Hot2Mol

This is a repository of our paper "Hot-spot-guided generative deep learning for drug-like PPI inhibitor design".

Protein–protein interactions (PPIs) are vital therapeutic targets. However, the large and flat PPI interfaces pose challenges for the development of small-molecule inhibitors. Traditional computer-aided drug design approaches typically rely on pre-existing libraries or expert knowledge, limiting the exploration of novel chemical spaces needed for effective PPI inhibition. To overcome these limitations, we introduce Hot2Mol, a deep learning framework for the de novo design of drug-like, target-specific PPI inhibitors. Hot2Mol generates small molecules by mimicking the pharmacophoric features of hot-spot residues, enabling precise targeting of PPI interfaces without the need for bioactive ligands.

Creating a new environment in conda

Installation typically takes within 5 minutes.

conda env create -f environment.yml
conda activate Hot2Mol

Training

Unzip the data.zip folder to get the training and validation data.

Run train.py using the following command:

python train.py <output_dir> --device cuda:0 --show_progressbar

Use a pre-trained Hot2Mol model to generate molecules

Prepare model weights

Download the model files from the release page. Save to Hot2Mol/pretrained_model folder

Unzip Test_PPI.zip to obtain all testing PPI target structures.

Prepare pharmacophore hypotheses

The input to Hot2Mol should be a pharmacophore hypothesis based on the "hot-spot" residues of a PPI target. The first step is to identify the top three hot-spot residues. This can be achieved using docking methods such as HawkDock, which calculates residue-specific binding energies via MM/GBSA. From these results, the residues contributing most to binding energy can be selected. Alternatively, hot-spot residues can be identified through literature research. For instance, in the MDM2/p53 interaction, the key hot-spots are Leu26, Trp23, and Phe19 on the p53 peptide.

A pharmacophore hypothesis should be provided in .posp format, which includes the type of pharmacophore feature in the first column and spatial coordinates in the last three columns. See 1z92_IL2R.posp in Test_PPI/IL-2:IL-2R/ for an example.

Supported pharmacophore types:

• AROM: aromatic ring
• POSC: cation
• HACC: hydrogen bond acceptor
• HDON: hydrogen bond donor
• HYBL: hydrophobic group (ring)
• LHYBL: hydrophobic group (non-ring)

Build the hypotheses

Use the pharma_extract.py to generate a pharmacophore hypothesis for given hot-spot residues.

usage:

python pharma_extract.py [-h] pdb_file residues output_file

positional arguments:
  pdb_file            Path to the PDB file of the PPI chain that includes the selected hot-spot residues.
  residues            List of residues in the format RESIDUE_NAME RESIDUE_NUMBER (e.g., LYS 7)
  output_file         Path to the output .posp file.

The output is a .posp file containing the pharmacophore hypotesis.

For example, to generate a pharmacophore hypothesis for the demo input (Arg36, Leu42, and Hie120 on IL-2R of the IL-2/IL-2R complex, PDB ID: 1Z92), use the following command:

python pharma_extract.py ./Test_PPI/IL-2:IL-2R/1z92_IL2R.pdb ARG 36 LEU 42 HIE 120 ./Test_PPI/IL-2:IL-2R/1z92_IL2R.posp

Generate

Use generate.py to generate molecules.

usage:

python generate.py [-h] [--n_mol N_MOL] [--device DEVICE] [--filter] [--batch_size BATCH_SIZE] [--seed SEED] input_path output_dir model_path tokenizer_path

positional arguments:
  input_path            the input file path. If it is a directory, then every file ends with `.posp` will be processed
  output_dir            the output directory
  model_path            the weights file (xxx.pth)
  tokenizer_path        the tokenizers (tokenizer_r_iso.pkl, tokenizer_delta_qeppi.pkl)

optional arguments:
  -h, --help            show this help message and exit
  --n_mol N_MOL         number of generated molecules for each pharmacophore file
  --device DEVICE       `cpu` or `cuda`, default:'cpu'
  --filter              whether to save only the unique valid molecules
  --batch_size BATCH_SIZE
  --seed SEED

The output is a .txt file containing the generated SMILES strings. Using a single 3090Ti, it takes approximately 30 seconds to generate 10,000 molecules.

To run generation on the demo input:

python generate.py ./Test_PPI/IL-2:IL-2R/1z92_IL2R.posp results pretrained_model/epoch32.pth pretrained_model --n_mol 1000 --filter --device cuda:0 --seed 123

The current model allows for a maximum of 8 pharmacophore features in a single hypothesis. To increase this limit, you can retrain the model with a greater number of randomly selected pharmacophore elements and a larger MAX_NUM_PP_GRAPHS.

Acknowledgements

This implementation is inspired and partially built on earlier works [1], [2].

References

• [1] Zhu, Huimin, et al. "A pharmacophore-guided deep learning approach for bioactive molecular generation." Nature Communications 14.1 (2023): 6234.

• [2] Yoshida, Shuhei, et al. "Peptide-to-small molecule: a pharmacophore-guided small molecule lead generation strategy from high-affinity macrocyclic peptides." Journal of Medicinal Chemistry 65.15 (2022): 10655-10673.

Citation

@article{
  title={Hot-spot-guided generative deep learning for drug-like PPI inhibitor design},
  author={Sun, Heqi and Li, Jiayi and Zhang, Yufang and Lin, Shenggeng and Chen, Junwei and Tan, Hong and Wang, Ruixuan and Mao, Xueying and Zhao, Jianwei and Li, Rongpei, Li and Wei, Dong-qing},
  year={2025},
  journal={Interdisciplinary Sciences},
  doi={10.1007/s12539-025-00756-w}
}