Valid, Plausible, and Diverse Retrosynthesis Using Tied Two-way Transformers with Latent Variables

This repository is an implementation of "Valid, Plausible, and Diverse Retrosynthesis Using Tied Two-way Transformers with Latent Variables" submitted to the Journal of Chemical Information and Modeling. The code is built on top of OpenNMT-py.

Requirements

• Python 3.6
• torch 1.4
• torchtext 0.4.0
• configargparse
• tqdm

Preprocess

The USPTO-50k dataset is located in ./data/. To train the model, the dataset needs to be preprocessed.

DATA=data

TRAIN_SRC=$DATA/src-train.txt
VALID_SRC=$DATA/src-val.txt
TRAIN_TGT=$DATA/tgt-train.txt
VALID_TGT=$DATA/tgt-val.txt

DATA_PREFIX=data/USPTO-50k_no_rxn_processed
python preprocess.py \
    -train_src $TRAIN_SRC \
    -train_tgt $TRAIN_TGT \
    -valid_src $VALID_SRC \
    -valid_tgt $VALID_TGT \
    -save_data $DATA_PREFIX \
    -share_vocab -overwrite

Train

Train the model using the preprocessed dataset. In the paper, we tested the dimension of multinomial latent variable for 1, 2, and 5. In the code below, the dimension, num_experts, is set to 1.

OUT="onmt-runs"
[ -d $OUT ] || mkdir -p $OUT
[ -d $OUT/model ] || mkdir -p $OUT/model

python train.py -data $DATA_PREFIX \
    -save_model $OUT/model/model -train_steps 500000 \
    -save_checkpoint_steps 5000 -keep_checkpoint 11 \
    -valid_step 5000 -report_every 5000 \
    -batch_size 4096 -batch_type tokens -normalization tokens \
    -max_relative_positions 4 -share_relative_pos_embeddings \
    -layers 6 -rnn_size 256 -word_vec_size 256 -heads 8 -transformer_ff 2048 \
    -num_experts 1 -position_encoding -share_embeddings \
    -dropout 0.3 -max_generator_batches 0 -early_stopping 40 \
    -gpu_ranks 0 -world_size 1 -accum_count 4 \
    -max_grad_norm 0 -optim adam -adam_beta1 0.9 -adam_beta2 0.998 \
    -decay_method noam -warmup_steps 8000 -learning_rate 2 \
    -param_init 0 -param_init_glorot -seed 2020 \
    2>&1 | tee -a $OUT/model/train.log

After finishing the training, we averaged the last 5 models. For example, below is the code for averaging 5 models.

models="$OUT/model/model_step_20000.pt \
        $OUT/model/model_step_25000.pt \
        $OUT/model/model_step_30000.pt \
        $OUT/model/model_step_35000.pt \
        $OUT/model/model_step_40000.pt" 
MODEL="$OUT/model/model_step_40000_avg5.pt"

python average_models.py -models $models -output $MODEL

Inference & Evaluation

You can use the code below to evaluate the model for a given test dataset.

DATA=data
TEST_SRC=$DATA/src-test.txt
TEST_TGT=$DATA/tgt-test.txt

MODEL="$OUT/model/model_step_*.pt"

TRANSLATE_OUT=$OUT/model/test/step_*
[ -d $TRANSLATE_OUT ] || mkdir -p $TRANSLATE_OUT

python translate.py -model $MODEL \
    -src $TEST_SRC -tgt $TEST_TGT \
    -output $TRANSLATE_OUT \
    -beam_size 20 -n_best 10 \
    -max_length 200 \
    -num_experts 1 \
    -batch_size 128 \
    -replace_unk -gpu 0

Reproducing the results & Choosing the number of latent dimension

For those who want to reproduce the results in the paper, we have released one of the trained models which can be downloaded from https://www.dropbox.com/s/57wdw8a937ruvwn/model_L2.pt?dl=0. This model is trained with the latent dimension of 2. After downloading the model in the ./onmt-runs/ directory, try the code below. Note that the num_experts option is set to 2 since the latent dimension of the model is 2. As the dimension increases, the diversity in results improved (please refer the paper for details). However, there is a trade-off between the dimension and speed. So, we have tested the dimension for 1, 2, and 5.

DATA=data
TEST_SRC=$DATA/src-test.txt
TEST_TGT=$DATA/tgt-test.txt

OUT="onmt-runs"
MODEL="$OUT/USPTO_L2.pt"

TRANSLATE_==$OUT/USPTO_L2_results
[ -d $TRANSLATE_OUT ] || mkdir -p $TRANSLATE_OUT

python translate.py -model $MODEL \
    -src $TEST_SRC -tgt $TEST_TGT \
    -output $TRANSLATE_OUT \
    -beam_size 10 -n_best 10 \
    -max_length 200 \
    -num_experts 2 \
    -batch_size 128 \
    -replace_unk -gpu 0

The translation and evaluation results can be found in onmt-runs/USPTO_L2_results/. Open onmt-runs/USPTO_L2_results/pred_cycle_lp2.txt.score to check the evaluation results. The top 10 rows show k (first column), top-k accuarcy (second column), and invalid SMILES rates (third column) while the last row is unique rate.

1,46.7226,0.059952
2,61.0312,0.069944
3,67.9856,0.113243
4,71.7426,0.194844
5,74.0008,0.335731
6,75.4396,0.516254
7,76.6787,0.845038
8,77.2982,1.46133
9,77.8377,2.62457
10,78.1775,4.60631
89.8321