pocketz.doi.bio/2026-01-11_200054_nature_s42256-025-01095-7

Electron-density-informed effective and reliable de novo molecular design and optimization with ED2Mol | Nature Machine Intelligence

URL: https://www.nature.com/articles/s42256-025-01095-7

Saved: 2026-01-11T04:00:59.665Z

Article Published: 20 August 2025 Electron-density-informed effective and reliable de novo molecular design and optimization with ED2Mol Mingyu Li, Kun Song, Jixiao He, Mingzhu Zhao, Gengshu You, Jie Zhong, Mengxi Zhao, Arong Li, Yu Chen, Guobin Li, Ying Kong, Jiacheng Wei, Zhaofu Wang, Jiamin Zhou, Hongbing Yang, Shichao Ma, Hailong Zhang, Irakoze Loïca Mélita, Weidong Lin, Yuhang Lu, Zhengtian Yu, Xun Lu, Yujun Zhao & Jian Zhang 

Nature Machine Intelligence volume  7,  pages 1355–1368 (2025)Cite this article

4710 Accesses

4 Citations

18 Altmetric

Metrics details

Abstract

Generative drug design opens avenues for discovering novel compounds within the vast chemical space rather than conventional screening against limited libraries. However, the practical utility of the generated molecules is frequently constrained, as many designs prioritize a narrow range of pharmacological properties and neglect physical reliability, which hinders the success rate of subsequent wet-laboratory evaluations. Here, to address this, we propose ED2Mol, a deep learning-based approach that leverages fundamental electron density information to improve de novo molecular generation and optimization. The extensive evaluations across multiple benchmarks demonstrate that ED2Mol surpasses existing methods in terms of the generation success rate and >97% physical reliability. It also facilitates automated hit optimization that is not fully implemented by other methods using fragment-based strategies. Furthermore, ED2Mol exhibits generalizability to more challenging, unseen allosteric pocket benchmarks, attaining consistent performance. More importantly, ED2Mol has been applied to various real-world essential targets, successfully identifying wet-laboratory-validated bioactive compounds, ranging from FGFR3 orthosteric inhibitors to CDC42 allosteric inhibitors, GCK and GPRC5A allosteric activators. The directly generated binding modes of these compounds are close to predictions through molecular docking and further validated via the X-ray co-crystal structure. All these results highlight ED2Mol’s potential as a useful tool in drug design with enhanced effectiveness, physical reliability and practical applicability.

You have full access to this article via your institution.

Download PDF Similar content being viewed by others A pocket-based 3D molecule generative model fueled by experimental electron density Article Open access 06 September 2022 PocketFlow is a data-and-knowledge-driven structure-based molecular generative model Article 11 March 2024 Generation of 3D molecules in pockets via a language model Article Open access 15 January 2024 Main

Innovative drug discovery has long been recognized as a time-intensive, costly and high-risk endeavour1. In general, this process is categorized into four stages: hit identification, lead optimization, preclinical evaluation and clinical trials. Among these, hit/lead compound discovery serves as a cornerstone, as identifying a potent active compound can substantially enhance the effectiveness of subsequent clinical evaluation/trials2. Over the past several decades, the majority of active compounds have been identified through high-throughput screening against established libraries, which typically consist of between 104 and 1010 compounds3,4. Nowadays, after the extensive and prolonged screening of such limited-diversity libraries through various campaigns, discovering active compounds with novel scaffolds has become increasingly challenging. Alternatively, de novo molecular design—in which active compounds are generated from scratch—offers a promising strategy to explore a much broader chemical space (~1060) beyond existing libraries5,6.

Recently, with the explosion of available protein structures and advancements in deep generative models (DGMs), structure-based DGMs have been empowered7,8. To some extent, they have indeed shown success in producing ligands that potentially bind tightly, with many exhibiting better computational binding scores than reference experimental ligands (generation success rate). However, a closer examination of the ligands generated by existing models reveals several challenges, outlined as follows. (1) In terms of intermolecular reliability, many generated molecules present binding poses that clearly clash with the receptor and/or undergo drastic reorientation on re-docking, indicating that directly generated poses often represent suboptimal binding modes9. (2) In terms of intramolecular reliability, a notable proportion of the generated molecules display physiochemically implausible topologies and/or problematic substructures, leading to poor synthetic accessibility (SA) and/or pharmacokinetics10. (3) The generated compounds often act as hit compounds, exhibiting moderate potency with the potential for optimization. However, most current de novo design models lack the implementation of hit optimization networks. (4) There is a non-negligible gap between in silico theoretical binding predictions and actual ligand binding; it is essential that the bioactivity of DGM-generated molecules is supported by wet-laboratory experiments.

Integrating domain-specific insights into DGMs is increasingly recognized as a promising approach to address these challenges11,12. LiGANN13 utilizes target pharmacophoric shapes for molecular generation, but the three-dimensional (3D) shapes are implicitly decoded into one-dimensional (1D) molecular SMILES strings. DeepICL14 explicitly incorporates prior interaction knowledge via reference ligands, which limits its applicability to unexplored binding sites. FLAG15 leverages the knowledge of fragment motifs, potentially offering efficiency in exploring synthetically accessible chemical space, whereas it frequently generated ligands with distorted geometry. More fundamentally, structural biologists utilize the ligand electron density (ED) map for accurate structure determination. The macroscopic shape of ED outlines spatial complementarity between the ligand and target16, while its microscopic density encapsulates one unified electromagnetic force governing target–ligand interactions17. Consequently, incorporating ED into DGMs would not only foster the understanding of the global complementary environment of the binding pocket but also enable more generalized and accurate modelling of interatomic interactions. Once a ligand’s ED is available, structural biologists widely use LigandFit, embedded in Phenix18, to fit its 3D structure into the map (Fig. 1a). LigandFit also adopts a fragment-based strategy. Unlike FLAG, it initially places the most rigid portion (namely, the core), and then iteratively attaches the remaining fragments by following the ED. This strategy has the potential to resolve intramolecular reliability issues and is seamlessly applied in hit optimization, where fragments can be replaced or extended.

Fig. 1: Architecture and work flow of ED2Mol.

a, Illustration of the process by which structural biologists fit a ligand binding mode into an experimentally determined ED map, utilizing tools such as LigandFit. b, Schematic of the ED2Mol de novo design process. The ED for the generated molecules is derived from the pocket structure via a VAE generator. A starting core is located and oriented near the largest contiguous region within the ED. Subsequently, fragments are iteratively extended using an EGNN generator, which is guided by the ED environment, enabling the molecule to evolve progressively from state 0 to state n. c, Schematic of the ED2Mol hit optimization process. A preplaced hit ligand serves as the initial core. Initially, unexplored ED is identified at state 0, and the ligand is iteratively refined through the replacement or addition of fragments following the ED. d, An illustrative example showcasing the practical overall 3D molecular evolution.

Full size image

In light of these observations, we propose ED2Mol, an ED-informed structure-aware DGM for de novo molecular design and optimization using the fragment-based strategy. In brief, ED2Mol leverages a variational autoencoder (VAE) to impute ligand ED from the pocket structure. Subsequently, it applies a peak density search algorithm to first fit the core and iteratively extend fragments into unoccupied ED through an equivariant graph neural network (EGNN; Fig. 1b). Thereupon, ED2Mol readily accommodates hit optimization by expanding a predefined core structure (active compound) into a more potent molecule fragment by fragment, guided by the ED (Fig. 1c). These features make ED2Mol versatile, allowing for the generation of hit/lead compounds tailored to various application scenarios in drug design (Fig. 1d). More details of ED2Mol are described in the Methods and Extended Data Fig. 1.

Our comprehensive evaluation results demonstrate that ED2Mol outperforms the compared state-of-the-art DGMs across two pivotal tasks in drug discovery: de novo generation and hit optimization. In addition to benchmarking on the widely used Directory of Useful Decoys—Enhanced (DUD-E) dataset, which dominantly consists of 94 conserved orthosteric sites19, we also introduce a more challenging dataset, the AlloSteric Benchmark—Enhanced (ASB-E) dataset, comprising 112 diverse allosteric pockets (Methods). Furthermore, ED2Mol was applied in five real-world drug design campaigns—three de novo design and two hit optimization efforts. Wet-laboratory experiments validated the capability of ED2Mol to generate a diverse range of bioactive molecules, including both orthosteric and allosteric types for four important drug targets: fibroblast growth factor receptor 3 (FGFR3), cell division control protein 42 homologue (CDC42), hexokinase-4 (GCK) and retinoic-acid-induced protein 3 (RAI3 or GPRC5A).

Results Model performance for de novo generation

A pocket-aware 3D DGM should satisfy two design criteria crucial for drug discovery. First, from an intermolecular perspective, it should learn the geometric distributions of ligands within binding pockets to generate binders with optimal binding modes. Second, from an intramolecular perspective, it should learn pharmacologically relevant distributions of ligands across diverse pockets to dictate essential drug-likeness properties. To quantify the first design criterion, we defined a reliable generated binder based on four aspects: binding site, efficiency, validity and stability, as detailed below. For the second, we assessed the pharmacological characteristics of the generated molecules, including synthesizability, drug-likeness and structural diversity.

Binding mode reliability of generated molecules

To begin with, despite the fact that the models are pocket aware, the binding sites of the generated molecules may vary (Extended Data Fig. 2). To offer a quantitative description, we introduced pocket occupancy density. Higher density signifies better concentration within the pocket. ED2Mol archives higher inner-pocket density than baselines. Conversely, ResGen and GraphBP generate molecules diffusely, often outside the pocket, as reflected by the following average rankings: ED2Mol (28.2) > TargetDiff (26.2) > Pocket2Mol (25.6) > FLAG (15.0) > ResGen (7.9) > GraphBP (4.5). FLAG’s intermediate score reflects moderate occupancy but with dispersion. Furthermore, we defined occupancy variance (distribution uniformity) and converge (spatial exploration). High variance reflects uneven, clustered distributions. Pocket2Mol exhibits higher variance, suggesting sparse, clustered atomic positions of generated molecules, probably due to overlapping similar substructures with limited diversity (see the data below). On the other hand, ED2Mol obtains higher occupancy coverage (84%) than TargetDiff (74%), indicating broader pocket exploration.

Next, DGMs have been observed to generate large molecules with fused rings (Supplementary Fig. 1), inflating binding affinity estimates. To alleviate this bias and allow smaller compounds to be also attractive for subsequent optimization, we used ligand efficiency (LE), which scales the binding affinity by molecular size20. As demonstrated in Fig. 2a and Supplementary Table 1, ED2Mol achieves the highest success rates: 67.3% on DUD-E and 68.9% on ASB-E. It is abnormal that ResGen and GraphBP produce poor binding molecules scored higher than zero. A careful examination of the generated molecules revealed that a substantial proportion suffer from irrational binding substructures and steric clashes with the receptor (see the data on binding validity). It is worth noting that TargetDiff, Pocket2Mol and ResGen exhibited a moderate decrease in the generation success rate on the ASB-E dataset compared with the DUD-E dataset, whereas ED2Mol maintained robust performance across both datasets. This robustness to diverse, unseen allosteric pockets probably stems from its fragment-based protocol, which generates fragments that bind more efficiently to small, superficial allosteric pockets, a characteristic also contributing to FLAG’s consistent performance. On the other hand, ED2Mol’s ability to capture external complementary shape and internal precise distribution of ED map, may further enhance its LE.

Fig. 2: Performance comparison of DGMs using the DUD-E (blue) and more challenging ASB-E (red) sets.

a, Metrics encompass LE, success rate (SR), SA, QED, RO5, Brenk, pan-assay interference compounds (PAINS) and molecular diversity (Div). The size of the circles corresponds to the mean normalized score for the respective metric (from 0% to 100%), whereas the darkness of the circles reflects the performance ranking (from 1 to 6). The original data are provided in Supplementary Table 1 for LE and SR, and Supplementary Table 3 lists the remaining. b, Distribution of RMSD between the original generated poses and re-docked poses (n = 94 for DUD-E, whereas n = 112 for ASB-E). The reference violin plot is derived from re-docking the ligand from the co-crystal structure. The box plots indicate the median and upper and lower quartiles, with whiskers extending to 1.5× the interquartile range and individual points denoting outliers. c, Comparative analysis of the generation success rates between ED2Mol and baselines. The striped bars (success rate) indicates the proportion of generations with LE higher than the reference ligand, whereas the solid bars (reliable success rate) signify the proportion of generations that also meet the binding validity and stability criteria.

Full size image

Furthermore, we conducted a binding validity examination utilizing 18 criteria of intramolecular and intermolecular validity outlined in PoseBusters10. Ligand poses that pass these criteria are termed PB-valid. ED2Mol substantially outperforms other models in generating PB-valid binding poses: ED2Mol (97.3 ± 0.4%) > Pocket2Mol (63.1 ± 0.4%) > FLAG (59.4 ± 1.4%) > TargetDiff (58.1 ± 0.5%) > ResGen (34.1 ± 3.9%) > GraphBP (32.2 ± 1.2%). In specific, as shown in Supplementary Tables 2 and 3, ED2Mol demonstrated the highest feasibility from all aspects. By contrast, intermolecularly, over half of the molecules generated by ResGen and GraphBP suffer from steric clashes (mainly T7 and T10), whereas this ratio is nearly 20% for Pocket2Mol and FLAG, and 10% for TargetDiff. Intramolecularly, all baseline models frequently encountered distorted geometries. FLAG and TargetDiff primarily struggle with predicting the correct bond angles (T2), whereas Pocket2Mol and ResGen often fail to predict valid bond lengths (T1), with GraphBP performing poorly in both respects.

An additional observation is that many of the generated ligand-binding poses undergo dramatic rearrangements after re-docking, suggesting that DGMs may not capture the globally optimal interactions9. To investigate the binding stability, we conducted re-docking experiments comparing the original poses to re-docked poses, calculating the minimum root-mean-squared deviation (RMSD) across the top ten docking poses. Remarkably, reference co-crystal ligands recovered 72% of their crystallized poses (RMSD < 2 Å) across both datasets, validating the use of re-docked poses, at least in benchmark datasets, as a surrogate ground truth for evaluating the fitness of directly generated poses. After we re-docked the molecules generated by ED2Mol and the baselines following the same pipeline, ED2Mol showed the highest fidelity in maintaining the binding poses (Fig. 2b), with 12.7% more stable poses than the second-best model, Pocket2Mol, on DUD-E (34.8% total) and 15.9% better stability than the second-best model, TargetDiff, on ASB-E (29.0% total).

Collectively, we define the reliable success rate as the proportion of generated ligands that not only exhibit greater binding efficiency (success rate) but also possess high-quality (PB-valid) and stable (RMSD < 2 Å on re-docking) binding poses (Fig. 2c). ED2Mol achieved a reliable success rate of 34.4 ± 0.5%, more than twice as high as the best-performing baseline models, TargetDiff (17.1 ± 1.0%) and Pocket2Mol (15.0 ± 2.0%). FLAG (6.7 ± 0.1%) and ResGen (4.5 ± 1.1%) produced very few reliable ligands, and GraphBP produced almost no reliable ligands in our evaluation. Furthermore, ED2Mol demonstrates more efficient binding mode generation for large orthosteric pockets on the DUD-E dataset, and maintains comparable binding mode quality between large and regular pockets on the ASB-E (Supplementary Fig. 2).

Pharmacologically desirable properties of generated molecules

To this end21,22, we accessed the generated molecules based on five widely recognized metrics: SA, QED, adherence to Lipinski’s rule of five (RO5), Brenk’s rules and pan-assay interference compounds (PAINS) filters. ED2Mol obtains the best scores across four metrics (Fig. 2a and Supplementary Table 4), highlighting its capability to generate synthesizable and drug-like molecules with fewer problematic functionalities. Moreover, it consistently achieves the best or near-best performance across crucial pharmacokinetic properties, including membrane permeability, metabolic stability, oral bioavailability and lead diversification (Supplementary Table 5). Additionally, molecules produced by ED2Mol show comparable diversity to those produced by the baselines.

From de novo design to hit optimization

Currently, it is virtually impossible to directly generate fully optimized drugs through automated de novo design using DGMs23. Hit optimization constitutes a crucial step towards fully de novo drug design. A primary strategy involves elaborating known active ligands by swapping and/or adding new moieties (fragments) to establish additional interactions with the receptor. This strategy is directly applicable to ED2Mol through a fine-tuned work flow, which bypasses the core searching procedure and instead utilizes a prelocated core scaffold as a template for further fragment elaboration. To demonstrate the utility of ED2Mol in identifying effective fragment replacements or additions, we present practical case studies from the literature: fragment replacements of Ski2-like RNA helicase (Brr2) inhibitors and fragment additions of peroxisome-proliferator-activated receptor gamma (PPARγ) activators.

Design of Brr2 inhibitors via fragment replacements

Brr2, a helicase crucial for spliceosome function and gene expression fidelity24, was recently targeted by an allosteric inhibitor reported in ref. 25. Initially, they identified a high-throughput screening hit (inhibitor 3) binding an allosteric pocket. By retaining the core scaffold (6-benzyl-4,6-dihydropyrido[4,3-d]pyrimidine-2,7(1H,3H)-dione) and substituting other fragments, they then developed lead inhibitor 9, boosting the inhibitory activity 67-fold.

To simulate a practical lead optimization scenario in which the lead co-crystal structure is unknown, we utilized the Brr2 pocket conformation from the hit-bound structure (PDB ID: 5URJ). Keeping the core scaffold of inhibitor 3, ED2Mol was integrated for fragment replacements. The ligand ED predicted by ED2Mol effectively guided exploration into vacant subpocket regions near residues T1197, F1255 and K1716. The generated molecules were assessed by binding affinity, the critical metric during the optimization stage, rather than LE. ED2Mol produced molecules with higher predicted affinities than baseline methods (Fig. 3a). Importantly, it achieved this optimization within the relevant chemical space of the reference lead (inhibitor 9), avoiding simple increases in molecular weight (MW) or lipophilicity (Supplementary Fig. 3a,b).

Fig. 3: Retrospective hit optimization case studies.

Left, fragment replacements of Brr2 inhibitors. Right, fragment additions of PPARγ activators. a,c, Distribution of docking affinity scores for the optimized samples of Brr2 (a) and PPARγ (c) generated by ED2Mol and other methods. The black solid lines denote the scores of computationally re-docked (for Brr2) and experimentally determined (for PPARγ) lead compounds. The red dashed lines represent the scores of compounds generated by ED2Mol that recover the lead compounds. b,d, Top, the reported lead compound (top right, red) can only be recovered by ED2Mol (top right, blue) from the original hit compound (top left, grey). The elaboration moieties involved are highlighted in blue. Bottom, two representative samples with superior affinity compared with the reported lead compound are shown: one with high chemical similarity to the reported lead, and the other with low similarity. The diverse elaboration moieties are highlighted in orange. Intermolecular interactions are analysed using protein-ligand interaction profiler (PLIP).

Full size image

Among all the methods tested, only ED2Mol was able to recover lead inhibitor 9. Meanwhile, its generated binding pose closely aligned with the molecular docking result (RMSD = 0.56 Å), underscoring its capability to produce reliable binding modes (Fig. 3b). A closer inspection revealed two distinct groups of molecules (Fig. 3b). The first, termed ‘high similarity & good affinity’, comprised molecules topologically similar to inhibitor 9 (fingerprint similarity = 0.71) but with better affinity. This improvement stems from preserving key interactions (like the H bond to F1255) and establishing new hydrogen bonds via the replacement fragment to residues T1197 and K1716 in the subpocket. The second group, ‘low similarity & good affinity’, represents potential new scaffolds. For example, one molecule relies on a sulfonamide nitrogen for maintaining polar interactions with F1255 and its phenyl ring for broader non-polar contacts. These computational results suggest that ED2Mol can effectively identify moiety replacements optimizing diverse, favourable interactions.

Design of PPARγ activators via fragment additions

PPARγ activators represent a promising therapeutic strategy for treating type 2 diabetes and associated metabolic syndromes26. Through biochemical screening, a potent PPARγ activator, activator 1, was identified and its co-crystal structure was solved (PDB ID: 3ET0)27. Utilizing a structure-guided fragment-growing strategy, they obtained lead activator 3 (PDB ID: 3ET3), which exhibited over a 70-fold potency increase.

Here we investigate ED2Mol’s ability to detect subpockets and optimize a more challenging activator. Using the co-crystal structure of activator 1 as the input, ED2Mol captured additional growing space surrounding residues R288, Y327, L330 and S342. Also, we computationally scored the binding affinity of the generated molecules. The affinity distribution for ED2Mol-generated molecules was more left-skewed compared with both experimentally confirmed lead compounds and other methods (Fig. 3c and Supplementary Figs. 3c,d). Likewise, ED2Mol was the only method to recover lead activator 3 with a reliable binding pose consistent with its co-crystal structure (RMSD = 0.95 Å; Fig. 3d). A representative ligand from the ‘high similarity & good affinity’ group retains the hydrogen bond to Y327 via the same sulfonyl group as activator 1. Instead, a ligand from the ��low similarity & good affinity’ group archives this interaction via an alternative amide group. Moreover, both ligands gained additional polar interactions with the main chain of residue S342 and non-polar interactions with residues R288, L330 and others, probably accounting for their enhanced binding potency (Fig. 3d).

From benchmarking evaluation to wet-laboratory verification

To substantiate the effectiveness of ED2Mol, we conducted wet-laboratory experiments on five real-world scenarios: FGFR3 orthosteric inhibitor generation and optimization, CDC42 allosteric inhibitor de novo generation, GCK allosteric activator optimization and GPRC5A allosteric modulator de novo generation. Despite the fact that these proteins are well-established targets for various diseases, particularly in antitumour and antidiabetic therapies, there remains a paucity of efficient small-molecule intervention agents. Expanding the chemical space to identify potent compounds, thus, holds promise.

FGFR3 orthosteric inhibitor generation and optimization

FGFR3, a member of the tyrosine kinase superfamily, propagates FGF-related signals that regulate numerous biological processes28. Since the dysregulation of FGFR3 is implicated in a wide range of diseases, particularly cervical and bladder carcinomas29, blocking it has been established as a therapeutic approach for these malignancies.

To address this, ED2Mol was used to generate novel FGFR3 inhibitors. Using the orthosteric pocket defined by an ATP-analogue bound structure (PDB ID: 4K33; Fig. 4a)30, ED2Mol generated 50,000 molecules (100 min, single RTX 4080 GPU). Recognizing that orthosteric kinase inhibitors typically exhibit drug-like properties and mimic ATP’s hydrogen bonding to the hinge region31,32, we applied stringent filters: (1) MW > 250, (2) RO5 compliance, (3) QED > 0.8, (4) SA < 4.5 and (5) ≥1 hinge H bond, reducing the pool to 687 candidates. Clustering these by two-dimensional similarity yielded five representative molecules (Supplementary Fig. 4). Among these, F4 was selected for synthesis (Supplementary Scheme 1) because its predicted pose showed the closest hinge-binding pattern to ATP (H bonds with E556 and A558). The surface plasmon resonance (SPR) assay confirmed that F4 binds reversibly to FGFR3, yielding a dissociation constant (KD) of 599.0 μM (Fig. 4b and Supplementary Fig. 5).

Fig. 4: Discovery of new FGFR3 orthosteric inhibitors through practical applications of ED2Mol.

a, The generated F4 compound at the orthosteric pocket of FGFR3. Bottom, the input structure (PDB ID: 4K33) of FGFR3 (grey surface) in complex with orthosteric ligand (green stick). Top, a magnified view of the generated binding mode of the F4 compound, where the 2-amino-6,7-dihydrothiazolo[4,5-cb]pyridin-5(4H)-one moiety forms critical hydrogen bonds with hinge residues E556 and A558. b, Dose-dependent curve for F4 measured by the SPR assay. RU, response units. c, Co-crystal structure of FGFR3 in complex with F4 (PDB ID: 9KFU). Left, superimposition of the binding mode of F4 with FGFR3 as generated by ED2Mol (grey) and that determined experimentally by X-ray crystallography (magenta). The arrows indicate conformational changes in the Gly-rich loop and adjacent β1 and β2 strands, as well as the fluorobenzene moiety of the F4 compound. Right, based on the newly resolved complex structure, the second-round ED extraction captures an additional subpocket, indicated by the black dashed circle. d, Dose-dependent curve for F42 measured by the SPR assay. e, Hit optimization leading to the F42 compound. Left, FGFR3–F42 interaction analysis as determined by PLIP. Right, the binding mode of the generated F42 fully occupies the newly identified subpocket, with the chemical structure shown at the bottom, and the swapped portion highlighted in orange.

Source data

Full size image

To precisely understand the binding interactions, we determined the co-crystal structure of the F4–FGFR3 complex at a high resolution of 1.4 Å. The experimentally observed binding mode closely matched that generated directly via ED2Mol (RMSD = 2.00 Å). Intriguingly, the crystal structure revealed that F4’s fluorobenzene moiety adopted an upward orientation, driving conformational changes in the Gly-rich loop and adjacent β strands, leading to a widened orthosteric site. This observation is consistent with the induced-fit binding mechanism frequently observed in kinase inhibitors, and presents new opportunities for optimization33 (Fig. 4c).

Indeed, ED2Mol captured an extended hydrophobic subpocket near residues L624, and we performed the same fragment replacement protocol as Brr2 inhibitors, generating 10,000 molecules (20 min). After applying adjusted filters, that is, (1) MW > 250, (2) RO5 compliance, (3) QED > 0.7 and (4) SA < 4, and clustering the resulting 361 candidates into 3 groups (Supplementary Fig. 6), candidate F42 (Supplementary Scheme 2) was chosen based on the predicted SA (3.6). The SPR assay determined that F42 possessed a KD of 61.4 μM, achieving a 9.8-fold increase in potency over the initial F4 (Fig. 4d and Supplementary Fig. 7). This improvement can be largely attributed to F42’s efficient occupation of the newly resolved subpocket, promoting additional hydrophobic contacts with surrounding residues (Fig. 4e). These results showed that F42 could be a starting potential hit compound for later studies.

CDC42 allosteric inhibitor de novo generation

CDC42, a small GTPase, serves as a binary switch in a variety of downstream signalling pathways, including cell migration and angiogenesis34. The overexpression of CDC42 is frequently observed in multiple human cancers, rendering it an attractive target for antitumour therapies35. Recently, an allosteric pocket was identified at the protein–protein interaction site between CDC42 and its downstream partner p21-activated kinase (PAK), posing a new avenue for designing allosteric inhibitors36.

Again, we deployed ED2Mol taking the CDC42-PAK (PDB ID: 2ODB) crystal structure as the input (Fig. 5a). First, ED2Mol generated a set of 50,000 molecules (88 min). Second, these molecules were filtered: (1) MW > 250, (2) RO5 compliance, (3) QED > 0.8 and (4) SA < 4.5, yielding 1,547 candidates. Third, five representative molecules were selected based on clustering (Supplementary Fig. 8). Among the representatives, compound C1 exhibited a stronger π–π-stacking interaction with residue F37 and was deemed more synthetically accessible. Also, its generated binding pose reveals considerable similarity to that obtained via molecular docking (RMSD = 0.82 Å), reinforcing the reliability of ED2Mol’s predictions. C1 and its analogue C11 (from the same cluster) were synthesized via comparable synthesis routes (Supplementary Schemes 3 and 4). The in vivo bioluminescence resonance energy transfer assay validated the inhibitory activity against CDC42 for both C1 (IC50 = 47.58 ± 3.71 μM) and C11 (IC50 = 111.63 ± 0.90 μM; Supplementary Fig. 9). Furthermore, the bioactivity of C1 was confirmed by an in vitro SPR assay, with a direct KD of 5.35 μM (Fig. 5b). These results establish C1 as a valuable starting hit compound for subsequent investigations.

Fig. 5: Discovery of new CDC42 allosteric inhibitors and GCK allosteric activators through practical applications of ED2Mol.

a, The generated C1 compound at the allosteric pocket of CDC42. Left, the input structure (PDB ID: 2ODB) of CDC42 (grey cartoon) in complex with orthosteric ligand and allosteric PAK6 (red cartoon). The W40 (red)-mediated hydrophobic allosteric protein–protein interactions with surrounding residues (tint) are highlighted. Right, superimposition of the binding mode of C1 with CDC42 as generated by ED2Mol (grey) and that predicted by molecular docking (magenta). b, Dose-dependent curve for C1 measured by the SPR assay. c, Two previously identified activators at the allosteric pocket of GCK. Top, the input structure (PDB ID: 3GOI) of GCK (pink surface) in complex with orthosteric ligand and allosteric activator. Bottom, a magnified view of the superimposed binding modes of the activators, which shows that the added methyl group on the aniline NH group induces rotamer changes in Y215, as indicated by the black dashed arrow. d, Binding mode of hit compound G0. Left, ED extraction reveals an additional subpocket, as indicated by the black dashed circle. Right, dose-dependent curve for G0 measured by a DEC assay. The experiments were performed independently three times. All data are shown as mean ± s.d. e, Binding mode of the generated lead compound G11. Left, GCK–G11 interaction analysis suggests the grown methylbenzene group forms several hydrophobic interactions with residues Q98, I211, Y214 and L451. Right, dose-dependent curve for G11 measured by a DEC assay. The experiments were performed independently three times. All data are shown as mean ± s.d.

Source data

Full size image GCK allosteric activator optimization

GCK is an enzyme that catalyses the phosphorylation of glucose to glucose-6-phosphate, playing an important role in glucose metabolism37. Small-molecule activators of GCK have garnered attention as therapeutic agents for type 2 diabetes38. For instance, an allosteric activator with a thiazole-amide scaffold (PDB ID: 3FR0)39 was discovered. Later structural analysis showed a methyl addition on the aniline moiety enabled the emergence of a new subpocket via Y215 outwards reorientation (PDB ID: 3GOI)40 (Fig. 5c).

Building on this observation, we applied the same ED2Mol work flow as the PPARγ case, targeting the newly formed subpocket in the 3GOI structure, starting from analogue G0 (Fig. 5d)40. ED2Mol was initially set to propose 10,000 molecules (25 min). To narrow down the selection, four filters were implemented: (1) MW > 250, (2) RO5 compliance, (3) QED > 0.4 and (4) SA < 4. These filters yield 995 molecules, leaving three categories after clustering (Supplementary Fig. 10). From these, compound G1 and its analogue G11 (from the same cluster) were preferred due to its ease of synthesis (Supplementary Schemes 6 and 7) and the effective occupation of the subpocket. Bioassays demonstrated that G1 had an improved potency (EC50 = 290 nM, 6.1-fold over G0 (Supplementary Fig. 11)), whereas G11 proved even more potent (EC50 = 150 nM), representing an 11.9-fold improvement (Fig. 5e). These bioassay results suggest that G11 could be a strong candidate for later evaluation and development.

GPRC5A allosteric modulator de novo generation

GPRC5A, a member of class C GPCRs, was initially identified as a tumour suppressor in lung tumourigenesis but has gained increasing interest due to its association with a wide range of cancers41. Modulating GPRC5A has emerged as a new antitumour therapeutic strategy. However, these efforts are hampered by the orphan status of GPRC5A, as its ligand recognition, activation mechanism and structural characteristics remain completely elusive.

To address this challenge, we explored a sequence-to-hit framework by integrating sequence-to-structure modelling, structure-to-pocket detection, pocket-to-candidate generation and candidate-to-hit verification methodologies (Fig. 6). We first modelled the GPRC5A structure from its sequence using AlphaFold 3 (ref. 42; Fig. 6a), and subsequently detected a potential druggable allosteric pocket via AlloSite43 within the AlphaFold 3 high-confidence region (Fig. 6b). Targeting this site, ED2Mol generated 50,000 molecules from scratch (78 min). Next, applying four filters—(1) MW > 150, (2) RO5 compliance, (3) QED > 0.6 and (4) SA < 4.5—resulted in 1,603 candidates. The minimum MW threshold was set to 150 due to the small pocket. Clustering analysis produced five representative molecules (Supplementary Fig. 12), with A4 standing out as the top candidate owing to its stronger interaction profile and favourable SA. Here the reliability of ED2Mol was supported by a consistent binding mode (RMSD = 1.48 Å) observed between the ED2Mol-generated and the top-ranked AlphaFold 3-predicted holo structure from extensive samplings (Fig. 6c). Following chemical synthesis (Supplementary Scheme 8), bioassays confirmed A4 functions as a GPRC5A activator with an EC50 of 8.94 ± 2.30 μM (Fig. 6d). Our discovery of the A4 compound establishes a foundational step towards the development of effective GPRC5A activators.

Fig. 6: Discovery of new GPRC5A allosteric modulators through the integrated applications of ED2Mol.

a, AlphaFold to model the GPRC5A apo structure from its sequence. The top-ranked model (score = 0.82) is presented, coloured by per-residue confidence. The ranking score was calculated as 0.8 × ipTM + 0.2 × pTM + 0.5 × disorder – 100 × clash, where pTM represents the predicted template modelling score for the entire structure, ipTM reflects the interfacial accuracy, disorder indicates the proportion of disordered residues and clash penalizes structures with >50% clashing atoms per chain or chains with >100 clashing atoms. b, AlloSite to detect potential allosteric pockets (grey surface) from the apo structure. The most druggable pocket, based on the druggability score, is highlighted in orange. c, ED2Mol to generate candidate small molecules targeting the identified allosteric pocket. The representative generated compound A4 (green stick) demonstrates a consistent binding mode with the top-ranked A4–GPRC5A complex predicted by AlphaFold (pink stick); additional sampled holo structures are shown in grey. d, Dose-dependent curve for A4 measured by PRESTO-Tango assay. The experiments were performed independently three times. All data are shown as mean ± s.d.

Source data

Full size image Discussion

In this study, we describe ED2Mol, leveraging the ED map as generative cues for structure-based de novo design and optimization. ED2Mol offers a versatile and reliable tool to help accelerate the discovery and development of therapeutic agents.

It is essential to validate the practical utility of DGMs rigorously through wet-laboratory experiments. PocketFlow generated and confirmed orthosteric inhibitors against two drug targets, although one of them was post-scored with the help of molecular docking11. Dragonfly successfully identified orthosteric activators for PPARγ, with activity scoring reliant on tailored QSAR models44. ED2Mol successfully designed de novo or optimized small-molecule intervention agents for four appealing drug targets. These real-world scenarios encompass diverse binding pocket types, including orthosteric and allosteric sites, and involve different ligand types, such as inhibitors and activators, demonstrating its robustness across diverse drug discovery contexts. As a result, ED2Mol identified two FGFR3 orthosteric inhibitors, two CDC42 allosteric inhibitors, two GCK and one GPRC5A allosteric activator, all of which are valuable for further systematic research.

However, several avenues remain open for further enhancement. Although the current fragment libraries balance efficiency and generative effectiveness, as demonstrated by benchmark results, it may favour well‑trodden scaffolds and underscoring synthetically uncommon yet biologically relevant structures, such as natural product-like scaffolds45. The impact of expanding the library, potentially at the expense of efficiency, should be assessed in future work. Moreover, one may optimize the binding affinity or other preclinical desirable pharmalogical properties (for example, low toxicity) individually or concurrently through explicit guidance or reinforcement learning46,47.

Methods Details of ED2Mol

Inspired by LigandFit18, ED2Mol adopts a similar fragment-wise sampling strategy comprising three main components: (1) extracting the ligand ED from the pocket structure; (2) placing cores within the ED; and (3) generating a molecule by iterative fragment extension into the ED.

Ligand ED extraction

Constructing an ED map for a crystal structure involves first calculating the structure factors for the reflections in reciprocal space and then applying an inverse Fourier transform to convert these factors into a real-space map. According to Bragg’s law, the diffraction pattern of a crystal reflection can be characterized by a series of parallel, hypothetical crystalline planes that mathematically intersect the axes of the crystal unit cell. These planes are described by Miller indices (hkl), which correspond to the reciprocals of the intercepts along the unit-cell axes. Thereafter, the structure factor, F(hkl), can be obtained by summing the contributions of waves scattered by N atoms in the crystal along the direction of hkl reflection:

$${F}(hkl)={\mathop{\sum}\limits{j=0}^{N}}{f}{j}{\rm{e}}^{2\uppi{\rm{i}}({hx}+{ky}+{lz})},$$ (1)

where each term involves the atomic scattering factor fj, representing the X-ray scattering power of the atom j. Once the structure factors are known, the ED ρxyz at any point (x, y, z) in the unit cell can be computed via the inverse Fourier transform:

$${\rho }{{xyz}}=\frac{1}{V}\mathop{\sum }\limits{-\infty }^{+\infty }{|F}({hkl})|{\rm{e}}^{-2\uppi{\rm{i}}({hx}+{ky}+{lz})},$$ (2)

where V is the unit-cell volume.

In this study, the structure factors and the corresponding ED values for each complex structure were calculated using the CCTBX Python package with the predefined same unit-cell parameters48. The binding pocket, including both protein and ligand atoms, was represented as an ED map on a cubic grid of side length 24 Å, with a spatial resolution of 0.5 Å.

To reconstruct the ligand ED from the pocket ED, we used a VAE framework. This model consists of an encoder, which compresses the pocket ED into a one-dimensional latent variable Z that approximates a predefined prior distribution p(Z), and a decoder, which reconstructs the ligand ED from this latent representation. The loss function used during model training comprised two weighted components: a mean-squared error term that quantified the similarity between the actual and predicted ligand ED and a Kullback–Leibler (KL) divergence term, which measured the deviation of the learned latent distribution from the true prior distribution. The loss function is formulated as

$$L1=\alpha {{||x}-\hat{x}{||}}^{2}+\beta {{\rm{KL}}}(q({Z|x}),\,p\left(Z\,\right)).$$ (3) Initial core placement

Our molecular generation strategy begins with fitting initial cores in 3D space portrayed by the generated ligand ED. To mitigate the computational cost associated with exhaustive searches across the entire map, we reduced the number of grid points by hierarchically applying a max-pooling operation followed by spatial clustering to identify a limited number of local density peaks. These peaks serve as candidate positions for core placement. The cores from the library are enumerated at each identified position and systematically rotated around the XYZ axes in increments of 40°. To efficiently evaluate the compatibility of various core fragment positions and orientations with the ED map, we introduced the following Q score18:

$$Q=\,\frac{M\mathop{\sum }\nolimits{i}^{M}{\rho }{i}{Z}{i}}{\mathop{\sum }\nolimits{i}^{M}{Z}_{i}},$$ (4)

where Zi denotes to the atomic number of the atom i, ρi indicates the density value at the closest grid point and M is the number of heavy atoms in the core. The Q score favours conformations that exhibit strong overlap with the ED map and align with the density shape in terms of atomic number distribution.

Iterative fragment extension

ED space E is represented as a set of points {〈c1, ρ1〉, 〈c2, ρ2〉…〈cm, ρm〉}, where c and ρ denote the Cartesian coordinate and the density value of grid points, respectively, and m is the total number of grid points. To sample a new state si+1 from current state si, we decomposed the process into two steps (Extended Data Fig. 1). First, the current state graph ({G}{{s}{i}}) is updated to the intermediate graph ({G}{{s}{i}}^{* }) by EGNN generator part I ({{\emptyset}_{1}}), which identifies the growth site for fragment attachment. Formally,

$${G}{{s}{i}}^{* }=\,{{{\emptyset}}}{1}({G}{{s}_{i}},\,E\,).$$ (5)

Second, given the updated intermediate state, EGNN generator part II ∅2 determines how to attach a new fragment graph Gf conditioning on E as follows:

$${s}{i+1}=\,{{{\emptyset}}}{2}\left({G}{{s}{i}}^{* },\,{G}_{\rm{f}},\,E\right).$$ (6)

Specifically, in the first step, we encode the ED environment of the atom nodes for graph embedding to assess the potential of each ligand hydrogen to serve as a growing point. For each atom node a, five concentric spheres with radii ranging from 1 Å to 3 Å are defined to capture the local density environment ({{{f}{a}^{1},f}{a}^{1.5},{f}{a}^{2},{f}{a}^{2.5},{f}{a}^{3}\,}). ({f}{a}^{r}) computes the number of occupied and vacant density grid points scaled by the total grid points within a sphere of radius r, which is expressed as

$${f}{a}^{\;r}=\left(\frac{{p}{{{\rm{occ}}}}^{{ar}}}{{p}{{{\rm{total}}}}^{{ar}}},\,\frac{{p}{{{\rm{vac}}}}^{{ar}}}{{p}_{{{\rm{total}}}}^{{ar}}}\right).$$ (7)

Using the local density environment as the initial atom embedding, it captures the atomic-level growing space information through equivariant graph convolutional layers (EGCLs). Given the atom embedding hl and the corresponding coordinate xl in the lth layer, message passing is described as

$${m}{{ij}}=\,{{{\emptyset}}}{e}\left({h}{i}^{l},{h}{j}^{l},\left\Vert{x}{i}^{l}-{x}{j}^{l}\right\Vert^{2},{e}{{ij}}\right),$$ (8) $${x}{i}^{\,l+1}={x}_{i}^{l}+\sum _{j\epsilon N\left(i\right)}\left({x}{i}^{l}-{x}{j}^{l}\right){{{\emptyset}}}{x}({m}{{ij}}),$$ (9) $${h}{i}^{l+1}={{{\emptyset}}}{h}\left({h}{i}^{l},{\sum}{j\epsilon N\left(i\right)}{m}_{{ij}}\right),$$ (10)

where eij is the edge feature (one-hot encoding of the bond types), and ({{\emptyset}{\rm{e}}}), ({\emptyset}{{x}}) and ({\emptyset}_{{h}}) are multilayer perceptrons. Subsequent multilayer perceptrons output the probabilities for hydrogen atoms, indicating whether they can provide space for new fragments. The hydrogen with the highest score is selected to attach the fragment.

In the second step, we translate the problem of determining the next state si+1 into predicting the torsion angle of the newly formed bond connecting the current state si and the target fragment. To this end, we adopt the normalized torsion angle α introduced in ref. 49 as the model output, which considers invariances with respect to the full conformer. The computation of the torsion angle α between bonded atoms a and b is summarized using the following equations:

$${\bf{s}}{\rm{mn}}=\,{\left[\frac{\cos {\Delta}{\rm{mn}}}{\sin {\Delta}{\rm{mn}}}\right]}{{\rm{m}}\in {\rm{N}}\left({\rm{a}}\right){\backslash}{{\rm{b}}},{\rm{n}}\in {\rm{N}}({\rm{b}}){\backslash}{{\rm{a}}}},$$ (11) $${\bf{s}}=\sum {{m,n}}{c{\bf{s}}{\rm{mn}}},$$ (12) $$\alpha ={{\rm{anta}}}2\left(\frac{\bf{s}}{{\parallel{\bf{s}}\parallel}}\right),$$ (13)

where m and n represent the local structure, the additional graph neighbours, associated with nodes a and b, respectively, and c refers to the real coefficients to avoid s to be a null vector. The EGNN generator part II incorporates three blocks, called fenc1, fenc2 and fenc3 for separately processing different inputs. The block fenc1 utilizes convolutional layers to map the features of ED E, whereas fenc2 and fenc3 encode the intermediate-state graph ({G}{{s}{i}}^{* }) and the target fragment graph Gf, respectively, extracting embeddings using EGCL layers. Supplementary Table 4 lists the details on the initial feature representations of two graphs ({G}{{s}{i}}^{* }) and Gf. For ({G}{{s}{i}}^{* }), the relative atom position with the ED cube is additionally considered:

$${{\rm{pos}}}=\,\frac{{x}_{i}-c}{S}+0.5,$$ (14)

where xi and c are the atom coordinate and centre of ED, respectively, and S is its side length. The final torsion angle is predicted as

$$\alpha ={{\rm{MLP}}}(\;{f}{{{\rm{enc}}}1}(E\,)\oplus {f}{{{\rm{enc}}}2}({G}{{s}{i}}^{* })\oplus {f}{{{\rm{enc}}}3}({G}{{\rm{f}}})),$$ (15)

where (\oplus) refers to concatenation. Each fragment in the library is fed into the model and rotated according to the predicted torsion angles. To encourage the exploration of a broader chemical space, the molecules at state si+1 are grouped by their originated molecules at state si and an assigned number of structurally dissimilar molecules are selected from each group based on their Q-score rankings. This strategy encourages the exploration of a broader chemical space.

Datasets construction

To construct the dataset for training, the ligand ED extraction model initially collected approximately 780,000 ligand-binding pockets from the PDB database. A subset of high-quality 49,987 pocket–ligand pairs was selected based on the following criteria:

1.

deletion of duplicated ligands appearing in multiple chains;

2.

retention of ligands with a real-space correlation coefficient greater than 0.7;

3.

exclusion of non-functional ligands and crystallization additives;

4.

selection of ligands with MW greater than 250.

To prevent data leakage, we partitioned the dataset into training, validation and test sets (7:1.5:1.5) such that the proteins in the test set share less than 70% sequence identity with any protein in the training or validation sets, as determined by CD-Hit (https://github.com/weizhongli/cdhit)50. In addition, we excluded from the training set any protein pockets that share more than 70% structural similarity with pockets in either the DUD-E or ASB-E datasets, as measured by the APOC software (https://sites.gatech.edu/cssb/apoc/)51.

A subset of drug-like molecules (log[P] < 5 and 250 ≤ MW ≤ 500) were sourced from the ZINC database for producing the corresponding ED maps4. The ETKDG algorithm from the RDKit package was used to convert the two-dimensional structures into 3D conformations52. Each molecule was then decomposed into fragments along rotatable bonds, ensuring that each fragment contained at least two atoms. The largest fragment was treated as the core component, whereas the remaining fragments were retained. Next, to train the models for iterative fragment extension, each ZINC compound was reconstructed by starting from its core structure and the neighbouring fragments were sequentially attached, thereby generating a series of intermediate states. From these processed data, two datasets are derived—the GP and TA datasets. The GP dataset is designed to evaluate the potential of hydrogen atoms on an intermediate-state molecule to serve as fragment growth points. Each molecular ED map is paired with a sequence of intermediate states, labelling each hydrogen as a growth point or not. The other TA dataset is constructed to predict the normalized torsion angles of the bonds connecting target fragments to intermediate states. The normalized torsion angle was discretized into 36 bins over the range [–π, π], enabling a classification-based prediction task. In total, the GP dataset contains 2,986,278 samples derived from 94,519 molecules, whereas the TA dataset includes 17,978,368 samples derived from 4,497,302 molecules. Both datasets were randomly split into training, validation and test sets using a 7:1.5:1.5 ratio. To balance the generative efficacy and effectiveness, during model inference, only cores and fragments with a frequency greater than 25 were retained, resulting in final libraries comprising 362 cores and 169 fragments, respectively.

Benchmarks and baselines

We found that most DGMs are typically assessed using highly evolutionarily conserved endogenous pockets (also termed as orthosteric pockets)19,53. Rare attention has been paid to assess their performance on non-canonical pockets, which predominantly include a broader range of less-conserved allosteric pockets that are topologically distinct from orthosteric sites and can theoretically be located anywhere on the protein surface. Besides, allosteric pockets tend to be more superficial and smaller than orthosteric pockets, posing greater challenges for designing active compounds54,55.

To address these, two benchmark datasets were used in this study. The first dataset, DUD-E, is widely used and comprises 94 pharmacological targets after retaining only orthosteric pockets. The second dataset, ASB-E, meticulously curates 112 high-quality and diverse allosteric pockets. The DUD-E dataset originally contains 102 protein–ligand co-crystal structures, primarily spanning three protein classes: GPCRs, kinases and nuclear receptors19. Of these, eight allosteric ligands, as annotated in the Allosteric Database (https://mdl.shsmu.edu.cn/ASD/), were excluded from the analysis, resulting in a final set of 94 orthosteric pockets3. On the other hand, the ASB-E dataset was curated from ASBench (https://mdl.shsmu.edu.cn/asbench/) for validating DGMs in the more challenging scenario of discovering allosteric modulators56. ASBench contains 480 high-quality allosteric protein–ligand pairs, of which 296 allosteric pockets bind small endogenous ligands (for example, ions) and were excluded due to their limited utility for structure-based allosteric modulator design. Among the remaining 184 allosteric pockets that bind exogenous ligands, 112 representative allosteric pockets were selected based on their similarity scores computed by APOC and used to benchmark the DGMs.

The de novo design baselines in our study for comparison comprise five state-of-the-art pocket-conditional DGMs, spanning from the atom-wise autoregressive models like Pocket2Mol57, GraphBP58 and ResGen59 to fragment-wise autoregressive model FLAG15 and a non-autoregressive diffusion-based model TargetDiff60. Additionally, the baselines used in the hit optimization range from the autoregressive fragment-wise FLAG to the diffusion-based DiffDec61 and the traditional method, FragGrow62 (Supplementary Methods).

Training details

We implemented the models using PyTorch and PyTorch Geometric. The Adam optimizer63 was utilized for model training with an initial learning rate of 0.001, a patience of 5 and a minimum learning rate of 1 × 10−5. Training was conducted on an NVIDIA V100 32-GB GPU, with a batch size of 64 and a total of 1,000 training steps. Specifically, for the VAE model, we used a latent space of 128 dimensions. The loss L1 was computed with weight parameters α = 1 and β = 1 × 10−4, respectively. The EGNN generator consists of two parts: part I includes 2 EGCL layers with 10 initial features, and part II contains 8 EGCL layers, with 35 initial features for the intermediate-state graph and 32 initial features for the target fragment graph (Supplementary Table 6). The graph embedding dimensions for the hidden and output layers were set to 1,024 and 256, respectively. Besides, the ED embedding was compressed into 512 dimensions using a CNN architecture with residual blocks. Training the EGNN generator was formulated as a classification task, using cross-entropy as the loss function. As only small amounts of atoms connecting hydrogens were labelled as positive samples for the part I model, we assigned a weight of 11 to the less frequent class during training to address the class imbalance.

SPR assay for FGFR3 and CDC42 inhibitors

The SPR assay was conducted using a Biacore 8 K instrument (Cytiva). Human FGFR3 (residues 449–759, Biortus) bearing a C-terminal Strep-tag II (BP32837-03A) and human CDC42 (GenBank NM_044472.2) with an N-terminal 6×His-tag were immobilized onto separate flow cells of a CM5 sensor chip. Each protein (50 μg ml−1 in 10-mM sodium acetate, pH 4.5) was coupled via amine chemistry to achieve a surface density of 18,000 response units for FGFR3 (10,000 response units for CDC42). The surface was first activated with a 1:1 mixture of 1-ethyl-3-(3-dimethylaminopropyl)carbodiimide/N-hydroxysuccinimide (contact time, 600 s; flow rate, 10 μl min−1), followed by protein immobilization and deactivation with 1-M ethanolamine (pH 8.5). Binding experiments were carried out at 25 °C with a running buffer comprising 10-mM Na2HPO4, 1.8-mM KH2PO4, 137-mM NaCl, 2.7-mM KCl, 0.05% Tween 20 and 5% dimethyl sulfoxide (DMSO), pH 7.4. Compounds were threefold serially diluted using the running buffer and injected over the sensor surface at a flow rate of 30 μl min−1 at 25 °C with association/dissociation times of 120 s/120 s for FGFR3 and 60 s/300 s for CDC42. Solvent corrections were applied using running buffers containing 4.5%–5.5% DMSO. The data were fitted with the steady-state affinity analysis model for FGFR3 and kinetic analysis model for CDC42 in Biacore 8 K evaluation software, respectively.

Bioluminescence resonance energy transfer assay for CDC42 inhibitors

Human CDC42 (NM044472.2) and PAK1 (NM001128620.2) genes were separately synthesized and cloned into a pcDNA3.1(+) vector containing either an N-terminal Rluc or GFP tag (Sangon). HEK293T cells (obtained from ATCC and routinely screened for mycoplasma contamination) were cultured in Dulbecco’s modified Eagle’s medium (Thermo Fisher Scientific) supplemented with 10% foetal bovine serum. Next, the HEK293T cells were plated on 6-cm dishes at a density of 1 × 106 cells per dish. After incubation for 18–24 h, when the cells reached 60%–80% confluency, they were co-transfected with 0.4 μg of Rluc-Cdc42 and 1.6 μg of GFP-PAK1 using the ExFect Transfection Reagent (Vazyme) according to the manufacturer’s protocol.

The bioluminescence resonance energy transfer assay was utilized to detect the interaction between PAK1 and GTP-bound, the active form of CDC42. The transfected HEK293T cells were reseeded into white, opaque-bottom 96-well plates at a density of 10,000 cells per 100 μl per well and cultured in Dulbecco’s modified Eagle’s medium supplemented with 10% foetal bovine serum at 37 °C and 5% CO2 for 12 h. The test compounds were then added to the well, and the cells were further incubated for an additional 12 h. Afterwards, the culture medium was carefully removed and replaced with Hank’s balanced salt solution (Gibco) containing 5-μM coelenterazine 400a (Maokangbio). Fluorescence readings were taken immediately using a Synergy Neo plate reader (BioTek) with emission filters at 410 nm and 525 nm. The inhibition ratio of the test compounds was normalized to values obtained from DMSO-treated controls. The data were fitted with the standard dose–response curve in GraphPad Prism v. 8.0 software to determine the IC50 values.

DEC assay for GCK activators

Human GCK (residues 1–465) was subcloned into the pET-28a vector. The recombinant protein was overexpressed in Escherichia coli BL21(DE3), induced with 0.5-mM isopropyl-1-thio-d-galactopyranoside at 16 °C overnight. Cells were harvested and lysed in a buffer containing 50-mM Tris-HCl pH 7.5, 300-mM NaCl and 5% glycerol. The solution was clarified by centrifugation at 27,670g for 45 min. The resulting supernatant was applied to a Ni-NTA column (GE Healthcare) equilibrated with the lysis buffer and then eluted with the butter containing 250-mM iminazole. The eluate was subsequently dialysed into GCK storage buffer (25-mM HEPES, 25-mM KCl, 2-mM MgCl2 and 1-mM DTT, pH 7.5).

The double-enzyme-coupled (DEC) assay was performed using glucose-6-phosphate dehydrogenase in a final volume of 120 μl containing 25-mM HEPES pH 7.1, 25-mM KCl, 10-mM glucose, 2-mM MgCl2, 1-mM DTT, 1-mM NAD+, 0.1% bovine serum albumin, 5 U ml−1 glucose-6-phosphate dehydrogenase and 18.7 μg ml−1 GK protein. After incubation at 37 °C for 30 min, the reaction was initiated by adding 1-mM ATP, and the production of NADPH was monitored kinetically at 340 nm. The activation ratio of the test compounds was normalized to values obtained from DMSO-treated controls. The data were fitted with the standard dose–response curve in GraphPad Prism v. 8.0 software to determine the EC50 values.

PRESTO-Tango assay for GPRC5A activators

The PRESTO-Tango plasmids used in this study were obtained from the Bryan Roth laboratory (Addgene, catalogue no. 66382). Helper T lymphocyte cells were seeded in white 96-well plates at a density of 40,000 cells per well and incubated for 24 h. Transient transfection was performed using ExFect Transfection Reagent (Vazyme, T101-01/02) with both GPRC5A-Tango plasmid and β-arrestin2-TEV plasmid. After 24 h of transfection, cells were treated with test compounds at various concentrations and incubated for an additional 24 h. After that, the culture medium was replaced with 100 µl of D-luciferin reagent (150 µg kg−1), prediluted 1:100 in 1× Dulbecco’s phosphate-buffered saline buffer.

Subsequently, the cell culture medium was replaced with 100 µl of D-luciferin reagent (150 µg kg−1), which had been prediluted 100-fold with 1× Dulbecco’s phosphate-buffered saline buffer. After a 5-min incubation at 37 °C in the dark, the luminescence was measured using a BioTek Synergy Neo2 Multi-Mode Microplate Reader. The data were fitted with the standard dose–response curve in GraphPad Prism v. 8.0 software to determine the EC50 values.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Data availability

All input data are freely available from public sources. The training data for the ligand ED extraction model were originally obtained from the PDB database (https://www.rcsb.org), and the processed data are available via Zenodo (https://doi.org/10.5281/zenodo.16736826)64. Similarly, the training data for the iterative fragment extension model were originally obtained from the ZINC database (https://zinc.docking.org), with the processed data accessible via Zenodo (https://doi.org/10.5281/zenodo.16736826)64. Source data are provided with this paper.

Code availability

The source code of ED2Mol is freely available under an MIT License via GitHub (https://github.com/pineappleK/ED2Mol)65, Zenodo (https://doi.org/10.5281/zenodo.16736826)64 or Code Ocean (https://doi.org/10.24433/CO.7635060.v2)66.

References

Sadybekov, A. V. & Katritch, V. Computational approaches streamlining drug discovery. Nature 616, 673–685 (2023).

Article   Google Scholar

Hughes, J. P., Rees, S., Kalindjian, S. B. & Philpott, K. L. Principles of early drug discovery. Br. J. Pharmacol. 162, 1239–1249 (2011).

Article   Google Scholar

He, J. et al. ASD2023: towards the integrating landscapes of allosteric knowledgebase. Nucleic Acids Res. 52, D376–D383 (2024).

Article   Google Scholar

Irwin, J. J. et al. ZINC20—a free ultralarge-scale chemical database for ligand discovery. J. Chem. Inf. Model. 60, 6065–6073 (2020).

Article   Google Scholar

Reymond, J. L. The chemical space project. Acc. Chem. Res. 48, 722–730 (2015).

Article   Google Scholar

Schneider, G. & Fechner, U. Computer-based de novo design of drug-like molecules. Nat. Rev. Drug Discov. 4, 649–663 (2005).

Article   Google Scholar

Isert, C., Atz, K. & Schneider, G. Structure-based drug design with geometric deep learning. Curr. Opin. Struct. Biol. 79, 102548 (2023).

Article   Google Scholar

Li, M., Lan, X., Lu, X. & Zhang, J. A structure-based allosteric modulator design paradigm. Health Data Sci. 3, 0094 (2023).

Article   Google Scholar

Harris, C. et al. PoseCheck: Generative models for 3D structure-based drug design produce unrealistic poses. In NeurIPS 2023 Generative AI and Biology (GenBio) Workshop (2023); https://openreview.net/forum?id=Nf5BxVllgq

Buttenschoen, M., Morris, G. M. & Deane, C. M. PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences. Chem. Sci. 15, 3130–3139 (2024).

Article   Google Scholar

Jiang, Y. et al. PocketFlow is a data-and-knowledge-driven structure-based molecular generative model. Nat. Mach. Intell. 6, 326–337 (2024).

Article   Google Scholar

Pan, Y. Heading toward Artificial Intelligence 2.0. Engineering 2, 409–413 (2016).

Article   Google Scholar

Skalic, M., Sabbadin, D., Sattarov, B., Sciabola, S. & De Fabritiis, G. From target to drug: generative modeling for the multimodal structure-based ligand design. Mol. Pharm. 16, 4282–4291 (2019).

Article   Google Scholar

Zhung, W., Kim, H. & Kim, W. Y. 3D molecular generative framework for interaction-guided drug design. Nat. Commun. 15, 2688 (2024).

Article   Google Scholar

Zhang, Z., Min, Y., Zheng, S. & Liu, Q. Molecule generation for target protein binding with structural motifs. In Eleventh International Conference on Learning Representations (2023); https://openreview.net/forum?id=Rq13idF0F73

Nakatsuji, H. Electron-cloud following and preceding and the shapes of molecules. J. Am. Chem. Soc. 96, 30–37 (1974).

Article   Google Scholar

Rico, J. F., López, R., Ema, I. & Ramírez, G. Chemical forces in terms of the electron density. Theor. Chem. Acc. 118, 709–721 (2007).

Article   Google Scholar

Terwilliger, T. C., Klei, H., Adams, P. D., Moriarty, N. W. & Cohn, J. D. Automated ligand fitting by core-fragment fitting and extension into density. Acta Cryst. D62, 915–922 (2006).

Google Scholar

Mysinger, M. M., Carchia, M., Irwin, J. J. & Shoichet, B. K. Directory of useful decoys, enhanced (DUD-E): better ligands and decoys for better benchmarking. J. Med. Chem. 55, 6582–6594 (2012).

Article   Google Scholar

Hopkins, A. L., Keseru, G. M., Leeson, P. D., Rees, D. C. & Reynolds, C. H. The role of ligand efficiency metrics in drug discovery. Nat. Rev. Drug Discov. 13, 105–121 (2014).

Article   Google Scholar

Ertl, P. & Schuffenhauer, A. Estimation of synthetic accessibility score of drug-like molecules based on molecular complexity and fragment contributions. J. Cheminform. 1, 8 (2009).

Article   Google Scholar

Bickerton, G. R., Paolini, G. V., Besnard, J., Muresan, S. & Hopkins, A. L. Quantifying the chemical beauty of drugs. Nat. Chem. 4, 90–98 (2012).

Article   Google Scholar

Schneider, G. & Clark, D. E. Automated de novo drug design: are we nearly there yet? Angew. Chem. Int. Ed. 58, 10792–10803 (2019).

Article   Google Scholar

Mozaffari-Jovin, S. et al. Inhibition of RNA helicase Brr2 by the C-terminal tail of the spliceosomal protein Prp8. Science 341, 80–84 (2013).

Article   Google Scholar

Iwatani-Yoshihara, M. et al. Discovery of allosteric inhibitors targeting the spliceosomal RNA helicase Brr2. J. Med. Chem. 60, 5759–5771 (2017).

Article   Google Scholar

Choi, S.-S., Park, J. & Choi, J. H. Revisiting PPARγ as a target for the treatment of metabolic disorders. BMB Rep. 47, 599–608 (2014).

Article   Google Scholar

Artis, D. R. et al. Scaffold-based discovery of indeglitazar, a PPAR pan-active anti-diabetic agent. Proc. Natl Acad. Sci. USA 106, 262–267 (2009).

Article   Google Scholar

L’Hote, C. G. & Knowles, M. A. Cell responses to FGFR3 signalling: growth, differentiation and apoptosis. Exp. Cell Res. 304, 417–431 (2005).

Article   Google Scholar

Krook, M. A. et al. Fibroblast growth factor receptors in cancer: genetic alterations, diagnostics, therapeutic targets and mechanisms of resistance. Br. J. Cancer 124, 880–892 (2021).

Article   Google Scholar

Huang, Z. et al. Structural mimicry of a-loop tyrosine phosphorylation by a pathogenic FGF receptor 3 mutation. Structure 21, 1889–1896 (2013).

Article   Google Scholar

Li, M., Rehman, A. U., Liu, Y., Chen, K. & Lu, S. Dual roles of ATP-binding site in protein kinases: orthosteric inhibition and allosteric regulation. Adv. Protein Chem. Struct. Biol. 124, 87–119 (2021).

Article   Google Scholar

Attwood, M. M., Fabbro, D., Sokolov, A. V., Knapp, S. & Schioth, H. B. Trends in kinase drug discovery: targets, indications and inhibitor design. Nat. Rev. Drug Discov. 20, 839–861 (2021).

Article   Google Scholar

Chu, W. T., Chu, X. & Wang, J. Binding mechanism and dynamic conformational change of C subunit of PKA with different pathways. Proc. Natl Acad. Sci. USA 114, E7959–E7968 (2017).

Article   Google Scholar

Melendez, J., Grogg, M. & Zheng, Y. Signaling role of Cdc42 in regulating mammalian physiology. J. Biol. Chem. 286, 2375–2381 (2011).

Article   Google Scholar

Maldonado, M. D. M. & Dharmawardhane, S. Targeting Rac and Cdc42 GTPases in cancer. Cancer Res. 78, 3101–3111 (2018).

Article   Google Scholar

Jahid, S. et al. Structure-based design of CDC42 effector interaction inhibitors for the treatment of cancer. Cell Rep. 39, 110760 (2022).

Article   Google Scholar

Matschinsky, F. M., Glaser, B. & Magnuson, M. A. Pancreatic beta-cell glucokinase: closing the gap between theoretical concepts and experimental realities. Diabetes 47, 307–315 (1998).

Article   Google Scholar

Grimsby, J. et al. Allosteric activators of glucokinase: potential role in diabetes therapy. Science 301, 370–373 (2003).

Article   Google Scholar

Nishimura, T. et al. Identification of novel and potent 2-amino benzamide derivatives as allosteric glucokinase activators. Bioorg. Med. Chem. Lett. 19, 1357–1360 (2009).

Article   Google Scholar

Mitsuya, M. et al. Discovery of novel 3,6-disubstituted 2-pyridinecarboxamide derivatives as GK activators. Bioorg. Med. Chem. Lett. 19, 2718–2721 (2009).

Article   Google Scholar

Jiang, X. et al. GPRC5A: an emerging biomarker in human cancer. Biomed. Res. Int. 2018, 1823726 (2018).

Article   Google Scholar

Abramson, J. et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493–500 (2024).

Article   Google Scholar

Huang, W. K. et al. AlloSite: a method for predicting allosteric sites. Bioinformatics 29, 2357–2359 (2013).

Article   Google Scholar

Atz, K. et al. Prospective de novo drug design with deep interactome learning. Nat. Commun. 15, 3408 (2024).

Article   Google Scholar

Hert, J., Irwin, J. J., Laggner, C., Keiser, M. J. & Shoichet, B. K. Quantifying biogenic bias in screening libraries. Nat. Chem. Biol. 5, 479–483 (2009).

Article   Google Scholar

Mnih, V. et al. Human-level control through deep reinforcement learning. Nature 518, 529–533 (2015).

Article   Google Scholar

Cremer, J., Le, T., Noe, F., Clevert, D. A. & Schutt, K. T. PILOT: equivariant diffusion for pocket-conditioned de novo ligand generation with multi-objective guidance via importance sampling. Chem. Sci. 15, 14954–14967 (2024).

Article   Google Scholar

Grosse-Kunstleve, R. W., Sauter, N. K., Moriarty, N. W. & Adams, P. D. The Computational Crystallography Toolbox: crystallographic algorithms in a reusable software framework. J. Appl. Crystallogr. 35, 126–136 (2002).

Article   Google Scholar

Ganea, O. et al. GeoMol: torsional geometric generation of molecular 3D conformer ensembles. Adv. Neural Inf. Process. Syst. 34, 13757–13769 (2021).

Google Scholar

Li, W. & Godzik, A. Cd-hit: a fast program for clustering and comparing large sets of protein or nucleotide sequences. Bioinformatics 22, 1658–1659 (2006).

Article   Google Scholar

Gao, M. & Skolnick, J. APoc: large-scale identification of similar protein pockets. Bioinformatics 29, 597–604 (2013).

Article   Google Scholar

Riniker, S. & Landrum, G. A. Better informed distance geometry: using what we know to improve conformation generation. J. Chem. Inf. Model. 55, 2562–2574 (2015).

Article   Google Scholar

Francoeur, P. G. et al. Three-dimensional convolutional neural networks and a cross-docked data set for structure-based drug design. J. Chem. Inf. Model. 60, 4200–4215 (2020).

Article   Google Scholar

Zha, J., Li, M., Kong, R., Lu, S. & Zhang, J. Explaining and predicting allostery with allosteric database and modern analytical techniques. J. Mol. Biol. 434, 167481 (2022).

Article   Google Scholar

Akbar, R. & Helms, V. ALLO: a tool to discriminate and prioritize allosteric pockets. Chem. Biol. Drug Des. 91, 845–853 (2018).

Article   Google Scholar

Huang, W. et al. ASBench: benchmarking sets for allosteric discovery. Bioinformatics 31, 2598–2600 (2015).

Article   Google Scholar

Peng, X. et al. Pocket2Mol: efficient molecular sampling based on 3D protein pockets. In Proc. 39th International Conference on Machine Learning 17644–17655 (PMLR, 2022).

Liu, M., Luo, Y., Uchino, K., Maruhashi, K. & Ji, S. Generating 3D molecules for target protein binding. In Proc. 39th International Conference on Machine Learning 13912–13924 (PMLR, 2022).

Zhang, O. et al. ResGen is a pocket-aware 3D molecular generation model based on parallel multiscale modelling. Nat. Mach. Intell. 5, 1020–1030 (2023).

Article   Google Scholar

Guan, J. et al. 3D Equivariant diffusion for target-aware molecule generation and affinity prediction. In Eleventh International Conference on Learning Representations (2023); https://openreview.net/forum?id=kJqXEPXMsE0

Xie, J., Chen, S., Lei, J. & Yang, Y. DiffDec: structure-aware scaffold decoration with an end-to-end diffusion model. J. Chem. Inf. Model. 64, 2554–2564 (2024).

Article   Google Scholar

Zhang, Y. et al. FragGrow: a web server for structure-based drug design by fragment growing within constraints. J. Chem. Inf. Model. 64, 3970–3976 (2024).

Article   Google Scholar

Kingma, D. P. Adam: a method for stochastic optimization. Preprint at https://arxiv.org/abs/1412.6980 (2014).

Li, M. Electron-density informed effective and reliable de novo molecular design and optimization with ED2Mol. Zenodo https://doi.org/10.5281/zenodo.16736826 (2025).

Li, M. ED2Mol: Electron-density-informed effective and reliable de novo molecular design and optimization with ED2Mol. Github https://github.com/pineappleK/ED2Mol (2025).

Li, M. ED2Mol: electron-density informed effective and reliable de novo molecular design and optimization. Code Ocean https://doi.org/10.24433/CO.7635060.v2 (2025).

Download references

Acknowledgements

This study was partly supported by grants from the National Key R&D Program of China (2023YFF1205103 to J. Zhang), the National Natural Science Foundation of China (81925034, 82441035 and 22237005 to J. Zhang; 824B2105 to M.L.), Innovative Research Team of High-Level Local Universities in Shanghai (SHSMU-ZDCX20212700 to J. Zhang), the Starry Night Science Fund of Zhejiang University Shanghai Institute for Advanced Study (SN-ZJU-SIAS-007 to J. Zhang), the Key Research and Development Program of Ningxia Hui Autonomous Region (2022BEG01002 to J. Zhang), Ningxia Peptide and Small Molecule Innovative Drug Research Science and Technology Innovation Team (2024CXTD013 to J. Zhang), Open Competition Mechanism to Select the Best Candidates for Key Research Projects of Ningxia Medical University (XJKF240322 to J. Zhang), Lingang Laboratory (LG8888 to J. Zhang), Zhiyuan Scholar Program (ZIRC2020-06, to M.L.), and Shanghai Jiao Tong University School of Medicine PhD Cultivation Fund for Science and Innovation (24KCPYZD001, to M.L.).

Author information Author notes

These authors contributed equally: Mingyu Li, Kun Song, Jixiao He.

Authors and Affiliations

State Key Laboratory of Medical Genomics, National Research Center for Translational Medicine at Shanghai, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China

Mingyu Li, Jixiao He, Mingzhu Zhao, Jie Zhong, Jiacheng Wei, Jiamin Zhou, Weidong Lin, Xun Lu & Jian Zhang

Department of Pharmaceutical and Artificial-Intelligence Sciences, Institute of Medical Artificial Intelligence, Shanghai Jiao Tong University School of Medicine, Shanghai, China

Mingyu Li, Jixiao He, Mingzhu Zhao, Jie Zhong, Jiacheng Wei, Jiamin Zhou, Weidong Lin, Xun Lu & Jian Zhang

College of Pharmacy, Ningxia Medical University, Yinchuan, China

Mingyu Li & Jian Zhang

Nutshell BioTech Co. Ltd, Shanghai, China

Kun Song, Gengshu You & Zhengtian Yu

Blueray Biopharma Co. Ltd, Shanghai, China

Mengxi Zhao, Arong Li, Yu Chen, Guobin Li, Ying Kong, Zhaofu Wang, Hongbing Yang, Shichao Ma & Hailong Zhang

State Key Laboratory of Drug Research, Shanghai Institute of Materia Medica, Chinese Academy of Sciences, Shanghai, China

Irakoze Loïca Mélita, Yuhang Lu & Yujun Zhao

University of Chinese Academy of Sciences, Beijing, China

Irakoze Loïca Mélita, Yuhang Lu & Yujun Zhao

Contributions

J. Zhang conceived and supervised the project. J. Zhang, M.L. and K.S. designed the experiments. M.L., K.S. and G.Y. established and validated the ED2Mol model. J.H., Mingzhu Zhao, J. Zhou, I.L.M., Y.L., X.L. and Y.Z. carried out the chemical synthesis. J. Zhong, J.W. and W.L. performed the bioactivity assays. Mengxi Zhao, A.L., Y.C., G.L., Y.K., Z.W., H.Y., S.M. and H.Z. solved the crystal structures. M.L. and K.S. drafted the manuscript, and all authors contributed to specific parts of the manuscript. J. Zhang assumed responsibility for the manuscript in its entirety. M.L., K.S., Z.Y. and J. Zhang discussed the results and revised the manuscript.

Corresponding author

Correspondence to Jian Zhang.

Ethics declarations Competing interests

The authors declare no competing interests.

Peer review Peer review information

Nature Machine Intelligence thanks Wanxiang Shen and the other, anonymous, reviewer(s) for their contribution to the peer review of this work.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Extended data Extended Data Fig. 1 Detailed frameworks of fragment attachment in ED2Mol.

The process of growing a molecule from state n to state n + 1 can be decomposed into two main steps: (1) assigning the growth point using the first equivariant graph neural network (EGNN) model, and (2) predicting the torsion angles of the attachment using the second EGNN model. The second EGNN model consists of three blocks designed to process distinct inputs. Block 1 accepts the ED map and extracts features using a residual convolutional neural network (CNN). Block 2 and Block 3 employ EGNNs to generate graph-level embeddings for the core and the attaching fragments, respectively. The embeddings from these three blocks are then aggregated and passed through multi-layer perceptrons (MLPs) to predict the torsion angles for adding the attaching fragment to the molecule.

Extended Data Fig. 2 Visualized distributions of binding positions for all generated molecules by different methods.

Three quintessential drug targets are shown in cartoon representation (white). The blue and red points correspond to the positions of heavy atoms in the generated molecules, based on orthosteric pockets in DUD-E and allosteric pockets in ASB-E sets, respectively. Owing to ResGen’s inability to produce an adequate number of molecules, the distribution appears notably sparse. To provide a quantitative description, we introduced pocket occupancy density to measure the overall atom density within the pocket, alongside occupancy variance and converge to capture the distsribution’s uniformity and extent. Specifically, the pocket volume was discretized into voxels with a grid spacing of 0.375 Å, consistent with AutoDock Vina, and the number of atoms from all generated molecules within each voxel were counted. Occupancy density and variance were computed as the average and variance of these counts, while occupancy coverage was defined as the fraction of pocket voxels containing at least one generated atom. The occupancy variance (↓) and Occupancy converge (↑) for ED2Mol, Pocket2Mol, and TargetDiff are 1260 and 84%, 1782 and 75%, and 1278 and 74%, respectively.

Supplementary information Supplementary Information

Supplementary Tables 1–7, Figs. 1–12, Schemes 1–8, Methods and chemical synthesis routes.

Reporting Summary Supplementary Data 1

Source data for bioassays in the Supplementary Information.

Source data Source Data Fig. 4

Source data for Fig. 4b,d.

Source Data Fig. 5

Source data for Fig. 5b,d,e.

Source Data Fig. 6

Source data for Fig. 6d.

Rights and permissions

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.

Reprints and permissions

About this article Cite this article

Li, M., Song, K., He, J. et al. Electron-density-informed effective and reliable de novo molecular design and optimization with ED2Mol. Nat Mach Intell 7, 1355–1368 (2025). https://doi.org/10.1038/s42256-025-01095-7

Download citation

Received 20 December 2024

Accepted 08 July 2025

Published 20 August 2025

Version of record 20 August 2025

Issue date August 2025

DOI https://doi.org/10.1038/s42256-025-01095-7

Share this article

Anyone you share the following link with will be able to read this content:

Get shareable link

Provided by the Springer Nature SharedIt content-sharing initiative

Subjects Drug discovery Machine learning This article is cited by Real-world validation of a structure-aware pipeline for molecular design Ana Laura DiasTiago Rodrigues

Nature Machine Intelligence (2025)