BatmanNet : Bi-branch Masked Graph Transformer Autoencoder for Molecular Representation

Requirements

• RDKit : Tested on 2019.09.3
• Python : Tested on 3.8.13
• PyTorch : Tested on 1.9.1

To install RDKit, please follow the instructions here http://www.rdkit.org/docs/Install.html

We highly recommend you to use conda for package management.

For the other packages, please refer to the requirements.txt.

Quick Start

Pretraining

Pretrain BatmanNet given the unlabelled molecular data.
Note: There are more hyper-parameters which can be tuned during pretraining. Please refer to add_pretrain_args/ in BatmanNet/util/parsing.py/.

python preprocess.py  --data_path data/zinc/all.txt \
                      --output_path data/processed 
                      

python main.py pretrain \
               --data_path data/processed \
               --save_dir model/pretrained_model/ \
               --batch_size 32 \
               --hidden_size 100
               --dropout 0.1 \
               --depth 3 \
               --num_attn_head 2 \
               --num_enc_mt_block 6\
               --num_dec_mt_block 2\
               --epochs 15 \
               --init_lr 0.0002 \
               --max_lr 0.0004 \
               --final_lr 0.0001 \
               --weight_decay 0.0000001 \
               --activation PReLU 

Finetuning

Finetuning on Molecular Property Prediction Datasets

Molecular Feature Extraction

python scripts/save_features.py --data_path data/finetune/bbbp.csv \
                                --save_path data/finetune/bbbp.npz \
                                --features_generator rdkit_2d_normalized \
                                --restart 

Note: There are more hyper-parameters which can be tuned during finetuning. Please refer to add_finetune_args inBatmanNet/util/parsing.py .

python main.py finetune --data_path data/finetune/bbbp.csv \
                        --features_path data/finetune/bbbp.npz \
                        --save_dir model/finetune/bbbp/ \
                        --checkpoint_path pretrained_model/model/model.ep15 \
                        --dataset_type classification \
                        --split_type scaffold_balanced \
                        --ensemble_size 1 \
                        --num_folds 3 \
                        --ffn_hidden_size 200 \
                        --ffn_num_layer 2\
                        --batch_size 32 \
                        --epochs 100 \
                        --init_lr 0.00015

The final finetuned model is stored in model/bbbp and will be used in the subsequent prediction and evaluation tasks.

Finetuning on DDI Prediction Datasets

python DDI/finetune_snap.py --parser_name ddi \
                            --data_path data/biosnap/raw/all.csv \
                            --save_dir model/biosnap/ \
                            --checkpoint_path pretrained_model/model/model.ep15 \
                            --dataset biosnap \
                            --ffn_hidden_size 200 \
                            --ffn_num_layer 2\
                            --batch_size 32 \
                            --epochs 100 \
                            --init_lr 0.00015

The final finetuned model is stored in DDI/runs

Finetuning on DTI Prediction Datasets

python DTI/cross_validate_human.py  --parser_name dti \
                                    --data_path data/human/raw/data.txt \
                                    --save_dir model/human/ \
                                    --checkpoint_path pretrained_model/model/model.ep15 \
                                    --dataset human \
                                    --model human \
                                    --ffn_hidden_size 200 \
                                    --ffn_num_layer 2 \
                                    --num_folds 3 \
                                    --num_iters 3 \
                                    --batch_size 8 \
                                    --epochs 30 \
                                    --lr 0.00005 \
                                    --hid 32 \
                                    --heads 4 \
                                    --deep 1 \
                                    --dropout 0.2 
                                    

The final finetuned model is stored in DTI/test