[Paper Review] Target Specific De Novo Design of Drug Candidate Molecules with Graph Transformer-based Generative Adversarial Networks

Atabey Ünlü, Elif Çevrim+9 more|arXiv (Cornell University)|Feb 15, 2023

Computational Drug Discovery Methods4 citations

TL;DR

This paper introduces DrugGEN, a graph transformer-based generative adversarial network for target-specific de novo drug design, representing molecules and proteins as graphs to generate novel, potent AKT1 inhibitors. The system achieves competitive performance on benchmarks and demonstrates high binding affinity and drug-likeness in in silico validation, with all models and data publicly shared for reproducibility and further development.

ABSTRACT

Discovering novel drug candidate molecules is one of the most fundamental and critical steps in drug development. Generative deep learning models, which create synthetic data given a probability distribution, offer a high potential for designing de novo molecules. However, to be utilisable in real life drug development pipelines, these models should be able to design drug like and target centric molecules. In this study, we propose an end to end generative system, DrugGEN, for the de novo design of drug candidate molecules that interact with intended target proteins. The proposed method represents molecules as graphs and processes them via a generative adversarial network comprising graph transformer layers. The system is trained using a large dataset of drug like compounds and target specific bioactive molecules to design effective inhibitory molecules against the AKT1 protein, which is critically important in developing treatments for various types of cancer. We conducted molecular docking and dynamics to assess the target centric generation performance of the model, as well as attention score visualisation to examine model interpretability. In parallel, selected compounds were chemically synthesised and evaluated in the context of in vitro enzymatic assays, which identified two bioactive molecules that inhibited AKT1 at low micromolar concentrations. These results indicate that DrugGEN's de novo molecules have a high potential for interacting with the AKT1 protein at the level of its native ligands. Using the open access DrugGEN codebase, it is possible to easily train models for other druggable proteins, given a dataset of experimentally known bioactive molecules.

Motivation & Objective

• To address the challenge of designing novel, structurally diverse drug candidates for understudied or undruggable targets.
• To overcome limitations in traditional drug development, including high failure rates due to toxicity and low efficacy.
• To develop a generative model that produces target-specific molecules without relying on structural templates.
• To integrate protein structure and compound graph representations for accurate, biologically relevant molecule generation.
• To provide a reproducible, open-source framework for community use in target-centric drug discovery.

Proposed method

• DrugGEN employs two serially connected generative adversarial networks (GANs) with graph transformer architectures to process molecular and protein structures as graphs.
• Molecules and target proteins are encoded as molecular graphs using SMILES and PDB-based representations, respectively.
• The generator network produces new molecular structures conditioned on target protein features, while the discriminator evaluates their realism and target-specificity.
• The model is trained on a large dataset combining ChEMBL compounds and target-specific bioactive molecules, focusing on AKT1 inhibition.
• Five independent model variants are trained with unique sampling routines to enhance diversity and robustness in molecule generation.
• Post-generation analysis includes molecular docking, deep learning-based bioactivity prediction (DEEPScreen), and physicochemical property evaluation (QED, SA, logP).

Experimental results

Research questions

• RQ1 Can a graph-based GAN with transformer architecture effectively generate novel, target-specific drug candidates for a given protein target?
• RQ2 How does DrugGEN's performance compare to state-of-the-art methods in generating molecules with desired physicochemical and biological properties?
• RQ3 To what extent do de novo generated molecules bind effectively to the target protein's active site, as predicted by molecular docking?
• RQ4 Can the generated molecules achieve binding affinities comparable to known inhibitors of AKT1?
• RQ5 How well do the generated molecules maintain drug-likeness and synthetic accessibility metrics?

Key findings

• DrugGEN models achieve competitive or superior performance on standard benchmarks compared to existing state-of-the-art methods in de novo molecule generation.
• The generated molecules exhibit high binding affinity to the AKT1 protein, with predicted binding free energies below -9 kcal/mol, indicating strong potential for interaction.
• Molecular docking results show that top-ranked de novo molecules bind effectively to the AKT1 active site, with poses closely resembling the native ligand's binding mode.
• The generated molecules maintain favorable drug-likeness properties, with QED, synthetic accessibility (SA), and logP values comparable to known AKT1 inhibitors.
• In silico analysis using DEEPScreen confirms high predicted bioactivity for the generated compounds, supporting their potential as effective inhibitors.
• The complete codebase, datasets, trained models, and results are publicly available at https://github.com/HUBioDataLab/DrugGEN for community use and reproducibility.