This is the github repo for the paper "In-silico 3D Molecular Editing through Physics-Informed and Peference-Aligned Generative Foundation Models". An early version is preprinted at doi:10.26434/chemrxiv-2023-j2n6l-v2.

MolEdit, developed by the Gao Group at Peking University, is a molecular generative artificial intelligence (GenAI) designed to address the complex challenges of molecular editing. This fundamental aspect of functional molecule design involves the generation, modification, and evolution of molecules with specified structural and chemical features. Our goal is to develop molecular GenAIs as powerful as image-based GenAIs, such as image denoising diffusion probabilistic models (DDPMs).

Using our approach, MolEdit scales 3D molecular generation from QM9 (≤9 heavy atoms, small molecules) to ZINC (≤64 heavy atoms, drug-like), and further to QMugs (≤100 heavy atoms, bioactive).

We propose an asynchronous multimodal diffusion schedule for molecular DDPMs. This approach makes it possible to rigorously handle molecular symmetries (various point groups), and facilitates robust generation across various molecular modalities, including SMILES, graphs, and 3D structures.

We develop Group-Optimized Denoising Diffusion (GODD), that explicitly integrates group actions into DDPMs. GODD functions through group-optimized labeling, which serves as a plug-and-play and non-invasive modification to DDPM training objectives. GODD enhances the model's awareness of the diverse molecular symmetries embedded in various molecular point groups.

To suppress hallucinations of physically unstable and chemically invalid structures (atom clashes, inappropriate bond lengths, and incorrect valences) generated by molecular GenAIs, we draw inspiration from preference alignment techniques used in mainstream GenAIs. We implement a Boltzmann-Gaussian mixture (BGM) kernel for physics alignment. This approach, grounded in statistical mechanics and molecular force fields, built upon a "Hamiltonian oracle" and significantly improves the stability of structures in the molecules generated.

Inspired by the success of image GenAIs, a foundational molecular GenAI should be able to respond to various specifications of molecular editing. Therefore, we equip MolEdit with flexible controls in context of multiple conditions. For instance, MolEdit is able to control the shape of molecules using the radius of gyration. The model can also be conditioned on (sub)molecular graphs and "inpaint" the missing motifs given predefined chemical contexts like fragments, groups, etc., and/or given geometric restraints like substructures, orientations, etc. This controllable generation strategy also facilitates MolEdit to render diverse high-quality structures from textual molecular representations like SMILES and graphs.

Detailed examples can be found in notebook moledit_ZINC.ipynb!

Before running notebooks, download and unzip pre-trained MolEdit checkpoints (checkpoints.zip) from here, and move folder moledit_dataset and folder params into your working directory.

Explore MolEdit's capabilities with:

• moledit_qm9.ipynb notebook, which contains demos of generation and editing of molecules (less than 9 non-hydrogen atoms), and physics-alignment via MolEdit trained on QM9 dataset.
• moledit_QMugs.ipynb notebook, which contains demos of property-guided sampling via MolEdit trained on QMugs dataset
• moledit_ZINC.ipynb notebook, which contains demos of de-novo design of drug-like molecules, and following applications via MolEdit trained on ZINC dataset.
  • High quality structure rendering and diverse conformational sampling for complex molecules
  • Structure editing and inpainting

    

  • Re-design of aromatic systems

    

  • Functional-core based design

    

  • Linker design

    

  • Structure building for free energy perturbation
  • Lead-imprinted binder design for protein pockets
  • ...

Running these notebooks requires:

• python==3.10
• jax==0.4.20, jaxlib==0.4.20
• flax==0.8.3
• e3nn_jax==0.20.6
• rdkit==2023.9.6
• ml_collections
• Xponge==1.5.0a6, installed via pip install git+https://gitee.com/gao_hyp_xyj_admin/xponge.git
• nglview==3.1.2 (for visualization in notebooks)

In theory, any environment compatible with the packages mentioned above should run successfully. Our configuration includes Ubuntu 22.04 (GNU/Linux x86_64), NVIDIA A100-SXM4-80GB, CUDA 11.8 and Anaconda 23.7.2. The complete notebook execution takes approximately 0.5 hours.

Details on calculating quantitative metrics, including validity, uniqueness, molecular physical stability, and conformational diversity, can be found in evaluation. Additionally, we provide useful scripts to evaluate energy/force under the General AMBER force field (without periodic boundary conditions). We also provide scripts used in quantitative benchmark study (linker design & binder design).

We provide an example training script for implementing MolEdit. To train your model:

1. Save each data point as individual .pkl files
2. Create a name_list.pkl file containing a dictionary with the following structure:

{
    (0, 8): {
        "size": ...,          # [description of this field]
        "name_list": [pkl_file_1, pkl_file_2, ...]
    }, 
    (8, 16): {
        "size": ...,          # [description of this field]
        "name_list": [pkl_file_1, pkl_file_2, ...]
    },
    ...
}

• Dictionary keys: Tuple ranges indicate atom count intervals (e.g., (0,8) for molecules with 0-8 atoms)
• Value structure:
  • size: [brief explanation of what this represents]
  • name_list: List of paths to your .pkl files

Each .pkl file should contain:

{
    'feature': {
        mask_type: {
            'n_atoms': ...,       # Number of atoms
            'atom_feat': ...,     # Atom features [describe format if needed]
            'bond_feat': ...      # Bond features [describe format if needed]
        },
        ...
    },
    'structure': {
        mask_type: {
            'rg': ...,             # Radius of gyration
            'atom_crd': ...,       # Ground truth 3D coordinates
            'perm_transrot_crd': ...,  # Group-optimized noisy structure
            'perm_transrot_label': ... # Group-optimized labels
        },
        ...
    }
}

You may alternatively implement your own data pipeline if this format doesn't suit your needs.

The dataset used in the development of MolEdit can be found in here.

@article{lin2023versatile,
    title={Versatile Molecular Editing via Multimodal and Group-optimized Generative Learning},
    author={Lin, Xiaohan and Xia, Yijie and Huang, Yupeng and Liu, Shuo and Zhang, Jun and Gao, Yi Qin and Zhang, Jun},
    year={2023},
}

For questions or further information, please contact jzhang@cpl.ac.cn.