MolSnapper: Conditioning Diffusion for Structure-Based Drug Design

Yael Ziv, Fergus Imrie, Brian D. Marsden, Charlotte M. DeaneJournal of Chemical Information and Modeling 2025 · DOI

TL;DR

MolSnapper is a new computational tool that uses diffusion models to generate drug-like molecules tailored to specific protein binding sites. By incorporating expert knowledge in the form of 3D pharmacophores, MolSnapper significantly improves the quality and validity of generated molecules compared to existing methods, producing more molecules that are likely to bind effectively and be synthesizable.

What Problem Does This Paper Address?

Designing new drugs is a complex and costly process. A major challenge lies in generating molecules that not only have the desired chemical properties but also bind effectively and specifically to a target protein's 3D structure. Traditional methods are often slow and expensive, while early AI approaches, though promising, struggled to incorporate crucial expert knowledge about how molecules interact with proteins. Specifically, existing generative models often produce molecules that are physically implausible, fail to form essential hydrogen bonds, or are difficult to synthesize. This paper addresses the need for a generative model that can be guided by prior knowledge to produce more realistic, effective, and synthesizable drug candidates for structure-based drug design (SBDD).

Methods

The core of MolSnapper is a diffusion model, a type of generative AI that learns to create data by gradually denoising random noise. The novelty of MolSnapper lies in its "conditioning" mechanism, which allows it to incorporate expert knowledge during the generation process. Instead of training the diffusion model on protein-ligand complexes (which are scarce), MolSnapper leverages models pre-trained on vast datasets of molecules. It then conditions these models using 3D pharmacophores – representations of key chemical interactions (like hydrogen bonds) crucial for binding.

MolSnapper achieves this conditioning in two main ways:

  • Constraining Atom Positions: It guides the diffusion process to ensure that specific atoms within the generated molecule occupy positions dictated by the pharmacophore. This is done by subtly adjusting the atom positions during the reverse diffusion (generation) process, anchoring them towards desired locations.
  • Preventing Clashes: To avoid generated molecules physically colliding with the protein, MolSnapper incorporates a "clash guidance" mechanism. This uses the protein's surface information to steer the molecule away from unfavorable interactions.

The method also allows for constraining atom types, offering further control over the generated molecular structure.

Datasets

The study evaluated MolSnapper using two established datasets:

  • CrossDocked2020: This dataset was created by redocking and cross-docking known binders with experimental structures. It was filtered to include high-quality docking poses and diverse proteins.
  • Binding MOAD: This dataset exclusively comprises experimentally determined protein-ligand complexes.

These datasets were used for both training (implicitly, as MolSnapper uses pre-trained models) and rigorous evaluation of the generated molecules.

Key Insights and Ideas

  • Expert knowledge is key: The paper strongly emphasizes that effectively incorporating prior knowledge (like pharmacophores) is crucial for moving beyond generating plausible molecules to generating useful drug candidates.
  • Diffusion models can be conditioned without protein-specific training: MolSnapper demonstrates that powerful diffusion models, pre-trained on general molecular data, can be effectively adapted for SBDD by conditioning them with structural information, bypassing the need for scarce protein-ligand complex training data.
  • Physically plausible interactions matter: The study highlights that many existing SBDD methods fail to generate molecules that form realistic hydrogen bonds. MolSnapper's conditioning directly addresses this by focusing on pharmacophore-guided interactions.
  • Direct pose evaluation is critical: The authors argue against relying solely on redocking for evaluation, as it can mask fundamental issues in the initially generated poses. MolSnapper's evaluation focuses on the direct output of the generative process.

Key Results

MolSnapper demonstrated superior performance across several metrics compared to baseline methods like MolDiff (unconditioned) and SILVR (another conditioning method):

  • Higher Validity: MolSnapper generated approximately twice as many valid molecules that passed the PoseBusters checks (a measure of physical plausibility) compared to SILVR.
  • Improved Similarity: MolSnapper achieved significantly higher SCRDKit scores (measuring shape and color similarity to reference ligands) than both MolDiff and SILVR.
  • Better Interaction Recovery: MolSnapper was much more successful at recapitulating the hydrogen bond interactions between the ligand and the protein, achieving an average "Interaction Similarity" of 0.746 compared to 0.276 for MolDiff and 0.493 for SILVR.
  • Enhanced Synthesizability: MolSnapper generated molecules with better Synthetic Accessibility (SA) scores than SILVR, indicating they are more likely to be synthesizable.
  • Competitive with Pocket-Specific Training: When compared to DiffSBDD (a model trained specifically on protein-ligand complexes), MolSnapper showed comparable performance in similarity and interaction metrics, while offering the advantage of not requiring specialized training data.

Limitations and Future Work

  • Synthesizability: The authors acknowledge that, like many generative models, some molecules generated by MolSnapper might still be challenging to synthesize.
  • Tautomeric/Protomeric Forms: The current framework does not explicitly enforce specific tautomeric or protomeric forms of molecules.
  • Need for New Metrics: The paper calls for the development of new metrics to rigorously assess the capabilities of SBDD generative models that produce 3D poses.

Future work could explore integrating reaction-based generative models for simultaneous molecule and synthesis route generation, or developing more sophisticated methods for assessing generated poses. The authors also suggest that their framework is flexible and can be extended to incorporate other types of constraints beyond pharmacophores.

Summary updated

BibTeX
@article{ziv2025molsnapper,
  author = {Yael Ziv and Fergus Imrie and Brian D. Marsden and Charlotte M. Deane},
  title = {{MolSnapper: Conditioning Diffusion for Structure-Based Drug Design}},
  journal = {Journal of Chemical Information and Modeling},
  year = {2025},
  doi = {10.1021/acs.jcim.4c02008},
}
Figure