Pharmacophore-guided de novo drug design with diffusion bridge

TL;DR

PharmacoBridge tackles de novo drug design by explicitly guiding 3D molecular generation with pharmacophore constraints using an SE(3)-equivariant diffusion bridge. The method defines a denoising diffusion process that maps pharmacophore endpoints to molecular structures, parameterized with an EGNN-based score model to preserve geometric symmetries. It demonstrates competitive unconditional generation while delivering superior pharmacophore alignment and binding affinity in ligand-based and structure-based settings, validated on CrossDocked2020, and beats several state-of-the-art baselines. The approach offers a principled, geometry-aware path from pharmacophore hypotheses to 3D hit molecules, potentially accelerating hit identification and target-specific drug design.

Abstract

De novo design of bioactive drug molecules with potential to treat desired biological targets is a profound task in the drug discovery process. Existing approaches tend to leverage the pocket structure of the target protein to condition the molecule generation. However, even the pocket area of the target protein may contain redundant information since not all atoms in the pocket is responsible for the interaction with the ligand. In this work, we propose PharmacoBridge, a phamacophore-guided de novo design approach to generate drug candidates inducing desired bioactivity via diffusion bridge. Our method adapts the diffusion bridge to effectively convert pharmacophore arrangements in the spatial space into molecular structures under the manner of SE(3)-equivariant transformation, providing sophisticated control over optimal biochemical feature arrangements on the generated molecules. PharmacoBridge is demonstrated to generate hit candidates that exhibit high binding affinity with potential protein targets.

Figures (9)

• Figure 1: Overview of PharmacoBridge. The diffusion bridge process is devised to map the molecule data $\mathbf{G}_0$ to the extracted pharmacophore data $\mathbf{G}_T$ via Doob's $h$-transforms. Reversely, a score matching model is trained to estimate the score function, which composes the denoising bridge process that recovers molecule data from the pharmacophore data.
• Figure 2: SA score ($\downarrow$) distribution. SA scores of the molecules generated by EDM and GruM are concentrated between 6.0 and 8.0, which is obviously larger than the distribution of the original dataset. The SA scores achieved by our model are lower than the baselines and more evenly distributed like the original dataset distribution.
• Figure 3: QED score ($\uparrow$) distribution. Molecules generated by our model and EDM exhibit higher QED scores than GruM. The distribution achieved by our model shows similar pattern to the dataset distribution.
• Figure 4: Pharmacophore matching sore distribution. Compared with Unconstrained generation and TargetDiff, pharmacophore-guided generation significantly enhanced the matching of pharmacophores extracted from original ligands and generated molecules. Notably, TargetDiff suffered from low validity, which resulting in less available molecules.
• Figure 5: Distribution of Vina scores, with lower Vina score representing higher binding affinity. The reference scores given by the original ligands are indicated by red dashed lines in the figure. PharmacoBridge consistently generated molecules with higher binding affinities. Although Pocket2Mol sometimes outperformed, it produced molecules with a notably broader range of Vina scores in certain groups, e.g., 1EOC and 4H2Z. TargetDiff failed to generate molecules with binding affinities surpassing those of the original ligands in most cases.

Theorems & Definitions (7)

• Theorem 3.1: Denoising diffusion bridge
• Theorem 3.2: Equivariant denoising diffusion bridge
• Theorem 1.1: Denoising diffusion bridge
• proof
• Theorem 1.1: Equivariant denoising diffusion bridge
• Theorem 1.1: Equivariant denoising diffusion bridge
• proof