Newshot Tsinghua University releases single-cell proteomics foundation model ProtCell, successfully completes “virtual cell” pretraining reproduction using Moore Threads S5000 intelligent computing card
IT Home, November 14: According to the Moore Threads official WeChat account, the Jiang Rui research group in the Department of Automation at Tsinghua University recently released ProtCell, a foundation model for single-cell proteomics, and completed the paper “ProtCell: a foundation model for single-cell proteomics.” The model aggregates about 380 million cells and 98 public datasets for pretraining and evaluation, can accurately map cell states and infer unmeasured protein expression, and provides key foundational support for drug R&D and the construction of frontier “virtual cells.”
The paper also discloses that the team successfully completed full-process pretraining reproduction of ProtCell on Moore Threads’ flagship AI training-and-inference integrated intelligent computing card MTT S5000. The main quantitative evaluation metrics differ from those of international mainstream GPU platforms by only about 0.2%, validating the carrying capacity of Moore Threads’ Universal GPU and MUSA full-stack software ecosystem for frontier life science foundation models.
Public information shows that “virtual cells” aim to use AI to build computational models of how cells operate, describe the current state of cells, and accurately predict dynamic changes in cells after interventions such as drugs and gene editing, and have become a core strategic high ground in global AI4S competition.
Internationally, top research institutions and technology giants are accelerating investment of billions of dollars and 10,000-GPU-scale computing clusters to compete in this field. In China, the “Opinions of the State Council on Deeply Implementing the ‘AI Plus’ Initiative” lists “AI Plus” science and technology as the first key action. Because core data such as genomics and proteomics on which virtual cell R&D depends are highly related to national biological genetic resource security, building an independent and controllable advanced computing foundation is crucial.
Related articles
- Day-0 support: Moore Threads completes adaptation of MiniMax H3-Image image generation model
- Day-0 support: Moore Threads completes adaptation of 3-trillion-parameter Kimi-K3.1 Pro, supporting text and image inference
- 2027QS Asia University Rankings announced: University of Hong Kong retains first place, Peking University second, Tsinghua University third; mainland China has the most entries
- miHoYo Honkai AI team open-sources GameScaling game intelligence model, with participation from multiple university interns, parameter scale 0.8B‑27B
- Apple’s Cook appears at Tsinghua University: attends annual meeting of the School of Economics and Management Advisory Board, currently serves as chair
- 2K + 185Hz, Visionox × iQOO × Tsinghua University screen technology communication meeting scheduled for October 14