Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Unified Guidance for Geometry-Conditioned Molecular Generation

Sirine Ayadi, Leon Hetzel, Johanna Sommer, Fabian Theis, Stephan Günnemann

TL;DR

The paper addresses the challenge of geometry-conditioned molecular generation by introducing UniGuide, a unified self-guidance framework that steers unconditional diffusion models using a general condition map $C$ from geometric sources $\mathcal{S}$ to diffusion configurations $\mathcal{Z}$. By deriving a self-guided score update and ensuring equivariance through carefully designed condition maps, UniGuide enables conditioning on diverse geometric modalities (pockets, fragments, shapes) without additional training or external networks. The authors demonstrate broad applicability across ligand-based, structure-based, and fragment-based drug design, achieving competitive or superior performance to specialized baselines on MOSES, CrossDocked, Binding MOAD, and ZINC-based tasks. This separation of model training from conditioning, along with the ability to combine multiple geometric cues, promises a flexible and data-efficient path to versatile molecular generation in drug discovery. Overall, UniGuide advances unified, geometry-driven diffusion generation with tangible gains in practicality and transferability across multiple application scenarios.

Abstract

Effectively designing molecular geometries is essential to advancing pharmaceutical innovations, a domain, which has experienced great attention through the success of generative models and, in particular, diffusion models. However, current molecular diffusion models are tailored towards a specific downstream task and lack adaptability. We introduce UniGuide, a framework for controlled geometric guidance of unconditional diffusion models that allows flexible conditioning during inference without the requirement of extra training or networks. We show how applications such as structure-based, fragment-based, and ligand-based drug design are formulated in the UniGuide framework and demonstrate on-par or superior performance compared to specialised models. Offering a more versatile approach, UniGuide has the potential to streamline the development of molecular generative models, allowing them to be readily used in diverse application scenarios.

Unified Guidance for Geometry-Conditioned Molecular Generation

TL;DR

The paper addresses the challenge of geometry-conditioned molecular generation by introducing UniGuide, a unified self-guidance framework that steers unconditional diffusion models using a general condition map from geometric sources to diffusion configurations . By deriving a self-guided score update and ensuring equivariance through carefully designed condition maps, UniGuide enables conditioning on diverse geometric modalities (pockets, fragments, shapes) without additional training or external networks. The authors demonstrate broad applicability across ligand-based, structure-based, and fragment-based drug design, achieving competitive or superior performance to specialized baselines on MOSES, CrossDocked, Binding MOAD, and ZINC-based tasks. This separation of model training from conditioning, along with the ability to combine multiple geometric cues, promises a flexible and data-efficient path to versatile molecular generation in drug discovery. Overall, UniGuide advances unified, geometry-driven diffusion generation with tangible gains in practicality and transferability across multiple application scenarios.

Abstract

Effectively designing molecular geometries is essential to advancing pharmaceutical innovations, a domain, which has experienced great attention through the success of generative models and, in particular, diffusion models. However, current molecular diffusion models are tailored towards a specific downstream task and lack adaptability. We introduce UniGuide, a framework for controlled geometric guidance of unconditional diffusion models that allows flexible conditioning during inference without the requirement of extra training or networks. We show how applications such as structure-based, fragment-based, and ligand-based drug design are formulated in the UniGuide framework and demonstrate on-par or superior performance compared to specialised models. Offering a more versatile approach, UniGuide has the potential to streamline the development of molecular generative models, allowing them to be readily used in diverse application scenarios.
Paper Structure (51 sections, 2 theorems, 26 equations, 8 figures, 15 tables, 2 algorithms)

This paper contains 51 sections, 2 theorems, 26 equations, 8 figures, 15 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related work
  3. Diffusion models and controllable generation
  4. De novo molecule generation
  5. Conditional generation of molecules
  6. Controlling the generation of diffusion models
  7. Self-guiding diffusion models
  8. UniGuide
  9. Notation
  10. Unified self-guidance from geometric conditions $s\in\mathcal{S}$
  11. Guidance in the special case of $\mathcal{S}=\mathcal{Z}$
  12. UniGuide for drug discovery
  13. Ligand-based drug design
  14. Structure-based drug design
  15. Fragment-based drug design
  16. ...and 36 more sections

Key Result

Theorem 4.1

Consider a function $C: \mathcal{S} \times \mathcal{Z} \rightarrow \mathcal{Z}$. If $C(\bm{s}, \mathbf{z})$ is invariant to rigid transformations $G$ in the first argument and equivariant in the second argument, then the gradient $\nabla_\mathbf{z} \lVert \bm{v} \rVert_2^2$ of the vector $\bm{v} = \

Figures (8)

  • Figure 1: UniGuide handles diverse conditioning modalities for guidance, including: (i) a target receptor for SBDD, (ii) additional molecular fragments for FBDD, or (iii) a predefined 3D shape for LBDD. It combines a source condition $\bm{s} \in \mathcal{S}$ and the unconditional model $\bm{\epsilon}_\theta(\mathbf{z}_t,t)$ within its condition map to enable self-guidance. The flexible formulation of our approach can be generalised to new geometric tasks, for example, conditioning on atomic densities.
  • Figure 2: Surface condition map $C_{\partial V}$: For each atom coordinate $\bm{x}_i$, the closest surface points $\bm{y}_j$ are computed. The target condition $\bm{c}_{\mathbf{x},i}$ is the projection along the mean of neighbours $\bar{\bm{y}}_i$ to the inside of the volume by a margin $\alpha$, where $d=\lVert \bar{\bm{y}}_i - \hat{\bm{x}}_i\rVert_2$.
  • Figure 3: Examples of the two shape-conditioned ligands generated by UniGuide. The goal is to have low molecular graph similarity and high shape similarity.
  • Figure 4: Qualitative example of a test protein pocket (6c0b) from the Binding MOAD dataset. We show the reference ligand (grey) and samples generated by UniGuide (blue).
  • Figure 5: For various pocket-conditioned FBDD tasks, we show reference ligands (grey), desired fragments (magenta), and ligands generated by UniGuide (blue).
  • ...and 3 more figures

Theorems & Definitions (4)

  • Theorem 4.1
  • proof
  • Theorem 1
  • proof