Research map / Probabilistic and scientific modeling / Molecules and proteins
Piloting Structure-Based Drug Design via Modality-Specific Optimal Schedule
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 VLB-Optimal Scheduling (VOS), which trains a BFN with a decoupled generalized loss over a 2D time-rescaling plane and uses dynamic programming to find a VLB-optimal joint noise schedule for structure-based drug design (MolPilot).
Problem
Structure-based drug design generative models struggle with geometric fidelity because the joint probability path over continuous 3D positions and discrete 2D topology is poorly designed; default noise schedules make the path 3D-dominant.
Approach
Show that for multi-modal generation the VLB is a path integral dependent on the joint noise schedule, reduce the schedule design space to a 2D plane via time rescaling, train one model over that plane with a generalized loss, and search for the VLB-optimal path with dynamic programming.
Key results
- MolPilot reaches a PoseBusters passing rate of 95.9% on CrossDock (ID), versus 84.6% for MolCRAFT.
- On the OOD PoseBusters set, MolPilot achieves PB-Valid 79.1% and scRMSD<2Å 56.1%, compared with 57.8% and 46.4% for MolCRAFT (re-run).
- Applying VOS to TargetDiff raises its PB-Valid from 50.5% to 58.1% on CrossDock.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply VLB-Optimal Scheduling to MolPilot | MolPilot · CrossDock ID | PB-Valid: 84.6% -> 95.9% | Helped |
| Apply VLB-Optimal Scheduling to MolPilot | MolPilot · PoseBusters OOD | PB-Valid: 57.8% -> 79.1% | Helped |
| Apply VLB-Optimal Scheduling to TargetDiff | TargetDiff · CrossDock | PB-Valid: 50.5% -> 58.1% | Helped |
| Add generalized loss (ablation) | TargetDiff · CrossDock | PB-Valid: 50.5% -> 53.7% | Helped |
| Add the optimal schedule after generalized-loss training (ablation) | TargetDiff + generalized loss · CrossDock | PB-Valid: 53.7% -> 58.1% | Helped |
| Use the optimal schedule instead of the default schedule (ablation) | MolPilot · Validation | validation loss: 5.51 -> 4.23 | Helped |
| Learn the noise schedule directly during training instead of deriving it by search (ablation) | MolPilot · Validation | validation loss: 4.23 -> 5.39 | Hurt |
| Move from identity time-rescaling to the optimal schedule (ablation) | MolPilot · PoseBusters OOD | PB-Valid: 0.79 -> 0.80 | Helped |
| Move from identity time-rescaling to the optimal schedule (ablation) | MolPilot · CrossDock ID | Vina Score Avg.: -6.63 -> -6.92 | Helped |
| Use generalized loss instead of the default objective (ablation) | MolPilot · CrossDock ID | Vina Score Avg. | Helped |
| Train with generalized loss rather than a fixed-path VLB objective (ablation) | MolPilot · PoseBusters OOD | binding affinity | Helped |
| Use the optimal schedule rather than the default noise schedule (ablation) | MolPilot trained with generalized loss · PoseBusters OOD | binding pose quality | Helped |
Surprising findings
- Learning the noise schedule jointly with the generative model during training did not improve results; the learned schedule barely moved from initialization and gave a suboptimal VLB.
- The authors suggest PoseBusters is a more challenging OOD test than CrossDock, since CrossDock may have data leakage.
Limitations
- Optimal schedule relies on a grid of cost estimates and requires quadratic model evaluations over the M x M grid.
- Directly learning the schedule during training was unstable and gave suboptimal VLB; bilevel optimization left for future work.
- Likelihood-based VLB and sample quality diverge because the accuracy-level scaling puts more weight on the continuous modality.
Experimental setup
Datasets: CrossDock PoseBusters
Benchmarks: CrossDock PoseBusters
Metrics: PB-Valid passing rate Vina Score Vina Min Vina Dock scRMSD Strain energy passed Connected QED SA RMSD<2Å docking success
Baselines: AR Pocket2Mol FLAG TargetDiff DiffSBDD DecompDiff MolCRAFT
Builds on: Bayesian Flow Network MolCRAFT TargetDiff EquiFM
Closest papers by meaning
- FlexSBDD: Structure-Based Drug Design with Flexible Protein Modeling (NeurIPS 2024 · similarity 0.87)
- CBGBench: Fill in the Blank of Protein-Molecule Complex Binding Graph (ICLR 2025 · similarity 0.86)
- Integrating Protein Dynamics into Structure-Based Drug Design via Full-Atom Stochastic Flows (ICLR 2025 · similarity 0.86)
- MolCRAFT: Structure-Based Drug Design in Continuous Parameter Space (ICML 2024 · similarity 0.86)
- Empower Structure-Based Molecule Optimization with Gradient Guided Bayesian Flow Networks (ICML 2025 · similarity 0.86)
- DecompOpt: Controllable and Decomposed Diffusion Models for Structure-based Molecular Optimization (ICLR 2024 · similarity 0.86)
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.