The integration of large language models (LLMs) into drug design is gaining momentum; however, existing approaches often struggle to effectively incorporate three-dimensional molecular structures. Here, we present Token-Mol, a token-only 3D drug design model that encodes both 2D and 3D structural information, along with molecular properties, into discrete tokens. Built on a transformer decoder and trained with causal masking, Token-Mol introduces a Gaussian cross-entropy loss function tailored for regression tasks, enabling superior performance across multiple downstream applications. The model surpasses existing methods, improving molecular conformation generation by over 10% and 20% across two datasets, while outperforming token-only models by 30% in property prediction. In pocket-based molecular generation, it enhances drug-likeness and synthetic accessibility by approximately 11% and 14%, respectively. Notably, Token-Mol operates 35 times faster than expert diffusion models. In real-world validation, it improves success rates and, when combined with reinforcement learning, further optimizes affinity and drug-likeness, advancing AI-driven drug discovery.
The Adenosine A2B receptor (A2BAR) is considered a novel potential target for the immunotherapy of cancer, and A2BAR antagonists have an inhibitory effect on tumor growth, proliferation, and metastasis. In our previous studies, we identified a class of benzimidazole-pyrazine scaffolds whose derivatives exhibited the antagonistic effect but lacked subtype selectivity towards A2BAR. In this work, we developed a scaffold-based protocol that incorporates a deep generative model and multilayer virtual screening to design benzimidazole-pyrazine derivatives as potential selective A2BAR antagonists. By utilizing a generative model with reported A2BAR antagonists as the training set, we built up a scaffold-focused library of benzimidazole-pyrazine derivatives and processed a virtual screening protocol to discover potential A2BAR antagonists. Finally, five molecules with different Bemis-Murcko scaffolds were identified and exhibited higher binding free energies than the reference molecule 12o. Further computational analysis revealed that the 3-benzyl derivative ABA-1266 presented high selectivity toward A2BAR and showed preferred draggability, providing future potent development of selective A2BAR antagonists.Communicated by Ramaswamy H. Sarma.
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.
Target-aware molecular generation models have emerged as promising tools for structure-based drug discovery, yet it remains unclear whether they genuinely exploit target information or merely resemble the Texas Sharpshooter fallacy by retrospectively rationalizing outputs. To address this, we introduce TarPass, a benchmark comprising 18 well-characterized targets and reference actives, enabling systematic evaluation across three paradigms: 3D in-situ, non-3D, and optimization-based methods. We assessed 15 representative models from the perspectives of protein-ligand interactions (PLIs), molecular plausibility, and drug-likeness. Our results show that only a few models surpass random baselines in recovering PLIs, while pretrained approaches yield more chemically plausible and drug-like molecules with fewer structural alerts. Optimization-based methods effectively redirect outputs toward favorable regions of chemical space, but often at the expense of other properties. In terms of target specificity, existing models can distinguish at the protein-family level but struggle to recognize more fine-grained structural differences. Integrating these insights, we designed a multi-tier virtual screening workflow combining hard filters with refined selection, which can cluster of promising candidates for experimental validation. Overall, TarPass reveals that current models were still far from rational design tools, highlighting both their potential to explore novel chemical space and the need for future improvements in capturing the principles of protein-ligand interactions.