2025 Science and Technology Highlights No.43/ October 13-17
This Issue Highlights: Lin Junjie
Medical and Health
01 AI Insulin Designed from Scratch for Safer and More Effective Treatment
On October 13, a research team from the University of Washington, the University of Texas, and other institutions published a new achievement in the Cell sub-journal Molecular Cell, titled "Tuning insulin receptor signaling using de novo-designed agonists." The study used AI-based protein design to target two key binding sites (L1 and F1) of the insulin receptor, designing two high-affinity binding proteins (S1B and S2B). These were combined with different lengths and rigidities of connecting peptides to form novel agonists (AI insulin). Not only does AI insulin have better glucose-lowering effects than insulin, but it can also precisely regulate intracellular signaling pathways, avoiding the promotion of cancer cell growth by insulin. In mouse experiments, AI insulin showed better glucose-lowering effects than natural insulin (only half the dose required, with a longer duration), and could activate mutant receptors in patients with insulin resistance.
• Commentary: Insulin is the core drug for diabetes treatment, but it has limitations such as complex production, strict storage requirements, and potential carcinogenic risks. Insulin receptor activation triggers both glucose-lowering and growth-promoting pathways, the latter often posing safety risks. This study demonstrates a new method for precisely regulating insulin signaling pathways and lays the foundation for developing safer and more effective diabetes treatments. (Luo Xianxian)
02 Transplantation of Human Embryonic Stem Cell-Derived Dopamine Progenitor Cells Improves Motor Symptoms in Parkinson's Disease
On October 13, a research team from Yonsei University College of Medicine in South Korea published a paper in the journal Cell, titled "Phase 1/2a clinical trial of hESC-derived dopamine progenitors in Parkinson’s disease." This was a Phase 1/2a clinical trial data on the transplantation of human embryonic stem cell (hESC)-derived dopaminergic progenitor cells to treat Parkinson's disease. A total of 12 patients with moderate to severe Parkinson's disease (9 males) were recruited, and after enrollment, they received bilateral putamen transplants of high-purity dopaminergic progenitor cells (A9-DPC) generated from human embryonic stem cells. The low-dose group (6 patients) received 3.15 million cells, while the high-dose group (6 patients) received 6.3 million cells, and all patients simultaneously received immunosuppressive treatment. The study results showed that no dose-limiting toxicity or graft-related adverse events were observed during treatment. After 12 months of transplantation, the motor symptoms of the patients improved, with better efficacy observed in the high-dose group.
• Commentary: This study verified the safety and preliminary efficacy of hESC-derived dopaminergic precursor cells, providing clinical evidence for cell replacement therapy for Parkinson's disease. It is hoped that subsequent trials will further confirm the exact efficacy and long-term safety of cell therapy. (Luo Xianxian)
03 Revealing a New Mechanism by Which Cells Combat Viral Infections
On October 13, a collaborative research team from West China Hospital of Sichuan University, Fox Chase Cancer Center in the United States, Russian National Research University Higher School of Economics, and the University of Würzburg in Germany published a paper in Nature, titled "Host cell Z-RNAs activate ZBP1 during virus infections." The study revealed a new host cell antiviral defense mechanism: when herpes simplex virus type 1 (HSV-1) and influenza A virus (IAV) infect cells, the host cells produce Z-RNA (a rare left-handed double-helix RNA structure) to activate Z-type nucleic acid-binding protein-1 (ZBP1) protein, leading to cell death to limit viral replication and spread. These Z-RNAs mainly originate from viruses interfering with the host's transcription termination process, resulting in the accumulation of abnormal transcripts.
• Commentary: Currently, the molecular mechanism by which ZBP1 is activated during viral infections is unclear. It has long been believed that Z-RNA activating ZBP1 mainly comes from the virus itself, but its exact source and formation mechanism remain controversial. The above research findings challenge previous understanding of the mechanism by which viral infections activate cell death, deepening the understanding of virus-host interactions, and providing potential targets for developing new antiviral strategies. (Luo Xianxian)
04 AI Cancer Model C2S-Scale Released
On October 16, Google DeepMind and Yale University introduced a new foundational model designed specifically for single-cell analysis with 27 billion parameters — Cell2Sentence Scale 27B (abbreviated as C2S-Scale). The model treats single-cell RNA sequencing (scRNA-seq) data as a language processing task, converting high-dimensional scRNA-seq expression data into "cell sentences" — ordered sequences of gene names, thereby enabling extensive biological analysis. The C2S-Scale model discovered that the kinase inhibitor CK2 silmitasertib (CX-4945) has a significant "environmental differentiation effect" — it significantly enhances antigen presentation only in environments with active immune signals, and is almost ineffective in immune-neutral environments. During the experimental phase, researchers tested this hypothesis in a human neuroendocrine cell model, confirming the reliability of the model's predictions and revealing a potential new pathway for making the immune system more likely to recognize tumors.
The research team publicly released the preprint of the paper on bioRxiv, titled "Scaling Large Language Models for Next-Generation Single-Cell Analysis." The paper is currently in the "peer review" stage.
• Commentary: C2S-Scale integrates single-cell data with text annotations such as literature, disease labels, and pathway information, enabling the model to perform biological reasoning by combining contextual knowledge, significantly enhancing interpretability and semantic understanding depth. It is expected that the research team will further expand the achievements of this study, allowing large models to decipher more mysteries of life's language. (Luo Xianxian)
Digital Intelligence
01 OpenAI Collaborates with Broadcom to Develop Custom AI Accelerators
On October 13, OpenAI announced a collaboration with Broadcom to jointly develop 10-gigawatt custom AI accelerators. Both parties expect to deploy racks equipped with AI accelerators and network systems starting from the second half of 2026, completing deployment by the end of 2029. These racks will be expanded using Broadcom's Ethernet and other connectivity solutions.
OpenAI will be responsible for designing these accelerators and systems, and will jointly develop and deploy them with Broadcom. Through self-developed chips and systems, OpenAI hopes to directly integrate the experience accumulated in advanced model and product development into hardware design, releasing new performance and intelligence levels.
• Commentary: On September 22, 2025, OpenAI received a $10 billion investment from NVIDIA to deploy at least 10 gigawatts of computing power; on October 6, OpenAI also partnered with AMD, planning to deploy at least 6 gigawatts of computing power. With Broadcom, OpenAI is shifting towards a diversified computing strategy of "self-reliance + cooperation." (Cao Yan)
02 OpenAI Builds Latin America's First "Stargate" in Argentina
On October 14, OpenAI announced an intent letter with energy company Sur Energy, planning to build a data center hub in Argentina, which could become one of the largest technology infrastructure projects in the country's history. The Argentine government had previously announced the "Argentina Stargate" project, which involves up to $25 billion in investment, building an advanced AI computing facility with a capacity of 500 megawatts.
The data center will be located in the Patagonia region, known for its vast land and abundant renewable energy resources. Although the exact construction schedule has not been announced, local reports indicate that the project may start next year, with the initial 100-megawatt phase expected to be completed by 2027.
• Commentary: Following the United States, UAE, and Norway, Argentina has become the latest addition to OpenAI's global computing power layout. The strong AI development ecosystem in the region makes Argentina an ideal location for AI infrastructure construction. The official implementation of the Stargate project in Argentina is expected to create thousands of jobs, and whether Argentina's digitalization process will take off? Let's wait and see. (Wei Youyi)
03 Oracle and AMD Advance Strategic Cooperation
On October 14, Oracle and AMD announced further expansion of their cooperation. Oracle Cloud Infrastructure (OCI) will become the first AI supercluster partner driven by AMD Instinct™ MI450 series GPUs. This supercluster will deploy 50,000 GPUs by the third quarter of 2026, and gradually expand after 2027.
The OCI planned supercluster will use AMD's "Helios" rack design, integrating MI450 GPUs, next-generation EPYC CPUs, and PENSANDO networking, forming a synergistic computing architecture. This vertically optimized architecture aims to provide high-performance, scalable, and energy-efficient solutions for large-scale AI training and inference.
• Commentary: AMD has successively reached chip collaborations with OpenAI and Oracle, not only meeting the explosive growth of AI computing power demand, but also breaking through the industry's reliance on NVIDIA. For AMD, the latest MI450 chip has achieved large-scale commercial use in public cloud scenarios, helping to enhance its position in the data center market. (Cao Yan)
04 World's First Smart Breeding Robot Capable of Autonomous Navigation for Hybrid Pollination
On October 16, a research result from the Institute of Genetics and Developmental Biology, Chinese Academy of Sciences, and the Institute of Automation, Chinese Academy of Sciences, titled "Engineering crop flower morphology facilitates robotization of cross-pollination and speed breeding" was published on the cover of the journal Cell. The study integrated biotechnology and artificial intelligence (BT+AI) deeply, proposing the concept of crop-robot co-design for the first time. By gene editing, the study re-designed the flower shape of crops, rapidly and accurately created "robot-friendly" male sterile lines, and successfully developed the world's first smart breeding robot capable of autonomous navigation for hybrid pollination — "Jier" (Genome Editing combined with AI-based Robotics, GEAIR).
"Jier" consists of a mobile base, cameras, and pollination devices, among other parts. It uses AI algorithms to accurately identify flowers (accuracy rate 85.1%), and uses a robotic arm to gently perform pollination (taking 15 seconds per flower). One round of navigation pollination achieved a success rate of 77.6% ± 9.4%, and the robot can continuously and repeatedly navigate automatically for hybrid pollination to ensure that each flower is successfully pollinated and fruits. The team also applied the "Jier" system to soybeans, achieving the rapid creation of a structural male sterile line in soybeans, and using this sterile line can save 76.2% of manual pollination operation time.
????The figure shows the robotic arm of the GEAIR system precisely pollinating a genetically edited tomato flower. Image source: cell
• Commentary: The high cost and low efficiency of hybrid breeding and seed production are important bottlenecks limiting the utilization of heterosis. The above results break through this bottleneck, greatly reducing breeding costs, shortening the breeding cycle, improving breeding efficiency, accelerating the cultivation and promotion of superior varieties, thus increasing the yield and quality of crops. (Luo Xianxian)
Energy and Astronomy
01 Bloom Energy and Brookfield Reach $5 Billion Collaboration
On October 13, fuel cell company Bloom Energy and asset management company Brookfield announced an AI infrastructure collaboration, with Brookfield investing $5 billion to deploy Bloom Energy's advanced fuel cell technology. The two companies will build AI factories worldwide, including a European factory to be announced by the end of 2025.
https://investor.bloomenergy.com/press-releases/press-release-details/2025/Brookfield-and-Bloom-Energy-Announce-5-Billion-Strategic-AI-Infrastructure-Partnership/default.aspx
This collaboration is the first investment in Brookfield's AI infrastructure strategy. Bloom Energy is the preferred power supplier for Brookfield's global AI factories. Currently, Bloom has collaborated with American Electric Power (AEP), Equinix, Oracle, and other data center developers to deploy hundreds of megawatts of fuel cells.
• Commentary: The demand for power in AI data centers is surging rapidly, and fuel cells have become a key solution to meet this demand. This collaboration is expected to break through the bottleneck of insufficient traditional grid power supply, and the "computing power + power" integrated AI factory model may become a new model for the construction of large AI data centers in the future. (Cao Yan)
02 CFS and DeepMind Collaborate on Nuclear Fusion Research
On October 16, U.S. nuclear fusion commercial company Commonwealth Fusion Systems (CFS) and Google DeepMind announced a collaboration on nuclear fusion research, jointly exploring optimization plans to improve the operational efficiency of the SPARC fusion device.
Previously, DeepMind and the Plasma Center of École Polytechnique Fédérale de Lausanne (EPFL) demonstrated that deep reinforcement learning could accurately control tokamak magnets, stabilizing complex plasma forms, and in 2024, developed a fast differentiable plasma simulator TORAX based on the JAX framework.
Now, Google DeepMind will apply TORAX to CFS, and the two parties have started collaboration in three key areas: building a fast, accurate, differentiable fusion plasma simulation system; exploring efficient and robust paths to maximize fusion energy output; and using reinforcement learning to explore new real-time control strategies.
Google became a shareholder of CFS in 2021, and on June 30, 2025, continued to increase its stake in CFS and committed to purchasing 200 megawatts of fusion power from its first ARC power plant in the 2030s.
• Commentary: The combination of AI and fusion, two transformative technologies, will create what kind of exciting results? At present, the world's top AI research teams and the most advanced commercial controlled fusion hardware teams are accelerating fusion energy research together, which is indeed worth looking forward to. (Li Yiluo)
03 Global First Mid-Infrared Solar Magnetic Field Special Observation Equipment Put into Use
On October 17, the National Astronomical Observatory of the Chinese Academy of Sciences announced that the world's first solar magnetic field precise measurement mid-infrared observation system (AIMS telescope) led by the observatory has recently passed the national acceptance and officially started operation. This telescope filled the gap in international mid-infrared solar magnetic field observations. The AIMS telescope is built at the D observation platform of the Qinhai Cold Lake Shishiteng Mountain (elevation 4000 meters), as the world's first mid-infrared band solar magnetic field special observation equipment, since its development began in 2015, it has achieved several key technological breakthroughs. The research team achieved a breakthrough in direct measurement of solar magnetic fields through the 12.3-micron mid-infrared band, improving the magnetic field measurement accuracy to better than 10 Gauss level, solving a bottleneck problem in the century-long history of solar magnetic field measurement.
• Commentary: The launch of this telescope means that humans have gained a "mid-infrared eye" for observing the sun for the first time, directly overcoming a century-old measurement challenge, allowing more accurate exploration of the magnetic field origins driving solar storms, greatly enhancing the ability to forecast "space weather." This is not only a scientific breakthrough, but also genuinely protects the modern lives of everyone around the globe, building a stronger defense for our power grids and communication security. (Lin Junjie)
Southern Weekend Science and Technology Innovation Research Center
Editor: Huang Jinping
Original article: https://www.toutiao.com/article/7563334433588265535/
Statement: The article represents the views of the author. Please express your opinion by clicking the [top/foot] button below.