Tsinghua Releases ProtCell Model and Validates It on Moore Threads GPUs
Researchers at Tsinghua University have released the ProtCell single-cell proteomics foundation model and successfully reproduced its pre-training on domestic Moore Threads MTT S5000 AI accelerators, achieving a performance difference of only about 0.2% compared to mainstream international GPUs. This achievement marks an important milestone for AI for Science (AI4S) by demonstrating that domestic hardware can support advanced life science foundation models, which is crucial for national biological data security and computational independence. ProtCell was trained on approximately 380 million cells and 98 public datasets to map cellular states and infer unmeasured protein expression. Its full-process pre-training was successfully validated on the MTT S5000 card using its proprietary MUSA software stack.
## BACKGROUND
Single-cell proteomics and 'virtual cell' modeling use AI to simulate cellular dynamics, track current states, and predict responses to drugs or gene editing. The Moore Threads MTT S5000 is a full-function GPU AI accelerator designed for large-scale model training and high-performance computing.