Research map / Graph learning / Graph self-supervised learning
BindGPT: A Scalable Framework for 3D Molecular Design via Language Modeling and Reinforcement Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Pretrains a 108M GPT-NeoX-style language model on text-encoded 3D molecules (SMILES plus XYZ coordinates) and pockets, then SFTs on CrossDocked and fine-tunes with REINFORCE using Vina docking scores for pocket-conditioned 3D molecule generation.
Problem
Generating novel 3D molecules that bind a given protein pocket is hard; existing spatial generators depend on external tools (e.g., OpenBabel) for bond reconstruction and rely on specialized equivariant architectures trained on small datasets.
Approach
Represent molecules jointly as SMILES graph plus XYZ coordinates in plain text and pockets as CA-coordinate text, pretrain a decoder-only GPT-style model at scale, then finetune on CrossDocked and apply REINFORCE with docking rewards, avoiding equivariance assumptions and external graph reconstruction.
Key results
- BindGPT-RL reaches mean Vina score -8.60 ± 1.90 versus -7.80 for TargetDiff and -7.15 for Pocket2Mol in pocket-conditioned generation, with SA 0.84.
- Pretrained BindGPT achieves 98.58% validity versus 12.87% for XYZ-TF, and generates 1000 valid molecules in about 13 s versus 165 s for XYZ-TF.
- On GEOM-DRUGS, BindGPT attains the lowest JS divergences on bond lengths (0.029 vs 0.246 for EDM) but trails MolDiff on QED, SA and Lipinski.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use BindGPT instead of XYZ-TF for pretrained 3D molecule generation | 3D molecule generation · 3D molecule generation, pretraining | Valid: 12.87% -> 98.58% | Helped |
| generate explicit hydrogens (ablation) | BindGPT · 3D molecule generation with explicit hydrogens, pretraining | Valid: 98.58% -> 77.33% | Hurt |
| condition on the binding energy score during fine-tuning (ablation) | BindGPT-FT · Pocket-conditioned generation | Vina score: -5.44 ± 2.09 -> -7.24 ± 1.68 | Helped |
| fine-tune with reinforcement learning using docking rewards (ablation) | BindGPT-FT · Pocket-conditioned generation | Vina score: -5.44 ± 2.09 -> -8.60± 1.90 | Helped |
| fine-tune with reinforcement learning using docking rewards (ablation) | BindGPT-FT · Pocket-conditioned generation | QED: 0.50 ± 0.17 -> 0.43\pm 0.17 | Hurt |
Surprising findings
- Uni-Mol fails to generalize zero-shot to the Platinum conformer dataset even when assisted by RDKit.
- BindGPT-FT alone (-5.44 Vina) is weaker than TargetDiff and Pocket2Mol; the gains come mainly from RL finetuning.
Limitations
- Pocket-conditioned BindGPT-FT and BindGPT-RFT underperform baselines before RL finetuning.
- Vina docking is used as the sole RL reward, and QED drops for BindGPT-RL (0.43) relative to baselines.
- Code, data and weights are not stated as released.
Experimental setup
Datasets: Uni-Mol pretraining dataset CrossDocked 2020 GEOM-DRUGS Platinum ZINC-250k
Benchmarks: GEOM-DRUGS Platinum CrossDocked hold-out pockets ZINC-250k perplexity
Metrics: validity QED SA Lipinski JS divergences of 3D structure and bond distributions RMSD RMSD coverage Vina score generation time
Baselines: EDM MolDiff XYZ-Transformer Torsional Diffusion Uni-Mol TargetDiff Pocket2Mol
Builds on: GPT-NeoX REINFORCE Uni-Mol CrossDocked SMILES randomization
Closest papers by meaning
- UniMoMo: Unified Generative Modeling of 3D Molecules for De Novo Binder Design (ICML 2025 · similarity 0.90)
- CBGBench: Fill in the Blank of Protein-Molecule Complex Binding Graph (ICLR 2025 · similarity 0.90)
- Controllable 3D Molecular Generation for Structure-Based Drug Design Through Bayesian Flow Networks and Gradient Integration (NeurIPS 2025 · similarity 0.89)
- Swallowing the Bitter Pill: Simplified Scalable Conformer Generation (ICML 2024 · similarity 0.89)
- NExT-Mol: 3D Diffusion Meets 1D Language Modeling for 3D Molecule Generation (ICLR 2025 · similarity 0.88)
- Searching for High-Value Molecules Using Reinforcement Learning and Transformers (ICLR 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.