3bmGPT™: A Large Language Model for Systematic 3D Binding Mode Exploration

3bmGPT™: A Large Language Model for Systematic 3D Binding Mode Exploration

1. What is 3bmGPT?

3bmGPT (3D Binding Mode GPT) is a generative pre-trained transformer model that transforms 3D protein–ligand binding interactions into textual representations. It addresses the complexity of molecular binding by converting 3D geometric data into "binding words" and "binding sentences," enabling a Large Language Model (LLM) to learn chemical recognition patterns through contextual reasoning.

2. Distinctive Advantages Over Traditional Structure-Based Models

1. Dynamic Binding representation
   Unlike static docking snapshots, the model is trained on diverse interaction poses to capture broader binding contexts.

2. Linearization of 3D Data
   Traditional models struggle to process 3D coordinates; 3bmGPT introduces a method to tokenize spatial interactions into a stable, learnable string format.

3. Functional Context
   Beyond simple structure prediction, the model focuses on the "meaning" of the interaction, identifying functional similarities between different proteins.

3. Core Technology: How 3bmGPT Decodes Molecular Interactions

1. Encoding 3D Geometry into "Binding Words"
   Rather than simply reading spatial coordinates, 3bmGPT captures the underlying mechanism of molecular recognition. Each interaction is encoded into a "Binding Word"—a symbolic unit that integrates:

   • Pharmacophoric Roles: Defines chemical features such as H-bond donors/acceptors, hydrophobic interactions, and aromatic stacking.

   • Interaction Partners: Identifies specific amino acid residues and corresponding ligand atoms.

   • Geometric Descriptors: Distance, angles, and orientations are categorized into discrete bins, preserving the precision of 3D spatial arrangements within a linguistic framework.

2. Constructing "Binding Sentences" for Contextual Reasoning
   Individual binding words are sequenced into a "Binding Sentence," representing the comprehensive binding mode within a protein pocket. This allows the model to move beyond simple physical distances and instead interpret the "grammar of molecular recognition" and the functional flow of interactions.

3. Unrivaled Training Scale: 60 Million Binding Sentences
   To ensure the model’s robustness and versatility, the CrossDocked2020 dataset was extensively augmented:

   • Instead of learning from a single static pose, the model was trained on 120 diverse docking poses per protein-ligand pair.

   • This resulted in over 60 million binding sentences, providing a rich "Binding Context" that prevents the model from being biased toward specific structural snapshots and enables it to generalize across various binding environments.

4. Optimized Architecture with GPT-2 & tokenization
   3bmGPT utilizes a proven GPT-2-based architecture specifically optimized for molecular data. By implementing tokenization, the model breaks down complex interaction patterns into efficient tokens, enabling it to learn functional continuity and molecular similarities with high computational efficiency.

4. Scientific Validation & Performance

1. Chemical-Binding Correlation
   The model shows that the ratio of shared binding words correlates with chemical similarity (Tanimoto coefficient).

2. Comparative Performance
   In predicting Rebastinib–ABL1 interactions, 3bmGPT achieved an AUC of 0.8486, outperforming BERTbased models (AUC =0.66~0.77).

3. Functional Clustering
   UMAP visualization of 10,000 binding embeddings revealed 23 distinct clusters, accurately grouping proteins by their biological functions (e.g., Kinases, Proteases, Nuclear Receptors).

5. Case Study: EGFR (PDB ID: 5EDQ)

1. Target Classification
   When presented with a new EGFR structure, the model correctly identified it as a protein kinase and placed it within the Tyrosine Kinase (TK) family cluster.

2. Ligand Similarity & Repurposing
   The model identified ligands with similar binding contexts, such as Merestinib and Brigatinib, demonstrating its utility in identifying potential drug repurposing candidates based on binding-mode similarity.

6. Key Applications Highlighted in the Paper

1. Enhancing Virtual Screening
   Using the LLM’s embedding space to navigate large chemical libraries based on interaction context rather than just molecular shape.

2. Scaffold Hopping
   Identifying novel chemical scaffolds that maintain critical binding interactions (Binding Signatures).

3. Target-Agnostic Exploration
   Analyzing binding similarity even when explicit protein-target labels are absent.

4. Off-Target Identification
   Detecting potential cross-reactivity by finding similar binding sentences across unrelated protein families.

Conclusion: A New Paradigm for Drug Discovery

3bmGPT establishes a framework for treating 3D biological interactions as a language. By training on textualized binding data, the model captures the "grammar" of molecular recognition, providing a scientifically robust tool for lead discovery, safety profiling, and the systematic exploration of the protein–ligand interactome.