pocketz.doi.bio/2026-01-11_182008_nature_s42256-025-00997-w.md

Deep lead optimization enveloped in protein pocket and its application in designing potent and selective ligands targeting LTK protein | Nature Machine Intelligence

URL: https://www.nature.com/articles/s42256-025-00997-w

Saved: 2026-01-11T02:20:13.837Z

Article Published: 20 February 2025 Deep lead optimization enveloped in protein pocket and its application in designing potent and selective ligands targeting LTK protein Shicheng Chen, Odin Zhang, Chenran Jiang, Huifeng Zhao, Xujun Zhang, Mengting Chen, Yun Liu, Qun Su, Zhenxing Wu, Xinyue Wang, Wanglin Qu, Yuanyi Ye, Xin Chai, Ning Wang, Tianyue Wang, Yuan An, Guanlin Wu, Qianqian Yang, Jiean Chen, Wei Xie, Haitao Lin, Dan Li, Chang-Yu Hsieh, Yong Huang, …Peichen Pan Show authors

Nature Machine Intelligence volume  7,  pages 448–458 (2025)Cite this article

5929 Accesses

12 Citations

5 Altmetric

Metrics details

A preprint version of the article is available at arXiv.

Abstract

Optimizing the chemical structure of promising drug candidates through systematic modifications to improve potency and physiochemical properties is a vital step in the drug discovery pipeline. In contrast to the well-established de novo generation schemes, computational methods specifically tailored for lead optimization remain largely underexplored. Prior models are often limited to addressing specific subtasks, such as generating two-dimensional molecular structures, while neglecting crucial protein–ligand interactions in three-dimensional space. To overcome these challenges, we propose Delete (Deep lead optimization enveloped in protein pocket), a one-stop solution for lead optimization by combining generative artificial intelligence and structure-based approaches. Our model can handle all subtasks of lead optimization through a unified deleting (masking) strategy, and it accounts for intricate pocket–ligand interactions through an equivariant network design. Statistical assessments and retrospective studies across individual subtasks demonstrate that Delete has an outstanding ability to craft molecules with superior protein-binding energy and reasonable drug-likeness using given fragments or atoms. Subsequently, we utilize Delete to design inhibitors targeting the previously identified LTK protein. Among the ligands designed by Delete, CA-B-1 is successfully validated as a potent (1.36 nM) and selective inhibitor by in vitro and in vivo experiments. This work represents a successful implementation of the powerful structure-based lead optimization model, Delete, for rapid and controllable rational drug design.

You have full access to this article via your institution.

Download PDF Similar content being viewed by others Computational approaches streamlining drug discovery Article 26 April 2023 3D molecular generative framework for interaction-guided drug design Article Open access 27 March 2024 A geometric deep learning approach to predict binding conformations of bioactive molecules Article 02 December 2021 Main

Drug discovery has long been recognized as a time-consuming, costly and risky process1, and modern artificial intelligence-aided drug design (AIDD) methods aim to accelerate this process by facilitating the automatic discovery of drug candidates for chemists. The AIDD approach has shown its effectiveness in developing antibacterial drugs, as evidenced by the three studies led by Collins2,3,4. This approach is confined to well-studied pharmaceutic targets with sufficient molecular data for training reliable quantitative structure–activity relationship (QSAR) models. Another storyline of AIDD lies in molecular generation, exemplified by JAEJER5, GENTRL6 and GDL-RIPK1 (ref. 7), which explore molecules directly in a continuous chemical space by transcending discrete chemical libraries. However, current validated molecular generation efforts have primarily focused on ligand-based de novo design models. For one thing, the ligand-based approach is also constrained to well-studied targets, requiring finetuning of molecules proven to be active against the protein of interest. Additionally, chemists often come to the table with prior knowledge, aiming to modify lead compounds or scaffolds that exhibit low affinity or other less-than-ideal properties8. Unfortunately, the de novo design approach usually fails to integrate chemists’ prior knowledge into the generation process, thus reducing their inclination to rely on these ‘black box’ models.

To address these challenges, we tried to realize two foundational concepts in our approach: the structure-based paradigm and the lead optimization approach. This integration gave birth to an AI-driven model named Delete (Fig. 1), which more accurately aligns with the complexities of real-world requirements. Unlike the ligand-based paradigm that primarily focuses on recognizing chemical patterns through data memorization, the structure-based approach seeks to grasp the underlying physics9,10. However, it has only received attention in recent years, such as with DiffSBDD11, Pkt2Mol12 and ResGen9, which treat protein structures as hard conditions as they directly generate molecules within pockets. Inspired by these works, Delete utilized a geometric-aware and equivariant network as the backbone to comprehend detailed protein-induced interactions. In this way, the model can be effectively utilized to study unseen targets, because although three-dimensional (3D) structures vary depending on the case, the underlying fundamental physics remains the same. Another concept, lead optimization, has been widely employed to develop me-too/me-better drugs by using approved medications as starting points13,14 and to design first-in-class drugs through fragment-based drug discovery techniques15. However, despite its notable practical benefits, lead optimization has garnered less research and attention than de novo design methods in the field of AIDD. Current lead optimization models in the AIDD landscape tend to be single-task oriented and mainly ligand-based, such as DeepFrag16, DiffLinker17, FFLOM18 and ScaffoldGVAE19. These two characteristics of these models limit their applicability in real-world scenarios. To make Delete capable of handling all these tasks in lead optimization, we introduced a unified masking strategy. It should be noted that after our model preprint was published, two models, REINVENT4 (ref. 20) and DrugEX v.3 (ref. 21), which realize both the structure-based and lead optimization concepts, were released. However, they utilize docking scores as the reward function for reinforcement learning (RL) to guide molecular generation, which still cannot overcome the false-positive issues brought by the empirical docking scores22. Machine-learning-based scoring functions have the potential to mitigate these issues23, and we anticipate that future research will focus on integrating these approaches with classical scoring methods to create hybrid models that leverage the strengths of both techniques. Therefore, Delete enjoys the methodological advantages of its kind as an end-to-end structure-learning model that does not rely on training with noisy affinity labels.

Fig. 1: Delete is a one-for-all solution for lead optimization.

a, The workflow of Delete. b, Delete is a pocket-aware and 3D-generation model, which expands de novo design to four lead optimization tasks. c, The unified deleting (masking) strategies illustrated. The first three are enhanced strategies, and the last four are task-specific strategies.

Full size image

Our work with Delete represents a step forward in bridging the aforementioned gap, offering a pioneering structure-based, comprehensive solution for lead optimization. To be more specific, we summarized the lead optimization task into four distinct subtasks: linker design, fragment elaboration, scaffold hopping and side-chain decorations. To ensure that Delete adheres strictly to physical laws, our objective was to integrate the model with rotational equivariance and invariance24,25. In other words, Delete’s output geometry dynamically aligns with input rotation while maintaining a constant atomic and bonding topology, which eliminates the need for data augmentation and enables the concurrent generation of protein–ligand binding structures. To achieve this, we employed equivariant scalarization techniques within the model’s design. Inspired by masked graph modelling26 and pretraining strategies27 for Graph data structures, we introduced a unified deleting strategy, as shown in Fig. 1b, including three enhanced maskers (random, spatial and topological) and four task-specific maskers (linker, fragment, scaffold and side chains). Collectively, these elements contribute to the construction of a one-stop lead optimization model, Delete.

To validate its effectiveness, we conducted a comprehensive evaluation encompassing model comparisons and retrospective analyses, as illustrated in Fig. 2a, to swiftly assess its performance. Furthermore, Delete was applied in conceptualizing new inhibitors targeting the Leukocyte Tyrosine Kinase (LTK) protein, a previously identified target28, as illustrated in Fig. 2b. Specifically, in silico experiments demonstrated Delete’s superiority over seven baseline models across four subtasks, measured by predicted binding affinity. Additionally, Delete successfully reproduced the outcomes of five historical drug development cases in retrospective studies. In wet-lab experiments, we extended Delete to a hybrid version of the structure-based and ligand-based models by leveraging the transfer-learning technique (Fig. 2c). Finally, Delete successfully designed a series of active ligands targeting the kinase domain of LTK within the CLIP1–LTK fusion protein, where the most promising candidate, CA-B-1, shows excellent nanomolar activity (half-maximal inhibitory concentration (IC50) = 1.36 nM). Subsequent in vitro and in vivo tests on CA-B-1 confirmed its antitumor efficacy and the mechanism by which it impedes cancer cell proliferation. In summary, this work underscores the transformative potential of structure-based molecular generative techniques in drug discovery, which can accelerate drug design within a month and a half. We believe the Delete methodology will serve as a potential tool in future targeted therapy development.

Fig. 2: Retrospective studies and transfer-learning technique.

a, Illustration of five retrospective studies. b, Integrating ligand-based approach into Causal-Delete, using Delete to design inhibitors targeting the previously identified LTK target. c, Integrating ALK inhibitors’ chemical space into Delete. The transfer-learning technique used in real-world application, aiming to leverage ALK’s chemical space to help LTK ligand design.

Full size image Results and discussion

The construction of Delete relies on a unified masking strategy, an interaction framework, a molecule generation framework and a loss function. Following the input of the protein and lead compound fragments, the protein is modelled using a topological surface model ({{\mathcal{G}}}{\rm{p}}), and the molecule is modelled using a geometric graph model ({{\mathcal{G}}}{\rm{l}}). The interaction characterization framework learns the topological and geometric interactions of protein pockets and lead fragments, resulting in a node representation with geometric and energy information. Molecular generation takes on an autoregressive form that allows it to completely interact with topological pockets before using the pocket information as a strong constraint to direct the molecular generation process. The workflow of Delete is simplified and shown for understanding in Fig. 1.

Delete offers a one-stop solution for lead optimization, which fully considers the chemical environments and geometric profiles within protein pockets. To demonstrate its potential, we conducted a comparative analysis of Delete against seven other task-specific models across four subtasks. However, structure-based generation arts have only received attention in recent years, limiting the pool of directly comparable models to a select few, including DeepFrag15, DiffLinker17 and Fragmenstein29. To ensure a robust comparison, we also included four additional models as two-dimensional (2D) baselines across these subtasks, including DeLinker30, FFLOM18, ScaffoldGVAE19 and SCMG31. It’s important to acknowledge that computational metrics in model comparison have their own biases and sometimes can obscure the generation quality, as highlighted in prior studies32. As part of this, we curated five historical drug development cases for each of the four tasks (including two instances of linker design) to demonstrate Delete’s performance across different scenarios. Building on these retrospective analyses, Delete was then utilized to explore drug design for the LTK protein, which was recently identified as a promising avenue for targeted therapy in non-small cell lung cancer. Through extensive experimentation, Delete’s effectiveness was thoroughly validated across model comparisons, historical case reproductions and the pursuit of new drug candidates.

Delete outperformed baselines across four tasks

When evaluating a structure-based model that treats protein pockets as stringent constraints, the primary metric is the binding tightness between the designed molecule and the input target. This binding tightness is most intuitively reflected by the in situ binding energy, referred to as the Scoring metric, which represents the summation of the interactions between the directly generated molecule and the protein; the Docking metric recalculates the binding energy after optimizing the molecular pose within the pocket. The former metric is more sensitive to the model’s ability to generate accurate geometry, whereas the latter is a close approximation of the actual binding potential of the model-generated molecules (as a molecule in nature dynamically seeks the most suitable pose for protein binding). Among these metrics, Delete surpassed the baseline models, achieving a binding energy gain of nearly 1 kcal mol−1 over the second-best method in each subtask, as shown in Table 1 (more statistical data are shown in Supplementary Table 1).

Table 1 Model comparison on four subtasks Full size table

Notably, DeLinker, FFLOM, ScaffoldGVAE and SCMG are 2D generation methods that do not consider protein structures, thus lacking the Scoring metric. Except for SCMG, the average docking binding energies of the three models were lower than those of the molecules in the test set, highlighting the importance of considering structures in designing targeted drugs. For the other three models (that is, DiffLinker, DeepFrag and Fragmenstein) that can perceive protein structures, although DiffLinker exhibited a docking binding energy close to the average of the test set, the other two models lagged by about 0.5 kcal mol−1. The inferior performance of DeepFrag and Fragmenstein can be attributed to their limited generation capabilities. To explain further, DeepFrag scores potential fragments from a predefined library and adds high-scoring ones to the lead compound, effectively simplifying the generation problem to a classification problem. Hence, its exploration space is constrained by the preprepared fragment library, potentially hindering the generation of molecules that match the protein pocket well. Of note, Fragmenstein can generate only one result per iteration, failing to explore sufficient chemical space. In summary, Delete has demonstrated state-of-the-art performance across all subtasks compared with task-specific baselines, measured by the Docking, Scoring and Hit metrics.

Besides the most directly related energy metrics, Table 1 also lists several common molecular property metrics, including quantitative estimate of drug-likeness (QED)33, synthetic accessibility34, Lipinski’s Rule of Five35 and LogP36. These metrics are typically used as conditions in molecular property optimization models: that is, conditional generation models. However, in our structure-based model, the condition is the input protein structure, rather than molecular properties. Consequently, we believe that our structure-based model aims to generate molecules similar to those in the test set, especially considering their relatively indirect relevance to the metrics. For the 2D baselines, they cumulatively achieved the best performance eight times across four tasks regarding these molecular property metrics, possibly because 2D models are generally trained on larger compound libraries such as CHEMBL37 and ZINC38, which are considerably larger by 1–2 orders of magnitude than the 100,000 molecules in CrossDock39, naturally allowing 2D models to have a wide range of choices in the chemical space during generation. In addition, the lack of protein structure constraints allows 2D models to freely navigate the chemical space. In other words, 2D models only need to ensure the rationality of the generated molecules, leading to impressive results in molecular property metrics. Regarding the 3D baselines (that is, DeepFrag, Fragmenstein and DiffLinker), Delete scored the best once for a specific molecular property (synthetic accessibility in linker design), whereas DeepFrag emerged as the top performer three times. DeepFrag’s good performance in molecular properties is mainly attributed to its limited fragment library, where small fragments naturally contribute to improvements in computational property metrics. Overall, although Delete cannot always achieve state-of-the-art results in molecular properties due to the inherent limitations of its structure-based methodology, its generated outcomes are still satisfactory, closely resembling the molecules in the test set. Furthermore, our analysis of the generated geometry quality of Delete demonstrated that the Delete-generated binding patterns are near-native, as detailed in Supplementary Information Part 2.

Delete reproduced historical drug design cases

Due to the inherent biases of computational metrics, energy metrics calculated by docking software40 and molecular properties may not be accurate enough for model assessment. For example, the binding energy metric tends to favour molecules with a higher number of atoms due to the additive nature of binding energy contributions. Conversely, for synthesizability, molecules with fewer atoms are naturally easier to synthesize. Therefore, retrospective studies, which simulate historical drug discovery cases to check whether the models can reproduce the corresponding output results, provide another intuitive and qualitative perspective. Here we conducted five retrospective studies, including linker design for inosine 5’-monophosphate dehydrogenase (IMPDH) and SMARCA2 (PROTAC), fragment elaboration for β1-adrenergic receptor (Adrb1), scaffold hopping for kinesin Eg5 and side-chain decoration for D2 dopamine receptor (DRD2). For clarity, we only discuss the case of Eg5 in the main text, with the other four examples detailed in Supplementary Information Part 3. The input files used in these cases are available in the code repository for future reference. Kinesin Eg5 is a crucial target in the development of cancer chemotherapy drugs41. Ref. 41 discovered an allosteric pocket of Eg5, along with an inhibitor BI8, exhibiting nanomolar affinity (Fig. 3a–c). To accomplish the scaffold-hopping task, we employed the Bemis–Murcko scheme to identify the core scaffold with its associated side chains and then removed the scaffold of the BI8 compound, retaining only a few tiny fragments (TFs) for generation (Fig. 3d). The objective of this study is to investigate whether, using only TFs, Delete can construct reasonable scaffolds, generate compounds resembling the original structure and create compounds that retain the geometrical similarity while introducing a new scaffold. If Delete proves successful in this challenging scenario, it is expected to perform well in other smaller scaffold-hopping applications.

Fig. 3: Scaffold-hopping case study of Eg5.

a,b, 2D chemical structure of BI8 (a) and its scaffold (b). Red represents the Bemis–Murcko backbones, and the blue part represents the TFs. c, Crystal structure of BI8 (PDB: 3ZCW). d, The starting point for Delete generation, where the red cycle is the potential growing space. e, Analysis of four generated molecules. f,g, The scoring energy (f) and the docking energy (g). The blue curve is the predicting affinity distribution of generated compounds, and the orange line denotes the original compound. SA, synthetic accessibility.

Full size image

Delete successfully generated 131 structures, and surprisingly, the results reveal that these compounds closely resemble the original compound. The average 2D similarity of the generated structures is only 0.4700, indicating a substantial structural divergence from the original compound (with 0.6 typically considered a threshold in ref. 42), and the 3D shape similarity reached a high score of 0.7878, indicating that the spatial characteristics of the original compound was preserved in the generated molecules. By directly scoring the generated compounds using AutoDock-Vina, we found that the computed binding energies of the generated compounds were comparable or even superior to that of the original compound (Fig. 3f). Furthermore, we redocked the generated and original compounds and compared their docking scores. The results show that 40.5% of the generated compounds achieve better scores than the original one (Fig. 3g). Figure 3e exhibits the overlays of the four generated compounds, the interactions analysis and the corresponding 2D structures of the linker fragments. The first example generates a skeleton structure that closely resembles the original, differing only by a single fluorine atom, and exhibits a near-perfect 3D conformational overlay with the original and maintains most of its interactions. The remaining three examples illustrate three molecules generated by Delete, each with distinct scaffolds yet retaining comparable 3D conformational features. These molecules not only maintain most of the interactions observed in the crystal structure of the original compound but also explore additional interactions, such as hydrophobic interactions with W127, L160, L214 and F239 and hydrophilic interactions with G217.

Delete application on LTK target

CLIP1–LTK was recently discovered as an oncogenic gene fusion in non-small cell lung cancer, and its oncogenic potential depends on the kinase activity of LTK28. The research demonstrated that LTK is a therapeutically actionable target that can be suppressed by lorlatinib—a third-generation ALK/ROS1 inhibitor43. So far, there have been no publications on the design of LTK-targeted inhibitors. The successful treatment of ALK inhibitors implies that we can use some fragments on the ALK target as lead molecules to anchor compounds into the binding pocket and then use Delete to explore the in-pocket–ligand structures for selective inhibitors design. Thus, we finetuned the general Delete on the set of ALK inhibitors using transfer-learning techniques while evaluating single molecular impact with the in-house QASR model.

Delete generated 227 candidates; 13 molecules, highlighted in red in Fig. 4b, were selected from the top 50% of docking intervals, ensuring a wide representation of bioavailability and half-life properties. More detailed information can be found in the Supplementary Information. Ultimately, eight candidate compounds were successfully synthesized (Fig. 4d), and among them, CA-B-1 (Fig. 5a) displayed excellent growth-inhibitory activity against Ba/F3–CLIP1–LTK cells, achieving an inhibitory IC50 value of 6.60 nM (Fig. 5b). Therefore, we focused on CA-B-1 to further confirm its efficacy and mechanism in the following studies.

Fig. 4: Application of Causal-Delete on the LTK target.

a, Visualization of generated samples. The conformations are directly generated by Causal-Delete. b, Chemical space of generated molecules. c, The interaction pattern of CA-B-1 binding to LTK. d, Synthesized inhibitors designed by Causal-Delete.

Source data

Full size image Fig. 5: CA-B-1 exhibits antitumor efficacy in vitro and in vivo.

a, The structure of CA-B-1. b, The effects of CA-B-1 on the cell viability in Ba/F3–CLIP1–LTK cells determined by MTT assays (n = 3), and the inhibition potency of CA-B-1 on LTK through Mobility Shift assays (n = 2). Data are presented as mean values ± s.d. c, Plasma concentration–time profiles and in vivo pharmacokinetics parameters of CA-B-1 after administration intravenously (5 mg kg−1) in Sprague Dawley rats (n = 3). d, Inhibition of CLIP1–LTK kinase activity and the activation of PI3K/AKT and RAS/MAPK pathways downstream signalling by CA-B-1. β-Tublin was used as a loading control. e, The selectivity of CA-B-1 against a panel of 31 kinases. EIC50 represents the half-maximal inhibitory concentration of a compound against protein activity. f, The tumour volume of CA-B-1 treated group or vehicle control (n = 6 mice per group). All graph bars represent mean ± s.d. i.g., gavage administration; i.p. intraperitoneal injection.

Source data

Full size image Delete-designed CA-B-1 is a potent and selective LTK inhibitor

We next carried out an in vitro mobility shift assay to confirm whether CA-B-1’s growth inhibition on Ba/F3–CLIP1–LTK cells is due to its binding to LTK. The results demonstrated that CA-B-1 can inhibit LTK protein kinase activity at an IC50 of 1.36 nM (Fig. 5b), indicating the on-target mechanism of action of CA-B-1 in Ba/F3–CLIP1–LTK cells. The interaction pattern of CA-B-1 binding to LTK was displayed in Fig. 4c.

One of the benefits of structure-based drug design methodologies is that they design binding pocket-matching ligands, similar to designing keys to the lock, which inherently has benefits in ligands’ selectivity. Therefore, as a structure-based model, Delete is expected to generate selective ligands. To validate this claim, we evaluated the selectivity of CA-B-1 against a panel of 31 kinases from various families. As shown in Fig. 5e, CA-B-1 showed no inhibitory activity against most of the tested kinases (IC50 > 1 μM) and displayed notable inhibitory activity against only a few targets that shared a high degree of homology with LTK, namely ALK, ROS and KDR44,45. CA-B-1 exhibited dose-dependent inhibition of LTK phosphorylation in Ba/F3–CLIP1–LTK cells, as well as downstream p-AKT, p-MAPK and p-S6 levels, but it had no effects on total LTK, AKT, MAPK and S6 levels (Fig. 5d). These results support the potency and selectivity of CA-B-1 on the cellular level.

Due to the complexity of absorption and metabolism, molecules that are effective in vitro may not necessarily function in vivo. To further investigate the pharmaceutic effects of CA-B-1, we conducted a series of experiments in vivo. Figure 5c showed that the half-life (t1/2) of CA-B-1 was 8 h after injection intravenously at a single dose of 5 mg kg−1 and the area under the curve of CA-B-1 was 3,612 ng h ml−1, indicating its compound stability in vivo. The favourable pharmacokinetic properties of CA-B-1 echo with our selection criteria in the selection Delete-generated molecule. To further evaluate the antitumor efficacy of CA-B-1 in vivo, we established a subcutaneous xenograft model by injecting stably transformed NIH3T3–CLIP1–LTK cells into the flanks of nude mice. Mice were randomly allocated to groups and treated with CA-B-1 or vehicle control orally and intraperitoneally when tumours reached a volume of approximately 150 mm3. Remarkably, CA-B-1 treatment in each administration group considerably inhibited tumour growth compared to the vehicle group (Fig. 5f). In summary, the excellent selectivity and activity of Delete-designed CA-B-1 is a strong proof of Delete’s utility, highlighting the potential of generative AI combined with structure-based approach.

Conclusion

Delete is an all-in-one solution for lead optimization in drug discovery, made possible by the introduction of the 3D molecular generation framework and the unified deleting strategies. Compared with existing single-task models such as DeLinker and DeepFrag, this framework enables a single model to perform multiple tasks of lead optimization with excellent efficiency. Additionally, the ligand-based model only generates random-like molecules instead of functional ligands. The embedding of geometric neural networks in Delete allows for the simultaneous prediction of near-natural conformations of complete molecules, unifying molecular generation and conformation generation. Comprehensive in silico experiments conducted under four different lead optimization scenarios have confirmed the capabilities of Delete. It can be employed not only to optimize lead compounds for first-in-class drugs but also to assist medicinal chemists in developing me-too/me-better products by structural modification/substitution from existing drugs.

Although Delete has shown its power in the model comparison, retrospective studies and real-world application, demonstrating a successful validation of generative AI combined with structure-based methodology, it still faces some limitations. One notable challenge is its inability to account for conformational flexibility; the ligands generated are based solely on the rigid geometric and energetic properties of the binding pocket. Future improvements could involve employing dynamic graph neural networks to better capture the induced-fit effect between ligands and their target pockets. Additionally, scaling Delete to larger or more diverse datasets presents potential challenges, as increased computational demands may hinder performance and slow down the development process, particularly when applied to complex protein targets and extensive chemical libraries. The model’s efficacy in dealing with new or inadequately characterized protein structures remains to be established. To further enhance the generalizability of Delete across different protein targets, future research should prioritize refining training methodologies and integrating more effective data, as well as sampling techniques.

Methods Dataset construction

Delete was trained on the CrossDock dataset39, which is a recognized benchmark for pocket-aware 3D de novo design models12,46,47. CrossDock is curated from molecular cross-docking results. Although docked conformations do not share exactly the same patterns with X-ray crystallized structures in principle, it is acceptable to enrich the dataset using physical tools because crystal data are far from saturated. We used the MMSeq2 method to cluster proteins and ensure that the protein sequence similarity between the training set and the test set was less than 30%. After data processing, the final CrossDock contains 100,000 training sets and 100 test sets.

The pocket–ligand pairs need to be further processed to become training data for lead optimization. We comprehensively investigate the previous lead optimization methods and select the corresponding representative approach to obtain subtask data of lead optimization. In the linker design task, matched molecular pair analysis48 is performed to cut acyclic single bonds twice; in the fragment elaboration, functional fragments are obtained from one cut of acyclic single bonds; in scaffold hopping, the scaffold is derived as a Bemis–Murcko scaffold49; in side-chain decoration, the side chains are all terminal acyclic groups. We also provide some other decomposition methods in the program, such as using the BRICS50 rule to obtain fragments and using ScaffoldNetwork51 to obtain intermediate scaffolds.

Training

We utilized the PyTorch and PyTorch Geometric frameworks to develop the model. The model is trained with the Adam optimizer, with an initial learning rate of 2 × 10−4. A learning rate scheduler is employed to adjust the learning rate dynamically based on the validation loss. Gradient clipping is applied to enhance stability during training. The model was trained over 300 epochs on an NVIDIA 3090 GPU, with batch sizes of 256 and 4 for pretraining and finetuning, respectively. The training loss will be detailed in the following architecture sections.

Evaluation metrics

The primary indicator of model evaluation is the binding tightness between the generated molecule and the target protein, which is most directly manifested as in situ binding energy (Scoring): that is, calculating the interaction energy between the molecule and the protein through an empirical force field. In addition, docking binding energy (Docking) is used as a related indicator to refine the optimal conformation of the molecule in the protein pocket and calculate the binding energy. In addition to binding energy, Delete training also evaluates four common molecular attribute indicators: QED, synthetic difficulty, Lipinski’s five rules and LogP. These indicators are usually used to evaluate molecular optimization models.

Architecture of the Causal-Delete model

Delete is a tool designed to generate structures for chemists, drawing on the chemical and geometric profiles of protein pockets. To formulize the problem, the pocket is denoted as P, the existing structure is Lμ and the intact ligand is L. Consequently, the objective of Delete is to learn the distribution P(L|Lμ,P). To fully capture the intricate interaction between pockets and ligands, we choose to employ the geometric graph structures. These structures facilitate the extraction of detailed interactions as edge features, constructed through a k-nearest neighbourhood approach based on Euclidean distance. Additionally, the model’s equivariant design ensures the enforcement of necessary physical constraints. Moreover, to extract physical profiles of protein pockets, the proteins are represented by triangle patches52. Given protein pockets and the fragments of interest, the Interaction Module is learning the ‘state’ of the entire system, extending to each node and its adjacent geometric edges. Subsequently, the Generation Module suggests the placement of the next atom, drawing on the interaction features that have been learned. More details can be found in Supplementary Information Methods.

Transfer learning on the ALK active chemical space

The QSAR model characterizes an active chemical space defined by active molecules targeting to a specific structure. In principle, the chemical spaces of two similar structure displays a degree of overlap. Thus, we employed the causal-inspired transfer-learning technique to integrate the knowledge, mainly belonging to the ligand‐based drug design (LBDD) domain, into our structure-based drug design (SBDD) method. Such a composite approach could be effectively accomplished in the deep learning paradigm owing to its powerful representation learning. To achieve this, we need to incorporate the chemical space of active molecules into the model. Therefore, a QSAR model, which is a classical method to quantify the relationship between the molecular structure and affinity to a specific target, was trained as an active chemical space agent. Specifically, this QSAR model was trained on a collection of active ALK molecules, which were collected from BindingDB53, with 1,612 active molecules and 608 inactive ones. Then the QSAR model predictions, instead of biological labels, were utilized in the following finetuning, allowing a robust characterization of active inhibitors, balancing the random and systematic noise in different affinity values from various experiment environment.

The QSAR model, particularly, is constructed using a comprehensive ensemble method54. To elaborate, molecular representation is expressed as ECFP4 fingerprint55. With characterized molecules, a set of machine-learning methods, including Random Forest56, XGBoost57, CatBoost58, LightGBM59, KNN60 and Neural Network61, serve as the base layer, were trained on raw features, ECFP4, as base models. The trained embeddings from these models were then combined with raw features and inputted into the Stack Layer. Models similar to those used in the Base Layer were then trained on these concatenated features. Following this, the models trained in the Stack Layer were weighted according to learnable weights. The final binding affinity predictions were thus derived. This approach maximizes the utilization of raw features from active molecules. Moreover, each model was cross-validated fivefold on n separate random partitions of input data, a process referred to as n-repeated k-bagging62,63.

Except for the binding affinity, ADMET properties are another essential aspect of drug candidates that chemists care about. Another model, Quantitative Structure Property Relationship, is trained aiming to determine the regions of chemical space associated with specific properties. Of note, unlike affinity, we have no access to ADMET profiles of the 2,220 collected molecules, implying seeking a general model to achieve property prediction. ADMET Lab, which is trained on the most current data available, can be used directly. Therefore, the absorption and excretion properties—that is, bioavailability and half-life—were obtained from ADMET Lab for the following finetuning.

The Delete model was finetuned on relabelled affinity and two other ADMET properties to incorporate the active areas and good ADMET profile regions into the search space of Delete. The finetuning loss is

$$\mathcal{L} = -\frac{1}{m}\left(\mathop{\sum }\limits{i=1}^{m} l{i} \times \mathcal{L}{\rm{i}} + \beta{i} \times \mathcal{L}{\rm{i}} + \gamma{i} \times \mathcal{L}_{\rm{i}}\right)$$

where (m) is the number of active molecules and ({l}{i},{\beta }{i},\,{\gamma }_{i}) represents the relabelled affinity, bioavailability and half-life of molecule i.

Reporting summary

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

Data availability

The datasets used in this study are available via Zenodo at https://doi.org/10.5281/zenodo.14586176 (ref. 64). Source data are provided with this paper.

Code availability

The source code of this study is freely available at GitHub (https://github.com/HaotianZhangAI4Science/Delete) to allow replication of the results.

References

Grabley, S. & Thiericke, R. Drug Discovery from Nature (Springer Science & Business Media, 1998).

Wong, F. et al. Discovery of a structural class of antibiotics with explainable deep learning. Nature 626, 177–185 (2024).

Article   MATH   Google Scholar

Liu, G. et al. Deep learning-guided discovery of an antibiotic targeting Acinetobacter baumannii. Nat. Chem. Biol. 19, 1342–1350 (2023).

Article   MATH   Google Scholar

Stokes, J. M. et al. A deep learning approach to antibiotic discovery. Cell 180, 688–702.e613 (2020).

Article   MATH   Google Scholar

Godinez, W. J. et al. Design of potent antimalarials with generative chemistry. Nat. Mach. Intell. 4, 180–186 (2022).

Article   MATH   Google Scholar

Zhavoronkov, A. et al. Deep learning enables rapid identification of potent DDR1 kinase inhibitors. Nat. Biotechnol. 37, 1038–1040 (2019).

Article   MATH   Google Scholar

Li, Y. et al. Generative deep learning enables the discovery of a potent and selective RIPK1 inhibitor. Nat.Com. 13, 6891 (2022).

Article   MATH   Google Scholar

Aronson, J. K. & Green, A. R. Me-too pharmaceutical products: History, definitions, examples, and relevance to drug shortages and essential medicines lists. Br. J. Clin. Pharmacol. 86, 2114–2122 (2020).

Article   MATH   Google Scholar

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   MATH   Google Scholar

Lin, H. et al. DiffBP: generative diffusion of 3D molecules for target protein binding. Chem. Sci. 16, 1417–1431 (2025).

Article   MATH   Google Scholar

Schneuing, A. et al. Structure-based drug design with equivariant diffusion models. Nat. Comput. Sci. 4, 899–909 (2024).

Article   MATH   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).

Davidson, M. et al. Comparison of one-year efficacy and safety of atorvastatin versus lovastatin in primary hypercholesterolemia. Am. J. Cardiol. 79, 1475–1481 (1997).

Article   MATH   Google Scholar

Becnel Boyd, L. et al. Relationships among ciprofloxacin, gatifloxacin, levofloxacin, and norfloxacin MICs for fluoroquinolone-resistant escherichia coli clinical isolates. Antimicrob. Agents Chemother. 53, 229–234 (2009).

Article   Google Scholar

Bethke, E. et al. From type I to type II: design, synthesis, and characterization of potent pyrazin-2-ones as DFG-out inhibitors of PDGFRβ. ChemMedChem 11, 2664–2674 (2016).

Article   MATH   Google Scholar

Green, H., Koes, D. R. & Durrant, J. D. DeepFrag: a deep convolutional neural network for fragment-based lead optimization. Chem. Sci. 12, 8036–8047 (2021).

Article   MATH   Google Scholar

Igashov, I. et al. Equivariant 3D-conditional diffusion model for molecular linker design. Nat. Mach. Intell. 6, 417–427 (2024).

Article   MATH   Google Scholar

Jin, J. et al. FFLOM: a flow-based autoregressive model for fragment-to-lead optimization. J. Med. Chem. 66, 10808–10823 (2023).

Article   Google Scholar

Hu, C. et al. ScaffoldGVAE: scaffold generation and hopping of drug molecules via a variational autoencoder based on multi-view graph neural networks. J. Cheminform. 15, 91 (2023).

Article   MATH   Google Scholar

Loeffler, H. H. et al. Reinvent 4: modern AI–driven generative molecule design. J. Cheminform. 16, 20 (2024).

Article   MATH   Google Scholar

Liu, X., Ye, K., van Vlijmen, H. W. T., Ijzerman, A. P. & van Westen, G. J. P. DrugEx v3: scaffold-constrained drug design with graph transformer-based reinforcement learning. J. Cheminf. 15, 24 (2023).

Article   Google Scholar

Neves, M. A. C., Totrov, M. & Abagyan, R. Docking and scoring with ICM: the benchmarking results and strategies for improvement. J. Comput. Aided Mol. Des. 26, 675–686 (2012).

Article   MATH   Google Scholar

Adeshina, Y. O., Deeds, E. J. & Karanicolas, J. Machine learning classification can reduce false positives in structure-based virtual screening. Proc. Natl Acad. Sci. USA 117, 18477–18488 (2020).

Article   Google Scholar

Deng, C. et al. Vector neurons: a general framework for SO(3)-equivariant networks. In Proc. IEEE/CVF International Conference on Computer Vision 12200–12209 (IEEE, 2021).

Jing, B., Eismann, S., Suriana, P., Townshend, R. J. & Dror, R. Learning from protein structure with geometric vector perceptrons. Preprint at https://arxiv.org/abs/2009.01411 (2020).

Mahmood, O., Mansimov, E., Bonneau, R. & Cho, K. Masked graph modeling for molecule generation. Nat. Commun. 12, 3156 (2021).

Article   MATH   Google Scholar

Hu, Z., Dong, Y., Wang, K., Chang, K.-W. & Sun Y. Gpt-gnn: generative pre-training of graph neural networks. In Proc. 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining 1857–1867 (ACM, 2020).

Izumi, H. et al. The CLIP1–LTK fusion is an oncogenic driver in non‐small‐cell lung cancer. Nature 600, 319–323 (2021).

Article   MATH   Google Scholar

Ferla, M. P. et al. Fragmenstein: predicting protein–ligand structures of compounds derived from known crystallographic fragment hits using a strict conserved-binding–based methodology. J. Cheminform. 17, 4 (2025).

Article   Google Scholar

Imrie, F., Bradley, A. R., van der Schaar, M. & Deane, C. M. Deep generative models for 3D linker design. J. Chem. Inf. Model. 60, 1983–1995 (2020).

Article   Google Scholar

Langevin, M., Minoux, H., Levesque, M. & Bianciotto, M. Scaffold-constrained molecular generation. J. Chem. Inf. Model. 60, 5637–5646 (2020).

Article   Google Scholar

Zhang, O. et al. Learning on topological surface and geometric structure for 3D molecular generation. Nat. Comput. Sci. 3, 849–859 (2023).

Article   MATH   Google Scholar

Clark, D. E. & Pickett, S. D. Computational methods for the prediction of ‘drug-likeness’. Drug Discov. Today 5, 49–58 (2000).

Article   MATH   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   MATH   Google Scholar

Ganesan, A. The impact of natural products upon modern drug discovery. Curr. Opin. Chem. Biol. 12, 306–317 (2008).

Article   MATH   Google Scholar

Sangster, J. Octanol‐water partition coefficients of simple organic compounds. J. Phy. Chem. Ref. Data 18, 1111–1229 (1989).

Article   MATH   Google Scholar

Gaulton, A. et al. ChEMBL: a large-scale bioactivity database for drug discovery. Nucleic Acids Res. 40, D1100–D1107 (2012).

Article   Google Scholar

Irwin, J. J. & Shoichet, B. K. ZINC - a free database of commercially available compounds for virtual screening. J. Chem. Inf. Model. 45, 177–182 (2005).

Article   MATH   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   MATH   Google Scholar

Trott, O. & Olson, A. J. AutoDock Vina: Improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading. J. Comput. Chem. 31, 455–461 (2010).

Article   MATH   Google Scholar

Alexandar, S. P., Yennamalli, R. M. & Ulaganathan, V. Coarse grained modelling highlights the binding differences in the two different allosteric sites of the human kinesin EG5 and its implications in inhibitor design. Comput. Biol. Chem. 99, 107708 (2022).

Article   MATH   Google Scholar

Zheng, S. et al. Deep scaffold hopping with multimodal transformer neural networks. J. Cheminform. 13, 87 (2021).

Article   MATH   Google Scholar

Cooper, A. J., Sequist, L. V. & Lin, J. J. Third-generation EGFR and ALK inhibitors: mechanisms of resistance and management. Nat. Rev. Clin. Oncol. 19, 499–514 (2022).

Article   Google Scholar

Roll, J. D. & Reuther, G. W. ALK-activating homologous mutations in LTK induce cellular transformation. PLoS ONE 7, e31733 (2012).

Article   MATH   Google Scholar

Roskoski, R. ROS1 protein-tyrosine kinase inhibitors in the treatment of ROS1 fusion protein-driven non-small cell lung cancers. Pharmacol. Res. 121, 202–212 (2017).

Article   Google Scholar

Luo, S., Guan, J., Ma, J. & Peng, J. A 3D generative model for structure-based drug design. Preprint at https://arxiv.org/abs/2203.10446 (2022).

Ragoza, M., Masuda, T. & Koes, D. R. Generating 3D molecules conditional on receptor binding sites with deep generative models. Chem. Sci. 13, 2701–2713 (2022).

Article   Google Scholar

Hussain, J. & Rea, C. Computationally efficient algorithm to identify matched molecular pairs (MMPs) in large data sets. J. Chem. Inf. Model. 50, 339–348 (2010).

Article   MATH   Google Scholar

Bemis, G. W. & Murcko, M. A. The properties of known drugs. 1. Molecular frameworks. J. Med. Chem. 39, 2887–2893 (1996).

Article   MATH   Google Scholar

Degen, J., Wegscheid-Gerlach, C., Zaliani, A. & Rarey, M. On the art of compiling and using ‘drug-like’ chemical fragment spaces. ChemMedChem 3, 1503–1507 (2008).

Article   MATH   Google Scholar

Varin, T., Schuffenhauer, A., Ertl, P. & Renner, S. Mining for bioactive scaffolds with scaffold networks: improved compound set enrichment from primary screening data. J. Chem. Inf. Model. 51, 1528–1538 (2011).

Article   Google Scholar

Gainza, P. et al. Deciphering interaction fingerprints from protein molecular surfaces using geometric deep learning. Nat. Methods 17, 184–192 (2020).

Article   Google Scholar

Liu, T., Lin, Y., Wen, X., Jorissen, R. N. & Gilson, M. K. BindingDB: a web-accessible database of experimentally determined protein–ligand binding affinities. Nucleic Acids Res. 35, D198–D201 (2007).

Article   Google Scholar

Svetnik, V. et al. Boosting: an ensemble learning tool for compound classification and QSAR modeling. J. Chem. Inf. Model. 45, 786–799 (2005).

Article   MATH   Google Scholar

Rogers, D. & Hahn, M. Extended-connectivity fingerprints. J. Chem. Inf. Model. 50, 742–754 (2010).

Article   MATH   Google Scholar

Ho, T. K. Random decision forests. In Proc. 3rd International Conference on Document Analysis and Recognition 278–282 (IEEE, 1995).

Chen, T. Xgboost: extreme gradient boosting. R package version 0.4-2 (2015).

Dorogush, A. V., Ershov, V. & Gulin, A. CatBoost: gradient boosting with categorical features support. Preprint at https://arxiv.org/abs/1810.11363 (2018).

Ke, G. et al. Lightgbm: a highly efficient gradient boosting decision tree. In Proc. Advances in Neural Information Processing Systems 3149–3157 (ACM, 2017).

Guo, G., Wang, H., Bell, D., Bi, Y. & Greer, K. KNN model-based approach in classification. In Proc. On The Move to Meaningful Internet Systems 2003: CoopIS, DOA, and ODBASE (eds Meersman, R. et al.) 986–996 (Springer, 2003).

Devillers, J. Neural Networks in QSAR and Drug Design (Academic, 1996).

Mordelet, F. & Vert, J. P. A bagging SVM to learn from positive and unlabeled examples. Pattern Recog. Lett. 37, 201–209 (2014).

Article   MATH   Google Scholar

Hu, Y. et al. Silicon photonic MEMS switches based on split waveguide crossings. Nat.Commun. 16, 331 (2025).

Article   MATH   Google Scholar

Zhang, H. Delete code. Zenodo https://doi.org/10.5281/zenodo.14586175 (2025).

Download references

Acknowledgements

This study was supported by the National Key Research and Development Program of China (grant no. 2021YFE0206400) (to T.H.), the National Natural Science Foundation of China (grant nos. 92370130 (to T.H.), 22220102001 (to T.H.), 82204279 (to P.P.) and 82373791 (to Y.K.)), the Fundamental Research Funds for the Central Universities (grant no. 226-2022-00220) (to P.P.) and the Shenzhen Science and Technology Innovation Commission (grant no. KCXFZ20201221173404013) (to J.C.). This work was financially supported by Baidu Scholarship. We thank C. Guo from the Core Facilities, Zhejiang University School of Medicine for their technical support.

Author information Author notes

These authors contributed equally: Shicheng Chen, Odin Zhang, Chenran Jiang, Huifeng Zhao, Xujun Zhang.

Authors and Affiliations

College of Pharmaceutical Sciences, Zhejiang University, Hangzhou, China

Shicheng Chen, Odin Zhang, Huifeng Zhao, Xujun Zhang, Qun Su, Zhenxing Wu, Xinyue Wang, Wanglin Qu, Yuanyi Ye, Xin Chai, Ning Wang, Tianyue Wang, Wei Xie, Dan Li, Chang-Yu Hsieh, Yu Kang, Tingjun Hou & Peichen Pan

Pingshan Translational Medicine Center, Shenzhen Bay Laboratory, Shenzhen, Guangdong, China

Chenran Jiang, Yun Liu, Yuan An, Guanlin Wu, Qianqian Yang & Jiean Chen

Department of Chemistry, The Hong Kong University of Science and Technology, Hong Kong, China

Mengting Chen & Yong Huang

Westlake University, Hangzhou, China

Haitao Lin

Contributions

S.C. designed and performed the in vitro and in vivo experiments. O.Z. contributed to the Delete idea and code. H.Z. contributed to the retrospective studies. C.J., M.C., Y.L., Y.A., G.W., Q.Y., J.C. and Y.H. contributed to the synthesis of molecules. X.Z. contributed to the comparative model reproduction. Z.W. contributed to the ADMET analysis of generated molecules. Q.S., X.W., W.Q., Y.Y., X.C., N.W., T.W., H.L. and D.L. provided advice on experimental design. W.X., C.-Y.H. and Y.K. contributed to the manuscript revision and experimental design. T.H. and P.P. contributed to the essential financial support and conception and were responsible for the overall quality.

Corresponding authors

Correspondence to Chang-Yu Hsieh, Yong Huang, Yu Kang, Tingjun Hou or Peichen Pan.

Ethics declarations Competing interests

The authors declare no competing interests.

Peer review Peer review information

Nature Machine Intelligence thanks Ramil Nugmanov 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.

Supplementary information Supplementary Information

Supplementary Discussion, Figs. 1–11, Tables 1–3 and Methods.

Reporting Summary Source data Source Data Figs. 4 and 5

Souce data, unprocessed western blots, statistical source data.

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

Chen, S., Zhang, O., Jiang, C. et al. Deep lead optimization enveloped in protein pocket and its application in designing potent and selective ligands targeting LTK protein. Nat Mach Intell 7, 448–458 (2025). https://doi.org/10.1038/s42256-025-00997-w

Download citation

Received 22 April 2024

Accepted 16 January 2025

Published 20 February 2025

Version of record 20 February 2025

Issue date March 2025

DOI https://doi.org/10.1038/s42256-025-00997-w

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 Cheminformatics Machine learning This article is cited by ECloudGen: leveraging electron clouds as a latent variable to scale up structure-based molecular design Odin ZhangJieyu JinTingjun Hou

Nature Computational Science (2025)

Bridging chemical space and biological efficacy: advances and challenges in applying generative models in structural modification of natural products Chuan-Su LiuBing-Chao YanPema-Tenzin Puno

Natural Products and Bioprospecting (2025)