You can also find my publications on my Google Scholar profile.
Published in NeurIPS, 2024
Employing a 2D evolutionary positional encoding scheme and learning from AlphaFold2 Feedback, MSAGPT generates constructive virtual MSA that enables accurate protein structure predictions in situations where natural co-evolutionary information is scarce.
Published in Nature Methods, 2025
We propose a 100-billion-parameter unified protein language model, xTrimoPGLM, optimizing both protein understanding and generation objectives simultaneously through an innovative pre-training framework, demonstrating State-Of-The-Art performance on 18 protein understanding benchmarks, and powerful de novo and conditional sequence generation abilities.
Preprint on bioRxiv, 2025
Mimyr is a generative framework for reconstructing missing regions and virtual slices in spatial transcriptomics data caused by tissue damage or incomplete sampling. It jointly predicts cell locations, cell types, and gene expression profiles using a three-staged pipeline. Mimyr produces realistic, high-fidelity reconstructions across experimental settings and outperforms existing methods in both accuracy, efficiency and robustness.