LaMGen: LLM-based 3D molecular generation for multi-target drug design

Abstract

Multi-target drugs hold great promise for treating complex diseases, yet existing methodologies predominantly rely on ligand-based approaches, which lack sufficient biological context and are often confined to specific target pairs, resulting in limited generalizability. Here, we introduce LaMGen, a general-purpose multi-target drug design framework powered by large language models (LLMs). Built on MTD2025, a dataset comprising over 600,000 quantum-accurate molecular conformations and 700,000 multi-target associations, LaMGen directly yields energy-favorable conformations with quantum-level accuracy. The framework integrates ESM-C protein embeddings, rotation-aware ligand tokens, and a TriCoupleAttention module to capture multi-level target–ligand interactions. Across independent benchmarks, LaMGen outperforms diffusion-based model across multiple properties, generating molecules in an average of 0.44 s, while preserving high conformational plausibility. Retrospective analyses demonstrate that LaMGen not only can reproduce molecules identical to known actives, but also consistently produces structurally novel candidates with conserved core scaffolds and superior binding affinities.

Designing effective multi-target therapeutics remains a major challenge, as existing ligand- or protein-centric methods struggle to generate biologically contextualized, spatially valid 3D molecules, particularly for triple-target systems. This study introduces LaMGen, an LLM-powered framework that leverages large-scale protein-ligand data and rotation-aware molecular encoding to rapidly produce chemically plausible multi-target candidates, achieving strong zero-shot generalization, superior molecular quality, and robust performance across dual- and triple-target design tasks.

Introduction

Targeted drug design has long been a central task in drug discovery. However, therapies relying on single-target modulation often fall short when addressing complex, multifactorial diseases such as cancer, neurological disorders, and autoimmune conditions^1,2. A major limitation of these single-target approaches is their susceptibility to drug resistance and adverse side effects, limiting their ability to achieve sustained therapeutic benefits. In contrast, multi-target drug design strategy, also known as polypharmacology, has emerged as a promising paradigm³. By simultaneously modulating multiple disease-relevant targets, this strategy can induce synergistic pharmacological effects, thereby enhancing efficacy, reducing side effects, and improving pharmacokinetic and safety profiles⁴. Despite its considerable potential for treating complex diseases, multi-target drug design remains exceptionally challenging. A core requirement is that candidate compounds must simultaneously meet the structural constraints of multiple binding sites while maintaining favorable pharmacokinetic and drug-like properties⁵. Consequently, the development of efficient, accurate, and generalizable approaches for multi-target drug design has become a critical and urgent research direction in modern drug discovery.

In recent years, deep learning-based approaches, particularly molecular generative models, have made significant advances in single-target drug design^6,7. Current mainstream approaches can be broadly classified into two categories. The first relies on SMILES-based representations, employing architectures such as Recurrent Neural Networks (RNNs), Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), and Transformers to capture molecular syntax and structural rules^8–10. The second category focuses on three-dimensional (3D) conformations, utilizing Graph Neural Networks (GNNs), diffusion models, and pharmacophore combination strategies to more accurately capture spatial interactions governing molecular-target binding^11–14. These methods have demonstrated strong capabilities in generating novel compounds with desired target affinities and physicochemical properties. However, extending them to multi-target drug design introduces substantial challenges, including data scarcity and increased modeling complexity, as they must account for the synergistic interactions among multiple different protein targets, significantly raising the difficulty of the generative process. In 2024, Munson et al. proposed POLYGON¹⁵, a dual-target molecular generation framework that integrates a pretrained VAE with reinforcement learning (RL). While validated through the synthesis and testing of 32 compounds targeting MEK1 and mTOR, this approach remains heavily RL-dependent due to the VAE’s inability to intrinsically capture target-specific information. In the same year, Chen et al. introduced AIxFuse¹⁶, a structure-aware dual-target drug design method that integrates pharmacophore fusion based on predefined fragment library, RL, and active learning. AIxFuse was evaluated on two target pairs (GSK3β-JNK3 and RORγt-DHODH), with performance assessed via the predicted binding modes and estimated binding free energies of generated molecules. Zhou et al. subsequently proposed DualDiff¹⁷, a diffusion-based general framework for dual-target molecule generation, which utilized 3D ligand-sharing for knowledge transfer and was evaluated on a dataset of 12,917 target pairs (438 unique targets). Despite these advances, existing multi-target generative models exhibit three principal limitations. First, most multi-target molecular generation methods are ligand-based and heavily rely on activity data from specific target pairs, which limits their generalization to target pairs with few known reference molecules. Second, some ligand-based methods incorporate reinforcement learning or similar strategies, implicitly introducing target information through scoring functions to reduce reliance on labeled activity data. However, achieving a balance between maximizing target-specific rewards and maintaining chemical diversity remains a critical challenge, as over-optimization can lead to narrow scaffold exploitation. Third, general-purpose frameworks, particularly those based on diffusion architectures, are computationally intensive and lack comprehensively benchmarked against task-specific baselines. Moreover, they struggle to scale to more complex multi-target scenarios, where combinatorial complexity and the difficulty of simultaneously satisfying multiple binding constraints lead to declines in generation quality and controllability. Consequently, although generative models have achieved considerable success in single-target drug discovery, their extension to dual-target or triple-target settings remains hindered by challenges in generation quality, computational efficiency, and generalizability, highlighting an urgent need for more effective and scalable solutions.

Recent advances in large language models (LLMs), particularly Transformer-based architectures¹⁸ such as GPT-4¹⁹, have revolutionized computational drug discovery^20–22 by enabling remarkable capabilities in predicting molecular properties and generating novel drug-like compounds²³. These models are typically pretrained on large-scale molecular datasets to capture the fundamental rules of chemical structures, followed by fine-tuning on specific datasets to adapt to downstream tasks. By utilizing SMILES as both input and output, LLMs leverage sequence-based molecular representations, which a study by Flam-Shepherd et al.²⁴ demonstrated to be effective for modeling complex molecular distributions. To leverage text-based molecular representations to propose novel compounds, recent research has increasingly explored GPT-based models for single-target drug design²⁵. However, these models generally lack an understanding of molecular 3D information, and the application of LLMs to multi-target drug generation remains largely unexplored. Unlike traditional single-target design methods, multi-target approaches demand model complexity and generalization capabilities to address the combinatorial challenges in optimizing binding modes across diverse targets⁵.

To overcome these limitations, we propose LaMGen, a universal 3D molecular generation framework for multi-target drug design built upon LLMs. LaMGen encodes the internal degrees of freedom of ligands, such as torsion angles, into discrete, 3D rotation-aware tokens, pretrained on the GEOM dataset²⁶ to establish precise mappings between SMILES and torsional conformations. This is followed by multi-target adaptation on the MTD2025 dataset, which comprises over 4000 protein targets, 100,000 ligands represented by more than 600,000 quantum-precision 3D conformations, and over 700,000 dual- and triple-target associations. For each ligand, conformer searches were performed using CREST, and the resulting conformations were further optimized with LiTEN-FF²⁷, a deep potential-based force field that achieves quantum-level accuracy. This enables the LLM to serve as a generative decoder over conformational space, directly outputting physically plausible 3D molecular structures that respect spatial constraints. By decoupling from traditional single-modality reliance, LaMGen improves adaptability and accuracy in complex drug design. Crucially, it incorporates protein sequence information (instead of 3D structures) into generation, reducing dependency on structural data while leveraging large-scale sequence databases to learn diverse ligand-target interactions. For nuanced interaction modeling, we propose TriCoupleAttention, a module that jointly captures self-attention and cross-attention across target-target, ligand-target, and ligand-ligand pairs, enabling comprehensive multi-target binding representation. Designed for extensibility, LaMGen integrates task-specific fine-tuning to optimize drug-likeness, pharmacokinetics, and target specificity under multi-objective constraints. While initially validated for dual- and triple-binding scenarios, the framework has the potential to scale to more complex multi-target applications.

In terms of binding affinity, LaMGen outperforms DualDiff, which is currently the only open-source universal dual-target molecular generation model, on 17/20 benchmark target pairs. This advantage is consistent across both low (<0.4) and high (>0.8) sequence identity with training targets. LaMGen generates each molecule in 0.44 s, approximately 30 times faster than DualDiff (12.3 s). Regarding structural fidelity, LaMGen produces 3D conformations that closely align with LiTEN-FF-optimized structures, demonstrating accurate modeling of realistic molecular geometries and conformational preferences. In terms of generalization, LaMGen performs comparably to task-specific baselines such as AIxFuse¹⁶ and REINVENT2.0²⁸ under zero-shot evaluation. With few-shot fine-tuning, it further surpasses these methods by improving drug likeness, multi-target binding, and the proportion of molecules satisfying multiple pharmacological and binding-mode constraints. Under zero-shot conditions, retrospective analyses on real-world tasks showcase the model’s ability to generate plausible active molecules with high structural fidelity and preservation of key pharmacophores. Leveraging its flexible architecture and enriched modeling of multi-target interactions, LaMGen easily extends to triple-target design. Collectively, these results demonstrate the potential of LaMGen for applications in real-world complex multi-target drug design.

Results

Dataset construction and model architecture

We present LaMGen, a general-purpose framework for multi-target bioactive molecule generation based on LLMs, as illustrated in Fig. 1. To address the limited data availability in the field of multi-target drug design, we constructed MTD2025 (Multi-Target Dataset 2025), a high-quality dataset specifically designed for multi-target molecule generation, as shown in Fig. 1A. MTD2025 is built upon the Papyrus bioactivity database (version 5.5)²⁹, integrating experimental activity data from authoritative sources such as ChEMBL³⁰ and ExCAPE-DB³¹. While Papyrus is originally organized in a single-target format, we systematically filtered, paired, and reconstructed the data to create a multi-target resource. The final dataset contains 4011 unique proteins and 123,024 unique small molecules, together with more than 600,000 quantum-precision 3D conformations, resulting in approximately 446,000 dual-target and 283,000 triple-target associations. The distribution of association counts for dual- and triple-target pairs is shown in Supplementary Fig. 1. For each molecule, we firstly performed conformational search using CREST, and then optimized the resulting structures with the quantum-accurate LiTEN-FF to obtain the locally lowest-energy 3D conformation, ensuring high structural quality and physical plausibility. Detailed procedures are provided in the “Methods” section. MTD2025 thus offers a unique combination of curated bioactivity labels and reliable molecular structures, making it a valuable public resource for multi-target drug design.

Fig. 1. Overview of the LaMGen framework.

A Construction of MTD2025 dataset. B Ligand processing via the LiTEN-FF potential model, followed by feature encoding using a model pretrained on the GEOM dataset to capture conformational priors. C Architecture of the multi-target generation model, integrating ligand and ESM-C-based protein encoders with the TriCoupleAttention module for joint representation learning. D Inference workflow for dual- and triple-target generation, utilizing solely input protein sequences to directly yield 3D molecular conformations. Source data are provided as a Source data file.

LaMGen’s architecture consists of three main phases: ligand pretraining, multi-target fine-tuning, and sequence-based molecule generation. In the pretraining phase (Fig. 1B), LaMGen is pretrained on over 8 million high-quality molecular conformations from the GEOM dataset²⁶. It learns the mapping between SMILES and corresponding torsional angles in 3D conformational space. Unlike raw Cartesian coordinates, we represent these torsions using a rotation-aware discrete encoding scheme, which reduces sequence length, ensures rotational invariance, and significantly lowers computational complexity^32,33. To further improve generalization, a random masking strategy is employed to improve the model’s robustness and its ability to accurately infill missing torsion information.

In the fine-tuning phase (Fig. 1C), LaMGen is trained on MTD2025. To circumvent the difficulty of obtaining protein 3D structures, amino acid sequences are used as input. These sequences are encoded using ESM-C³⁴, a powerful protein language model pretrained on millions of sequences, to embed protein sequences into latent representations that effectively capture both functional and structural characteristics. These embeddings enable LaMGen to reason about target-specific spatial constraints even in the absence of explicit 3D protein structures. To further strengthen interaction modeling between molecules and dual-target proteins, we design the TriCoupleAttention module, which jointly models self-attention and cross-attention among protein-protein, protein-ligand, and ligand-ligand interactions. This mechanism simulates fine-grained molecular interactions by learning latent attention patterns, enabling the model to handle intricate multi-target binding scenarios.

LaMGen also supports supervised fine-tuning using activity data, allowing the model to optimize for drug-like properties and multi-target binding simultaneously. In the generation phase (Fig. 1D), LaMGen requires only the amino acid sequences of the protein targets of interest as input to rapidly generate 3D molecules with multi-target activity.

To further characterize MTD2025 and validate the effectiveness of our protein representations, we conducted statistical analyses and visualization (Supplementary Fig. 2). The results demonstrate that sequence length constraints reduce long-tail effects, thereby improving data balance and training stability. The dataset exhibits a rich multi-target distribution, dominated by dual-target molecules (61.5%) yet extending to broad-spectrum agents active against over 10 targets (~3%), thus providing a foundation for modeling complex ligand-target interactions. Embeddings extracted from ESM-C exhibit distinct clustering patterns at both protein function and enzyme classification levels under UMAP³⁵ projection (Supplementary Fig. 2C), indicating that these representations capture biologically relevant features and offer informative structural context for multi-target molecule generation. Furthermore, to characterize the distribution of molecular conformations, we performed UMAP visualization on over 600,000 conformers (Supplementary Fig. 3), revealing the diversity and structural patterns of the sampled 3D geometries, particularly in their torsional degrees of freedom.

Zero-shot generation of high-affinity 3D molecules across multi-targets with diverse sequence identity

We first constructed two independent test sets with distinct levels of target similarity to systematically assess the model’s generalization in multi-target molecular design, specifically evaluating its robustness in out-of-distribution (OOD) scenarios. Test Set 1, curated from external real-world datasets, including DrugCombDB³⁶ and PDBBind³⁷, comprises 20 dual-target combinations known to exhibit synergistic effects. Notably, all targets in this set share low sequence identity (<0.4) with the training data, thereby presenting a challenging out-of-distribution generalization scenario involving entirely unseen protein pairs. In contrast, Test Set 2 is a held-out subset of the MTD2025 dataset, comprising 10 dual-target combinations in a hybrid configuration: each pair integrates one target with low sequence identity (<0.4) and another with high sequence identity (>0.8) relative to the training proteins. This setup tests the model’s ability to generalize to partially novel target pairs. Together, these two test sets allow evaluation of the model’s performance across varying target novelty and task complexity.

To ensure rigorous comparison, we conducted benchmark experiments exclusively on the external Test Set 1, selecting DualDiff as the baseline. Adhering to the original DualDiff methodology, generating 100 candidate molecules per target pair. Given the model’s intrinsic invalid molecule filtering, this process yielded approximately 80 valid molecules per pair. Evaluation of key molecular properties of the generated compounds revealed that LaMGen achieved superior performance over DualDiff in both Drug-likeness (QED) and Synthetic Accessibility Score (SAScore). Furthermore, all generated molecules were docked against their corresponding targets, and the mean docking score of the top 50 molecules per target pair was used as the evaluation metric. For a more detailed view, the score distributions of these top 50 molecules for each of the 20 target pairs are provided in the Supplementary Information (Supplementary Fig. 4), along with a summary table reporting both the mean and median docking scores for every target pair (Supplementary Table 1). In addition, an overview of physicochemical property distributions for LaMGen- and DualDiff-generated molecules is provided (Supplementary Fig. 5). Molecules generated by LaMGen exhibit superior distributions across key physicochemical properties compared to those generated by DualDiff. In particular, for LogP and SAScore, LaMGen produces more concentrated distributions that are overall located within druggable ranges (SAScore <4; −0.4 <LogP <5.6), indicating control over the balance between hydrophobicity and synthetic feasibility. Meanwhile, the distributions of QED and TPSA remain within the typical drug-like chemical space, demonstrating that under the complex constraints of dual-target generation, LaMGen is capable of coordinated generation of multiple physicochemical properties. In addition, the docking score distributions on the two targets are obviously shifted toward higher-affinity regions, further validating the overall advantage of LaMGen in jointly maintaining favorable property distributions and binding propensity. As illustrated in Fig. 2A, LaMGen achieved superior binding affinity in over 85% of the target combinations, demonstrating its notable advantages in structural quality and target adaptability. In terms of generation efficiency, LaMGen required only 0.44 seconds per molecule on average, approximately 30 times faster than DualDiff (12.30 s per molecule), underscoring its suitability for large-scale applications. We further performed a systematic ablation study by removing torsion information during training to evaluate the contribution of the 3D torsion token. As summarized in Supplementary Table 2, the variant without the torsion token (LaMGenwo.3D) exhibits markedly lower average binding affinities across nearly all target pairs compared to the full model. This performance degradation demonstrates that explicitly encoding 3D torsional information is crucial for capturing conformational geometry and spatial orientation, thereby enhancing the quality of the generated molecules.

Fig. 2. Performance benchmarking and property analysis of generated molecules on Test Sets 1 and 2.

A Performance comparison between LaMGen and DualDiff on Test Set 1 in terms of Vina docking scores, key drug-like properties, and generation speed. Each bar represents the average score for a target, a longer bar indicates a lower (better) docking score. Labels on each bar indicate the UniProt ID of the corresponding protein target. B Distributions of molecular properties and docking scores for 10,000 LaMGen-generated molecules on Test Set 2, including QED, LogP, SAScore, TPSA, molecular weight, and binding affinities against two targets. In the box plots, the center line indicates the median, the box bounds represent the 25th and 75th percentiles, and the whiskers extend to the minimum and maximum values within 1.5 times the interquartile range (IQR) from the box boundaries. C Proportion of generated molecules with affinities below threshold x (x-axis). Left: lower than x for both targets; Middle: lower than x for Target 1 but higher than x for Target 2; Right: lower than x for Target 2 but higher than x for Target 1. Lower x values correspond to higher binding affinities. Source data are provided as a Source data file.

In Test Set 2, the model generated 1000 candidate molecules per target pair, yielding 10,000 structurally plausible compounds. Importantly, the generation process relied solely on primary protein sequences without any experimental bioactivity labels or 3D protein structures, thereby facilitating task adaptability and independence from structural priors. The generated molecules achieved a 95% validity rate and a high internal structural diversity of 0.89, indicating diverse outputs without mode collapse. The average structural Tanimoto similarity to the training set was only 0.11, reflecting the model’s capacity for structural innovation while preserving drug-like characteristics. To further assess chemical rationality and drug-likeness, we analyzed key molecular properties of all candidates. Figure 2B shows distributions of key drug-like properties and binding affinities, with their mean values reported in Supplementary Table 3. Drug-likeness (QED) values clustered at 0.4–0.7, largely aligning with the favorable threshold (≥0.5) and indicating development potential. Octanol-water partition coefficients (LogP) ranged from 3 to 5, falling within the optimal range (−0.4 to 5.6) to ensure membrane permeability and oral bioavailability. Over 84% of molecules had a Synthetic Accessibility Score (SAScore)<4, meeting the low-complexity standard for scalable synthesis. Topological Polar Surface Area (TPSA) and molecular weight distributions adhered to classical guidelines, fitting the 60–140 Å range for oral absorption and the “Rule of Five” (300–500 Da), supporting metabolic stability and pharmacokinetics.

Docking evaluations revealed an average binding affinity of −9.0 and −8.7 kcal/mol across both targets, indicating robust binding potential. This demonstrates the model’s ability to learn implicit structural semantics from sequences, generalize to known-unknown target pairs, and reliably generate potentially active dual-target molecules. Notably, LaMGen’s performance on Test Set 1 (composed entirely of low sequence identity target pairs) remained comparable to that on Test Set 2, further substantiating its robust generalization ability and robust adaptability to out-of-distribution multi-target scenarios. To further assess the binding characteristics of LaMGen-generated molecules, Fig. 2C presents the distribution of binding affinity profiles under dual- and single-target scenarios. In the dual-target analysis (left panel), 12.29% of molecules exhibit affinities below −9 for both targets, increasing to 21.66% at −8, thereby identifying a meaningful subset capable of simultaneously achieving high-affinity binding to both targets. In single-target assessments, 52.93% of molecules had affinities below −9 toward Target 1, and 40.71% toward Target 2. These findings indicate that LaMGen can generate molecules against both targets simultaneously, as well as subsets exhibiting strong binding to each individually, supporting its application in multi-target drug design. To explore LaMGen’s potential for generating molecules targeting multiple binding pockets, we performed a case study with two dual-pocket configurations. In each configuration, the primary pocket was derived from CDK2³⁸ (ANS and ATP site) and paired with a pocket from a different protein (UniProt IDs P05981 and P29274, respectively), forming two distinct dual-target setups (Supplementary Fig. 6). Using only pocket-level amino acid sequences, LaMGen generated molecules adapted to individual binding sites, with key molecular properties remaining within reasonable ranges. Notably, without additional retraining, the model can accommodate pocket-sequence inputs, although performance was reduced compared with using full protein sequences, indicating a degree of generalization to sub-regions of target proteins.

In summary, LaMGen provides an LLM-based framework for multi-target 3D molecular generation. Across dual-target pairs with sequence identities ranging from <40% to >80% relative to the training set, it generates chemically diverse molecules with high docking scores directly from protein sequences, without requiring prior bioactivity labels or structural data. It also maintains consistent generation efficiency and produces molecules with favorable drug-likeness and synthetic accessibility, supporting the discovery of novel candidates in multi-target drug design. Collectively, these results demonstrate that LaMGen serves as a promising computational asset for navigating complex chemical space in multi-target design tasks.

LaMGen can directly generate conformations with quantum-level accuracy

Accurate molecular conformation generation is critical for downstream applications such as molecular docking and quantum-level property prediction. To systematically evaluate the effectiveness of LaMGen in conformation generation, we selected five dual-target protein pairs from the DualDiff dataset¹⁷, ensuring they were strictly excluded from the training set. For each pair, the model generated 100 candidate molecules. These molecules were subsequently redocked using conformations derived from three distinct protocols: (1) LaMGen-predicted conformations, (2) RDKit-generated conformations optimized with the MMFF94 force field, and (3) High-accuracy conformations refined by LiTEN-FF. By comparing the vina docking scores across these strategies, we quantitatively assessed the impact of conformation quality on molecule-target binding performance.

Figure 3A, B presents vina docking score distributions (including mean scores and distribution profiles) across five dual-target protein pairs under different conformational regimes. As shown in Fig. 3A, the vina docking score distribution of LaMGen-generated conformations closely aligns with that of LiTEN-FF-optimized structures and generally outperforms the MMFF94-optimized conformations, albeit with modest differences. These findings suggest that for docking tasks where moderate geometric accuracy is typically sufficient, LaMGen can directly generate conformations that meet practical requirements without additional refinement. Complementing Fig. 3A, B displays the corresponding docking analysis for Target 2 in each dual-target pair. These consistent results across two target subsets reinforce the superiority of LaMGen- generated conformations in docking performance.

Fig. 3. Docking score distributions and geometric fidelity of generated conformations across dual-target protein pairs.

A Vina docking score distributions for Target 1 across five dual-target protein pairs under different conformational regimes: MMFF94-optimized, LiTEN-FF-optimized, and LaMGen-generated conformations. B Vina docking score distributions for Target 2 across the same five dual-target protein pairs under the three conformational regimes. C Conformational validity comparison of molecules generated by LaMGen and DualDiff. RMSD is calculated with respect to conformations optimized by MMFF94, LiTEN-FF, and MACE-OFF, respectively. D RMSD distributions of generated conformations across different heavy-atom count ranges. Source data are provided as a Source data file.

We consider DualDiff as a comparative baseline to further evaluate the geometric fidelity. Built on a diffusion-based architecture that directly generates 3D conformations within aligned binding pockets of dual targets, DualDiff is therefore expected to yield physically reasonable conformations that closely resemble docked structures. We employed DualDiff to generate a dataset of 500 molecules (100 per protein pair). Subsequently, we calculated ligand RMSD for both LaMGen- and DualDiff-generated conformations and those optimized by three different methods. To ensure a comprehensive assessment, we additionally employed MACE-OFF³⁹ as a physics-informed reference. As shown in Fig. 3C, more than 98% of LaMGen-generated conformations exhibit RMSD values below 2 Å relative to the LiTEN-FF-optimized structures, with a mean RMSD of merely 0.5 Å, indicating a high degree of structural consistency before and after optimization. In contrast, DualDiff-generated conformations display slightly higher RMSD values relative to the optimized structures. Overall, LaMGen consistently achieves lower RMSD values than DualDiff across all three reference methods, highlighting its superior capability in preserving geometric accuracy.

In addition, we further analyzed the similarity between generated and docked conformations. As shown in the upper panels of Fig. 3D, we calculated the RMSD of LaMGen- and DualDiff-generated conformations relative to the docked structures across different heavy-atom count ranges. Before performing any optimizations, all generated conformations were translated and aligned to the centroid of their corresponding docked structures. The results indicate that LaMGen-generated molecules undergo substantially smaller conformational changes before and after docking than those generated by DualDiff. Most RMSD values are close to 2 Å, a commonly used metric⁴⁰ for docking success, suggesting that LaMGen-generated conformations can effectively reduce the docking search space and improve docking accuracy. Meanwhile, the lower panel of Fig. 3D shows that, compared with DualDiff, LaMGen-generated molecules cover a broader range of heavy-atom counts, further demonstrating its advantage in chemical space exploration. Beyond assessing global spatial alignment with docking poses, we further investigated the intrinsic geometric stability and physical realism of the generated structures. To strictly quantify the structural fidelity of the generated molecules, we analyzed the stability of torsion angles upon optimization. As shown in Supplementary Fig. 7A, LaMGen-generated conformers exhibit minimal torsional deviations compared to DualDiff, indicating that the initial geometries are intrinsically close to their energetic minima. Furthermore, a holistic assessment using PoseBusters (Supplementary Fig. 7B) confirms superior physical plausibility, with LaMGen achieving an overall pass rate of 92%—significantly outperforming the 68% rate of DualDiff. Together, these metrics demonstrate that LaMGen directly generates physically valid 3D structures without requiring extensive post-correction.

These results not only confirm the model’s capacity for directly generating high-quality conformations but also emphasize the pivotal role of the training dataset. The dataset’s high geometric precision, diverse distribution, and structural standardization collectively enabled the model to learn accurate and generalizable conformation generation strategies. This “generate-and-use” capability streamlines downstream computational workflows; additional optimization using classical or AI-based force fields becomes redundant, thereby substantially reducing computational overhead. Overall, these findings highlight the model’s practical potential for multi-target drug design.

LaMGen demonstrates zero-shot generalization and fine-tuning potential in JNK3 and GSK3β dual-target design

Dual-target molecular design has emerged as a pivotal strategy in modern drug discovery. Although numerous computational approaches have been developed, such as structure-based virtual screening and pharmacophore fusion. However, these methods often rely heavily on existing structural or activity data, making them less effective in data-scarce or zero-shot scenarios, where models must generalize to entirely unseen target combinations. Consequently, the ability to generalize beyond the training distribution has emerged as a key requirement for the real-world applicability of dual-target generation models. To further assess the practical applicability and generalization capability of LaMGen in real-world dual-target molecular design scenarios, we performed a zero-shot and few-shot transfer learning case study focused on a representative dual-target challenge: the simultaneous inhibition of c-Jun N-terminal Kinase 3 (JNK3) and Glycogen Synthase Kinase 3 Beta (GSK3β). This case was chosen because it has been widely used as a benchmark in prior dual-target drug design studies^16,28,41,42, providing a basis for comparison with established baseline methods. Both kinases play critical regulatory roles in neurodegenerative diseases such as Alzheimer’s Disease (AD), mediating tau protein hyperphosphorylation and neuronal apoptosis pathways, respectively^43,44. Therefore, developing small molecules that can concurrently inhibit JNK3 and GSK3β is considered a key strategy for multi-target intervention.

In this experiment, LaMGen was conditioned solely on the primary amino acid sequences of JNK3 and GSK3β, generating 10,000 candidate molecules with well-defined structural topologies. Notably, no known small-molecule ligands, 3D protein structures, or experimental activity data were used during generation. This highlights LaMGen’s capacity for conditional molecular design in low-data conditions. We collected known active inhibitors of JNK3 and GSK3β from the literature. To guide dual-target molecular design, these single-target inhibitors were subjected to similarity screening, retaining only molecules with high similarity across both targets as a reference set. The generated molecules showed a 94.2% validity rate, confirming chemical soundness and synthetic feasibility. The internal diversity score was 0.863, while average similarity to training set remained low at 0.12, demonstrating effective exploration of novel chemical space beyond training data. To ensure the stability of docking-based evaluations, we performed three parallel docking runs for both targets. The results showed consistently high correlations and minimal deviations among replicates (JNK3: Pearson = 0.96, MAE = 0.13; GSK3β: Pearson = 0.98, MAE = 0.10), confirming the high reproducibility of the docking scoring protocol. Crucially, we assessed the binding affinity of generated molecules by docking them to JNK3 and GSK3β using Vina⁴⁵. LaMGen-generated molecules cluster around −9 kcal/mol for both targets, comparable to the average docking score of known inhibitors (Supplementary Fig. 8). Remarkably, 45.54% (JNK3) and 54.50% (GSK3β) of generated molecules achieved docking scores exceeding the mean scores of reference inhibitors, demonstrating high target affinity even without task-specific adaptation.

We further benchmarked LaMGen against several dual-target molecular generation frameworks, including AIxFuse¹⁶, MARS⁴¹, REINVENT2.0²⁸, and RationaleRL⁴². All models were trained on the same dual-target benchmark dataset introduced by Yang et al.¹⁶. Among these, AIxFuse is widely regarded as the current state-of-the-art (SOTA), employing an active learning pipeline coupled with reinforcement learning to iteratively optimize molecular docking affinities. Because MARS yielded only 2414 unique molecules out of 10,000 samples, we adopted a standardized evaluation based on the top 2000 unique molecules ranked by docking affinity to ensure fairness across all methods. As shown in Fig. 4, LaMGen achieved comparable or superior performance on most key metrics, despite operating in a zero-shot setting without prior exposure to any active compounds targeting these protein pairs. This highlights LaMGen’s inherent generalization capability and its ability to generate high-quality, target-relevant molecules using only primary protein sequences.

Fig. 4. Comparison of JNK3 and GSK3β inhibitors (green shading), four baseline models (AIxFuse, MARS, REINVENT, and RationaleRL), and LaMGen/LaMGen-FT (red shading) in terms of docking scores and key molecular properties.

For each model, the top 2000 molecules were selected based on docking scores. A Vina docking score comparison for JNK3 (left) and GSK3β (right) inhibitors across different models. For each plot, the left semicircle shows the mean vina score (absolute value), and the right violin plot illustrates the score distribution. Error bars represent the standard deviation (SD). B Distributions of five key molecular properties: QED, LogP, SAScore, TPSA, and Molecular Weight. In the box plots, the center line indicates the median, the box bounds represent the 25th and 75th percentiles, and the whiskers extend to the minimum and maximum values within 1.5 times the interquartile range (IQR) from the box boundaries. Source data are provided as a Source Data file.

To further evaluate LaMGen’s capability in transfer learning, we conducted systematic fine-tuning on the dual-target benchmark dataset while keeping hyperparameters and evaluation protocols consistent to ensure direct comparability. The resulting model, LaMGen-FT, preserved molecular diversity (0.864), exhibited a modest gain in validity (96.3%), and achieved consistently enhanced docking scores and property profiles for both targets (Fig. 4). The proportion of generated molecules surpassing the mean docking scores of reference inhibitors further increased to 49.00% for JNK3 and 58.74% for GSK3β (Supplementary Fig. 8), underscoring the benefit of lightweight transfer learning. As summarized in Table 1, under zero-shot conditions, 21.02% of LaMGen-generated molecules exhibited predicted binding affinities above the reference thresholds for both JNK3 and GSK3β, with docking pose centers located within 4 Å of the corresponding reference active molecules. LaMGen and LaMGen-FT also achieved docking centroid distances that were comparable to, or better than, those of baseline models (Supplementary Fig. 9). This metric is important because viable lead candidates must adopt binding poses that fall within a reasonable spatial range of the reference ligand’s centroid in the pocket; molecules that score below the reference or bind far from the canonical site are rarely considered for synthesis. Although the zero-shot proportion was slightly lower than that of four baseline models, likely reflecting the absence of task-specific fine-tuning, it increased substantially after fine-tuning, reaching 33.01%, surpassing the 31.80% achieved by the AlxFuse model. Overall, these results indicate that LaMGen has a clear advantage in dual-target molecular design, effectively identifying candidate compounds with high predicted affinities across multiple targets.

Table 1.

Comparison of LaMGen (including its fine-tuned variant, LaMGen-FT with baseline models on the JNK3 and GSK3β targets

Metrics	AIxFuse	MARS	REINVENT2.0	RationaleRL	LaMGen	LaMGen-FT
Validity (%)	100	100	98.9	99.9	94.2	96.3
Diversity	0.719	0.597	0.722	0.656	0.863	0.864
USRa QEDa (%)	9.11	3.87	2.50	16.40	8.40	12.87
USR SA (%)	15.63	2.20	17.91	2.41	11.90	21.30
USR dockingb	31.80	3.92	28.73	17.61	21.02	33.01
Success ratec (%)	5.40	2.15	1.73	2.32	4.84	8.46

^aUSR (Unique Success Rate): Proportion of non-redundant molecules that exceed the average single-property values of active compounds.

^bUSR docking: satisfying both the Vina score constraints for JNK3 and GSK3β, as well as the constraint on the distance (d < 4 Å) between the docking pose center and the reference pose center.

^cSuccess rate: proportion of generated molecules that simultaneously surpass the average multi-property values of active compounds.

Key evaluation metrics include molecular docking success rate (USR Docking) and overall optimization success rate (Success Rate). Bold values indicate the highest success rate among the compared methods.

For a comprehensive assessment, we used the mean property values of known active inhibitors as quantitative thresholds and evaluated the generated molecules’ success rates in surpassing these benchmarks across multiple attributes. As summarized in Table 1, LaMGen-FT outperforms most baseline methods in critical areas such as synthetic accessibility and dual-target docking success, culminating in an overall multi-property success rate of 8.46%. In comparison, AIxFuse achieved a 5.40% success rate by fusing pharmacophores from known inhibitors and incorporating docking scores, which improved success rate at the expense of molecular diversity (0.719). In contrast, LaMGen maintains high success rate while preserving diversity, demonstrating a more balanced multi-objective performance. Supplementary Table 4 further highlights that LaMGen is substantially faster than baseline methods, with a generation speed of 0.44 s/molecule and combined generation and docking speed of 4.84 s/molecule, compared with 1.64/6.98 s for AIxFuse and 12.30/16.50 s for DualDiff. These results indicate that LaMGen is effective at generating molecules with high predicted binding potential against both JNK3 and GSK3β while remaining synthetically feasible.

Notably, LaMGen exhibits competitive performance in zero-shot scenarios relative to existing baseline models, generating molecules with favorable docking scores and physicochemical properties. Subsequent fine-tuning further elevates these attributes, yielding superior binding affinities and properties, indicating that LaMGen has robust generalization and transfer capabilities in multi-target generation tasks. Its ability to generate structurally novel, chemically valid, and biologically relevant molecules without extensive prior ligand data makes it an ideal tool for exploring orphan targets and scaffold hopping. Coupled with transfer learning, LaMGen bridges de novo design and target-specific optimization, offering a practical computational framework to facilitate AI-driven drug discovery and navigate the complexities of multi-target therapeutic development.

Retrospective validation in realistic dual-target and triple-target tasks

We established a hierarchical benchmark including three dual-target pairs relevant to cancer therapy: EGFR/HER2, PI3K/mTOR, and LSD1/HDAC6, with the aim of evaluating the model’s practical applicability and ability to generalize to real-world scenarios. These pairs were selected to represent diverse target relationships: from high sequence identity and structural features (EGFR/HER2), to shared fold but distinct pocket geometry (PI3K/mTOR), to no structural similarity but functional complementarity (LSD1/HDAC6). Building on this foundation, we further escalated the complexity by integrating VEGFR into the EGFR/HER2 pair, creating the EGFR/HER2/VEGFR triple-target challenge. This selection holds substantial biological relevance, as VEGFR synergizes with EGFR and HER2 to modulate tumor angiogenesis and cell proliferation pathways, and their simultaneous inhibition has demonstrated promising clinical potential in cancer therapy. This triple-target scenario not only enhances the biological significance of the benchmark but also mirrors the complexity of real-world polypharmacology, imposing stringent demands on the model’s generalization capacity and ability to satisfy multiple pharmacological objectives simultaneously.

For each dual-target and triple-target task, we generated 10,000 candidate molecules using LaMGen. To evaluate the model’s capacity for capturing essential pharmacophoric scaffolds and achieving multi-target compatibility, we collected all accessible experimentally verified dual and triple inhibitors from recent literature, which were excluded from the training data. In total, 37 experimentally validated molecules were obtained, including 11 for EGFR/HER2, 17 for PI3K/mTOR, 8 for LSD1/HDAC6, and 1 for EGFR/HER2/VEGFR (Supplementary Table 5). Scaffold matching and structural analyses were performed to assess overlap and novelty. To simulate realistic drug discovery scenarios, we further selected the top 500 generated molecules—ranked by structural similarity to these known actives—for detailed molecular docking and downstream evaluations. This multi-layered validation framework probes the model’s ability to explore pertinent chemical space under multi-target guidance, generate task-specific molecules, and recapitulate key pharmacophoric elements, offering support for LaMGen’s generalization capability and real-world utility in multi-target drug discovery.

LaMGen facilitates structure–activity relationship reconstruction and guided chemical space exploration in EGFR/HER2 dual-target drug design

EGFR and HER2 share a highly homologous kinase domain, including conserved hinge residues and a key hydrophobic pocket⁴⁶, providing a common structural framework for dual inhibition. As depicted in Fig. 5A, the docking score distribution of LaMGen-generated molecules encompasses that of the experimentally validated inhibitors, with a notable subset achieving improved affinities toward both targets simultaneously. Representative generated molecules (Gen. Compound 173 and Gen. Compound 9) precisely recapitulate the structures of reference inhibitors (Exp. ref. Compound 3⁴⁷ and 4⁴⁸), achieving a Tanimoto similarity of 1.0 (indicating structural identity). Both sets adopt the canonical quinazoline or quinoline-nitrile scaffold—a privileged core widely used in EGFR/HER2 inhibitors. The ring nitrogen atoms form critical hydrogen bonds with hinge-region residues, anchoring the ligands within the active site. In addition, the generated compounds incorporate characteristic halogenated aromatic moieties (e.g., 3-chloro-4-fluoroaniline) that occupy the deep hydrophobic ATP-binding pocket, forming π-π stacking and halogen-π interactions that closely mirror the binding modes of the reference inhibitors.

Fig. 5. Representative results for three dual-target systems.

For each system, the left panel displays scatter plots of docking scores on both targets for the top 500 model-generated molecules (dots), ranked by structural similarity to the experimentally validated active compound (stars). The right panel shows the known dual-target active compound and its closest LaMGen-generated molecules, with ‘D’ indicating the Vina docking score. A Results for the EGFR/HER2 dual-target system. B Results for the PI3K/mTOR dual-target system. C Results for the LSD1/HDAC6 dual-target system. Source data are provided as a Source data file.

LaMGen also demonstrates a robust balance between exploitation and exploration. Analysis of structural similarity versus generation frequency reveals distinct algorithmic convergence toward experimentally validated scaffolds, as shown in Supplementary Fig. 10. Notably, the model successfully rediscovered Reference Scaffold 2 (generated 23 times) and frequently reproduced Reference Scaffold 1. This recurrence is indicative of high-probability modes within the learned chemical distribution, reflecting the model’s identification of these privileged scaffolds as near-optimal solutions for the dual-target objective. Interestingly, for Reference Scaffold 3, the most frequently generated molecule was not the reference compound itself but a structurally related analogue (Similarity = 0.55) generated 26 times. This observation suggests that LaMGen does not merely memorizing training patterns; instead, it actively recombines pharmacophores to propose potentially superior analogues—exhibiting capabilities for scaffold hopping and automated side-chain optimization. Beyond these high-frequency modes, each reference scaffold is accompanied by a long-tailed distribution of lower-frequency, medium-similarity variants (Similarity 0.20-0.70), ensuring substantial structural diversity. This distribution pattern mitigates mode collapse, preserves exploration of the broader chemical space, and provides a rich reservoir of novel scaffold variants with potential advantages in potency and selectivity.

De novo generation of dual PI3K/mTOR inhibitors: from precise reconstruction to structural innovation

PI3K and mTOR are critical regulators of cell growth, metabolism, and survival, and their dual inhibition has demonstrated significant therapeutic potential in overcoming resistance to single-target treatments. However, although these two proteins share a common fold, the distinct geometries of their binding pockets present substantial challenges for designing selective dual inhibitors⁴⁹. Leveraging the LaMGen for de novo generation (Fig. 5B), we observed that the model autonomously rediscovered molecules structurally identical to experimental reference compounds 13 and 15^50,51 (Similarity = 1.0), specifically Gen. 2463 and Gen. 3811. These candidates not only achieved structural identity with known drugs but also exhibited binding conformations that superimposed highly with the reference molecules in docking studies. This phenomenon underscores that, even in a template-free regime, the model successfully captures key pharmacophoric features of dual-target active molecules, as shown in Supplementary Fig. 11.

To further investigate whether the model’s unconstrained exploration yielded superior candidates, we screened six candidates from the generated library exhibiting varying degrees of structural similarity (Similarity 0.37-0.71) to reference compounds, and performed MM/GBSA for binding free energy estimation⁵². As illustrated in Supplementary Fig. 12, these LaMGen-generated molecules exhibited binding potentials that are comparable to or exceed those of the reference compounds. Especially for the PI3K target, all LaMGen-generated molecules displayed superior binding affinities, consistently outperforming known compounds. For the more challenging mTOR target constrained by distinct pocket geometry, half of the candidates still exhibited improved affinity. This suggests that LaMGen transcends mere imitation of existing molecules, as it autonomously identifies structural patterns capable of forming stronger interactions by exploring a broader chemical space. As shown in Supplementary Fig. 13, structural superposition of docked conformations against high-confidence reference active poses (derived from experimental crystal structures or AlphaFold 3 predictions) reveals remarkable alignment. The rigid heterocyclic scaffolds, critical for target binding, exhibit a high degree of spatial overlap, while deviations are predominantly localized to solvent-exposed flexible regions.

LaMGen generates pharmacophore-conserved analogues with distinct binding priorities by targeting hydrophobic and metal sites

LSD1 and HDAC6 are two well-studied targets in cancer and neurological disorders, involved respectively in epigenetic regulation and deacetylation processes, and often considered potential targets for combination therapy⁵³. Given the substantial structural differences between LSD1 and HDAC6, the eight experimentally validated compounds generally achieve higher scores on LSD1. The molecules generated by LaMGen exhibit a similar trend, with most showing comparatively higher binding affinities toward LSD1, as shown in Supplementary Fig. 14 and the left of Fig. 5C. For example, Gen. Compound 2635 retains the p-methoxyphenyl head group of reference compound 35⁵⁴ but undergoes extensive structural reconstruction in the remaining regions. As shown in the right of Fig. 5C, the original rigid linker is replaced with a more flexible long-chain structure, and the hydrophilic hydroxamic acid moiety at the tail is substituted with a hydrophobic biphenyl fragment. Although this modification removes the key zinc-binding group (ZBG) and consequently reduces its chelation ability toward HDAC6, the bulkier biphenyl tail effectively fills the hydrophobic pocket of LSD1, and the enhanced hydrophobic packing markedly improves its binding affinity to LSD1. However, some generated molecules are capable of simultaneously enhancing affinity toward both targets. Taking Gen. Compound 4929 as an example, it fully retains the hydroxamic acid ZBG and its linked aliphatic chain from the reference compound 36⁵⁵, which are responsible for HDAC inhibition, allowing the molecule to penetrate the tubular active site of HDAC6 and form stable chelation with the zinc ion at the bottom, thereby maintaining and slightly improving HDAC6 binding affinity (binding affinity optimized from −7.7 to −8.5 kcal/mol). More importantly, the tail region was reconstructed: the original flexible, small amine side chain in the reference compound was replaced with a bulkier bicyclic tetrahydroisoquinoline scaffold extended by a phenylacetylene moiety. This rigid scaffold effectively minimizes conformational entropy loss upon binding, while the extended tail exhibits optimal steric complementarity with the expansive hydrophobic substrate pocket of LSD1. This synergy enhances binding affinity by augmenting van der Waals and hydrophobic interactions (binding affinity improved from −9.0 to −11.1). At the same time, comparative analysis shows that the generated molecules exhibit high consistency with the reference compounds in terms of QED and SA scores. This conservation of property profiles indicates that, while optimizing LSD1/HDAC binding affinity, the model is able to precisely navigate chemical space, effectively avoiding issues typically associated with activity enhancement, such as uncontrolled molecular weight increase or a significant rise in synthetic complexity.

EGFR/HER2/VEGFR: LaMGen generates novel scaffolds with enhanced multi-target affinity while preserving key pharmacophoric features

To evaluate the performance of the LaMGen model in polypharmacology tasks, we extended our study to the EGFR/HER2/VEGFR triple-target system. The 3D scatter plot in the left panel of Fig. 6 illustrates the binding score distribution across all three target dimensions for the top 500 generated molecules. The data reveal that the generated molecules are predominantly concentrated in the low-energy (high-affinity) regions along all three axes, suggesting that the candidates generally possess the potential to simultaneously bind all three kinase domains without exhibiting significant target bias or imbalance. In the specific molecular structure analysis (Fig. 6, right panel), the model successfully generated Gen. 828, a molecule structurally identical to the experimental reference Compound 37⁵⁶ (Similarity = 1.0). Furthermore, Gen. 1409, which shares a structural similarity of 0.82 with the reference, maintained docking scores comparable to the reference (EGFR: −9.4, HER2: −10.8, VEGFR: −11.1) despite structural variations in its side chains. In Supplementary Fig. 15, the reference scaffold (similarity = 1.0) was generated 32 times, and a highly similar scaffold (similarity = 0.69) 95 times, demonstrating that LaMGen reliably reproduces known scaffolds while exploring structurally related analogs. This result indicates that the generated library covers a chemical space ranging from known active scaffolds to structural analogs exhibiting continuous structure-activity relationships (SAR).

Fig. 6. Representative results for the triple-target system EGFR/HER2/VEGFR.

Left: Scatter plots of docking scores across all three targets for the top 200 model-generated molecules ranked by structural similarity to known actives. Right: Reference triple-target active compound, its closest generated molecules, and the top-scoring compounds produced by LaMGen, with D indicating the Vina docking score. Source data are provided as a Source data file.

To investigate the generative logic of the model when handling increasing complexity, we further compared the dual-target (EGFR/HER2) and triple-target (EGFR/HER2/VEGFR) tasks, as shown in Supplementary Fig. 16. The t-SNE visualization of the chemical space distribution reveals a significant overlap between dual-target and triple-target molecules, suggesting that both groups share common structural features. However, the centroid of the triple-target distribution shows a distinct shift relative to the dual-target molecules, forming specific clusters. This indicates a directional migration of the chemical space from the original basis to accommodate the specific binding pocket of VEGFR.

Statistical analysis of physicochemical properties (Supplementary Fig. 17) shows that the distribution curves for LogP (lipophilicity) and HBD (hydrogen bond donors) highly overlap between the two groups, demonstrating that the model maintained fundamental drug-likeness and permeability characteristics during target expansion. In contrast, the distributions for HBA (hydrogen bond acceptors), TPSA (topological polar surface area), and RotBonds (rotatable bonds) in triple-target molecules exhibit a clear shift toward higher values. Collectively, these data demonstrate that when handling the EGFR/HER2/VEGFR triple-target task, LaMGen retains core pharmacophores (reflected by chemical space overlap and LogP consistency) while adaptively increasing molecular flexibility (RotBonds) and polar interaction sites (HBA, TPSA) to meet the binding requirements of the third target (VEGFR). This process reflects the model’s capacity for adaptive adjustment of molecular complexity under multi-objective constraints.

Through retrospective validation on multiple representative dual-target systems, LaMGen demonstrates generalizability and adaptability across diverse target combinations. Under zero-shot conditions, the model is able to precisely reproduce reference compounds in approximately two-thirds of the dual-target systems, while the majority of the generated candidates occupy high-affinity regions. Without relying on any prior information about active compounds, LaMGen can autonomously generate molecules with high affinity toward both targets, reflecting robust target transferability. Whether the targets are highly homologous kinases or structurally distinct proteins connected via functional signaling networks, the model effectively captures key structure-activity relationship features. The generated compounds exhibit reasonable conformations, favorable binding free energies, good synthetic accessibility, and promising novelty. Across multiple evaluation metrics, including docking scores, scaffold similarity, QED, and SAScore, the generated molecules consistently match or surpass reference compounds. Furthermore, LaMGen can be extended to triple-target systems, demonstrating broad applicability and potential value in multi-target drug design.

Discussion

LaMGen is a large language model (LLM) framework designed for multi-target molecule generation, demonstrating the potential of LLMs to address the complex demands of modern drug discovery. By leveraging ESM-C embeddings, LaMGen effectively represents multiple protein sequences and encodes ligand conformational flexibility through rotation-aware tokens, enabling the direct generation of spatially and chemically plausible 3D molecular structures. A key innovation of LaMGen is the TriCoupleAttention module, which captures multi-level interactions among multiple targets and ligands, substantially improving the docking scores, structural diversity, and drug-likeness of the generated molecules compared to existing dual-target molecular generation models. Notably, LaMGen demonstrates robust generalization across diverse target combinations without requiring task-specific retraining, consistently achieving high performance even on protein pairs with sequence identities below 10%. It surpasses established baselines such as DualDiff in dual-target benchmarks, offering superior conformational accuracy and nearly 30-fold faster generation. In retrospective validation, LaMGen reproduces known active molecules while generating variants that preserve key functional groups yet exhibit enhanced binding affinity. LaMGen can be easily extended to triple-target molecular design, producing candidate molecules with favorable physicochemical properties and high binding affinity. This advancement propels multi-target drug design from conceptual exploration toward practical, high-throughput applications. By integrating structural, bioactivity, and dynamic information with AI-guided energy constraints, LaMGen offers a powerful and efficient solution, paving the way for faster, more reliable, and innovative multi-target drug discovery. With the continued growth of biological datasets and computational capabilities, future iterations of LaMGen hold promise for tackling even more complex multi-objective drug design challenges.

Methods

Dataset

Multi-target dataset

The original data were sourced from the Papyrus V5.5 bioactivity dataset published by Béquignon et al.²⁹. This dataset integrates high-quality bioactivity data from multiple sources, including ChEMBL³⁰ and ExCAPE-DB³¹, and features molecular processing alongside rigorous quality assessment of measurements. Each entry contains a bioactive molecule, its target protein, and associated bioactivity information.

Building upon this, we developed MTD2025 (Multi-Target Dataset 2025), a curated dataset specifically designed for multi-target molecule generation. Although the original Papyrus data are organized for single targets, we applied systematic filtering and reconstruction to convert it into a structurally standardized multi-target dataset. Our filtering criteria included: retaining molecules with pChEMBL values ≥ 6.5 to ensure high bioactivity; selecting entries with protein sequences ≤ 1536 residues to avoid long-tail distributions and improve data quality; and including only molecules observed against at least two targets to guarantee multi-target activity.

After processing, MTD2025 comprises 4,011 unique proteins and 123,024 unique small molecules. For molecules with multiple targets, we handled target combinations as follows: For molecules with n targets (2 ≤ n < 10), all possible pairwise target combinations were exhaustively generated. For molecules with n targets (n ≥ 10), we randomly generated 5n pairs per molecule, balancing occurrence frequencies for each target to enhance diversity and dataset balance.

The final dataset comprises 446,388 dual-target data points and 283,244 triple-target associations. To ensure accurate structural representation, we performed a systematic conformer search for all small molecules in the MTD2025 dataset. Specifically, CREST v2.12 was employed with the GFN-FF force field for exhaustive exploration of the conformational space. Sampling was conducted with a timestep of 1.5 fs, Shake mode, and a maximum of 100 torsional flips (max_flip). For each molecule, dozens to hundreds of candidate conformers were generated, followed by energy-based filtering and clustering using CREST’s internal protocol to remove redundancies. After conformer generation, the five lowest-energy representative conformers per molecule were retained, yielding over 600,000 initial conformers. Subsequently, all candidate conformers were subjected to quantum-level geometry optimization using LiTEN-FF on an RTX 4090 GPU, allowing each conformer to relax fully to its local or global minimum. Finally, these optimized conformers were paired according to indices with the corresponding dual-target or triple-target entries.

For evaluation, the dual-target dataset was split into training and independent test sets based on targets: 3937 targets were assigned to training, and 74 to testing, ensuring no overlap between sets.

Pretraining dataset

The GEOM dataset is a comprehensive collection of molecular conformations designed to support research in molecular geometry, conformational analysis, and machine learning-based molecular modeling²⁶. It contains multiple 3D conformations for a large number of molecules, which are typically generated through high-precision quantum chemical calculations or advanced sampling methods. Each molecule in the GEOM dataset includes not only multiple energy-related conformations but also associated relative energies and structural feature information.

In this study, we utilized the small molecule SMILES and their corresponding high-precision conformations from the GEOM dataset as an important data source for model pretraining. Leveraging these high-quality conformational data helps enhance the model’s understanding of the distribution of molecular 3D structures.

Data preprocessing

To obtain high-quality protein representations, we employed the pre-trained model ESM-C to encode protein sequences. ESM-C (Evolutionary Scale Modeling of Proteins with Context) is a pre-trained language model trained on large-scale protein sequence data³⁴. It utilizes a Transformer architecture and leverages self-supervised learning to capture both short- and long-range dependencies among residues as well as rich evolutionary information from vast protein sequences, thereby partially reflecting the 3D structural and functional features of proteins. The embeddings generated by ESM-C provide sequence representations for downstream tasks such as protein-ligand interaction prediction, enhancing the model’s generalization and expressiveness when target information is limited.

Since the model training relies on the SMILES representation of ligands and their corresponding 3D conformations, constructing precise docking conformations for each molecule with two protein targets is computationally expensive and thus impractical for obtaining the true binding conformations of all molecules for their respective targets. To address this challenge, we adopted an approximate strategy: using tools like RDKit⁵⁷, we generated 100 possible 3D conformations for each molecule and optimized their energies using LiTEN-FF. The lowest-energy conformation was then selected as the representative structure to extract the molecule’s torsion angle information. This strategy builds upon the fundamental assumption that, while the selected conformation may not exactly correspond to the true binding pose for a specific target, its low-energy configuration sufficiently captures the molecule’s intrinsic geometric preferences, thereby serving as an informative structural prior for the generative modeling process.

To evaluate the effectiveness of this strategy in practical tasks, we selected five pairs of out-of-dataset dual-target proteins. For each pair, 1000 candidate molecules were generated, and molecular docking was performed using three types of input conformations: (i) predicted conformations, (ii) RDKit default conformations optimized with MMFF94, and (iii) conformations refined using an AI-based force field. Docking scores were used to assess the impact of different conformational strategies. All docking calculations were performed using Qvina2 with a cubic search box of 20 Å per side centered on the binding site. For each ligand, up to 10 exhaustive search runs were carried out. Default scoring functions were used to rank docking poses, and the top five scoring conformers per ligand were retained for analysis.

Model architecture

Backbone

LaMGen is a 3D molecular generation method designed for dual-target drug discovery tasks, integrating advanced protein language modeling with molecular structure modeling techniques. It aims to generate biologically active and spatially plausible molecules for unknown target combinations. The framework consists of two core modules: a target protein encoder and a molecule generator, built respectively upon the protein language model ESM-C and a large language model Transformer decoder architecture³⁴. Target Protein Encoding Module (ESM-C): To extract spatially relevant structural semantic information from raw amino acid sequences, LaMGen employs ESM-C as the encoder for dual-target proteins. ESM-C is a structure-aware protein language model trained via self-supervised learning on large-scale protein sequence data, with residue contact prediction as an auxiliary task. This enhances its ability to model long-range dependencies and latent 3D structures within sequences.

In LaMGen, the sequences of the two protein targets are fed separately into a shared-parameter ESM-C encoder, producing per-residue embeddings as well as global per-protein embeddings. These representations not only contain primary sequence information but also incorporate contact map features and potential functional site information, providing structural awareness for subsequent molecular generation.

Molecule Generation Module: LaMGen’s molecule generator is based on a large language model architecture consisting of a 12-layer Transformer decoder, each layer containing 8 independent multi-head attention heads. It generates 2D topology and 3D conformational information of molecules autoregressively, step by step. The input sequence is initiated by a special start token “<|beginoftext|>”, and the model sequentially outputs serialized fragments such as atom types, bond connections, and atomic coordinates until it reaches the end token “<|endofmask|>”. To ensure strict generation order during training and inference, the model employs a masking matrix mechanism that masks unpredicted positions, effectively preventing information leakage about target tokens. The training objective is to minimize the negative log-likelihood loss.

$$\mathcal{L}=-{\sum }_{\mathrm{i}=1}^{\mathrm{n}}\mathrm{log\; p}\left(x_i\mid x_{<i}\right)$$ 1

This autoregressive mechanism enables the model to learn complex molecular construction rules, allowing it to capture both chemical validity and spatial plausibility at the sequence level. The multi-head attention mechanism⁵⁸ is a key component of this architecture. Each attention head independently learns a set of Query (Q), Key (K), and Value (V) weights to compute dependencies between different positions in the sequence. The attention calculation is as follows:

$$\text{Attention}\left(Q,K,V\right)=\text{softmax}\left(\frac{Q{K}^T}{\sqrt{d_k}}\right)V$$ 2

Through this mechanism, the model can parallelly capture both local and global relationships in the input sequence from multiple perspectives, thereby enhancing structural representation ability and generalization performance⁵⁹.

Protein-Molecule Conditional Modeling Mechanism: during the actual generation process, LaMGen concatenates the embeddings of the two target proteins with the molecular generation input, guiding the molecule to be generated from scratch. The protein embeddings serve as global contextual control signals and are injected into each layer of the Transformer in the molecular generator via cross-attention. This allows the molecular generation model to perceive the structural constraints and functional requirements of the dual targets, enabling the targeted synthesis of dual-target candidate molecules with potential affinity and spatial compatibility.

Random causal masking

During the training phase of LaMGen, we introduce a Random Causal Masking strategy to enhance the model’s flexibility and generalization ability^60,61. Traditional causal masking, widely used in autoregressive tasks, prevents the model from accessing future tokens when predicting the current one, thus maintaining causal consistency in the generation process and ensuring that the output sequence is constructed step-by-step in order. However, fixed mask positions may limit the model’s ability to learn diverse generation paths, especially in complex tasks such as molecular structure generation, where the assembly order of molecules is inherently uncertain and variable.

To address this, LaMGen adopts an improved random causal masking strategy during training: for each training sample, a truncation position is randomly selected and used as the boundary between the known context and the prediction target. The model is only allowed to use the portion before this point for prediction, while all subsequent information is completely masked. This design not only preserves the requirement of causality but also enables the model to learn effective representations and prediction strategies across various stages of generation, mimicking the natural process of constructing a molecule step-by-step from fragments into a complete structure.

Through this mechanism, LaMGen is better adapted to molecular sequences of varying lengths and construction stages, enhancing the model’s robustness during both training and inference. It improves the generated molecules in terms of structural completeness, diversity, and spatial plausibility. This strategy plays a crucial role in enabling LaMGen to generate complex 3D molecular structures with biological activity and spatial feasibility.

TriCoupleAttention

We have carefully designed the TriCoupleAttention module, which is built upon a distinctive joint sequence concatenation strategy. Specifically, the feature sequence of target 1 is seamlessly fused with that of the ligand to construct a composite input representation $X_1=\left[P_1;L\right]$. Similarly, target 2 and the ligand are combined in the same manner to form $X_2=\left[P_2;L\right]$.

Based on these two highly integrated sequences, the module independently applies a multi-head self-attention mechanism to each, mathematically expressed as:

$$A_1=\text{softmax}\left(\frac{Q\left(X_1\right)K{\left(X_1\right)}^{\top }}{\sqrt{d_k}}\right)V\left(X_1\right)$$ 3

$$A_2=\text{softmax}\left(\frac{Q\left(X_2\right)K{\left(X_2\right)}^{\top }}{\sqrt{d_k}}\right)V\left(X_2\right)$$ 4

Here, $Q\left(\cdot \right)$、$K\left(\cdot \right)$、$V\left(\cdot \right)$ represent independent linear projections, and $d_k$ denotes the dimensionality of the key vectors. This design cleverly integrates the multi-level and multi-dimensional interactions among protein-protein, protein-ligand, and ligand-ligand pairs within a unified self-attention framework. By implicitly modeling all heterogeneous relationships in a shared attention structure, it simplifies the overall architecture while enhancing the expressiveness and granularity of the interaction representations.

To balance the individual impact and cooperative effects of the two target proteins on the ligand in a dual-target system, the TriCoupleAttention module further introduces a weighted fusion mechanism, where the outputs of the two attention branches are smoothly integrated using a linear weighting coefficient $\mathrm{\alpha }$:

$$A=\alpha A_1+\left(1-\mathrm{\alpha }\right)A_2$$ 5

Here, $\mathrm{\alpha }\in \left[\mathrm{0,\; 1}\right]$ is a tunable fusion weight, which can either be set based on prior knowledge or learned dynamically through end-to-end training. This flexibility enables the model to adaptively account for the varying contributions of each target protein. The fused attention matrix $A$ serves as a highly abstract and globally integrated representation of the complex binding patterns between dual targets and the ligand in the latent space. Its powerful expressive capacity greatly enhances the model’s ability to perceive spatial constraints and multibody cooperative effects within molecular systems.

With its refined design philosophy and exceptional computational efficiency, the TriCoupleAttention module overcomes the performance bottlenecks often caused by traditional multi-module coupling. By innovatively combining sequence fusion and weighted integration, it enables a unified and deep representation of both global and local interactions in dual-target molecular systems. Through this mechanism, LaMGen is able to precisely capture the intricate and dynamic 3D interaction details between protein targets and ligand molecules, providing a theoretical and algorithmic foundation for multi-objective optimization and spatial conformation control in dual-target drug design.

Benchmark metrics

To comprehensively evaluate the quality of the generated molecules, we employed ten metrics: validity, novelty, internal diversity, similarity to the training set, quantitative estimate of drug-likeness (QED)⁶², octanol-water partition coefficient (LogP)⁶³, synthetic accessibility score (SAScore)⁶⁴, molecular weight (MW)⁶⁵, topological polar surface area (TPSA)⁶⁶, and Vina score⁴⁵. Validity refers to the proportion of generated 3D structures that can be successfully converted into canonical SMILES. Internal diversity (IntDiv) measures the structural variability within a molecular set, calculated using pairwise Tanimoto distances based on ECFP4 fingerprints⁶⁷. Similarity to the training set reflects the degree of resemblance between the generated molecules and those in the training data, using the same calculation method as IntDiv. All computational experiments were conducted using Python 3.10.19. Key libraries included NumPy (v2.2.6) for numerical computations, pandas (v2.3.3) for data manipulation, and PyTorch (v2.4.0+cu118) with CUDA 11.8 for deep learning model training. The pretrained language models were implemented via the Transformers library (v4.24.0), and tokenization was performed using Tokenizers (v0.13.3). The Vina score estimates the binding affinity between the generated ligands and protein pockets, obtained using QVina2⁴⁵.

Reporting summary

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

Author contributions

Y.K., T.H., and H.L. designed the research study. Q.S. led the study, designed the algorithms, and performed the analyses. Q.G. performed all visualizations. Q.S. and Q.G. wrote the manuscript. Q.S., Q.G., R.H., H.Z., H.S., J.W., R.Q., Y.K., T.H., and H.L. revised the manuscript. Q.S., Q.G., Y.K., T.H., and H.L. wrote the review and editing. All authors read and approved the final manuscript.

Peer review

Peer review information

Nature Communications thanks Amr Alhossary and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Data availability

The data generated in this study are provided as Source Data files associated with this paper. The MTD2025 dataset has been deposited in the Zenodo repository and is publicly accessible at https://zenodo.org/records/18297683. The data is available upon request. Source data are provided with this paper.

Code availability

The code⁶⁸ used in the study is publicly available from the GitHub repository: https://github.com/cholin01/LaMGen/tree/master.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-71737-w.