Official Implementation of BounDr.E in Python

This repository contains the implementation of our paper, BounDr.E: Predicting Drug-likeness via Biomedical Knowledge Alignment and EM-like One-Class Boundary Optimization: a novel framework for drug-likeness boundary prediction. Our approach integrates multi-modal alignment between structural and knowledge-based embeddings to enhance the identification of drug-like compounds. [ICLR 2025 MLGenX Workshop, ICML 2025]

Drug-likeness prediction dataset

The benchmark data, data splits, and external validation sets are provided in benchmark folder.

Model description

Our full framework is two-step;

1. Multi-modal Mixup of drugs: The first step aligns the biomedical knowledge graph embedding vectors and Morgan fingerprint vectors into a unified space.
2. EM-like boundary construction: Then, using the structural encoder trained in step 1, we encode all the compounds and drugs into the unified space, then construct the 'drug-likeness boundary' through an EM-like optimization.

Requirements

To run the code, you will need the following dependencies:

• pytorch==1.8.1 or above
• numpy==1.24.1
• pandas==2.0.1
• rdkit==2022.9.5 or above

You can install the dependencies using:

pip install -r requirements.txt

Inference: predicting drug-likeness

You can perform drug-likeness inference by running the following command. The outputs indicate the relative distance of the input from the center of the drug boundary, divided by the boundary radius (inference value > 1 means the compound is outside the drug boundary).

python score.py --input_smiles test.smi --output output.csv \
                --project_dir [project_dir] --device cuda:0

Input Descriptions:
--input_smiles: SMILES format file for compounds to be evaluated.
--project_dir: directory to save and load checkpoint files under projects folder.
--output: name of outputfile

Demo files are provided in the demo folder. You can perform test runs using the demo file (test.smi) with:

python score.py --input_smiles demo/test.smi

which will save results in projects/test_run/output.csv.

Model Training

Here, we train the multi-modal alignment encoder and the drug-like boundary encoders and save the checkpoint .pt files in the project directory.

Step 1. Knowledge-integrated multi-modal alignment

To train the model with knowledge-guided multi-modal alignment, use the following command:

python train_multimodal_alignment.py --input_fingerprint drugs_fp.npy --input_knowledgegraph drugs_kg.npy  \
                                    --input_similarity ATC_sim.npy --project_dir [project_dir] \
                                    --device cuda:0 --seed 42

Input Descriptions:
--input_smiles: SMILES format file for compounds to be evaluated.
--input_fingerprint: Numpy file containing structural fingerprint embeddings.
--input_knowledgegraph: Numpy file containing biomedical knowledge graph embeddings.
--input_similarity: Numpy file containing ATC code similarity matrix for knowledge alignment.
--project_dir: directory to save and load checkpoint files under projects folder.
--device: GPU index/cpu to run the code on
--seed: Random seed, default 42

Step 2. Drug-Like Boundary Optimization

Using the structural encoder trained in step 1, we encode the structural representation (Morgan fingerprint) into the unified space then perform the EM-like optimization of drug-likeness boundary.

python train_boundary.py --train_drugs train_drugs.smi --train_compoonds train_comps.smi \
                        --valid_drugs valid_drugs.smi --valid_compounds valid_compounds.smi \
                        --test_drugs test_drugs.smi --test_compounds test_compounds.smi \
                        --project_dir [project_dir] --device cuda:0 --seed 42

Input Descriptions:
--train_drugs: SMILES format file for drugs in trainset
--train_comps: SMILES format file for compounds (non-drugs) in trainset
--valid_drugs: SMILES format file for drugs in validset
--valid_comps: SMILES format file for compounds (non-drugs) in validset
--test_drugs: SMILES format file for drugs in testset
--test_comps: SMILES format file for compounds (non-drugs) in testset
--project_dir: directory to save and load checkpoint files under projects folder.

Demo files are provided in the demo folder. You can perform test runs using the demo files with:

python train_boundary.py 

License

This project is licensed under the GPL-3.0 License.

Citation

@inproceedings{bang2025predicting,
  title={Predicting Drug-likeness via Biomedical Knowledge Alignment and EM-like One-Class Boundary Optimization},
  author={Bang, Dongmin and Sung, Inyoung and Piao, Yinhua and Lee, Sangseon and Kim, Sun},
  booktitle={ICLR 2025 Workshop on Machine Learning for Genomics Explorations}
}