Home

Awesome

license

Knowledge graph-enhanced molecular contrastive learning with functional prompt

This repository is the official implementation of KANO, which is model proposed in a paper: Knowledge graph-enhanced molecular contrastive learning with functional prompt.

šŸ”” News

Brief introduction

We propose a Knowledge graph-enhanced molecular contrAstive learning with fuNctional prOmpt (KANO), exploiting fundamental domain knowledge in both pre-training and fine-tuning.

šŸ¤– Model

Firstly, we construct a Chemical Element Knowledge Graph (ElementKG) based on the Periodic Table and Wikipedia pages to summarize the class hierarchy, relations and chemical attributes of elements and functional groups.

Second, we propose an element-guided graph augmentation in contrastive-based pre-training to capture deeper associations inside molecular graphs.

Third, to bridge the gap between the pre-training contrastive tasks and downstream molecular property prediction tasks, we propose functional prompts to evoke the downstream task-related knowledge acquired by the pre-trained model.

<div align=center><img src="./fig/overview.png" style="zoom:60%;" /> </div>

šŸ”¬ Requirements

To run our code, please install dependency packages.

python          3.7
torch           1.13.1
rdkit           2018.09.3
numpy           1.20.3
gensim          4.2.0
nltk            3.4.5
owl2vec-star    0.2.1
Owlready2       0.37
torch-scatter   2.0.9

šŸ“š Overview

This project mainly contains the following parts.

ā”œā”€ā”€ chemprop                        # molecular graph preprocessing, data splitting, loss function and graph encoder
ā”œā”€ā”€ data                            # sore the molecular datasets for pre-training and fine-tuning
ā”‚Ā Ā  ā”œā”€ā”€ bace.csv                    # downstream dataset BACE
ā”‚Ā Ā  ā”œā”€ā”€ bbbp.csv                    # downstream dataset BBBP
ā”‚Ā Ā  ā”œā”€ā”€ clintox.csv                 # downstream dataset ClinTox
ā”‚Ā Ā  ā”œā”€ā”€ esol.csv                    # downstream dataset ESOL
ā”‚Ā Ā  ā”œā”€ā”€ freesolv.csv                # downstream dataset FreeSolv
ā”‚Ā Ā  ā”œā”€ā”€ hiv.csv                     # downstream dataset HIV
ā”‚Ā Ā  ā”œā”€ā”€ lipo.csv                    # downstream dataset Lipophilicity
ā”‚Ā Ā  ā”œā”€ā”€ muv.csv                     # downstream dataset MUV
ā”‚Ā Ā  ā”œā”€ā”€ qm7.csv                     # downstream dataset QM7
ā”‚Ā Ā  ā”œā”€ā”€ qm8.csv                     # downstream dataset QM8
ā”‚Ā Ā  ā”œā”€ā”€ qm9.csv                     # downstream dataset QM9
ā”‚Ā Ā  ā”œā”€ā”€ sider.csv                   # downstream dataset SIDER
ā”‚Ā Ā  ā”œā”€ā”€ tox21.csv                   # downstream dataset Tox21
ā”‚Ā Ā  ā”œā”€ā”€ toxcast.csv                 # downstream dataset ToxCast
ā”‚Ā Ā  ā””ā”€ā”€ zinc15_250K.csv             # pre-train dataset ZINC250K
ā”œā”€ā”€ dumped                          # store the training log and checkpoints of the model 
ā”‚Ā Ā  ā””ā”€ā”€ pretrained_graph_encoder    # the pre-trained model
ā”œā”€ā”€ finetune.sh                     # conduct fine-tuning
ā”œā”€ā”€ initial                         # store the embeddings of ElementKG, and preprocess it for the model
ā”œā”€ā”€ KGembedding                     # store ElementKG, and get the embeddings of eneities and relations in ElementKG
ā”œā”€ā”€ pretrain.py                     # conduct pre-training
ā””ā”€ā”€ train.py                        # training code for fine-tuning

šŸš€ Quick start

If you want to use our pre-trained model directly for molecular property prediction, please run the following command:

>> bash finetune.sh
ParameterDescriptionDefault Value
data_pathPath to downstream tasks data files (.csv)None
metricMetric to use during evaluation.Defaults to "auc" for classification and "rmse" for regression.
dataset_typeType of dataset, e.g. classification or regression, this determines the loss function used during training.'regression'
epochsNumber of epochs to run30
num_foldsNumber of folds when performing cross validation1
gpuWhich GPU to useNone
batch_sizeBatch size50
seedRandom seed to use when splitting data into train/val/test sets. When num_folds > 1, the first fold uses this seed and all subsequent folds add 1 to the seed.1
init_lrInitial learning rate1e-4
split_typeMethod of splitting the data into train/val/test (random/ scaffold splitting/ cluster splitting)'random'
stepTraining phases (pre-training, fine-tuning with functional prompts or with other architectures)'functional_prompt'
exp_nameExperiment nameNone
exp_idExperiment IDNone
checkpoint_pathPath to pre-trained model checkpoint (.pt file)None

Note that if you change the data_path, don't forget to change the corresponding metric, dataset_type and split_type! For example:

>> python train.py \
    --data_path ./data/qm7.csv \
    --metric 'mae' \
    --dataset_type regression \
    --epochs 100 \
    --num_runs 20 \
    --gpu 1 \
    --batch_size 256 \
    --seed 43 \
    --init_lr 1e-4  \
    --split_type 'scaffold_balanced' \
    --step 'functional_prompt' \
    --exp_name finetune \
    --exp_id qm7 \
    --checkpoint_path "./dumped/pretrained_graph_encoder/original_CMPN_0623_1350_14000th_epoch.pkl"

āš™ Step-by-step guidelines

ElementKG and its embedding

ElementKG is stored in KGembedding/elementkg.owl. If you want to train the model yourself to obtain the embeddings of eneities and relations in ElementKG, please run $ python run.py. This may take a few minutes to complete. For your convenience, we provide the trained representaions, stored in initial/elementkgontology.embeddings.txt

After obtaining the embeddings of ElementKG, we need to preprocess it in order to utilize it in pre-training. Please excute cd KANO/initial and run $ python get_dict.py to get the processed file. Of course, we also provide processed files in initial, so that you can directly proceed to the next step.

Contrastive-based pre-training

We collect 250K unlabeled molecules sampled from the ZINC 15 datasets to pre-train KANO. The pre-training data can be found in data/zinc15_250K.csv. If you want to pre-train the model with the pre-training data, please run:

>> python pretrain.py --exp_name 'pre-train' --exp_id 1 --step pretrain
ParameterDescriptionDefault Value
data_pathPath to pre-training data files (.csv)None
epochsNumber of epochs to run30
gpuWhich GPU to useNone
batch_sizeBatch size50

You can change these parameters directly in pretrain.py. In our setting, we set epochs and batch_size to 50 and 1024, respectively. We also provided pre-trained models, which you can download from dumped/pretrained_graph_encoder/original_CMPN_0623_1350_14000th_epoch.pkl.

Prompt-enhanced fine-tuning

The operational details of this part are the same as the section Quick start.

šŸ’” Other functions

We also provide other options in this code repository.

Cluster splitting

Our code supports using cluster splitting to split downstream datasets, as detailed in the paper. You can set thesplit_type parameter to cluster_balanced to perform cluster splitting.

Other ways to incorporate functional group knowledge

Besides functional prompts, we also support testing other ways of incorporating functional group knowledge. By setting the step parameter to finetune_add or finetune_concat, you achieve adding or concatenating functional group knowledge with the original molecular representation, respectively.

Conducting experiments on a specified dataset

We also support specifying a dataset as the input for the train/val/test sets by setting the parameters data_path, separate_test_path and separate_val_path to the location of the specified train/val/test data.

Making predictions with fine-tuned models

We now support making predictions with fine-tuned models. Use the command python predict.py --exp_name pred --exp_id pred. Remember to specify the checkpoint_path (with a .pt suffix) and the path for the prediction data (with the header as 'smiles').

šŸ«±šŸ»ā€šŸ«²šŸ¾ Acknowledgements

Thanks for the following released code bases:

chemprop, torchlight, RDKit, KCL

About

Should you have any questions, please feel free to contact Miss Yin Fang at fangyin@zju.edu.cn.

References

If you use or extend our work, please cite the paper as follows:

@article{fang2023knowledge,
  title={Knowledge graph-enhanced molecular contrastive learning with functional prompt},
  author={Fang, Yin and Zhang, Qiang and Zhang, Ningyu and Chen, Zhuo and Zhuang, Xiang and Shao, Xin and Fan, Xiaohui and Chen, Huajun},
  journal={Nature Machine Intelligence},
  pages={1--12},
  year={2023},
  publisher={Nature Publishing Group UK London}
}