Pharmacophore-Conditioned Diffusion Model for Ligand-Based De Novo Drug Design
Amira Alakhdar, Barnabas Poczos, Newell Washburn
TL;DR
PharmaDiff is a novel transformer-based diffusion model that generates 3D molecules conditioned on pharmacophore features, improving ligand-based drug design and docking performance without requiring target structures.
Contribution
It introduces PharmaDiff, the first pharmacophore-conditioned diffusion model for 3D molecular generation, integrating pharmacophore hypotheses into the generative process.
Findings
Outperforms ligand-based methods in pharmacophore matching
Achieves higher docking scores across multiple proteins
Generates molecules aligning with predefined pharmacophore features
Abstract
Developing bioactive molecules remains a central, time- and cost-heavy challenge in drug discovery, particularly for novel targets lacking structural or functional data. Pharmacophore modeling presents an alternative for capturing the key features required for molecular bioactivity against a biological target. In this work, we present PharmaDiff, a pharmacophore-conditioned diffusion model for 3D molecular generation. PharmaDiff employs a transformer-based architecture to integrate an atom-based representation of the 3D pharmacophore into the generative process, enabling the precise generation of 3D molecular graphs that align with predefined pharmacophore hypotheses. Through comprehensive testing, PharmaDiff demonstrates superior performance in matching 3D pharmacophore constraints compared to ligand-based drug design methods. Additionally, it achieves higher docking scores across a…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
Clear motivation: bridges the gap between ligand-based and structure-based diffusion models by using pharmacophoric cues that are experimentally accessible yet geometry-aware. Architectural novelty: the combination of inpainting, cross-attention, and equivariance-preserving alignment is well-engineered and principled. Strong empirical evidence: PharmaDiff consistently improves pharmacophore-matching metrics (MS ≈ 0.90, PMR ≈ 0.70) and achieves better or comparable docking scores and synthetic
Ambiguity in conditioning signal: while pharmacophore features serve as conditioning inputs, it is unclear how stochastic or deterministic their placement is at inference—i.e., whether small coordinate perturbations affect generation stability. Scalability limits: the method assumes a small number (3–7) of pharmacophoric points; it is uncertain how performance scales with larger or flexible hypotheses. Ablation clarity: contributions of individual modules (cross-attention vs inpainting vs loss
All metrics are reported with error bars, which is highly appreciated. I appreciate the well-documented instructions on github to replicate the results in the paper.
Comparisons to other methods are weak. Only other method compared to is DiffSBDD for structure-conditioned ligand generation. Additional comparisons to newer methods are needed, and on more than just 4 targets. Methodological novelty is weak, as pharmacophore-conditioned generative ligand design has been shown (see “Pharmacophore-guided de novo drug design with diffusion bridge” by Wang and Rajapakse). Could the authors explain how their method differs from the one above?
Paper is overall clear and easy to read, and follows the track of how molecule generative models were built. The figure explains the model architecture clearly and how generation is done. Pharmacophore based generation is a useful and practical approach in drug design.
1.the concept itself is not novel nor new, there has been many attempts following similar tactics. What makes your model special and different compared to the previous approaches( refer to the weakness 2) 2.there are many pharmacophore based generative models that has the codes to compare. an in-depth comparison between these models is necessary: what is strong and special with your model? what makes your model unique? some models that i've found are: ShEPhERD (ICLR ’25), PhoreGen (Nat. Comput
* This paper introduces pharmacophore conditioned generation for molecular design, which is crucial when analyzing interaction patterns between drugs and targets. * PharmaDiff shows improvements in terms of pharmacophore matching and competetive docking performance compared to DiffSBDD.
* One primary concern of this work is the relationship between 'pharmacophore' and 'ligand-based drug deisgn'. Pharmacophore contains information like hydrogen bond donor (HBD), hydrogen bond acceptor (HBA), etc., which are important parts when forming H-bonds between ligands and targets. However, this paper claims a setting where target protein is not needed. In such a setting, the pharmacophore becomes an abstract spatial pattern inferred solely from ligand conformations rather than a true int
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsSynthesis and Biological Evaluation · Computational Drug Discovery Methods · Analytical Chemistry and Chromatography
MethodsDiffusion · ALIGN