A Generalist Cross-Domain Molecular Learning Framework for Structure-Based Drug Discovery

TL;DR

The paper introduces BIT, a generalist cross-domain molecular learning framework that unifies small molecules, proteins, and protein–ligand complexes within a single Transformer backbone. Via Mixture-of-Domain-Experts (MoDE) and Mixture-of-Structure-Experts (MoSE), BIT achieves domain-specific encoding while capturing cross-domain interactions, trained with unified coordinate and token denoising objectives on 2D and 3D data. Empirical results across binding affinity prediction, structure-based virtual screening, and molecular property prediction demonstrate state-of-the-art performance and favorable inference efficiency, with real-world validation in NMDA receptor screening. The work highlights BIT’s potential to accelerate structure-based drug discovery by enabling cross-domain representation learning, scalable pre-training, and versatile fine-tuning for diverse SBDD tasks.

Abstract

Structure-based drug discovery (SBDD) is a systematic scientific process that develops new drugs by leveraging the detailed physical structure of the target protein. Recent advancements in pre-trained models for biomolecules have demonstrated remarkable success across various biochemical applications, including drug discovery and protein engineering. However, in most approaches, the pre-trained models primarily focus on the characteristics of either small molecules or proteins, without delving into their binding interactions which are essential cross-domain relationships pivotal to SBDD. To fill this gap, we propose a general-purpose foundation model named BIT (an abbreviation for Biomolecular Interaction Transformer), which is capable of encoding a range of biochemical entities, including small molecules, proteins, and protein-ligand complexes, as well as various data formats, encompassing both 2D and 3D structures. Specifically, we introduce Mixture-of-Domain-Experts (MoDE) to handle the biomolecules from diverse biochemical domains and Mixture-of-Structure-Experts (MoSE) to capture positional dependencies in the molecular structures. The proposed mixture-of-experts approach enables BIT to achieve both deep fusion and domain-specific encoding, effectively capturing fine-grained molecular interactions within protein-ligand complexes. Then, we perform cross-domain pre-training on the shared Transformer backbone via several unified self-supervised denoising tasks. Experimental results on various benchmarks demonstrate that BIT achieves exceptional performance in downstream tasks, including binding affinity prediction, structure-based virtual screening, and molecular property prediction.

Figures (4)

• Figure 1: BIT overview. a) We employ a general-purpose Transformer model as the backbone network to carry out masked token denoising and coordinate denoising tasks on protein-ligand complex data, as well as on unbound small molecule and pocket datasets. Additionally, we introduce Mixture-of-Domain-Experts (MoDE) and Mixture-of-Structure-Experts (MoSE) within specific modules to capture multi-domain specificity and inter-domain relationships. b) BIT services as a foundation model with diverse functionalities, including a fusion encoder for binding affinity prediction, a dual encoder for virtual screening, or a molecule encoder for molecular property prediction. c) We introduce the MoSE, which utilizes specialized pairwise bias expert networks tailored for different domains. In the 2D-MoSE, we transition from a shared pairwise bias expert for small molecules and pockets to an independent pairwise bias expert for each entity. Conversely, in the 3D-MoSE, we maintain the shared expert for each entity, while introducing independent bias experts specifically for the protein-ligand interaction modeling. Distinct parameters within these networks are denoted by varying colors. d) We propose unified corrupt-then-denoise objectives (i.e., coordinate denoising and masked token denoising) applicable to various domain data.
• Figure 2: Illustration of the BIT-driven pipeline for virtual screening. BIT serves as a dual encoder for efficient coarse structure-based virtual screening and as a unimodal molecular encoder for fine ligand-based virtual screening.
• Figure 3: Visualization and experimental validation on identified hit compounds.
• Figure 4: The distribution of 2D structure between molecule and pocket. The gray area indicates the overlapping part. (a) The shorest path distance, none represents that there is no path between the atoms; (b) Out-degree, 0 represents that the atom has no edge connected to other atoms.