RetroExplainer

Overview

RetroExplainer: A chemical knowledge guided deep-learning framework for retrosynthesis prediction with molecular assembly reasoning and quantitative interpretability

Installation

Download the repository

git clone git@github.com:wangyu-sd/RetroExplainer
cd RetroExplainer

Install required packages

We recommend to use anaconda to get the dependencies. If you don't already have anaconda, install it by following instructions at this link: https://docs.anaconda.com/anaconda/install/.

Then, open the terminal on Linux and enter the following commands to create a new conda environment and install the required packages:

conda create -n mechretro --file requirement.txt
conda activate mechretro

If the previous commands don't work, you can install the packages separately:

1. Create a new virtual environment with a 3.7.13 version of python:

   conda create -n RetroExplainer python=3.7.13
   conda activate RetroExplainer

2. Install pytroch with correct CUDA version. To find your suitable version, see https://pytorch.org/get-started/locally/

   conda install pytorch torchvision torchaudio pytorch-cuda=11.6 -c pytorch -c nvidia

3. Install pytorch-lightning:

   conda install pytorch-lightning -c conda-forge

4. Install torch-geometric and its corresponding dependencies, see https://pytorch-geometric.readthedocs.io/en/latest/notes/installation.html.

   conda install pyg -c pyg

5. Install rdkit:

    conda install -c conda-forge rdkit

Other packages can be easily installed by calling pip install xxx command.

Model Training

The USPTO-50K datasets can be download here , and should be put under the folders /data .

You can run the training stages by following command for reaction type known conditions:

sh scripts/train_for_uspto50k.sh

And for reaction type unknown:

sh scripts/train_for_uspto50k_rxn_type_unknown.sh

As for additional datasets, e.g., USPTO-50K using t-split method, USPTO-FULL, USPTO-MIT , and your customized datasets own datasets, you can put them in the directory data/your_datasets/raw , and set the argument to data/your_datasets .

The above mentioned dataset also can be obtained in the following entries:

• USPTO-50K (random split): https://github.com/Hanjun-Dai/GLN ,
• USPTO-50K (Using t-split method): https://drive.google.com/file/d/1_cko4Wym0W_UaRp74KvbEJCrfQBfrVKQ/view?usp=sharing
• USPTO-MIT : https://github.com/wengong-jin/nips17-rexgen/blob/master/USPTO/data.zip,
• USPTO-FULL: https://github.com/Hanjun-Dai/GLN.

Re-ranking

The data correspond to re-ranking experiments can be found in https://drive.google.com/file/d/1SMK1P2IXcWh0T2KHxresxu6_dKkgA7jd/view?usp=sharing.

Reproduce results

Model performance

At first, you can download the checkpoints here for reaction-type-known model (USPTO-50K) and here for unknown model. Or you can train your own model according to the section Model Training .

Then, put the checkpoints under the model_saved folder, and run the following command:

sh scripts/test_for_uspto50k.sh

Similarly, for reaction type unknown conditions, you can:

sh scripts/test_for_uspto50k_rxn_type_unknown.sh

Other check points can be found in the following URLs and the reproduction can be easily run upon modifying the corresponding commands of the scripts.

• USPTO-50K (Using t-split method): https://drive.google.com/file/d/1_cko4Wym0W_UaRp74HJHUSrVKQ/view?usp=drive_link
• USPTO-MIT : https://drive.google.com/file/d/1lSy1zbrzMgQypEft5f9_4s3nvWbL5mx7/view?usp=sharing
• USPTO-FULL: https://drive.google.com/file/d/1iV40xTstPHSRCzNiYLXX_zird1I1Q6tN/view?usp=sharing

Re-ranking ability

Checkpoint of re-ranking model can be downloaded here: https://drive.google.com/file/d/1a7loRClJtF2t8Vd-4wN_-2HqM_wCvI18/view?usp=sharing.

The results can be reproduced through the following commands:

python reranking_and_explaining.py \
  --batch_size 64 \
  --cuda 2 \
  --dataset GLN_200topk_200maxk_noGT_77777777_test

Interpretability

We introduced an energy-based molecular assembly process that offers transparent decision-making and interpretable retrosynthesis predictions. This process can generate an energy decision curve that breaks down predictions into multiple stages and allows substructure-level attributions; the former can help understand the "counterfactual" predictions to discover potential biases in the dataset, and the latter can provide more granular references (such as the confidence of a certain chemical bond being broken) to inspire researchers to design customized reactants.

Generated explanations through molecular assembly process decision process. a. The decision process of two predictions, including reactions with and without leaving groups. b. The decision curve of top-12 predictions by RetroExplainer. The same reaction patterns have the same energy change. c-h. Nine representative instances for substructure attributions, which allows a granular insight. i-k. Performance on re-ranking results compared with three effective retrosynthesis models.

Interpretable Multi-Step Planing

For multi-step predictions, you can click here to get purchasable molecule set, and put it under the folder /data/multi-step/retro_data/dataset/.

We provide a more robust checkpoint trained on larger scale datasets, which takes about a week on three 3090 GPUs. You can download the checkpoint here. And don't forget put it under the folder model_saved. Note that this version need a larger vocabulary of leaving groups, you can click here to get the vocabulary. And put the leaving_group.pt under the folder data/multi-step/.

You can get the reasoning process and their energy scores by modifying smi_list in api_for_multisetp.py:

if __name__ == '__main__':
    planner = RSPlanner(
        gpu=-1,
        use_value_fn=False,
        iterations=500,
        expansion_topk=10,
        viz=True,
        viz_dir='data/multi-step/viz'
    )
    smi_list = [
        "CC(CC1=CC=C2OCOC2=C1)NCC(O)C1=CC=C(O)C(O)=C1"
    ]
    for smi in smi_list:
        result = planner.plan(smi, need_action=True)
        print(result)

Then you can get following retrosynthetic routes:

{'succ': True, 'time': 48.523579359054565, 'iter': 18, 'routes': "CC(CC1=CC=C2OCOC2=C1)NCC(O)C1=CC=C(O)C(O)=C1>3.2358>NCC(O)c1ccc(O)c(O)c1.CC(=O)Cc1ccc2c(c1)OCO2|NCC(O)c1ccc(O)c(O)c1>7.8090>NC(=O)C(O)c1ccc(O)c(O)c1|CC(=O)Cc1ccc2c(c1)OCO2>10.1133>CC(O)Cc1ccc2c(c1)OCO2|NC(=O)C(O)c1ccc(O)c(O)c1>9.0024>N.O=C(O)C(O)c1ccc(O)c(O)c1[['Select Leaving Group with Index 6 and Cost 0.43', 'Initial Cost:10.37', 'Add Bonds: between 1 and 24 with Bond Type 2.0 and Cost -6.05', 'Remove Bonds: between 1 and 12, with Bond Type 1.0 and Cost -1.35', 'H number change -1 cost of atom 2: 0.18', 'H number change 1 cost of atom 13: 0.09'], ['Select Leaving Group with Index 6 and Cost 3.49', 'Initial Cost:7.66', 'Add Bonds: between 1 and 12 with Bond Type 2.0 and Cost -1.01', 'H number change -2 cost of atom 2: 1.16'], ['Select Leaving Group with Index 20 and Cost 2.84', 'Initial Cost:7.17', 'Replace Bonds: between 1 and 2, from Bond Type 2.0 to Bond Type1 with Cost -0.52', 'H number change 1 cost of atom 2: 2.22', 'H number change 1 cost of atom 3: 1.24'], ['Select Leaving Group with Index 3 and Cost 2.58', 'Initial Cost:11.32', 'Add Bonds: between 1 and 13 with Bond Type 1.0 and Cost -2.01', 'Remove Bonds: between 0 and 1, with Bond Type 1.0 and Cost -1.42', 'H number change 1 cost of atom 1: 1.12'], None, None, None]", 'route_cost': 30.16051983833313, 'route_len': 4}