PoCo

Overview

PoCo is a contrastive learning framework for polymer representation learning, with applications to property prediction and interpretability. This repository contains the source code for PoCo. For more details, please see our paper: Contrastive representation learning for polymer informatics.

Environment setup

Install a PyTorch build that matches your local CUDA environment, then install the dependencies:

pip install -r requirements.txt

Pretraining

1. PoCo was pretrained on ~1M polymer SMILES from the PI1M dataset. The raw dataset contains invalid SMILES; please clean it with src/pretrain/preprocess.py before pretraining.
2. Use src/pretrain/train_tokenizer.py to train a tokenizer.
3. Run the pretraining script ./pretrain.sh.

We provide pretrained PoCo weights at https://huggingface.co/CremaX/PoCo.

Transfer learning

The training entry point for downstream tasks is src/finetune/finetune.py.

Use src/finetune/run_benchmark.py to reproduce the benchmark results. The adaptation code for the baseline models largely follows the original implementations. When running the benchmarks, additional dependencies or minor code modifications may be required to accommodate different local environments.

Benchmark results

Model	Params (M)a	drepb	Khazana-MTL	PolyOmics	RadonPy	OPC	Gas
polyBERT	25	600	0.794 ± 0.019	0.790 ± 0.002	0.817 ± 0.028	0.787 ± 0.020	0.750 ± 0.029
TransPolymer	82	768	0.792 ± 0.017	0.795 ± 0.002	0.809 ± 0.026	0.786 ± 0.027	0.762 ± 0.031
PolyCL	25	600	0.794 ± 0.023	0.792 ± 0.001	0.813 ± 0.025	0.766 ± 0.025	0.765 ± 0.023
MMPolymer	129	1280	0.799 ± 0.020	0.795 ± 0.002	0.815 ± 0.026	0.779 ± 0.026	0.761 ± 0.022
PerioGT	91	2304	0.810 ± 0.025	0.805 ± 0.002	0.827 ± 0.021	0.792 ± 0.022	0.779 ± 0.022
PoCo	10	512	0.815 ± 0.016	0.800 ± 0.001	0.830 ± 0.026	0.792 ± 0.023	0.786 ± 0.025
PoCoconcat	10	1536	0.822 ± 0.015	0.806 ± 0.001	0.842 ± 0.024	0.803 ± 0.022	0.786 ± 0.025

^a Number of parameters in millions.
^b Representation dimension.

Citation

If you use PoCo in your research, please cite our paper:

@article{wang2026poco,
  title = {Contrastive representation learning for polymer informatics},
  author = {Wang, Lida and Long, Donghui},
  journal = {ChemRxiv},
  year = {2026},
  doi = {10.26434/chemrxiv.15003645/v1}
}