Research map / Probabilistic and scientific modeling / Molecules and proteins

TurboHopp: Accelerated Molecule Scaffold Hopping with Consistency Models

NeurIPS 2024 · Probabilistic and scientific modeling › Molecules and proteins · approach: Structure-based 3D molecule generation · 2 citations (OpenAlex)

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.

Introduces TurboHopp, an E(3)-equivariant pocket-conditioned consistency model for 3D scaffold hopping that samples up to ~30x faster than DiffHopp, plus RLCM fine-tuning for docking score and steric clashes.

Problem

Diffusion-based 3D structure-based drug design models for scaffold hopping are too slow for practical use and for RL fine-tuning, and often produce biophysically implausible poses with steric clashes.

Approach

Replace the iterative DDPM sampler of scaffold-hopping models with an SE(3)-equivariant consistency model (GVP-based, skip-connection parameterization) conditioned on pocket and functional groups, add metric-based sampling that scores intermediate outputs, and apply RLCM (PPO) to optimize docking score and steric clashes.

Key results

• TurboHopp-100 with metric sampling reaches QED 0.664, SA 0.737 and QVina -8.319 kcal/mol, with 7.14 s per batch versus 442.48 s for DiffHopp with metric sampling (up to ~30x faster inference claimed).
• TurboHoppRL-50 with metric sampling achieves Vina -9.804 kcal/mol (vs -8.298 for TurboHopp-100) and diversity 0.869, at 3.69 s, though novelty drops to 0.936.
• Against inpainted TargetDiff and DecompDiff on CrossDocked, TurboHopp-100 gives QVina -7.051 kcal/mol in 6.17 s versus 740.33 s and 1263.72 s for the baselines, but with lower diversity (0.484 vs 0.841 and 0.856).

Measured effects reported in the paper

Surprising findings

• Consistency-model outputs show lower diversity than the DDPM baseline, which the authors attribute to training for consistency along trajectories.
• Geometric fidelity is mixed: TurboHopp is worse than DecompDiff on bond and torsion distributions, and the authors note it does not learn bond properties.

Limitations

• Lower diversity than DDPM-based baselines.
• Does not model bond diffusion, so bond and torsion geometry is less faithful than baselines.
• Uses AutoDockGPU instead of QVina2 for RL training and RL models are not expected to generalize to unseen molecules.

Experimental setup

Datasets: PDBBind CrossDocked

Benchmarks: PDBBind test set CrossDocked test set

Metrics: connectivity diversity novelty QED SA QVina score Vina score inference time validity JS divergence of bond/angle/torsion distributions

Baselines: DiffHopp DiffHopp_EGNN TargetDiff DecompDiff

Builds on: DiffHopp consistency models RLCM GVP

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)
• ShEPhERD: Diffusing shape, electrostatics, and pharmacophores for bioisosteric drug design (ICLR 2025 · similarity 0.88)
• 3DMolFormer: A Dual-channel Framework for Structure-based Drug Discovery (ICLR 2025 · similarity 0.88)
• Navigating the Design Space of Equivariant Diffusion-Based Generative Models for De Novo 3D Molecule Generation (ICLR 2024 · similarity 0.88)
• DecompOpt: Controllable and Decomposed Diffusion Models for Structure-based Molecular Optimization (ICLR 2024 · similarity 0.87)
• BindGPT: A Scalable Framework for 3D Molecular Design via Language Modeling and Reinforcement Learning (AAAI 2025 · similarity 0.87)