Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

3DMolFormer: A Dual-channel Framework for Structure-based Drug Discovery

Xiuyuan Hu, Guoqing Liu, Can Chen, Yang Zhao, Hao Zhang, Xue Liu

TL;DR

3DMolFormer introduces a unified dual-channel transformer to address two core structure-based drug discovery tasks—protein-ligand docking and pocket-aware 3D drug design—by representing pocket-ligand complexes as parallel token and coordinate sequences. The model uses a GPT-2–style token channel coupled with a numerical channel, enabling autoregressive generation and learning from large-scale 3D data through a composite token-CE and coordinate-MSE objective, followed by task-specific fine-tuning: supervised docking on PDBbind and reinforcement learning–based design guided by multi-objective rewards. Empirically, it outperforms traditional docking baselines and recent 3D drug-design methods, achieving higher pose accuracy and better multi-objective design metrics, with notable advantages in robustness and inference speed. This work demonstrates the feasibility and value of a single architecture that leverages dual-task duality to accelerate and improve structure-based drug discovery, offering practical potential for high-throughput screening and targeted drug design.

Abstract

Structure-based drug discovery, encompassing the tasks of protein-ligand docking and pocket-aware 3D drug design, represents a core challenge in drug discovery. However, no existing work can deal with both tasks to effectively leverage the duality between them, and current methods for each task are hindered by challenges in modeling 3D information and the limitations of available data. To address these issues, we propose 3DMolFormer, a unified dual-channel transformer-based framework applicable to both docking and 3D drug design tasks, which exploits their duality by utilizing docking functionalities within the drug design process. Specifically, we represent 3D pocket-ligand complexes using parallel sequences of discrete tokens and continuous numbers, and we design a corresponding dual-channel transformer model to handle this format, thereby overcoming the challenges of 3D information modeling. Additionally, we alleviate data limitations through large-scale pre-training on a mixed dataset, followed by supervised and reinforcement learning fine-tuning techniques respectively tailored for the two tasks. Experimental results demonstrate that 3DMolFormer outperforms previous approaches in both protein-ligand docking and pocket-aware 3D drug design, highlighting its promising application in structure-based drug discovery. The code is available at: https://github.com/HXYfighter/3DMolFormer .

3DMolFormer: A Dual-channel Framework for Structure-based Drug Discovery

TL;DR

3DMolFormer introduces a unified dual-channel transformer to address two core structure-based drug discovery tasks—protein-ligand docking and pocket-aware 3D drug design—by representing pocket-ligand complexes as parallel token and coordinate sequences. The model uses a GPT-2–style token channel coupled with a numerical channel, enabling autoregressive generation and learning from large-scale 3D data through a composite token-CE and coordinate-MSE objective, followed by task-specific fine-tuning: supervised docking on PDBbind and reinforcement learning–based design guided by multi-objective rewards. Empirically, it outperforms traditional docking baselines and recent 3D drug-design methods, achieving higher pose accuracy and better multi-objective design metrics, with notable advantages in robustness and inference speed. This work demonstrates the feasibility and value of a single architecture that leverages dual-task duality to accelerate and improve structure-based drug discovery, offering practical potential for high-throughput screening and targeted drug design.

Abstract

Structure-based drug discovery, encompassing the tasks of protein-ligand docking and pocket-aware 3D drug design, represents a core challenge in drug discovery. However, no existing work can deal with both tasks to effectively leverage the duality between them, and current methods for each task are hindered by challenges in modeling 3D information and the limitations of available data. To address these issues, we propose 3DMolFormer, a unified dual-channel transformer-based framework applicable to both docking and 3D drug design tasks, which exploits their duality by utilizing docking functionalities within the drug design process. Specifically, we represent 3D pocket-ligand complexes using parallel sequences of discrete tokens and continuous numbers, and we design a corresponding dual-channel transformer model to handle this format, thereby overcoming the challenges of 3D information modeling. Additionally, we alleviate data limitations through large-scale pre-training on a mixed dataset, followed by supervised and reinforcement learning fine-tuning techniques respectively tailored for the two tasks. Experimental results demonstrate that 3DMolFormer outperforms previous approaches in both protein-ligand docking and pocket-aware 3D drug design, highlighting its promising application in structure-based drug discovery. The code is available at: https://github.com/HXYfighter/3DMolFormer .

Paper Structure

This paper contains 46 sections, 7 equations, 5 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Molecular Pre-training
  4. Protein-ligand Docking
  5. Pocket-aware 3D Drug Design
  6. 3DMolFormer
  7. Format of Pocket and Ligand Sequences with 3D Coordinates
  8. Model Architecture
  9. Self-supervised Pre-training
  10. Fine-tuning
  11. Supervised Fine-tuning for Protein-ligand Binding Pose Prediction
  12. RL Fine-tuning for Pocket-aware 3D Drug Design
  13. Experiments
  14. Protein-ligand Binding Pose Prediction
  15. Data
  16. ...and 31 more sections

Figures (5)

  • Figure 1: The parallel sequence of a protein pocket with 3D coordinates.
  • Figure 2: The parallel sequence of a small molecule ligand with 3D coordinates.
  • Figure 3: Overview of 3DMolFormer. The left shows the dual-channel model architecture, the top right illustrates the input and output of the two SBDD tasks in a parallel sequence, and the bottom right outlines the pre-training and fine-tuning process.
  • Figure 4: The distributions of molecular weights, logP values, and the number of rotatable bonds of the molecules designed by 3DMolFormer.
  • Figure :