TurboHopp: Accelerated Molecule Scaffold Hopping with Consistency Models

TL;DR

TurboHopp introduces an SE(3)-equivariant, pocket-conditioned consistency-model for fast 3D scaffold hopping, addressing the slow inference of diffusion-based 3D-SBDD methods. By combining a PF-ODE-inspired consistency framework with target-aware conditioning and GVP-based architectures, it achieves up to 30× faster generation while maintaining or improving key drug-design metrics. The work further integrates Reinforcement Learning for Consistency Models (RLCM) to fine-tune outputs toward docking scores, reduced steric clashes, and overall drug-likeness without redocking. Extensive experiments on PDBBind and CrossDocked demonstrate substantial speedups, competitive quality across metrics, and the feasibility of RL-guided 3D-SBDD with consistency models, highlighting practical implications for rapid, interactive drug-design workflows.

Abstract

Navigating the vast chemical space of druggable compounds is a formidable challenge in drug discovery, where generative models are increasingly employed to identify viable candidates. Conditional 3D structure-based drug design (3D-SBDD) models, which take into account complex three-dimensional interactions and molecular geometries, are particularly promising. Scaffold hopping is an efficient strategy that facilitates the identification of similar active compounds by strategically modifying the core structure of molecules, effectively narrowing the wide chemical space and enhancing the discovery of drug-like products. However, the practical application of 3D-SBDD generative models is hampered by their slow processing speeds. To address this bottleneck, we introduce TurboHopp, an accelerated pocket-conditioned 3D scaffold hopping model that merges the strategic effectiveness of traditional scaffold hopping with rapid generation capabilities of consistency models. This synergy not only enhances efficiency but also significantly boosts generation speeds, achieving up to 30 times faster inference speed as well as superior generation quality compared to existing diffusion-based models, establishing TurboHopp as a powerful tool in drug discovery. Supported by faster inference speed, we further optimize our model, using Reinforcement Learning for Consistency Models (RLCM), to output desirable molecules. We demonstrate the broad applicability of TurboHopp across multiple drug discovery scenarios, underscoring its potential in diverse molecular settings.

Figures (11)

• Figure 1: (a) Previous diffusion-based SBDD models methodically explore vast chemical space for pocket-active molecules, with short arrows symbolizing a gradual, step-wise inference process. (b) TurboHopp efficiently generates active ligands using a consistency model, which accelerates inference and reduces the number of steps, as illustrated by the longer arrows. Moreover, it strategically leverages the functional groups of high-activity reference molecules, shown as colored areas on the diagram, to optimize the exploration within targeted chemical space.
• Figure 2: (a) Comparison of QED scores, with TurboHopp (blue) reaching peak values faster compared to DiffHopp (red), throughout the generation process. (b) Progression of generated outputs of both models. Final steps are highlighted with red boxes.
• Figure 3: Model Architecture of TurboHopp. Given a reference ligand and its corresponding protein pocket, an equivariant consistency model samples scaffolds conditioned on pocket substructure and functional groups. Models are trained to map points on the same PF-ODE path to the original data given context.
• Figure 4: Compared to TurboHopp-100 (Red), TurboHoppRL-50 (Blue) has enhancing binding affinity without losing performance in other metrics.
• Figure 5: Training curves for the metrics which compose of the loss function. Notice that all either increase or maintain approximately the same value. Connectivity and QED score slightly decrease because we start from a previously RL finetuned checkpoint which optimizes only for connectivity, SA, and QED score.