365 polymers with experimental Tg values) — Zenodo DOI: 10.5281/zenodo.14980914
• Benchmark dataset: PI1M — Ma, R.; Luo, T. J. Chem. Inf. Model. 2020, 60, 4684–4690
• polyBERT model: kuelumbus/polyBERT on HuggingFace
• Group contribution parameters: Van Krevelen & Te Nijenhuis, Properties of Polymers, 4th Ed., 2009

Web Application

POLY-X is freely accessible as a web application at https://insilicosigma.com/poly-x/, where users can submit polymer structures in PSMILES notation and obtain multi-tier predictions with all enhanced analytical features.

Citation

If you use POLY-X in your research, please cite:

@article{Jebril2026POLYX,
  title   = {POLY-X: A Multi-Tier Computational Platform for Polymer
             Thermophysical Property Prediction Integrating Group
             Contribution Theory, Machine Learning Ensembles, and
             Transformer Embeddings},
  author  = {Jebril, Iqbal H. and Alshehade, Salah Addin A.},
  journal = {Journal of Chemical Information and Modeling},
  year    = {2026},
  note    = {Submitted}
}

License

This software is free for academic, research, educational, and personal use. Commercial use requires written permission from the corresponding author. See LICENSE for details.