Mujie Lin

Mujie Lin

M.Phil. in Computer Science @ PKU. Building AI systems for molecular discovery, protein dynamics, and scientific foundation models.

Publications

You can also find my articles on my Google Scholar profile.

Tikhonov-Stabilized Bézier Representation Forecasting for Training-free Diffusion Acceleration

Published in Conference on Neural Information Processing Systems (NeurIPS 2026), poster, 2026

A training-free diffusion acceleration framework that forecasts output-proximal denoising representations along low-order Bézier trajectories, with Tikhonov-stabilized control-point fitting and a convex-hull-inspired guardrail against aggressive extrapolation. Accelerates FLUX.1 by up to 4.79× and HunyuanVideo by up to 4.11× while preserving generation quality.

Recommended citation: Lei Zhu#, Mujie Lin#, Ruochong Zheng, Guangyi Wang, Hao Li, Peng Jin, Chang Liu†, Jie Chen†. "Tikhonov-Stabilized Bézier Representation Forecasting for Training-free Diffusion Acceleration." NeurIPS, 2026.
Download Paper

PhiFold: Towards Dynamic Protein Design with Physics-Structured Covariance Modeling

Published in arXiv preprint (arXiv:2609.32309), 2026

A framework for jointly generating protein backbones and their second-order dynamics. Decomposes residue-displacement covariance into local flexibility, low-rank collective motion, and residue-wise collective participation, assembled into a positive-definite covariance with exact marginal consistency.

Recommended citation: Yutian Liu#, Mujie Lin#, Lanqian Zhang#, Meng Fan, Chang Liu†, Zhiwei Nie†, Siwei Ma†. "PhiFold: Towards Dynamic Protein Design with Physics-Structured Covariance Modeling." arXiv preprint arXiv:2609.32309, 2026.
Download Paper

OpenAI4S: Code as Action, Science as Sessions

Published in arXiv preprint (arXiv:2609.15096), 2026

An open-source scientific research agent built around persistent computational state and provenance: scientific actions are complete code cells executed in persistent Python and R kernels, with an append-only Action Ledger, versioned artifacts, and workspace checkpoints enabling inspectable, resumable, and reproducible long-running studies.

Recommended citation: Gongbo Zhang*, Hao Li*, Yu Wang, Mujie Lin, Liuzhenghao Lv, Yicheng Mao, Yimi Wang, Jun Zhu, Minhan Tang, Zhengxiang Jiang, Yusong Wang, Jiayu Yao, Kunpeng Ning, Dawei Pang, Yonghong Tian, OpenAI4S Community, Yuyang Liu, Li Yuan. "OpenAI4S: Code as Action, Science as Sessions." arXiv preprint arXiv:2609.15096, 2026.
Download Paper

Structure Prediction of Peptides Containing Non-Canonical Amino Acids: A Comprehensive Benchmark of Full-Atom Prediction Algorithms

Published in Genomics, Proteomics & Bioinformatics (SCI Q1 TOP, IF = 13.9), 2026

A benchmark study evaluating leading all-atom structure prediction models on canonical and modified peptide–protein complexes, providing practical guidance for protocol selection across diverse peptide design scenarios.

Recommended citation: Yu Wang, Yifan Deng, Fuming Zeng, Mujie Lin, Tao Guo, Bin Cong, Shiwei Sun, Xin Gao. "Structure Prediction of Peptides Containing Non-Canonical Amino Acids: A Comprehensive Benchmark of Full-Atom Prediction Algorithms." Genomics, Proteomics & Bioinformatics, accepted, 2026.
Download Paper

BioDynaSpec: Harmonic-Guided Spatio-Spectral Autoregressive Diffusion for Protein Dynamics Generation

Published in International Conference on Machine Learning (ICML 2026), 2026

A spatio-spectral generative framework for long-horizon protein molecular dynamics. Introduces IWFD frequency decomposition, hybrid autoregressive-diffusion generation, and IRFC structural priors. Reduces long-horizon trajectory error by over 60% on ATLAS vs. state-of-the-art methods.

Recommended citation: Mujie Lin, Yutian Liu, Yudi Guo, Yanzhen Hou, Yiheng Tao, Ruochong Zheng, Kaiwen Cheng, Xin Shan, Youdong Mao, Jie Chen. "BioDynaSpec: Harmonic-Guided Spatio-Spectral Autoregressive Diffusion for Protein Dynamics Generation." ICML, 2026.
Download Paper

ProAR: Probabilistic Autoregressive Modeling for Molecular Dynamics

Published in AAAI Conference on Artificial Intelligence (AAAI 2026), 2026

Probabilistic autoregressive framework for long-horizon MD generation with multivariate Gaussian timesteps, dual-network architecture, and anti-drifting sampling. Achieves 7.5% RMSE reduction and 25.8% improvement in dynamic pattern accuracy.

Recommended citation: Kaiwen Cheng, Yutian Liu, Zhiwei Nie, Mujie Lin, Yanzhen Hou, Yiheng Tao, Chang Liu, Jie Chen, Youdong Mao, Yonghong Tian. "ProAR: Probabilistic Autoregressive Modeling for Molecular Dynamics." AAAI, 2026.
Download Paper

MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention

Published in Technical Report (arXiv preprint), 2025

World’s first open-weight large-scale hybrid-attention reasoning model (456B MoE, 1M context). Foundation Model Team contributor — built automated evaluation and reasoning benchmarks. 3.2k+ GitHub stars; Hugging Face #1 Paper of the Day.

Recommended citation: Aili Chen, Aonian Li, ..., Mujie Lin, ..., et al. "MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention." arXiv:2506.13585, 2025.
Download Paper

ADCNet: a unified framework for predicting the activity of antibody-drug conjugates

Published in Briefings in Bioinformatics (SCI Q1, IF = 7.7), 2025

A unified deep learning framework for antibody-drug conjugate activity prediction, integrating antigen, antibody, linker, payload, and DAR representations with ESM-2 and FG-BERT. Published in Briefings in Bioinformatics.

Recommended citation: Liye Chen#, Biaoshun Li#, Yihao Chen#, Mujie Lin, Shipeng Zhang, Chenxin Li, Yu Pang, Ling Wang. "ADCNet: a unified framework for predicting the activity of antibody-drug conjugates." Brief. Bioinform., 2025.
Download Paper

MalariaFlow: A Comprehensive Deep Learning Platform for Multistage Phenotypic Antimalarial Drug Discovery

Published in European Journal of Medicinal Chemistry (SCI Q1 TOP, IF = 7.1), 2024

Considering the limitation of previous antimalarial activity prediction models in neglecting the multi-stage/multi-phenotype activity of Plasmodium, this study constructs a comprehensive data benchmark covering the three major life cycles of Plasmodium falciparum and activity test data from ten common mutant strains. Building upon this foundation, we propose an adaptive neural network, FP-GNN, which integrates molecular fingerprints and molecular graphs, effectively leveraging the advantages of chemical prior knowledge and graph representation learning. Compared to mainstream algorithms, FP-GNN achieves a 9.5% and 7.4% improvement in AUC and BA metrics, respectively, significantly enhancing the accuracy and robustness of Plasmodium activity prediction.

Recommended citation: Mujie Lin, Junxi Cai, Yuancheng Wei, Xinru Peng, Qianhui Luo, Biaoshun Li, Yihao Chen, Ling Wang. "MalariaFlow: A Comprehensive Deep Learning Platform for Multistage Phenotypic Antimalarial Drug Discovery." Eur. J. Med. Chem., 2024.
Download Paper

Uni-SMART: Universal Science Multimodal Analysis and Research Transformer

Published in arXiv preprint, 2024

Multimodal scientific literature understanding model for molecules, tables, and charts. arXiv preprint; featured on Hugging Face Papers as Paper of the Day.

Recommended citation: Hengxing Cai, Xiaochen Cai, Shuwen Yang, Jiankun Wang, Lin Yao, Zhifeng Gao, Junhan Chang, Sihang Li, Mingjun Xu, Changxin Wang, Hongshuai Wang, Yongge Li, Mujie Lin, Yaqi Li, Yuqi Yin, Zheng Cheng, Zifeng Zhao, Linfeng Zhang, Guolin Ke. "Uni-SMART: Universal Science Multimodal Analysis and Research Transformer." arXiv:2403.10301, 2024.
Download Paper

FG-BERT: a generalized and self-supervised functional group-based molecular representation learning framework for properties prediction

Published in Briefings in Bioinformatics (SCI Q1, IF = 7.7), 2023

Functional-group-based molecular language model with self-supervised pre-training for transfer learning on 44 molecular property benchmarks, outperforming state-of-the-art methods.

Recommended citation: Biaoshun Li, Mujie Lin, Tiegen Chen, Ling Wang. "FG-BERT." Brief. Bioinform., 2023.
Download Paper

DeepCancerMap: A versatile deep learning platform for target- and cell-based anticancer drug discovery

Published in European Journal of Medicinal Chemistry (SCI Q1 TOP, IF = 7.1), 2023

Large-scale anticancer dataset (25.63M records) and 832 FP-GNN models with web platform for virtual screening and similarity search. AUC up to 0.91 on benchmark test sets with time-split external validation.

Recommended citation: Jingxing Wu#, Yi Xiao#, Mujie Lin#, Hanxuan Cai, Duancheng Zhao, Yirui Li, Hailin Luo, Chuanqi Tang, Ling Wang. "DeepCancerMap." Eur. J. Med. Chem., 2023.
Download Paper