China Securities Co.,Ltd.: Dry-wet closed loop builds data barriers; large molecules expected to raise the industry ceiling

China Securities Co.,Ltd. released a research report stating that AI4S is moving from a competition in model capabilities toward an industrialization inflection point, driven by the fourfold resonance of technology, industry, demand, and policy. The value anchor is upgrading from model capabilities to scientific research productivity, data assets, and systematic biological design. Scientific research Agents are upgrading from answering questions to decomposing goals, invoking specialized models, and connecting with experiments, driving AI to leap from a single-point tool to scientific research productivity capable of executing complex R&D tasks. Pay-per-use models such as Science Token open up space for platform-based monetization; the speed at which models generate candidates has outpaced experimental validation capacity, making experimental capacity and feedback cycles the main constraints on R&D efficiency. Those who control real data entry points leverage the dry-wet closed loop to accumulate a data flywheel, and the focus shifts from model performance to experimental throughput, data quality, and feedback efficiency; AI capabilities are extending from small molecules to large molecules such as antibodies, proteins, and nucleic acids, and R&D is upgrading from generating a single candidate to systematic druggability design that integrates druggability optimization, delivery, and in vivo validation, which is expected to open up industrial value space. The main views of China Securities Co.,Ltd. are as follows: Fourfold resonance drives AI4S toward an industrialization inflection point On the technology side, generative biological design is expanding from single-molecule prediction to complete genomes; on the industry side, general-purpose model vendors are extending into wet labs, and MNCs are forming horizontal alliances to jointly build R&D infrastructure; on the demand side, the amplification of candidate supply is driving demand for synthesis and high-throughput validation, with upstream service providers realizing returns first; on the policy side, AI pharma, high-quality datasets, and scientific research infrastructure are being incorporated into systematic planning, with technology, capital, orders, and policy forming a combined force. The industrial value anchor is upgrading from single model capabilities to scientific research productivity, data assets, and systematic biological design, mainly reflected in three aspects. Paradigm evolution: Agents extend the R&D chain, TaaS opens up value space Agents are upgrading from answering questions to decomposing scientific research goals, invoking specialized models, and connecting with real experiments, enabling algorithms, data, and equipment to form a continuous workflow. Science Token packages models, computing power, task orchestration, and Siasun Robot&Automation experiments into standardized services, with customers paying by usage volume, and the platform can convert reusable scientific research capabilities into sustained, high-frequency platform revenue. Deepening barriers: The dry-wet closed loop builds a data flywheel, and wet labs precipitate data assets The speed at which models generate candidates has already outpaced experimental validation capacity, and experimental capacity, cost, and feedback cycles have become the main constraints on R&D efficiency. Tech giants, by building their own wet labs and connecting CROs and automated experimental platforms, control the entry points for real data production; experimental results continuously flow back for training, forming a self-reinforcing flywheel of "designvalidationfeedbackiteration," and the focus of competition shifts from model performance to experimental throughput, data quality, and feedback efficiency. Automated wet labs are expected to become an important scenario for the implementation of physical AI. Boundary expansion: From "selecting molecules" to "building systems," large molecules are expected to raise the industry ceiling Small molecules remain the most mature main battlefield for data foundations, modeling tools, and clinical validation; the development of protein structure prediction, protein language models, and generative biology is driving antibody, protein, and nucleic acid drugs toward prediction-driven and programmable design. Because sequence, structure, druggability, delivery, and in vivo effects are highly coupled, large-molecule AI will upgrade from generating a single candidate to systematic druggability design with multi-objective collaborative optimization. Risk analysis (1) Risk that AI4S technical capabilities and scientific validation fall short of expectations; (2) risk related to high-quality data supply and compliance; (3) risk that the construction of the dry-wet closed loop and commercialization progress fall short of expectations; (4) risk of R&D translation and clinical trial failure; (5) risk of open-source competition, intellectual property, and biosafety regulation.