NEAT-POCKET: Pocket-Conditioned Autoregressive 3D Molecular Generation with a Neighborhood-Guided Set Transformer

Roxane Axel Jacob; Daniel Rose; Thierry Langer; Johannes Kirchmair

cs.LGSep 4, 2026

NEAT-POCKET: Pocket-Conditioned Autoregressive 3D Molecular Generation with a Neighborhood-Guided Set Transformer

Authors: Roxane Axel Jacob, Daniel Rose, Thierry Langer, Johannes Kirchmair

Organizations: Department of Pharmaceutical Sciences, Faculty of Life Sciences, University of Vienna, Josef-Holaubek-Platz 2, 1090 Vienna, Austria. · Christian Doppler Laboratory for Molecular Informatics in the Biosciences, Department of Pharmaceutical Sciences, Faculty of Life Sciences, Josef-Holaubek-Platz 2, 1090 Vienna, Austria. · Vienna Doctoral School of Pharmaceutical, Nutritional and Sport Sciences (PhaNuSpo), Josef-Holaubek-Platz 2, 1090 Vienna, Austria.

Abstract

AI-driven de novo molecular design offers a promising route to accelerate early-stage drug discovery by generating novel ligands directly within target protein binding pockets. We present NEAT-POCKET, a pocket-conditioned extension of the autoregressive NEAT model for 3D molecular generation. NEAT-POCKET generates molecules atom by atom in protein pocket environments while preserving atom permutation invariance and explicitly modeling hydrogen atoms. Benchmarks on the CrossDocked and SPINDR datasets show that NEAT-POCKET achieves competitive structure-based generation performance while sampling substantially faster than existing baselines. Beyond full-molecule generation, NEAT-POCKET naturally enables pocket-conditioned fragment completion, a task directly relevant to lead optimization and scaffold elaboration. These results position NEAT-POCKET as a fast, flexible, and practical framework for structure-based drug design.

Figures & tables

Fig. 1: Overview of the architecture of NEAT-POCKET.

	PB valid	PC clashes	PC SE	Vina score	Vina min	Mol. weight	Rank	Sampling time
	[%] $\uparrow$	[1] $\downarrow$	[kcal/mol] $\downarrow$	[kcal/mol] $\downarrow$	[kcal/mol] $\downarrow$	[g/mol]	[1] $\downarrow$	[s] $\downarrow$
CrossDocked	94.0 $\pm$ 4.7	6.9 $\pm$ 1.1	63 $\pm$ 23	-7.1 $\pm$ 0.6	-7.3 $\pm$ 0.5	332 $\pm$ 29	-	-
Pocket2Mol	71.1 $\pm$ 3.7	6.9 $\pm$ 0.9	65 $\pm$ 36	-5.4 $\pm$ 0.5	-6.9 $\pm$ 0.5	223 $\pm$ 21	3.00	845 $\pm$ 838
TargetDiff	46.4 $\pm$ 3.3	10.8 $\pm$ 1.4	1003 $\pm$ 378	-6.5 $\pm$ 0.4	-7.5 $\pm$ 0.4	324 $\pm$ 20	3.39	722 $\pm$ 201
DiffSBDD	49.6 $\pm$ 2.6	14.3 $\pm$ 2.0	891 $\pm$ 271	-3.6 $\pm$ 0.8	-6.3 $\pm$ 0.4	296 $\pm$ 14	3.32	88 $\pm$ 30
DrugFlow	61.7 $\pm$ 3.7	9.2 $\pm$ 1.4	255 $\pm$ 113	-5.8 $\pm$ 0.5	-6.9 $\pm$ 0.4	309 $\pm$ 20	2.81	112 $\pm$ 35

Table 1: Performance of NEAT-POCKET vs. other pocket-conditioned 3D molecule generators trained on CrossDocked or SPINDR (best in bold, second-best underlined; sampling times reported for the generation of 100 molecules)

Fig. 2: Average values of general physicochemical properties of molecules generated by Pocket2Mol, TargetDiff, DiffSBDD, DrugFlow and NEAT-POCKET trained on CrossDocked — deviation from CrossDocked.

Fig. 3: Distribution of ring sizes (top) and atom types (bottom) of molecules generated by Pocket2Mol, TargetDiff, DiffSBDD, DrugFlow and NEAT-POCKET trained on CrossDocked — deviation from CrossDocked.

Fig. 4: Representative examples of pocket-conditioned 3D molecule generation with NEAT-POCKET. Molecules were generated in the binding pocket of DNA photolyase from the cyanobacterium Anacystis nidulans (PDB: 1QNF). (A) Experimentally resolved reference ligand (7,8-didemethyl-8-hydroxy-5-deazariboflavin). (B–D) Representative generated ligands.

CFG	PB valid	PC clashes	PC SE	Vina score	Vina min	Mol. weight
	[%] $\uparrow$	[1] $\downarrow$	[kcal/mol] $\downarrow$	[kcal/mol] $\downarrow$	[kcal/mol] $\downarrow$	[g/mol]
0.0	71.0 $\pm$ 3.2	8.3 $\pm$ 1.0	155 $\pm$ 90	-4.4 $\pm$ 0.3	-6.5 $\pm$ 0.4	273 $\pm$ 19
0.5	69.2 $\pm$ 3.6	6.6 $\pm$ 0.9	265 $\pm$ 167	-5.1 $\pm$ 0.3	-6.6 $\pm$ 0.4	289 $\pm$ 22
1.0	64.1 $\pm$ 3.9	5.8 $\pm$ 0.8	382 $\pm$ 250	-5.3 $\pm$ 0.3	-6.7 $\pm$ 0.4	300 $\pm$ 23
1.5	58.4 $\pm$ 4.3	5.1 $\pm$ 0.7	511 $\pm$ 308	-5.5 $\pm$ 0.4	-6.7 $\pm$ 0.4	308 $\pm$ 24
2.0	53.5 $\pm$ 4.5	4.7 $\pm$ 0.7	544 $\pm$ 390	-5.6 $\pm$ 0.4	-6.7 $\pm$ 0.4	316 $\pm$ 25

Table 2: Impact of cfg on the quality of molecules generated by NEAT-POCKET trained on the CrossDocked dataset (best in bold, second-best underlined)

CFG	PB valid	PC clashes	PC SE	Vina score	Vina min	Mol. weight
	[%] $\uparrow$	[1] $\downarrow$	[kcal/mol] $\downarrow$	[kcal/mol] $\downarrow$	[kcal/mol] $\downarrow$	[g/mol]
0.0	80.5 $\pm$ 1.4	4.8 $\pm$ 0.3	32 $\pm$ 15	-4.8 $\pm$ 0.2	-6.4 $\pm$ 0.2	281 $\pm$ 10
0.5	80.3 $\pm$ 1.6	3.7 $\pm$ 0.3	62 $\pm$ 34	-5.5 $\pm$ 0.2	-6.8 $\pm$ 0.2	307 $\pm$ 12
1.0	75.2 $\pm$ 2.0	3.2 $\pm$ 0.2	92 $\pm$ 55	-5.8 $\pm$ 0.2	-6.9 $\pm$ 0.2	324 $\pm$ 13
1.5	68.6 $\pm$ 2.4	2.7 $\pm$ 0.2	138 $\pm$ 80	-5.9 $\pm$ 0.2	-7.0 $\pm$ 0.2	338 $\pm$ 13
2.0	62.1 $\pm$ 2.8	2.4 $\pm$ 0.2	181 $\pm$ 103	-6.0 $\pm$ 0.2	-7.0 $\pm$ 0.2	350 $\pm$ 14

Table 3: Impact of cfg on the quality of molecules generated by NEAT-POCKET trained on the SPINDR dataset (best in bold, second-best underlined)

Prefix	PB valid	PC clashes	PC SE	Vina score	Vina min	Molecular weight	Unique
	[%] $\uparrow$	[1] $\downarrow$	[kcal/mol] $\downarrow$	[kcal/mol] $\downarrow$	[kcal/mol] $\downarrow$	[g/mol]	[%] $\uparrow$
DiffSBDD
none	49.6 $\pm$ 2.6	14.3 $\pm$ 2.0	891 $\pm$ 271	-3.6 $\pm$ 0.8	-6.3 $\pm$ 0.4	296 $\pm$ 14	97.8 $\pm$ 0.3
smallest	29.2 $\pm$ 5.0	36.5 $\pm$ 5.9	262 $\pm$ 160	14.2 $\pm$ 3.7	-3.7 $\pm$ 0.5	280 $\pm$ 16	99.5 $\pm$ 0.5
second-largest	40.8 $\pm$ 4.8	22.3 $\pm$ 4.5	275 $\pm$ 177	4.5 $\pm$ 2.6	-4.7 $\pm$ 0.4	287 $\pm$ 16	98.3 $\pm$ 1.0
largest	51.8 $\pm$ 5.0	14.1 $\pm$ 2.5	299 $\pm$ 225	-2.2 $\pm$ 1.2	-6.1 $\pm$ 0.4	306 $\pm$ 16	86.9 $\pm$ 4.8

Table 4: Performance of NEAT-POCKET vs. DiffSBDD on completing fragments of different sizes derived from the CrossDocked reference ligands (best in bold, second-best underlined)

Prefix	PB valid	PC clashes	PC SE	Vina score	Vina min	Molecular weight	Unique
	[%] $\uparrow$	[1] $\downarrow$	[kcal/mol] $\downarrow$	[kcal/mol] $\downarrow$	[kcal/mol] $\downarrow$	[g/mol]	[%] $\uparrow$
FLOWR
none	75.8 $\pm$ 1.0	4.6 $\pm$ 0.4	35 $\pm$ 9	-7.0 $\pm$ 0.2	-7.7 $\pm$ 0.2	395 $\pm$ 6	99.9 $\pm$ 0.1
smallest	57.7 $\pm$ 3.5	4.5 $\pm$ 0.4	22 $\pm$ 9	-7.4 $\pm$ 0.3	-8.0 $\pm$ 0.3	381 $\pm$ 15	95.6 $\pm$ 1.6
second-largest	41.8 $\pm$ 4.0	4.4 $\pm$ 0.4	22 $\pm$ 10	-7.2 $\pm$ 0.3	-7.8 $\pm$ 0.3	383 $\pm$ 16	95.3 $\pm$ 1.7
largest	31.1 $\pm$ 3.7	4.2 $\pm$ 0.4	19 $\pm$ 9	-7.2 $\pm$ 0.3	-7.8 $\pm$ 0.3	391 $\pm$ 16	88.6 $\pm$ 3.1

Table 5: Performance of NEAT-POCKET vs. FLOWR on completing fragments of different sizes derived from the SPINDR reference ligands (best in bold, second-best underlined)

Fig. 5: Representative examples of pocket- and prefix-conditioned 3D molecule generation with NEAT-POCKET. Molecules were generated from a fixed prefix (highlighted in red) in the binding pocket of the human PRMT5:MEP50 (PDB: 7KID). (A) Experimentally resolved reference ligand (5,5-bicyclic inhibitor 72). (B–D) Representative generated ligands.

Parameter	Value
batch size	256
bias	true
CFG dropout	0.2
clash penalty	4.0
clash penalty margin	0.0
dropout	0.1

Table S1: Hyperparameters of NEAT-POCKET pretrained on GEOM-Drugs and finetuned on CrossDocked.

Fig. S1: Two representative CrossDocked ligands are shown before and after hydrogen addition. Panels A–D show the first ligand: (A) raw input structure, (B) RDKit default hydrogen addition, (C) Open Babel hydrogen addition, and (D) RDKit constrained force-field-based hydrogen placement. Panels E–H show the same sequence for a second ligand. The raw molecules are displayed as kekulized structures, as provided in the input data, to illustrate the representation errors that can lead to incorrect perceptions of aromaticity, protonation assignments, and hydrogen placement.

Fig. S2: Average values of general physicochemical properties of molecules generated by Pocket2Mol, TargetDiff, DiffSBDD, DrugFlow and NEAT-POCKET trained on CrossDocked.

Fig. S3: Distribution of ring sizes (top) and atom types (bottom) of molecules generated by Pocket2Mol, TargetDiff, DiffSBDD, DrugFlow and NEAT-POCKET trained on CrossDocked.

	PC HBA	PC HBD	PC hydrophobic	PC vdW	SA score	Lipinski RO5
	[1] $\uparrow$	[1] $\uparrow$	[1] $\uparrow$	[1] $\uparrow$	[1] $\downarrow$	[%] $\uparrow$
CrossDocked	2.13 $\pm$ 0.41	0.96 $\pm$ 0.21	1.49 $\pm$ 0.41	8.55 $\pm$ 0.71	7.50 $\pm$ 0.21	62.0 $\pm$ 9.7
Pocket2Mol	0.94 $\pm$ 0.16	0.35 $\pm$ 0.05	1.63 $\pm$ 0.28	6.60 $\pm$ 0.54	7.12 $\pm$ 0.07	89.3 $\pm$ 4.2
TargetDiff	1.68 $\pm$ 0.27	0.72 $\pm$ 0.09	1.51 $\pm$ 0.24	9.44 $\pm$ 0.64	8.04 $\pm$ 0.10	66.3 $\pm$ 4.7
DiffSBDD	1.37 $\pm$ 0.19	0.55 $\pm$ 0.06	1.59 $\pm$ 0.23	8.53 $\pm$ 0.57	7.84 $\pm$ 0.08	77.8 $\pm$ 3.0
DrugFlow	1.61 $\pm$ 0.29	0.69 $\pm$ 0.10	1.60 $\pm$ 0.32	8.31 $\pm$ 0.61	7.54 $\pm$ 0.09	69.0 $\pm$ 6.5

Table S2: Performance of NEAT-POCKET vs. other 3D molecule generators trained on CrossDocked and SPINDR – extended results

	Mean	Median	Top-1	Mean w/o top-1	Median w/o top-1	Num.outliers upper tail	Num.outliers lower tail	Mean w/o outliers	Median w/o outliers
CrossDocked
Pocket2Mol	156	65	6,538	91	65	1	0	91	65
TargetDiff	1,516	1,003	30,257	1,225	989	2	1	1,129	989
DiffSBDD	2,485	891	106,920	1,430	882	5	2	1,004	869
DrugFlow	$1.5\times 10^{8}$	255	$1.5\times 10^{10}$	14,422	254	10	1	306	226
NEAT-POCKET	411	265	2960	385	264	0	2	445	279

Table S3: Aggregated strain energies in kcal/mol across CrossDocked and SPINDR test pockets under different aggregation and outlier-removal strategies

Fig. S4: Average strain energies of generated molecules per CrossDocked and SPINDR test pockets.

Fig. S5: Distribution of absolute and relative fragment sizes obtained from the CrossDocked reference ligands.

Fig. S6: Distribution of absolute and relative fragment sizes obtained from the SPINDR reference ligands.

Fig. S7: Examples of molecules generated by Pocket2Mol (3D).

Fig. S8: Examples of molecules generated by Pocket2Mol (2D).

Fig. S9: Examples of molecules generated by TargetDiff (3D).

Fig. S10: Examples of molecules generated by TargetDiff (2D).

Fig. S11: Examples of molecules generated by DiffSBDD (3D).

Fig. S12: Examples of molecules generated by DiffSBDD (2D).

Fig. S13: Examples of molecules generated by DrugFlow (3D).

Fig. S14: Examples of molecules generated by DrugFlow (2D).

Fig. S15: Examples of molecules generated by NEAT-POCKET trained on CrossDocked (3D).

Fig. S16: Examples of molecules generated by NEAT-POCKET trained on CrossDocked (2D).

Fig. S17: Examples of molecules generated by FLOWR (3D).

Fig. S18: Examples of molecules generated by FLOWR (2D).

Fig. S19: Examples of molecules generated by NEAT-POCKET trained on SPINDR (3D).

Fig. S20: Examples of molecules generated by NEAT-POCKET trained on SPINDR (2D).

3D Molecular Generation

3D molecular generation creates molecules by modeling their chemical structures and three-dimensional geometries.

See all 53 papers

Sesame: Structure-Aware Molecular Generation via Spatial Density-Map Conditioning
Jun 22, 2026Konstantin Yatsenko, Arvind ThiagarajanMolecular Generation Structure-Based Drug Design
Tessel Biosciences, Inc., 750 Main Street, Cambridge, MA 02139
Fine-tuning Pocket-Aware Diffusion Models via Denoising Policy Optimization
May 17, 2026Yuan Xue, Daniel Kudenko, Megha KhoslaMolecular Optimization Diffusion Model Fine-Tuning
L3S Research Center, Hanover, Germany · Delft University of Technology, Delft, Netherlands
Generating Developable 3D Molecules via Pocket-Conditioned Diffusion and Property-Aware Optimization
Jul 14, 2026Ruoxi Gao, Jiangweizhi Peng, Ziqi Chen +10Molecular Optimization Structure-Based Drug Design
Computer Science and Engineering, The Ohio State University, Columbus, OH 43210. · Industrial and System Engineering, University of Minnesota, Minneapolis, MN 55455. · Google Cloud, Google LLC, Mountain View, CA 94043. +6

Jun 22, 2026J/K move · Enter open · S save

Sesame: Structure-Aware Molecular Generation via Spatial Density-Map Conditioning

Konstantin Yatsenko, Arvind Thiagarajan

Tessel Biosciences, Inc., 750 Main Street, Cambridge, MA 02139

Molecular Generation Structure-Based Drug Design 3D Molecular Generation Molecular Optimization Diffusion Models Conditional Generative Modeling+1

Open paper AskPDF arXiv

Generative molecular models for drug design are a promising direction with much active research. In the next phase of computational drug design, such models will need to understand small molecule structure and protein-ligand interactions, and they will need to possess the machinery to generate molecules de novo. Incorporating each feature poses a critical challenge. Equally important, yet often treated as secondary, is the ability to grow a molecule from a partial starting point -- a scaffold or fragment supplied by a chemist -- which is the central operation of lead optimization. We present Sesame (Spatial Evoformer for a Structure-Aware Molecular Engine), a diffusion-based molecular generation model that leverages a novel spatial pairformer module to condition on partial molecular structure and the surrounding protein pocket, both expressed as continuous spatial density maps. This single conditioning mechanism supports both de novo generation and fragment-conditioned lead optimization, letting a medicinal chemist prune a hit to a scaffold and have Sesame grow it in productive ways. In addition to this module, we also introduce a diffusion framework for joint denoising of atom types, bond types, and positions, along with a trajectory finetuning scheme that trains on the model's own sampling rollouts to improve generation quality. Sesame is trained on a large corpus of ligand-only and protein-ligand datasets.

Abstract

3D Molecular Generation

Explore similar work

Sesame: Structure-Aware Molecular Generation via Spatial Density-Map Conditioning

Fine-tuning Pocket-Aware Diffusion Models via Denoising Policy Optimization

Generating Developable 3D Molecules via Pocket-Conditioned Diffusion and Property-Aware Optimization