Lei Jiang

In the real world, a molecule is a 3D geometric structure. Compared to 1D SMILES sequences and 2D molecular graphs, 3D molecules represent the most informative molecular modality. Despite the rapid progress of autoregressive-based language models, they cannot handle the generation of 3D molecular conformation due to several challenges: 1) 3D molecular structures are incompatible with LLMs' discrete token space, 2) integrating heterogeneous inputs like proteins, ligands, and text remains difficult within a unified model, and 3) LLMs lack essential scientific priors, hindering the enforcement of physical and chemical constraints during generation. To tackle these issues, we present Chem3DLLM, a unified protein-conditioned multimodal large language model. Our approach designs a novel reversible text encoding for 3D molecular structures using run-length compression, achieving 3x size reduction while preserving complete structural information. This enables seamless integration of molecular geometry with protein pocket features in a single LLM architecture. We employ reinforcement learning with stability-based rewards to optimize chemical validity and incorporate a lightweight protein embedding projector for end-to-end training. Experimental results on structure-based drug design demonstrate state-of-the-art performance with a Vina score of -7.21, validating our unified multimodal approach for practical drug discovery applications.

Yuting Zeng, Weizhe Huang, Lei Jiang, Tongxuan Liu, XiTai Jin, Chen Tianying Tiana, Jing Li, Xiaohua Xu. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Large Vision-Language Models (LVLMs) represent a significant advancement toward achieving superior multimodal capabilities by enabling powerful Large Language Models (LLMs) to understand visual input. Typically, LVLMs utilize visual encoders, such as CLIP, to transform images into visual tokens, which are then aligned with textual tokens through projection layers before being input into the LLM for inference. Although existing LVLMs have achieved significant success, their inference efficiency is still limited by the substantial number of visual tokens and the potential redundancy among them. To mitigate this issue, we propose Focal Pruning (FoPru), a training-free method that prunes visual tokens based on the attention-based token significance derived from the vision encoder. Specifically, we introduce two alternative pruning strategies: 1) the rank strategy, which leverages all token significance scores to retain more critical tokens in a global view; 2) the row strategy, which focuses on preserving continuous key information in images from a local perspective. Finally, the selected tokens are reordered to maintain their original positional relationships. Extensive experiments across various LVLMs and multimodal datasets demonstrate that our method can prune a large number of redundant tokens while maintaining high accuracy, leading to significant improvements in inference efficiency.