Research map / Probabilistic and scientific modeling / Molecules and proteins
CBGBench: Fill in the Blank of Protein-Molecule Complex Binding Graph
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.
placeholder
Problem
Structure-based drug design (SBDD) lacks a standardized, fair comparison: training/test splits, evaluators (Gnina vs AutoDock Vina) and architectures differ across methods, and most work only evaluates de novo generation.
Approach
Unify SBDD and lead-optimization tasks (de novo, linker, fragment, side chain, scaffold) as generative graph completion of the 3D complex binding graph, categorize methods along three dichotomies, and evaluate them with a unified four-aspect protocol (substructure, chemical property, interaction, geometry) plus a Friedman-style mean ranking.
Key results
- MolCraft achieves the best overall performance, with TargetDiff closely following, in the de novo generation benchmark (Friedman-style weighted mean ranking).
- CNN-based LiGAN and VoxBind remain highly competitive on interaction metrics such as Vina Score and Vina Min.
- Scaffold hopping is the most challenging lead-optimization task, while linker design is relatively the easiest; MPBG values are usually negative, indicating little optimization.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use diffusion-based generation instead of autoregressive generation | SBDD molecule generation · CrossDocked2020 de novo generation | overall performance | Helped |
| Incorporate domain knowledge priors | TargetDiff-derived methods · CrossDocked2020 de novo generation | overall performance | Hurt |
| Model chemical bond patterns during autoregressive generation | Autoregressive SBDD methods · CrossDocked2020 de novo generation | overall generation performance | Helped |
| Generate molecules with more than 20 atoms rather than 10–20 atoms | SBDD molecule generation · CrossDocked2020 | LBE: LiGAN: 0.44 -> 0.35; 3DSBDD: 0.41 -> 0.33; GraphBP: 0.33 -> 0.21; Pocket2Mol: 0.44 -> 0.33; TargetDiff: 0.44 -> 0.31; Di | Hurt |
| Use voxelized CNN-based generation | SBDD molecule generation · CrossDocked2020 de novo generation | interaction performance | Helped |
| Transfer the methods to scaffold hopping rather than linker design | Lead-optimization methods · Scaffold hopping and linker design | MPBG | Hurt |
Surprising findings
- Domain-knowledge-based variants (DecompDiff, D3FG) show some performance degradation relative to their base TargetDiff, suggesting current physicochemical priors hardly improve generated molecule quality.
- Autoregressive methods other than Pocket2Mol capture pocket conditions poorly, with very low IMP% in all Vina modes, and frequently generate protein-molecule clashes.
Limitations
- Codebase covers only GNN/EGNN-based methods; voxel-based CNN methods such as LiGAN and 3DSBDD are not integrated.
- Metrics rely on computational proxies (Vina energy, SA) without wet-lab validation.
- Lead-optimization comparisons include only six continuous-position GNN methods and omit domain-knowledge methods.
Experimental setup
Datasets: CrossDocked2020 GEOM-DRUG
Benchmarks: CrossDocked2020 test split GEOM-DRUG (random control set) ADRB1 and DRD3 actives
Metrics: JSD_at MAE_at JSD_rt MAE_rt JSD_fg MAE_fg QED SA LogP LPSK
Baselines: LiGAN 3DSBDD GraphBP Pocket2Mol DiffSBDD DiffBP TargetDiff FLAG D3FG DecompDiff MolCraft VoxBind
Builds on: TargetDiff Pocket2Mol GraphBP DiffSBDD DecompDiff
Closest papers by meaning
- BindGPT: A Scalable Framework for 3D Molecular Design via Language Modeling and Reinforcement Learning (AAAI 2025 · similarity 0.90)
- Controllable 3D Molecular Generation for Structure-Based Drug Design Through Bayesian Flow Networks and Gradient Integration (NeurIPS 2025 · similarity 0.90)
- FlexSBDD: Structure-Based Drug Design with Flexible Protein Modeling (NeurIPS 2024 · similarity 0.89)
- Binding-Adaptive Diffusion Models for Structure-Based Drug Design (AAAI 2024 · similarity 0.88)
- UniMoMo: Unified Generative Modeling of 3D Molecules for De Novo Binder Design (ICML 2025 · similarity 0.88)
- ProteinBench: A Holistic Evaluation of Protein Foundation Models (ICLR 2025 · 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.