Research map / Probabilistic and scientific modeling / Molecules and proteins
Enhancing Ligand Validity and Affinity in Structure-Based Drug Design with Multi-Reward Optimization
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.
Fine-tunes the MolCRAFT Bayesian flow network for structure-based drug design with multi-reward DPO (BFN-DPO), using softmax reward normalization and a variance-penalized ensemble to jointly improve Vina docking score, strain energy, and QED.
Problem
Structure-based drug design generative models mainly fit the training chemical distribution and optimize binding affinity, often producing ligands with poor validity (high strain energy, clashes) and low synthetic accessibility, so the affinity-property Pareto front is not expanded.
Approach
Derives a direct preference optimization objective for Bayesian flow networks (BFN-DPO), combines multiple rewards via per-batch softmax normalization and a variance-penalized score, and stabilizes training with E2PO, so a pretrained BFN can be steered toward higher affinity and validity together.
Key results
- Vina Dock median -7.89 vs -7.27 for MolCRAFT and -7.87 for AliDiff, with SA avg 0.74 and QED 0.55 (median), the best QED among compared methods.
- Generated ligands pass PoseBuster validity at 85% vs 84% for MolCRAFT and 25% for AliDiff (from Figure 3a).
- Optimizing all seven rewards together underperforms the three-reward (Dock+SE+QED) choice, and gradient surgery baselines fall below the reference in most metrics.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune MolCRAFT with multi-reward optimization using Vina Score, strain energy, and QED | MolCRAFT · CrossDocked test targets | Vina Score (average; lower is better): -6.59 -> -7.18 | Helped |
| Fine-tune MolCRAFT with multi-reward optimization using Vina Score, strain energy, and QED | MolCRAFT · CrossDocked test targets | SA (average; higher is better): 0.69 -> 0.74 | Helped |
| Fine-tune MolCRAFT with multi-reward optimization using Vina Score, strain energy, and QED | MolCRAFT · CrossDocked test targets | QED (average; higher is better): 0.50 -> 0.55 | Helped |
| Use gradient surgery on three reward losses instead of uncertainty-based reward optimization (ablation) | MolCRAFT multi-reward optimization · CrossDocked test targets | Vina Score (median; lower is better): -7.38 -> -6.80 | Hurt |
| Apply gradient surgery separately to six atom and type losses instead of uncertainty-based reward optimization (ablation) | MolCRAFT multi-reward optimization · CrossDocked test targets | Vina Score (median; lower is better): -7.38 -> -6.87 | Hurt |
| Use two-stage iterative fine-tuning instead of uncertainty-based reward optimization (ablation) | MolCRAFT multi-reward optimization · CrossDocked test targets | Vina Score (median; lower is better): -7.38 -> -7.16 | Hurt |
| Optimize all seven rewards instead of Dock+SE+QED (ablation) | MolCRAFT multi-reward optimization · CrossDocked test targets | Vina Score (average; lower is better): -7.18 -> -6.50 | Hurt |
| Optimize Vina Score alone instead of the three-reward combination (ablation) | MolCRAFT · CrossDocked test targets | Vina Score (average; lower is better): -7.18 -> -6.63 | Hurt |
| Optimize Vina Min alone instead of the three-reward combination (ablation) | MolCRAFT · CrossDocked test targets | Vina Score (average; lower is better): -7.18 -> -6.61 | Hurt |
| Optimize Vina Dock alone instead of the three-reward combination (ablation) | MolCRAFT · CrossDocked test targets | Vina Score (average; lower is better): -7.18 -> -6.63 | Hurt |
| Optimize strain energy alone instead of the three-reward combination (ablation) | MolCRAFT · CrossDocked test targets | Vina Score (average; lower is better): -7.18 -> -6.51 | Hurt |
| Optimize clash alone instead of the three-reward combination (ablation) | MolCRAFT · CrossDocked test targets | Vina Score (average; lower is better): -7.18 -> -6.44 | Hurt |
Surprising findings
- Optimizing all seven reward metrics jointly degraded performance relative to a selected subset of three rewards.
- Strain energy and synthetic accessibility are strongly correlated in the results, with no clear trade-off observed between them.
Limitations
- Clash metric was not improved even with its single reward, left for future work.
- Optimizing all seven rewards jointly degraded performance, so reward selection is needed.
- Evaluation relies on docking-based scores (AutoDock Vina) and computed proxies rather than experimental validation.
Experimental setup
Datasets: CrossDocked
Benchmarks: CrossDocked test set (100 targets)
Metrics: Vina Score Vina Min Vina Dock strain energy clash synthetic accessibility QED PoseBuster validity pass rate
Baselines: AR Pocket2Mol FLAG TargetDiff DecompDiff DecompOpt IPDiff AliDiff MolCRAFT
Builds on: Direct Preference Optimization Diffusion-DPO Bayesian flow networks MolCRAFT E2PO
Closest papers by meaning
- Controllable 3D Molecular Generation for Structure-Based Drug Design Through Bayesian Flow Networks and Gradient Integration (NeurIPS 2025 · similarity 0.90)
- Multi-domain Distribution Learning for De Novo Drug Design (ICLR 2025 · similarity 0.88)
- Aligning Target-Aware Molecule Diffusion Models with Exact Energy Optimization (NeurIPS 2024 · similarity 0.88)
- Prior-Guided Flow Matching for Target-Aware Molecule Design with Learnable Atom Number (NeurIPS 2025 · similarity 0.88)
- Multi-Objective Molecular Design Through Learning Latent Pareto Set (AAAI 2025 · similarity 0.88)
- Rethinking Specificity in SBDD: Leveraging Delta Score and Energy-Guided Diffusion (ICML 2024 · similarity 0.87)
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.