PaperlayerView exact revision on arXiv

Source-linked AI summary

DecompDiff: Diffusion Models with Decomposed Priors for Structure-Based Drug Design

Jiaqi Guan, Xiangxin Zhou, Yuwei Yang, Yu Bao, Jian Peng, Jianzhu Ma, Qiang Liu, Liang Wang, Quanquan Gu

arXiv:2403.07902v1q-bio.BMcs.LG

TL;DR

Structure-based drug design seeks 3D ligands conditioned on a target binding site, but existing methods can ignore differing atom roles and omit bonds from diffusion modeling. DecompDiff uses decomposed priors for ligand arms and scaffold, with atom-and-bond diffusion and validity guidance during sampling. On CrossDocked2020, it reports state-of-the-art performance, including a −8.39 Avg. Vina Dock score and 24.5% Success Rate.

  • Problem

    Existing diffusion-based SBDD methods omit bond modeling and treat ligand atoms alike, despite their different design roles and the large drug-like space to explore with limited high-quality complexes.

    Source: §1 · p. 1

  • Method

    DecompDiff uses data-dependent priors for ligand arms and scaffold, jointly diffuses atoms and bonds, and applies validity guidance during sampling.

    Source: §1 · p. 2

  • Results

    −8.39 Avg. Vina Dock score and 24.5% Success Rate: DecompDiff reports state-of-the-art performance on CrossDocked2020.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The authors conclude that decomposed priors improve training and sampling efficiency and binding affinity, while bond diffusion and validity guidance support more realistic molecules.

    Source: §5 · p. 9

    The model assumes protein atoms remain fixed as context.

    Source: Equivariant Network with Nodes and… · p. 5

Abstract

from arXiv · show

Designing 3D ligands within a target binding site is a fundamental task in drug discovery. Existing structured-based drug design methods treat all ligand atoms equally, which ignores different roles of atoms in the ligand for drug design and can be less efficient for exploring the large drug-like molecule space. In this paper, inspired by the convention in pharmaceutical practice, we decompose the ligand molecule into two parts, namely arms and scaffold, and propose a new diffusion model, DecompDiff, with decomposed priors over arms and scaffold. In order to facilitate the decomposed generation and improve the properties of the generated molecules, we incorporate both bond diffusion in the model and additional validity guidance in the sampling phase. Extensive experiments on CrossDocked2020 show that our approach achieves state-of-the-art performance in generating high-affinity molecules while maintaining proper molecular properties and conformational stability, with up to -8.39 Avg. Vina Dock score and 24.5 Success Rate. The code is provided at https://github.com/bytedance/DecompDiff

1 Introduction

DecompDiff uses ligand decomposition to address limitations of treating all atoms alike and omitting bonds in diffusion-based structure-based drug design. It combines decomposed priors, atom-and-bond generation, and sampling guidance, with reported state-of-the-art CrossDocked2020 results.

1 Introduction

  • Existing diffusion approaches omit bond modeling and treat the ligand as a whole, despite differing atom functions and a large drug-like space with limited high-quality complexes.Earlier autoregressive methods generate atoms and bonds iteratively but require a nontrivial generation order and can accumulate errors.Source: §1 · p. 1
  • In conventional design, arms contact the target while the scaffold positions them; lead optimization changes arms around a chosen scaffold, whereas scaffold hopping connects arms with a scaffold.The figure illustrates arm clusters with shape complementarity to local subpockets.Source: §1 · p. 2, Figure 1 · p. 2
  • DecompDiff introduces data-dependent decomposed priors over ligand arms and scaffold to reflect their natural roles in target interactions.The approach also diffuses bonds rather than assigning them only in post-processing.Source: §1 · p. 2
  • The model jointly diffuses atoms and bonds, while sampling guidance promotes scaffold–arm connections and avoids geometric clashes with the target.The stated aims are improved drug-likeness, synthesizability, and molecular validity.Source: §1 · p. 2
  • −8.39 Avg. Vina Dock score and 24.5% Success Rate: DecompDiff reports state-of-the-art performance on CrossDocked2020.The authors attribute the method's molecular validity guidance to promoting scaffold–arm connections and avoiding geometric clashes with the target.Source: §1 · p. 2

2 Related Work

Related work progresses from docking and shape-based ligand generation to models trained on target–ligand complexes. DecompDiff differs by using ligand decomposition and informative priors within diffusion-based design.

Structure-Based Drug Design

  • SBDD methods moved from docking-based optimization and shape-only generation toward directly modeling target–ligand correspondence, including density-grid variational autoencoders, autoregressive models, and diffusion models.Earlier diffusion approaches denoise atom types and positions sampled from a Gaussian prior, whereas DecompDiff decomposes ligands into arms and scaffold.Source: §2 · p. 2

Decomposed Molecular Generation

  • Fragment-based design includes scaffold hopping to connect existing arms and lead optimization to decorate a starting scaffold, but the cited methods require a reliable initial structure.DecompDiff instead supports de novo generation from scratch while retaining molecular decomposition.Source: §2 · pp. 2–3

Diffusion Models

  • Prior work on adaptive priors and closer-to-data distributions supports data-dependent diffusion priors, which this paper applies to complex 3D structure data.The passage characterizes such 3D prior design as less studied in previous work.Source: §2 · p. 3

3 Method

DecompDiff conditions ligand generation on a protein binding site and uses decomposed priors over ligand arms and scaffold. It combines atom and bond diffusion with guided sampling to address molecular connectivity and protein-ligand clashes.

3.1 Preliminaries

  • SBDD models ligand generation conditioned on a protein binding site, using a diffusion process that gradually adds noise and learns to recover molecules by denoising.The ligand is represented by atom positions and types, and the prior for sampling is induced by the forward process.Source: §3.1 · p. 3

3.2 Diffusion Model with Decomposed Priors

  • DecompDiff assigns Gaussian priors to ligand fragments representing arms and scaffold, with Proposition 3.1 bounding its minimum ELBO loss by the standard-prior loss when the prior aligns with the data.The fragments provide atom clusters whose positions are used to estimate the Gaussian priors.Source: §3.2 · pp. 3–4

3.3 Bond Diffusion and Model Architecture

  • Bond diffusion models bond types alongside atom coordinates and types, rather than relying on post-processing to infer bonds from generated coordinates.The bond dynamics cover bonds within arms or scaffold and between arms and scaffold.Source: §3.3 · p. 5

Equivariant Network with Nodes and Edges Update

  • The equivariant network maintains atom- and bond-level states, updates them through protein-ligand and ligand graphs, and keeps protein atoms fixed as context.Directional message passing updates bond states, while atom positions are updated for ligand atoms.Source: Equivariant Network with Nodes and… · p. 5

3.4 Validity Guidance

  • Sampling-time validity guidance promotes arms-scaffold connections and avoids protein-surface clashes, using a 1.2–1.9 Å range to represent bond lengths.The method describes separate guidance for connections and clashes.Source: §3.4 · pp. 5–6

4 Experiments

Experiments evaluate DECOMPDIFF on CrossDocked2020 for molecular conformation, binding affinity, and molecular properties. The reported results and ablations examine decomposed priors, bond diffusion, validity guidance, and performance with fewer diffusion steps.

4.1 Experimental Setup

  • The evaluation uses CrossDocked2020 complexes and compares representative baselines, measuring distance-distribution fit, Vina affinity metrics, QED, SA, diversity, and Success Rate.The dataset is refined for docking-pose quality and protein diversity; 100,000 complexes are selected for training and 100 novel proteins for testing.Source: §4.1 · pp. 6–7

4.2 Main Results

  • 0.23 and 0.06: DECOMPDIFF has the lowest JSD for carbon-carbon bond-distance and all-atom pairwise-distance distributions, fitting reference distances better than Pocket2Mol and TargetDiff.The baselines show biased carbon-carbon distance estimates and long-range distance estimates, especially at 10–12 Å.Source: Figure 3 · p. 6, §4.2 · p. 7
  • 6% and 12%: DECOMPDIFF exceeds TargetDiff in Avg. and Med. High Affinity; Vina Min and Vina Dock improve around 0.40, while 24.5% Success Rate is comparable to Pocket2Mol and exceeds TargetDiff.DECOMPDIFF trails Pocket2Mol in QED and SA.Source: §4.2 · p. 7, Table 3 · p. 8

4.3 Ablation Studies

  • With fewer diffusion steps, DECOMPDIFF achieves better validation loss than TargetDiff and can generate high-quality ligands; at 200 sampling steps, TargetDiff examples tend to contain unrealistic local structures such as messy rings.The comparison considers validation loss, QED, SA, and Vina Score across step settings.Source: §4.3 · pp. 7–8, Figure 4 · p. 8, Figure 5 · p. 8
  • Pocket Prior performs better than Ref Prior in the atom-only comparison, and both outperform TargetDiff when using appropriate prior knowledge.Pocket Prior uses target subpockets, while Ref Prior estimates a Gaussian prior from the reference molecule; the authors suggest the reference molecule may not be ideal for the target.Source: Table 4 · p. 8, §4.3 · p. 8
  • Bond diffusion substantially improves QED, SA, and Success Rate, supporting simultaneous atom-and-bond generation as a way to produce more reasonable 2D structures.Source: Table 5 · p. 9, §4.3 · p. 9
  • Combining arm-scaffold connection drift and clash drift improves both complete rate and Vina score compared with using the guidance drifts separately.Source: Table 6 · p. 9, §4.3 · p. 9

5 Conclusions

DecompDiff decomposes drug space and uses informative priors, joint atom-and-bond generation, and sampling-time validity guidance. The authors report improved training and sampling efficiency and Vina binding affinity, with more realistic molecules as a benefit of joint generation.

5 Conclusions

  • Informative decomposed priors improve training and sampling efficiency and significantly improve binding affinity measured by Vina.The priors reflect the decomposition of drug space.Source: §5 · p. 9
  • DecompDiff simultaneously generates atom coordinates, atom types, and bond types to produce more realistic molecules.Source: §5 · p. 9
  • Sampling-time validity guidance is described as simple but effective under the decomposition framework.Source: §5 · p. 9

A The Generation Details of Pocket Prior Centers

The method identifies candidate pocket regions from AlphaSpace atoms, clusters beta-atoms, and filters the resulting clusters to assign arm and scaffold prior centers. Cluster scores and graph-geodesic distances guide those assignments.

A The Generation Details of Pocket Prior Centers

  • Alpha-atoms represent virtual points contacting the protein surface, while beta-atoms group alpha-atoms and receive scores indicating pocket ligandability.Only atoms within 10 Å of the reference ligand or region of interest are considered.Source: A The Generation Details of Pocket… · p. 12
  • Beta-atoms within 5.5 Å are grouped into clusters, whose scores are the average beta-scores of their member atoms.Source: A The Generation Details of Pocket… · p. 12
  • Filtering selects promising clusters as arm priors and a cluster supporting scaffold-arm geometry as the scaffold prior.Figure 6 depicts the alpha-atoms, beta-atoms and scores, alpha-graph, beta-clusters, filtered clusters, and resulting priors.Source: A The Generation Details of Pocket… · pp. 12–13, Figure 6 · p. 12
  • Beta-cluster distances for filtering are graph-geodesic distances on the alpha-graph, with unconnected centers assigned infinite distance.The alpha-graph connects alpha-atoms to other alpha-atoms and beta-cluster centers.Source: A The Generation Details of Pocket… · p. 12

B The Details of Arms-Scaffold Fragmentation

Training ligands are fragmented into arms and scaffold by matching terminal ligand fragments to protein subpockets and clustering the fragments. The resulting arms are terminal fragments, while the scaffold is assigned using distance from arm centers.

B The Details of Arms-Scaffold Fragmentation

  • AlphaSpace2 subpockets are extracted as potential arm and scaffold clustering centers, and BRICS decomposes ligands into fragments for assignment.The method sets beta-atom and pocket clustering distances to 1.6 Å and 6.0 Å, respectively.Source: B The Details of Arms-Scaffold… · p. 13
  • Terminal fragments are assigned to subpockets, their centroids and remaining subpockets form arm centers, and the farthest fragment centroid defines the scaffold center.Nearest-neighbor clustering is performed while keeping arm fragments terminal.Source: B The Details of Arms-Scaffold… · p. 13
  • The procedure extracts arms and scaffold from ligands; Figure 7 illustrates fragmented examples with their original ligands and protein pockets.Source: B The Details of Arms-Scaffold… · p. 13, Figure 7 · p. 14

C Theoretical Analysis

The analysis compares decomposed priors with a standard Gaussian prior under stated assumptions. It concludes that the decomposed prior yields a lower minimum negative ELBO in the analyzed setting, and discusses covariance alignment as a separate potential advantage.

C Theoretical Analysis

  • The analysis treats decomposed priors as multiple Gaussian distributions with different means, rather than one distribution with a single mean.Source: C Theoretical Analysis · p. 14
  • Under the stated assumptions, the minimum negative ELBO with decomposed priors is lower than with standard priors.The derivation focuses on comparing the expected denoising-error term under the two priors.Source: C Theoretical Analysis · pp. 14–16
  • For the covariance aspect, the paper cites tighter ELBOs when prior covariance aligns with data covariance and notes that MLE-derived decomposed priors are expected to align better.This claim relies on Lee et al. (2021) and is stated under assumptions.Source: C Theoretical Analysis · pp. 14, 16

D Derivation of Validity Guidance

The section describes validity drifts, input featurization, and the model configuration used for decomposed generation. Protein and ligand atoms carry region and structural features, while graph attention updates atoms, bonds, and positions.

D Derivation of Validity Guidance

  • The arms-scaffold drift promotes their connection, while the clash drift guides generated molecules outside the protein surface.To derive both drifts, the authors use ξ − max(0, ξ − y) as a surrogate for the discontinuous indicator and find constant ζ1, ζ2, ζ3 work well.Source: D Derivation of Validity Guidance · p. 17

E.1 Featurization

  • Protein atoms include element, amino-acid, backbone, and arm/scaffold features; ligand atoms include element and predetermined arm/scaffold features.Protein atoms within 10 Å of any arm prior center are labeled arm-region atoms; others are labeled scaffold-region atoms.Source: §E.1 · p. 17
  • The model uses ligand and protein k-nearest-neighbor message passing with k = 32, plus a fully connected ligand graph.Its six blocks combine atom, bond, and position updates with graph attention; each block has hidden dimension 128 and 16 heads.Source: §E.1 · p. 17, §E.2 · p. 18

E.2 Model Details

  • Atom coordinates use a sigmoid β schedule, while atom and bond types use a cosine β schedule across 1,000 diffusion steps.For coordinates, β1 = 1e-7 and βT = 2e-3; the cosine schedule uses s = 0.01.Source: §E.2 · p. 18

E.3 Training Details

Training uses Adam with a scheduled learning-rate decay, protein-coordinate noise augmentation, and atom and bond loss reweighting. The model converged in 300k steps within 36 hours on one NVIDIA GeForce GTX A100 GPU.

E.3 Training Details

  • The model was trained with Adam at an initial learning rate of 0.001, batch size 4, and gradient-norm clipping at 8.The authors set Adam betas to (0.95, 0.999) and multiply atom-type and bond-type losses by α = 100.Source: §E.3 · p. 18
  • Training adds Gaussian noise with standard deviation 0.1 to protein atom coordinates and decays the learning rate by a factor of 0.6.Decay occurs after 10 consecutive evaluations without validation-loss improvement, with a minimum learning rate of 1e-6; evaluation runs every 2,000 steps.Source: §E.3 · p. 18
  • The model converged within 36 hours and 300k steps on one NVIDIA GeForce GTX A100 GPU.Source: §E.3 · p. 18

F.1 Full Evaluation Results

The authors report that bond diffusion improves QED and SA across model variants, while bond diffusion and Opt Prior together yield a better balance of molecular rationality and binding affinity. Bond diffusion and validity guidance increase sampling time.

F.1 Full Evaluation Results

  • Bond diffusion consistently improves QED and SA across the evaluated model variants.Variants with bond diffusion and Opt Prior show a better balance between 2D structure rationality and binding affinity, leading to a higher success rate.Source: §F.1 · p. 18

F.2 Time Complexity

  • The increase in training time is mainly due to bond diffusion, while the decomposed prior has negligible impact on training time.Source: §F.2 · pp. 18–19, Table 8 · p. 19
  • Bond diffusion increases sampling time to 1.62x TargetDiff, and validity guidance adds a further 10%.Generating 100 valid molecules takes DecompDiff 5570s without validity guidance and 6189s with it, on average.Source: §F.2 · p. 19

G Examples of Generated Ligands

The examples compare reference ligands with ligands generated by TargetDiff and DecompDiff across protein targets. The authors report that generated ligands can occupy more of concave binding sites and achieve better Vina scores while retaining reasonable QED and SA.

G Examples of Generated Ligands

  • The authors report that DecompDiff-generated ligands can occupy more space in concave target binding sites due to the decomposed prior.Source: G Examples of Generated Ligands · p. 19
  • The authors report that bond diffusion and validity guidance promote ligand quality, with better Vina scores alongside reasonable QED and SA.Source: G Examples of Generated Ligands · p. 19
  • Figure 8 colors reference-ligand carbons green, TargetDiff carbons yellow, and DecompDiff carbons cyan; each row shows one protein.Source: Figure 8 · p. 20

Related papers

Loading 2403.07902v1…