Generalizing property prediction of ionic liquids from limited labeled data: a one-stop framework empowered by transfer learning

Guzhong Chen^a,b, Zhen Song^a,, Zhiwen Qi^a,, Kai Sundmacher^b,c

^a State Key laboratory of Chemical Engineering, School of Chemical Engineering, East China University of Science and Technology, 130 Meilong Road, Shanghai 200237, China

^b Process Systems Engineering, Max Planck Institute for Dynamics of Complex Technical Systems, Sandtorstr. 1, D-39106 Magdeburg, Germany

^c Process Systems Engineering, Otto-von-Guericke University Magdeburg, Universitätsplatz 2, D-39106 Magdeburg, Germany

Corresponding authors: [email protected] (Z. S.); [email protected] (Z. Qi.)

This is the official implementation of "Generalizing property prediction of ionic liquids from limited labeled data: a one-stop framework empowered by transfer learning". In this work, we present a one-stop ILTransR (IL transfer learning of Representations), a pre-training fine-tuning paradigm that leverages large-scale unlabeled data for generalizing IL property prediction from limited labeled data.

Getting Started

Installation

Set up conda environment and clone the github repo

# create a new environment
$ conda create --name iltransr python=3.8
$ conda activate iltransr

# install requirements
$ pip install mxnet-cu112
$ pip install gluonnlp
$ pip install PyYAML
$ pip install rdkit-pypi==2022.3.2.1

# clone the source code of ILTransR
$ git clone https://github.com/GuzhongChen/ILTransR.git
$ cd ILTransR

Dataset

You can download the original Pubchem coumpound database (access date: 2020.12.17) here. The IL-like dataset for pre-training can be found in data/pubchem_illike_100.csv. All the IL propertes datasets for fine-tuning are saved in datasets/il properties.

Pre-training

First, download 'pubchem_illike_100.csv' from here or from Zenodo repository and put it under ILTransR/datasets/pubchem.

Then, preprocess the IL-like dataset, run

$ python preprocess_illike.py

To pre-train the IL Transformer on the IL-like dataset,run

$ python pre-train_IL_Transformer.py

Pre-training log can be found in pretraining/il_transformer_128_1024.log

Fine-tuning

To fine-tune the ILTransR on downstream IL properties datasets, run jupyter notebooks in examples.

More IL properties will be added in the future.

Pre-trained models

Pre-trained IL Transformer can be found in pre-training/valid_best.params.

We also provide pre-trained models for IL properties prediction, which can be found in iltransr/pre-trained params.

Application example of ILTransR: CO2 absorbent screening

First, download 'predicted_IL_properties.csv' from here and put it under ILTransR/IL screening.

One can also do predictions with IL SMILES provieded in 'ONLY_SMILES.csv' using the notebooks provided in ILTransR/examples.

Then, run the jupyter notebook ILTransR/IL screening/screening.ipynb to reprodece the screening results.