FuseDiff: Symmetry-Preserving Joint Diffusion for Dual-Target Structure-Based Drug Design

Jianliang Wu; Anjie Qiao; Zhen Wang; Zhewei Wei; Sheng Chen

cs.LGMar 5, 2026

FuseDiff: Symmetry-Preserving Joint Diffusion for Dual-Target Structure-Based Drug Design

Authors: Jianliang Wu, Anjie Qiao, Zhen Wang, Zhewei Wei, Sheng Chen

Organizations: School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China · School of Computer Science and Engineering, Guangdong Province Key Laboratory of Computational Science, Sun Yat-sen University, Guangzhou, China · Gaoling School of Artificial Intelligence, Renmin University of China, Beijing, China · School of Life Science, Tsinghua University, Beijing, China

Abstract

Dual-target structure-based drug design aims to generate a single ligand together with two pocket-specific binding poses, each compatible with a corresponding target pocket, enabling polypharmacological therapies with improved efficacy and reduced resistance. Existing approaches typically rely on staged pipelines, which either decouple the two poses via conditional-independence assumptions or enforce overly rigid correlations, and therefore fail to jointly generate two target-specific binding modes. To address this, we propose FuseDiff, an end-to-end diffusion model that jointly generates a ligand molecular graph and two pocket-specific binding poses conditioned on both pockets. FuseDiff features a message-passing backbone with Dual-target Local Context Fusion (DLCF), which fuses each ligand atom's local context from both pockets to enable expressive joint modeling while preserving the desired symmetries. Together with explicit bond generation, FuseDiff enforces topological consistency across the two poses under a shared graph while allowing target-specific geometric adaptation in each pocket. To support principled training and evaluation, we derive a dual-target training set and use an independent held-out test set for evaluation. Experiments on the benchmark and a real-world dual-target system show that FuseDiff achieves state-of-the-art docking performance and enables the first systematic assessment of dual-target pose quality prior to docking-based pose search.

Figures & tables

Figure 1. Overview of FuseDiff . FuseDiff defines a conditional diffusion model over $M=(G,X_{1},X_{2})$ given a pocket pair $P=(P_{1},P_{2})$ , where the two target-specific complexes share the same molecular graph $G$ while maintaining separate poses in $P_{1}$ and $P_{2}$ . A fixed forward (diffusion) process $q$ (red arrows) corrupts $(G^{(0)},X_{1}^{(0)},X_{2}^{(0)})$ into $(G^{(T)},X_{1}^{(T)},X_{2}^{(T)})$ for training. For generation, the learned reverse (denoising) process $p_{\theta}$ (green arrows) starts from $M^{(T)}\!\sim\!p_{\text{base}}(M^{(T)}\mid P)$ and iteratively samples $M^{(t-1)}$ using a denoiser $f_{\theta}$ . At step $t$ , $f_{\theta}$ takes $(G^{(t)},X_{1}^{(t)},X_{2}^{(t)})$ , $t$ , and $P$ as input and predicts $(\widehat{X}_{1}^{(0)},\widehat{X}_{2}^{(0)},\widehat{V}^{(0)},\widehat{B}^{(0)})$ to parameterize the reverse transition.

Figure 2. Dual-target Local Context Fusion (DLCF) within one denoising layer. (a) Two pocket–ligand point-cloud graphs for

P_{1}

and

P_{2}

, each containing a fully-connected ligand subgraph. (b) The augmented graph induced from the two point-cloud graphs, on which dual-target message passing updates node/edge embeddings. (c) Coordinate update block that separately updates ligand node coordinates in the two point-cloud graphs (with pocket atoms fixed). dlcf

Methods	Metrics
	QED( $\uparrow$ )	SA( $\uparrow$ )	Div.( $\uparrow$ )	logP	LPSK( $\uparrow$ )
Ref.	0.51	0.77	0.74	2.89	4.50
TargetDiff	0.54	0.55	0.67	2.32	4.69
CompDiff	0.59	0.57	0.69	3.48	4.74
DualDiff	0.59	0.58	0.69	3.54	4.88
LinkerNet	0.61	0.61	0.29	3.21	4.71

Table 1. Molecular property statistics of generated ligands and reference ligands (Avg.). (

\uparrow

) indicates higher is better. Top 2 results (excluding Ref.) are highlighted with bold text and underlined text , respectively.

Methods	Metrics
	P1-Vina Score ( $\downarrow$ )			P2-Vina Score ( $\downarrow$ )			P1-Vina Min ( $\downarrow$ )		P2-Vina Min ( $\downarrow$ )
	Avg.	T-Avg.	Med.	Avg.	T-Avg.	Med.	Avg.	Med.	Avg.	Med.
Ref.	-7.92	-7.97	-7.83	-7.86	-7.92	-7.92	-8.02	-7.89	-7.93	-7.96
TargetDiff	-6.36	-7.03	-7.26	-	-	-	-7.78	-7.95	-	-
CompDiff	OOR	-1.68	-3.46	-	-	-	OOR	-6.26	-	-
DualDiff	OOR	-2.47	-3.63	-	-	-	OOR	-6.27	-	-

Table 2. Dual-pose quality before docking-based pose search. “–” denotes not applicable (N/A) when a method does not produce target-specific binding poses for the corresponding pocket, and “OOR” denotes out-of-range values caused by extreme/invalid poses that fall outside the reported scale.

Methods	Metrics
	P1-Vina Dock ( $\downarrow$ )			P2-Vina Dock ( $\downarrow$ )			Max Vina Dock ( $\downarrow$ )		Dual High Aff. ( $\uparrow$ )
	Avg.	Med.	LBE	Avg.	Med.	LBE	Avg.	Med.	Avg.	Med.
Ref.	-8.23	-7.98	-0.38	-8.14	-8.23	-0.37	-6.15	-6.73	-	-
TargetDiff	-8.95	-8.86	-0.35	-7.14	-7.99	-0.28	-6.88	-7.36	31.1%	23.4%
CompDiff	-8.75	-8.79	-0.36	-8.73	-8.65	-0.36	-7.92	-7.98	39.8%	33.0%
DualDiff	-8.80	-8.82	-0.37	-8.72	-8.78	-0.37	-7.99	-8.09	40.5%	33.4%

Table 3. Docking-based interaction evaluation and dual-target compatibility. Top 2 results (excluding Ref.) are highlighted with bold text and underlined text , respectively.

Figure 3. QED–SA density of filtered molecules ( $\mathrm{LPSK}\!=\!5$ , $\mathrm{logP}\!\in\![-0.4,5.6]$ ). KDE heatmaps show $\mathrm{QED}$ (x-axis) versus $\mathrm{SA}$ (y-axis) for FuseDiff (left) and DualDiff (right). The upper-right region corresponds to DL ( $\mathrm{QED}\!>\!0.6$ , $\mathrm{SA}\!>\!0.7$ ), where FuseDiff shows higher density, consistent with its larger DL yield (78 vs. 15).

Figure 4. Vina Score and Vina Dock distributions on GSK3

\beta

and JNK3. Left: FuseDiff Vina Score on generated poses. Right: Vina Dock after Vina conformational search for FuseDiff and DualDiff .

Panel of Figure 5 — Figure 5. Qualitative visualization in two pockets. Top: GSK3 $\beta$ ; bottom: JNK3. Left: a reference dual-target ligand with poses obtained by Vina pose search. Middle/Right: two FuseDiff samples.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Methods	Metrics
	QED( $\uparrow$ )	SA( $\uparrow$ )	Div.( $\uparrow$ )	logP	LPSK( $\uparrow$ )
Ref.	0.51	0.77	0.74	2.89	4.50
TargetDiff -CD	0.53	0.54	0.70	2.48	4.63
TargetDiff -BN2	0.54	0.55	0.67	2.32	4.69
CompDiff -CD	0.57	0.58	0.72	3.67	4.54
CompDiff -BN2	0.59	0.57	0.69	3.48	4.74

Appendix

Table 4. Checkpoint comparison on DDF (molecular properties). We report molecule-level metrics for TargetDiff , CompDiff , and DualDiff when using the official CrossDocked2020 pretrained checkpoint ( -CD ) versus our BindingNet v2 pretrained checkpoint ( -BN2 ).

Methods	Metrics
	P1-Vina Dock ( $\downarrow$ )			P2-Vina Dock ( $\downarrow$ )			Max Vina Dock ( $\downarrow$ )		Dual High Aff. ( $\uparrow$ )
	Avg.	Med.	LBE	Avg.	Med.	LBE	Avg.	Med.	Avg.	Med.
Ref.	-8.23	-7.98	-0.38	-8.14	-8.23	-0.37	-6.15	-6.73	-	-
TargetDiff -CD	-8.89	-8.91	-0.37	-7.23	-7.68	-0.28	-6.84	-6.91	30.8%	23.3%
TargetDiff -BN2	-8.95	-8.86	-0.35	-7.14	-7.99	-0.28	-6.88	-7.36	31.1%	23.4%
CompDiff -CD	-8.35	-8.46	-0.36	-8.35	-8.37	-0.36	-7.53	-7.84	37.9%	31.4%

Appendix

Table 5. Checkpoint comparison on DDF (docking-based affinity). Docking metrics on the two pockets (P1/P2) and their aggregate (Max Vina Dock), together with Dual High Affinity, for baselines using -CD versus -BN2 pretrained TargetDiff checkpoints.

Structure-Based Drug Design

Structure-based drug design uses protein structures and binding-site geometry to find, rank, and create molecules that bind therapeutic targets.

See all 44 papers

Fused Bayesian Flow Networks for Dual-Target Molecular Design
Aug 2, 2026Jingyuan Zhou, Shikui Tu, Lei XuStructure-Based Drug Design Molecular Generation
School of Computer Science, Shanghai Jiao Tong University · Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ)
Don't Retrain, Just Reuse: Recovering Dual-Target Molecules from Single-Target Diffusion Models
May 25, 2026Qingyuan Zeng, Pengxiang Cai, Zixin Guan +5Molecular Optimization Molecular Generation
The Hong Kong University of Science and Technology (Guangzhou) · Guangzhou University of Chinese Medicine
Co-Diffusion: An Affinity-Aware Two-Stage Latent Diffusion Framework for Generalizable Drug-Target Affinity Prediction
Mar 11, 2026Yining Qian, Pengjie Wang, Yixiao Li +4Drug Discovery Latent Diffusion Models
School of Computer Science and Engineering, Northeastern University, Shenyang 110819, China · College of Information Science and Engineering, Northeastern University, Shenyang 110819, China · Westlake University, Hangzhou 310000, China +2

Aug 2, 2026J/K move · Enter open · S save

Fused Bayesian Flow Networks for Dual-Target Molecular Design

Jingyuan Zhou, Shikui Tu, Lei Xu

School of Computer Science, Shanghai Jiao Tong University · Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ)

Structure-Based Drug Design Molecular Generation Conditional Generative Modeling Drug Discovery

Open paper AskPDF arXiv

Dual-target drug design aims to generate 3D molecules that can simultaneously interact with two target proteins, offering a promising route for discovering polypharmacological compounds against complex diseases. While recent generative models have shown encouraging performance in single-target drug design, existing dual-target approaches either focus on sequence generation or introduce an additional predictive drift term into the diffusion-based generative trajectory, which limits their ability to fully integrate feature information from both targets. We propose FusedBFN, a fused Bayesian flow network (BFN) for dual-target molecular design. FusedBFN formulates dual-target generation as distribution fusion in a unified continuous parameter space and employs a product-of-experts formulation to incorporate dual-target information throughout the generative process. To address the scarcity of dual-target structural data, we leverage a pretrained target-aware BFN model as the shared backbone. We further introduce a chemically aware prior-based alignment method and a prior-free pocket alignment strategy to construct aligned dual-target contexts. Extensive experiments demonstrate that FusedBFN generates molecules with strong binding affinity toward dual targets while maintaining favorable molecular properties.

Abstract

Structure-Based Drug Design

Explore similar work

Fused Bayesian Flow Networks for Dual-Target Molecular Design

Don't Retrain, Just Reuse: Recovering Dual-Target Molecules from Single-Target Diffusion Models

Co-Diffusion: An Affinity-Aware Two-Stage Latent Diffusion Framework for Generalizable Drug-Target Affinity Prediction