Recently, Google DeepMind published an article titled “AlphaFold: Five Years of Impact,” looking back at the tremendous role of technological breakthroughs in protein structure prediction over the past five years in driving scientific progress.

Artificial intelligence is reshaping the landscape of scientific research at an unprecedented pace. Among numerous research fields, biological fields such as life sciences and biomedicine have become the most active and leading frontiers of AI + scientific research (hereinafter referred to as scientific intelligence) due to factors such as abundant data, clear application scenarios, and urgent social needs. AI models and tools have not only made breakthroughs in basic research such as predicting protein structures but are also propelling new drug pipelines into clinical trials and even starting to autonomously discover new biological pathways.

Under the leadership of tech companies like Google DeepMind that have been continuously delving into AI for Science, scientific intelligence, represented by biology, is entering a period of high output and rapid iteration in terms of application implementation. The AI-driven research paradigm of “basic model + scientific intelligent agent + autonomous laboratory” is gradually taking shape.

Google DeepMind Leads the Evolution of Scientific Intelligence Technology

Google has been deeply engaged in scientific research for over a decade. With AI computing infrastructure centered around TPUs and an AI model foundation based on Gemini, it continuously develops scientific intelligence technologies. It has created world-class scientific intelligence models and tool systems such as AlphaFold, leading the global evolution of scientific intelligence technology.

AlphaFold leads the leap in biological research from structure prediction to generative design. Biology is the field where DeepMind made its earliest strategic deployment and has the deepest moat. Its core logic lies in using deep learning to solve the conformational space problem of high-dimensional biological macromolecules. The advent of AlphaFold marked the substantive solution of the protein structure prediction problem. It not only won the 2024 Nobel Prize in Chemistry but also became the digital infrastructure of modern biology. AlphaProteo has propelled biological research into the era of generative biology. Combined with AlphaMissense’s accurate prediction of the pathogenicity of gene mutations, DeepMind has further opened up the entire chain from “target discovery – structure analysis – drug design.”

WeatherNext achieves a dimensionality reduction attack on numerical simulation in meteorology through data-driven methods. DeepMind’s latest WeatherNext 2 model (the successor of GraphCast) has comprehensively surpassed traditional physical models in terms of accuracy and efficiency. In 99.9% of the predicted variables and time spans, WeatherNext 2 outperforms the HRES system of the European Centre for Medium-Range Weather Forecasts (ECMWF), and its inference speed has increased by several orders of magnitude.

GNoME and AlphaQubit expand the application of AI in the fields of physics and materials science. GNoME (Graph Networks for Materials Exploration) has conducted a massive search in the inorganic crystal space using deep learning, predicting millions of stable new material structures, several times the sum of human experimental discoveries over the past few decades. This provides a vast candidate pool for the research and development of battery technologies and superconducting materials. In the field of quantum computing, the AlphaQubit model has successfully applied the Transformer architecture to quantum error correction, significantly reducing the qubit readout errors in quantum computing chips.

AlphaEvolve promotes the evolution of mathematics and computer science from logical reasoning to the self-evolution of algorithms. By introducing an evolutionary computing paradigm, AlphaEvolve is committed to breaking the limitations of human-designed algorithms, automatically searching for and discovering more efficient machine learning algorithms and loss functions, achieving a meta-level leap from “human design” to “automatic discovery.” On this basis, AlphaChip transforms chip design into a reinforcement learning problem and successfully optimizes the layout of Google’s TPU v6. AlphaGeometry and AlphaProof, on the other hand, demonstrate AI’s breakthroughs in formal mathematical proofs and logical reasoning.

Progress in the Biological Field Leads the Frontier of Scientific Intelligence Implementation

The technological breakthroughs led by Google DeepMind have ignited a global wave of technological research and development and industry application in scientific intelligence. Biology has become the research field with the fastest progress, followed by materials science, physics, meteorology, computer science, and mathematics.

(I) Scientific Intelligence Enters the Deep Waters of Basic Biological Research

New discoveries have been made in the AI-generated analysis of single-cell behavior. Google and Yale University jointly released the C2S-Scale, a 27-billion-parameter basic model for single-cell analysis, which generated new hypotheses about cancer cell behavior and verified them in multiple in vitro experiments. This demonstrates the potential of using AI to propose original scientific hypotheses and is expected to explore new ways to develop anti-cancer methods accordingly.

The protein generative simulation and prediction system has become more comprehensive. Microsoft’s BioEmu model fills the gap in protein dynamics simulation and achieves a simulation speed increase of up to 100,000 times. The Chinese Academy of Sciences team proposed a reverse-folding protein prediction model that integrates structural and evolutionary constraints, opening up a new path for protein engineering. The relevant results were published in the journal Cell.

An AI-assisted genomics research and development system has been initially established. Through ten years of continuous research and exploration, Google has gradually built an AI genomics research and application system covering gene sequencing, reading, and variant identification, gene expression prediction and pathogenic potential assessment, and disease gene detection and diagnosis, which helps promote the development of genetics and gene-based medicine.

(II) AI-Driven Medical Applications Are in Full Bloom

AI-assisted pathological detection expands into new disease scenarios. The DeepGEM pathological large model jointly developed by Tencent Life Sciences Laboratory, the First Affiliated Hospital of Guangzhou Medical University, and the Guangzhou Institute of Respiratory Health has completed large-scale validation in lung cancer gene mutation prediction. It can complete lung cancer gene mutation prediction within 1 minute using only conventional pathological section images, with an accuracy rate of 78% – 99%.

AI detection of gene mutations is further being turned into tools. Google released the DeepSomatic toolset for identifying gene variations in tumor cells, applicable to cancer types such as leukemia, breast cancer, and lung cancer, with an identification accuracy rate superior to existing solutions.

AI-driven drug research and development has crossed the clinical phase II stage. The AI-optimized candidate drug MTS-004 jointly conducted by Peking University Third Hospital and other hospitals and Jitai Technology has completed phase III clinical research, becoming the first AI-empowered formulation new drug in China to complete phase III clinical trials. This drug is expected to target neurological diseases such as amyotrophic lateral sclerosis and stroke. These progresses have broken through the bottleneck that in the past few years, AI-driven drug discovery in China and even internationally has rarely broken through clinical phase II, attracting attention and recognition both domestically and internationally.

(III) AI Accelerates Applications in Materials Science, Meteorology, Mathematics, and Other Fields

AI + materials science is expected to become the next frontier of scientific intelligence. Periodic Labs, founded by former members of OpenAI and DeepMind, is engaged in the AI-automated discovery of new superconducting materials. CuspAI, which has just secured $100 million in Series A funding, is developing an AI platform for discovering new materials for carbon capture. RhinoWise, a material innovation platform being built by Dingxi Zhichuang, which was incubated at the Shenzhen-Hong Kong Science and Technology Innovation Cooperation Zone of Peking University Shenzhen Graduate School, is carrying out key material innovations in fields such as new energy and semiconductors.

The applications of AI in meteorology and physics have yielded practical results. DeepMind’s hurricane AI model has successfully predicted the paths and intensity changes of super hurricanes such as “Melissa,” helping the United States and surrounding countries issue early warnings. Black hole theoretical physicist Alex Lupsasca used GPT-5 to derive new characteristics of black hole theory within half an hour. Nuclear fusion start-up CFS uses Google’s open-source TORAX tool to assist in the research and development of the SPARC nuclear fusion device.

The application potential of AI in mathematics and computer science is huge. Mathematical researchers used GPT-5 to explore solutions to the historical mathematical problem – the Erdős problem. Google is advancing mathematical and theoretical computer science research based on AlphaEvolve. NVIDIA’s open-source model system GenCluster won the gold medal in the IOI 2025 competition. Large models such as OpenAI’s internal model, Gemini Deep Think, and DeepSeek Math-V2 are also constantly刷新 (setting new records) AI’s gold medal scores in the International Mathematical Olympiad.

Technical Foundations, Collaboration Models, and Research Scales: Three Dimensions of AI Reshaping the Research Paradigm

From the progress of scientific intelligence represented by biology, it can be seen that AI’s reshaping of scientific research is systematic. It is changing the traditional approaches of scientific discovery from three dimensions: technical foundations, collaboration models, and research scales. The AI-driven research paradigm of “basic model + scientific intelligent agent + autonomous laboratory” is gradually taking shape.

(I) General Models and Specialized Models Build the Technical Foundations of Scientific Intelligence

General basic large models are expected to become the “operating systems” of scientific intelligence. General basic large models can provide powerful understanding, reasoning, analysis, and generation capabilities, as well as comprehensive scientific basic knowledge and general knowledge reserves, which can greatly improve the daily research efficiency of scientific researchers. At the same time, pioneering large model companies are also continuously improving the scientific professional capabilities of basic models. Anthropic’s Claude Sonnet 4.5 has significantly improved its understanding and application of life science task workflows and enhanced its ability to utilize scientific tools and resources based on intelligent agent capabilities and connectors.

Scientific specialized large models serve as the “specialized engines” for vertical scientific research fields and in-depth breakthroughs. These models usually integrate knowledge, research methods, and experience in specific fields. Google has a comprehensive leading position globally in scientific specialized large models, with its specialized models and algorithms covering various fields such as life sciences and biology, materials science and chemistry, Earth and climate science, and mathematics and basic sciences. The aforementioned C2S-Scale, BioEmu, DeepGEM, etc., also belong to this category of models. In addition, the Panshi Scientific Basic Large Model jointly developed by a team from the Chinese Academy of Sciences is also a beneficial practice of integrating basic models with specialized models.

(ii) Scientific Intelligent Agents Based on Human-Machine Collaboration Begin to Promote Active Scientific Discovery

AI handles trivial and time-consuming but indispensable research links, while human scientists control the research direction and evaluate research results. This will be a typical human-machine collaboration research model in the future.

With the accelerated development of intelligent agent technology, AI is transforming from a passive tool into a collaborator or even an active discoverer of scientists. Harvard and MIT jointly launched ToolUniverse, a research tool platform specifically designed for AI intelligent agents, which contains more than 600 scientific tools and is compatible with mainstream basic large models. This helps inspire more researchers to build intelligent agent scientists for specific scientific research fields. Google Deepmind’s AlphaEvolve is an evolutionary AI intelligent agent with coding capabilities that can actively discover and automatically optimize general algorithms in mathematics and computing. It has been applied in actual scenarios such as Google’s internal data center scheduling, chip design, and large model performance optimization. A joint team from the Shanghai Artificial Intelligence Laboratory and Zhejiang University proposed Agentic Science, aiming to build an AI system that can autonomously complete a closed-loop scientific research process.

(iii) Autonomous Laboratories Accelerate the Industrialization, Scaling, and Platformization of Scientific Intelligence

AI and robotics technologies are upgrading traditional “workshop-style” laboratories that rely on manual trial and error into automated, high-throughput, and closed-loop “scientific factories,” which are interconnected to form platforms serving the entire scientific research ecosystem.

Countries around the world attach great importance to the research and development of autonomous laboratories. Many research universities and national laboratories in the United States, such as MIT, have built autonomous laboratories. The Materials Innovation Factory (MIF) at the University of Liverpool in the UK is one of the most advanced autonomous laboratories in Europe. The IKTOS laboratory in France, Atinary SDLabs in Switzerland, and the FULL-MAP project in Germany are also quite capable autonomous laboratories, continuously contributing in fields such as chemistry and new materials. At the same time, international start-ups such as Lila Sciences and Periodic Labs, which have recently secured hundreds of millions of dollars in financing, are also targeting this field. Meanwhile, the United States recently launched the “Genesis Mission,” which lists advanced manufacturing technologies as one of its key technological breakthroughs, with one of its main goals being to accelerate the creation of new-generation scientific research infrastructure such as autonomous laboratories and improve the efficiency of AI-driven scientific discovery and industrial application transformation. This plan further integrates scientific research computing power, AI basic models, related datasets, and autonomous laboratory systems into a science and security platform as scientific intelligence infrastructure.

The construction of autonomous laboratories and scientific intelligence platforms in China is in full swing. The AI + robotics platform of Jingtai Technology has become its core competitiveness. The “ChemBrain intelligent agent + ChemBody robot” of the Chinese Academy of Sciences and the Uni-Lab-OS intelligent operating system of the Beijing Academy of Artificial Intelligence in Science are all aimed at accelerating the research, development, and promotion of domestic autonomous laboratories. The Panshi Scientific Basic Large Model developed by the Chinese Academy of Sciences is also an important practice of domestic scientific intelligence platforms. The platform can manage various resources such as data and models and schedule various scientific research tools. It has already been applied in fields such as life sciences, high-energy physics, and mechanics research.

AI for Science, Science for Humanity

In the next few years, the evolution speed of scientific intelligence technology and the conversion efficiency of application value will further improve with the continuous increase in the capabilities of AI basic large models, the continuous maturity and scaling of robotics technologies. The scientific intelligence research paradigm will further mature, and the scientific research ecosystem will also undergo reconstruction and upgrading. More major discoveries will emerge from AI-driven scientific research. Sam Altman predicted at Sequoia Capital’s AI Summit this year that AI large models will make scientific discoveries close to the level of the theory of relativity by 2028.

However, while technology is updating and iterating at a high speed, we cannot ignore the improvement of human beings as the main body of scientific discovery in terms of original scientific research capabilities and the renewal of scientific ethics and responsibilities. Scientists should always be the yardstick of scientific intelligence, ensuring that AI becomes a promoter of human technological evolution and a guardian of the continuation of human civilization.