FSGLD

A Full-Spectrum Generative Lead Discovery (FSGLD) Pipeline via DRUG-GAN: A Multiscale Method for Drug-like/Target-specific Compound Library Generation

Introduction

We present the Full-Spectrum Generative Lead Discovery (FSGLD), a deep learning-driven pipeline for efficient drug lead identification. FSGLD integrates generative modeling with molecular docking, molecular dynamics simulations, ligand-residue interaction profile, MM-PBSA, thermodynamic integration (TI), and experimental validation to bridge theoretical design and practical application. The core multiscale DRUG-GAN models enable de novo design for both drug-like and target-specific compounds across three scenarios: I. generation of random drug-like compounds, II. generation of target-specific compounds, III. generation of target-biased compound series featuring shared chemical structures. FSGLD significantly outperformed traditional computer-aided drug design methods in generating novel chemicals which specifically target the CB2 receptor.

System/software requirements

The source code developed in Python 3.10.8 and Tensorflow 2.10.0. The required python dependencies are given below. FSGLD is supported for cpu/gpu and there is no additional non-standard hardware requirements.

-tensorflow 2.10.0
(-tensorflow-gpu 2.10.0 #if gpu available
-cudatoolkit 11.2.2
-cudnn 8.4)
-matplotlib 3.10.1
-pandas 2.2.3
-numpy 1.26.4
-scikit-learn 1.6.1

Optional: You may need to install openbabel (https://openbabel.org/docs/Installation/install.html) to convert searched molecules (.sdf) to FP2 format to do the second-round similarity search.

Datasets

The training set for the discriminator is located in discriminator/maccs.json. We used 10-fold cross validation to validate the discriminator model for the determination of the optimal architecture.
The training sets for the dcGAN/cdcGAN models in the three scenarios are located in:

• DCGAN_s1/data (Scenario I: generic drug-like compounds)

• CDCGAN_s2/data (Scenario II: CB2-specific ligands)

• CDCGAN_s3/data (Scenario III: CB2 compound series)

  Each dataset is provided in structured CSV format with explicit fields (compound ID and MACCS fingerprints). For the discriminator, the training set is provided in JSON format. All compound records were curated from ChEMBL and ZINC database, and subsequently cleaned and reformatted into the CSV/JSON files provided here.

  For similarity search, due to size limitations, we only provide a template.bin file for demonstration, located in similarity_search/s1_s2 and similarity_search/s3.

Model training and molecular generation

To train the discriminator, we provide our source code for each of the discriminator architecture for users to reproduce the results (include the generation of ROC curves). You can easily run .ipynb files to reproduce the results.

DCGAN_s1, CDCGAN_s2 and CDCGAN_s3 respectively stand for generative model in Scenarios I, II and III. To train dcGAN/CDCGAN, generate and evaluate drug-like molucules/CB2 compounds/CB2 compound series, you can run the following command in all three folders:

$ python train.py
$ python generate.py

Note:
We also provide pre-trained generators, so users can skip training and directly reproduce the generation step. The pre-trained weights are located at: DCGAN_s1/bestmodel/saved_best_model.h5, CDCGAN_s2/bestmodel/saved_best_model.h5, CDCGAN_s3/bestmodel/saved_best_model.h5.

The evaluation metrics for the generated 10,000 samples will be printed after molecular generation, including uniqueness, diversity, novelty, average similarity and maximal similarity.

Similarity search

In our work, we calculated tanimoto similarity between our generated fingerprints (FPs) and compounds in ChEMBL/ZINC library. In similarity_search folder, we provide csh and ELF code for you to convert MACCS FPs to FP2 FPs and conducted similarity search. tanimoto_ss is used for simple similarity search, while tanimoto_ss_mcs is used for similarity search fixed with MCS features. We provide MCS information for all mcs i, mcs ii and mcs iii, which are located in similarity_search/mcs. For the searched ChEMBL/ZINC compounds using generated samples in MACCS format, you can first use openbabel to convert searched compounds (sdf format) to FP2 in hex format, then used our provided hex2bin2 to convert hex format to binary format. Take similarity search for molecules with MCS ii features as an example (similarity_search/s3)

1. Similarity search in similarity_search/s3/first_round

$ ./gen_bat

2. Analyze the searched molecules in output/ and selected top ones. Retrive ChEMBL/ZINC sdf files, Use openbabel to convert sdf to FP2 FPs.
3. Convert MACCS to FP2:

$ ./hex2bin2 -i maccs.bin -o fp2.bin

4. Similarity search (./gen_bat) for the second round.

Example notebook

For convenience, we provide Jupyter notebooks example_inference.ipynb (located in DCGAN_s1/, CDCGAN_s2/, CDCGAN_s3/) in each scenario folder that demonstrate the full workflow from model training to molecular generation:

Users can directly run these notebooks to:

• train the discriminator and generator, or skip training by loading the provided pre-trained weights

• generate molecular fingerprints

• compute evaluation metrics (uniqueness, diversity, novelty, similarity)

• save the generated samples to outputs/generated_samples.csv.

  These notebooks serve as a quick start and benchmarking examples to facilitate reproducibility.