CAGenMol: Condition-Aware Diffusion Language Model for Goal-Directed Molecular Generation

Yanting Li; Zhuoyang Jiang; Enyan Dai; Lei Wang; Wen-Cai Ye; Li Liu

cs.LGApr 13, 2026

CAGenMol: Condition-Aware Diffusion Language Model for Goal-Directed Molecular Generation

Authors: Yanting Li, Zhuoyang Jiang, Enyan Dai, Lei Wang, Wen-Cai Ye, Li Liu

Organizations: The Hong Kong University of Science and Technology (Guangzhou) · Jinan University, Guangzhou

Abstract

Goal-directed molecular generation requires satisfying heterogeneous constraints such as protein--ligand compatibility and multi-objective drug-like properties, yet existing methods often optimize these constraints in isolation, failing to reconcile conflicting objectives (e.g., affinity vs. safety), and struggle to navigate the non-differentiable chemical space without compromising structural validity. To address these challenges, we propose CAGenMol, a condition-aware discrete diffusion framework over molecular sequences that formulates molecular design as conditional denoising guided by heterogeneous structural and property signals. By coupling discrete diffusion with reinforcement learning, the model aligns the generation trajectory with non-differentiable objectives while preserving chemical validity and diversity. The non-autoregressive nature of diffusion language model further enables iterative refinement of molecular fragments at inference time. Experiments on structure-conditioned, property-conditioned, and dual-conditioned benchmarks demonstrate consistent improvements over state-of-the-art methods in binding affinity, drug-likeness, and success rate, highlighting the effectiveness of our framework.

Figures & tables

Figure 1: Overview of CAGenMol. UCA encodes either protein-pocket structure or target properties, which guides a condition-aware masked diffusion model on SAFE sequences. The model is further trained with Step-PPO, and perform inference with EFO refinement.

Methods	Vina Dock ( $\downarrow$ )	High Affinity ( $\uparrow$ )	QED ( $\uparrow$ )	SA ( $\uparrow$ )	Diversity ( $\uparrow$ )	Success Rate ( $\uparrow$ )
Reference Set	-7.45	-	0.48	0.73	-	25.0%
TargetDiff	-7.80	58.1%	0.48	0.58	0.72	10.5%
FLAG	-5.63	-	0.49	0.70	0.70	14.1%
Pocket2Mol	-7.15	48.4%	0.56	0.74	0.69	24.4%
DecompDiff	-8.39	64.4%	0.45	0.61	0.68	24.5%
MolCRAFT	-9.25	59.1%	0.46	0.62	0.61	36.1%

Table 1: Comparison on the CrossDocked2020 benchmark. We report the average metrics across 100 test pockets. The best results are highlighted in bold , and the second best are underlined .

Variant	Vina Dock ( $\downarrow$ )	QED ( $\uparrow$ )	SA ( $\uparrow$ )	Diversity ( $\uparrow$ )	Success Rate ( $\uparrow$ )
SFT (Base)	-6.47	0.53	0.77	0.88	14.3%
SFT (w/o Attn.)	-6.60	0.55	0.76	0.86	17.5%
SFT (w/o Phys.)	-6.55	0.54	0.78	0.87	15.8%
SFT (Full UCA)	-6.61	0.58	0.77	0.86	19.2%
Full Model (Step-PPO)	-8.41	0.70	0.89	0.80	69.7%

Table 2: Ablation study of UCA components and RL training. All variants except the final one are trained solely with Supervised Fine-Tuning (SFT).

Model	1IEP	3EML	3NY8	4RLU	4UNN	5MO4	7L11	Avg
3DSBDD	-9.05 $\pm$ 0.38	-10.02 $\pm$ 0.15	-10.10 $\pm$ 0.24	-9.80 $\pm$ 0.55	-8.23 $\pm$ 0.30	-8.71 $\pm$ 0.45	-8.47 $\pm$ 0.18	-9.20
AutoGrow4	-13.23 $\pm$ 0.11	-13.03 $\pm$ 0.09	-11.70 $\pm$ 0.00	-11.20 $\pm$ 0.00	-11.14 $\pm$ 0.12	-10.38 $\pm$ 0.27	-8.84 $\pm$ 0.33	-11.36
Pocket2Mol	-10.17 $\pm$ 0.53	-12.25 $\pm$ 0.27	-11.89 $\pm$ 0.16	-10.57 $\pm$ 0.12	-12.20 $\pm$ 0.34	-10.07 $\pm$ 0.62	-9.74 $\pm$ 0.38	-10.98
PocketFlow	-12.49 $\pm$ 0.70	-9.25 $\pm$ 0.29	-8.56 $\pm$ 0.35	-9.65 $\pm$ 0.25	-7.90 $\pm$ 0.78	-7.80 $\pm$ 0.42	-8.35 $\pm$ 0.31	-9.14
ResGen	-10.97 $\pm$ 0.29	-9.25 $\pm$ 0.95	-10.96 $\pm$ 0.42	-11.75 $\pm$ 0.42	-9.41 $\pm$ 0.23	-10.34 $\pm$ 0.39	-8.74 $\pm$ 0.24	-10.20
DST	-10.95 $\pm$ 0.57	-10.67 $\pm$ 0.24	-10.54 $\pm$ 0.22	-10.88 $\pm$ 0.37	-9.71 $\pm$ 0.19	-10.03 $\pm$ 0.36	-8.33 $\pm$ 0.41	-10.16

Table 3: Top-10 average docking scores on additional benchmark(lower is better).

Figure 2: Histograms of Distribution shift under ADMET constraints for three settings.

Setting	Property	De novo	Step-PPO	Step-PPO + EFO	Target
Setting 1	HIA	0.97	0.98	0.99	$\uparrow$ (1)
	BBB	0.53	0.89	0.89	$\uparrow$ (1)
	Ames	0.37	0.21	0.15	$\downarrow$ (0)
Setting 2	CYP3A4sub	0.48	0.64	0.69	$\uparrow$ (1)
	LogP	2.3	3.4	3.5	[3,5]
	DILI	0.55	0.30	0.29	$\downarrow$ (0)

Table 4: ADMET evaluation results under different settings.

Methods	Vina Dock ( $\downarrow$ )	QED ( $\uparrow$ )	SA ( $\uparrow$ )	Ames ( $\downarrow$ )
Pocket2Mol	-10.38	0.71	0.70	0.54
TargetDiff	-10.80	0.39	0.51	0.44
FLAG	-6.38	0.60	0.67	0.36
MolCRAFT	-11.33	0.43	0.66	0.49
DecompDiff	-12.33	0.26	0.54	0.57
w/o Ames Condition	-10.13	0.83	0.88	0.34

Table 5: Comparison under dual-conditioned generation on the 3o96_A pocket.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 3: Case Study. Two pockets are shown. For each pocket, we visualize the reference ligand and two CAGenMol-generated ligands with their docking poses in the pocket.

Figure 4: Setting 1 CNS Drugs

Figure 5: Setting 2 Hepatic Drugs

Figure 6: Setting 3 Peripheral Drugs

Method	Vina Dock ( $\downarrow$ )	QED ( $\uparrow$ )	SA ( $\uparrow$ )
SFT(Full)+Step-PPO	-10.13	0.83	0.88
SFT(Base)+Step-PPO	-9.97	0.84	0.88
Step-PPO only	-9.49	0.84	0.89
EFO only (3 iters)	-7.66	0.70	0.75
EFO only (5 iters)	-8.02	0.69	0.77

Appendix

Table 6: Ablation of initialization, RL fine-tuning, and EFO on pocket 3o96_A.

Method	Time (s)
3D-SBDD Luo et al. (2021)	15986.4
Pocket2Mol Peng et al. (2022)	2827.3
GraphBP Liu et al. (2022)	1162.8
TargetDiff Guan et al. (2023)	3428.0
DecompDiff Guan et al. (2024)	6189.0
DiffSBDD Schneuing et al. (2024)	629.9

Appendix

Table 7: Inference time (seconds) for generating 100 molecules per pocket. Lower is better.

Explore similar work

MolDA: Molecular Understanding and Generation via Large Language Diffusion Model
Apr 6, 2026Seohyeon Shin, HanJun Choi, Jun-Hyung Park +2Multimodal Large Language Models Molecular Graph Generation
Gwangju Institute of Science and Technology, Gwangju, South Korea · Hankuk University of Foreign Studies, Seoul, South Korea
Fine-Tuning Diffusion Models for Molecular Generation via Reinforcement Learning and Fast Sampling
May 31, 2026Guang Lin, Shikui Tu, Lei XuMolecular Optimization Diffusion Model Fine-Tuning
Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai, China · Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ), Shenzhen, Guangdong, China
One Sequence, Many Decodings: CAGenMol-2 Recasts Drug Design as Masked Molecular Inference
Sep 28, 2026Yanting Li, Enyan Dai, Lei Wang +2Molecular Optimization Masked Diffusion Models
The Hong Kong University of Science and Technology (Guangzhou) · Jinan University, Guangzhou

Apr 6, 2026J/K move · Enter open · S save

MolDA: Molecular Understanding and Generation via Large Language Diffusion Model

Seohyeon Shin, HanJun Choi, Jun-Hyung Park, Hong Kook Kim+1

Gwangju Institute of Science and Technology, Gwangju, South Korea · Hankuk University of Foreign Studies, Seoul, South Korea

Multimodal Large Language Models Molecular Graph Generation Masked Diffusion Models Molecular Representation Learning

Open paper AskPDF arXiv

Large Language Models (LLMs) have significantly advanced molecular discovery, but existing multimodal molecular architectures fundamentally rely on autoregressive (AR) backbones. This strict left-to-right inductive bias is sub-optimal for generating chemically valid molecules, as it struggles to account for non-local global constraints (e.g., ring closures) and often accumulates structural errors during sequential generation. To address these limitations, we propose MolDA (Molecular language model with masked Diffusion with mAsking), a novel multimodal framework that replaces the conventional AR backbone with a discrete Large Language Diffusion Model. MolDA extracts comprehensive structural representations using a hybrid graph encoder, which captures both local and global topologies, and aligns them into the language token space via a Q-Former. Furthermore, we mathematically reformulate Molecular Structure Preference Optimization specifically for the masked diffusion. Through bidirectional iterative denoising, MolDA ensures global structural coherence, chemical validity, and robust reasoning across molecule generation, captioning, and property prediction.

Abstract

Explore similar work

MolDA: Molecular Understanding and Generation via Large Language Diffusion Model

Fine-Tuning Diffusion Models for Molecular Generation via Reinforcement Learning and Fast Sampling

One Sequence, Many Decodings: CAGenMol-2 Recasts Drug Design as Masked Molecular Inference