The PyTorch implementation of MoMu, described in "Natural Language-informed Modeling of Molecule Graphs".
Our collected dataset consists of three folders holding molecular graphs, smiles and texts, respectively. The dataset can be downloaded in https://pan.baidu.com/s/1hYbVOMhr7pwLkUtl5FWqbg, the passward is 1234. For cross-modality retrieval, the PCdes dataset is available at https://github.com/thunlp/KV-PLM. For text-to-molecule generation, the pre-trained MoFlow is available at https://github.com/calvin-zcx/moflow. For molecule caption, the ChEBI-20 dataset is directly available in our repository. For molecule property prediction, the eight datasets from MoleculeNet are available at https://github.com/deepchem/deepchem/tree/master/datasets and the processed datasets are available at: http://snap.stanford.edu/gnn-pretrain/data/chem_dataset.zip. Please refer to the ReadMe in each part for the detailed data downloading and usage.
We will introduce our pretrain method and all the downstream tasks below. Please refer to the ReadMe in each part for the source code, system requirements, installation, demo, instructions for use, etc.
The pretrain code is available in the Pretrain/ folder.
Our pretrained models MoMu-S and MoMu-K can be downloaded on the Baidu Netdisk, the password is 1234. All the downstream tasks use these two models.
MoMu-K: checkpoints/littlegin=graphclinit_bert=kvplm_epoch=299-step=18300.ckpt
MoMu-S: checkpoints/littlegin=graphclinit_bert=scibert_epoch=299-step=18300.ckptSince our MoMu model is pre-trained by matching weakly-correlated texts to corresponding molecular graphs, it is able to process both the graph and text modalities of molecules. We evaluate its performance in cross-modality retrieval. Given a molecule graph, graph-to-text (G-T) retrieval aims to retrieve the most relevant text descriptions of this molecule. Conversely, given a text paragraph, text-to-graph (T-G) retrieval aims at retrieving the most relevant molecule graph it describes. The code for these two downstream tasks is available in the GraphTextRetrieval/ folder.
The molecule captioning task aims to generate texts to describe the given molecule. The code for this downstream task is available in the MoleculeCaption/ folder.
We propose a new task called zero-shot text-to-graph molecule generation. The goal is to design a cross modality molecule generator that takes as input the natural language description of the desired conditions and imagines new molecules that match the description. The code for this downstream task is available in the Text2graph/ folder.
Molecular property prediction is a graph-level prediction task that is usually used to evaluate the transfer ability of pre-trained graph encoders. The code for this downstream task is available in the MoleculePrediction/ folder.
@article{su2022molecular,
title={Natural Language-informed Modeling of Molecule Graphs},
author={Bing Su, Dazhao Du, Zhao Yang, Yujie Zhou, Jiangmeng Li, Anyi Rao, Zhiwu Lu, Hao Sun, Ji-Rong Wen},
year={2022}
}