Research map / Probabilistic and scientific modeling / Molecules and proteins
Rethinking Specificity in SBDD: Leveraging Delta Score and Energy-Guided Diffusion
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.
Proposes the Delta Score, a specificity metric for SBDD generated molecules derived from conditional binding probability, and SBE-Diff, an energy-guided diffusion model that uses a contrastive-trained specific binding energy to steer generation toward target-specific binders on CrossDocked2020.
Problem
Deep generative SBDD models achieve high docking scores, but these molecules often bind many unintended pockets with high affinity, so docking scores and current models lack specificity (conditional binding) assessment.
Approach
Derive a specificity metric from log p(y|x) using docking scores across pockets (Delta Score: score on target minus score on a randomly sampled other pocket), and guide a diffusion model with a contrastively trained energy (SBE) that uses other pockets' ligands as negatives, so the energy gradient steers sampling toward specific binders.
Key results
- SBE-Diff achieves mean Delta score 0.552 vs 0.335 for TargetDiff and 0.354 for DecompDiff, while reference ligands reach 1.158.
- SBE-Diff reports mean absolute docking score 6.815 (Glide, shown as magnitude) vs TargetDiff 6.665 and DecompDiff 7.102, and ratio better-than-reference 0.300 mean.
- Random training-set molecules achieve mean docking score close to reference (5.727 vs 6.632) with 30.83% of molecules better than reference, suggesting docking scores are weak specificity signals.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| adding SBE energy guidance (contrastive-trained energy, active compounds as negatives) to TargetDiff diffusion (ablation) | TargetDiff · CrossDocked2020 | delta score mean (lower better in Table 1 sense; higher in Table 2 reported): 0.335 -> 0.552 | Helped |
| adding SBE energy guidance to TargetDiff (ablation) | TargetDiff · CrossDocked2020 | absolute docking score mean: 6.665 -> 6.815 | Helped |
| adding SBE energy guidance to TargetDiff (ablation) | TargetDiff · CrossDocked2020 | ratio better than reference (mean): 0.259 -> 0.300 | Helped |
| adding SBE energy guidance to TargetDiff (ablation) | TargetDiff · CrossDocked2020 | QED, SA, diversity: QED 0.503->0.497; SA 0.587->0.586; diversity 0.880->0.879 | No clear effect |
| use of random sampling from the training set as a baseline | docking-score evaluation · CrossDocked2020 | fraction of molecules better than reference (Vina/Glide docking): 30.83% for Random vs 30.55% AR | No clear effect |
| DecompDiff and DrugGPS vs TargetDiff (prior methods) | baseline generative models · CrossDocked2020 | delta score mean: DrugGPS 0.490 and DecompDiff 0.354 vs TargetDiff 0.335; docking score rank differs | Mixed |
Surprising findings
- Randomly sampled training-set molecules achieve docking scores close to reference ligands, with about 31% outperforming them.
- Current SOTA generative models underperform reference ligands on the Delta Score despite surpassing them on absolute docking score.
Limitations
- Delta Score relies on docking scores as proxies for interaction energy, which the paper itself notes are biased.
- Delta Score is estimated with a single randomly sampled alternative pocket per target rather than the full minimum over pockets.
- Evaluation relies on Glide docking of 1D/2D-converted molecules rather than the generated 3D conformations.
Experimental setup
Datasets: CrossDocked2020
Benchmarks: CrossDocked2020
Metrics: Glide docking score (absolute) Delta score ratio better than reference QED SA diversity
Baselines: AR Pocket2Mol DrugGPS TargetDiff DecompDiff Random training-set sampling
Builds on: TargetDiff energy-guided diffusion model DecompDiff
Closest papers by meaning
- Controllable 3D Molecular Generation for Structure-Based Drug Design Through Bayesian Flow Networks and Gradient Integration (NeurIPS 2025 · similarity 0.88)
- Enhancing Ligand Validity and Affinity in Structure-Based Drug Design with Multi-Reward Optimization (ICML 2025 · similarity 0.87)
- DeltaDock: A Unified Framework for Accurate, Efficient, and Physically Reliable Molecular Docking (NeurIPS 2024 · similarity 0.87)
- FlexSBDD: Structure-Based Drug Design with Flexible Protein Modeling (NeurIPS 2024 · similarity 0.86)
- Aligning Target-Aware Molecule Diffusion Models with Exact Energy Optimization (NeurIPS 2024 · similarity 0.86)
- CBGBench: Fill in the Blank of Protein-Molecule Complex Binding Graph (ICLR 2025 · similarity 0.85)
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.