At Bund Conference, AI for Science Experts Map Path From Research Assistants to Autonomous Discovery

Wang Jun, a professor at University College London and founder of RL China, said in his opening speech that scientific research proceeds through hypotheses, tests and revisions, and that an important goal of AI for Science is to model that process so agents can participate in the research loop more completely. He distinguished exploration from discovery: exploration addresses known unknowns, while discovery must reach problems that are not yet part of the existing knowledge system or have not even been recognized. Wang proposed a 'big discovery model' direction intended to move AI beyond reasoning over existing knowledge toward finding new questions and knowledge. The key, he said, is a closed loop between the digital and physical worlds. His team has used self-learning and self-evolution methods so agents improve through repeated feedback, and in collaboration with chemists has conducted more than 900 experiments combining computational decisions with real experiments for new material discovery. He said the approach has also shown practical results in proteins, small molecules, chip processes, mining and financial factor mining. Wang also described engineering efforts around edge agents and an 'agent cloud' that connects computing power, data and tools for research tasks.

Zhang Tongyi, an academician of the Chinese Academy of Sciences and director of the Materials Genome Engineering Institute at Shanghai University, focused on the source of scientific data. He said AI's foundation depends not only on mastering existing data but also on the ability to continuously produce new data, which for materials science means real experiments, production processes and observation of natural phenomena. Materials research faces complex physical and chemical processes that are only partly understood; in nanomaterials, for example, stress, composition distribution and diffusion are coupled, and models are often built under simplified conditions. AI in materials research therefore still needs to be combined with mechanism studies, physical simulation and experimental validation. Zhang proposed building 'soft cornerstones,' covering computation, simulation and algorithms, and 'hard cornerstones,' connecting experiments and production, with both becoming more intelligent. AI laboratories, he said, require collaboration among materials specialists, simulation, control and AI, and real-time detection so experimental results can feed back into subsequent decisions. He also noted that generating candidate materials is only part of R&D; whether a material can actually be synthesized remains a real challenge. When experimental data are insufficient, simulation can supplement real experiments, but fitting performance cannot replace judgment about mechanisms. Zhang cited his own research showing that the same dataset can fit different mechanism models well, making it hard to decide which explanation is more reasonable from fitting alone.

Xie Zhi, a researcher at the Guangzhou National Laboratory, discussed progress and limits of AI in life sciences through virtual cell and molecular design research. He said drug R&D difficulties lie not only in molecular optimization efficiency but also in incomplete understanding of disease mechanisms, efficacy and toxicity, and patient-drug matching. In virtual cell work, Xie's team tried to link cell-line drug and gene perturbation data with human disease-related transcriptome data, exploring whether models can transfer from cell lines to organoids and human-related data. He said the team integrated more than 2 million cell-line perturbation datasets and about 500,000 human transcriptome datasets, and observed some transfer ability in predicting transcriptome changes, drug sensitivity and some patient response data. But there were boundaries: when the model was extended to broader disease-drug relationships, inference based only on transcriptome changes performed poorly. The team combined a medical knowledge graph to improve predictions and obtained preliminary validation in animal experiments for some new drug indications. In molecular design, Xie described work on mRNA sequence optimization, including incorporating evolutionary information, combining reinforcement learning for multi-objective optimization, and using automated experiments to supplement data. He proposed that next-generation life science research systems need models that understand biological mechanisms, a discovery engine that proposes and evaluates scientific hypotheses, treatment design capabilities, and feedback from experiments and clinical settings. He also said human participation is still needed for key decisions and that the stability of autonomous research needs improvement.

Qin Gang, co-founder and CEO of Qide Pharmaceutical, described efforts to combine automated experimental platforms with AI decision-making in bioconjugate drug development. Using antibody-drug conjugates as an example, Qin explained the idea of combining targeted delivery with drug payloads. As conjugates expand toward more XDC forms, combinations of targets, payloads and mechanisms of action become richer, creating a vast design and screening space. Qide built the iDiscovery platform, which combines high-throughput preparation with drug screening based on patient-derived organoids. The platform can prepare more than 1,000 samples per day, and with partners it has built an organoid library covering more than 30 indications. In this system, R&D begins with identifying unmet clinical needs, builds models from patient-derived materials, designs drug components with biological information, and uses automated platforms for combination, screening and evaluation. Qin said the high-quality structured data generated by continuous operation can provide a foundation for AI in subsequent drug discovery, forming a 'dry-wet loop' in which computational design and real experiments feed back to each other. As the number of available drug components increases, he said, an important value of AI will be in decision-making: deciding what to prioritize, what to screen, and how to adjust the next round of experiments based on results. The team plans, under the same clinical needs, to let AI models direct the experimental platform in R&D and compare the results with experienced R&D teams, exploring an XDC drug decision laboratory. Qin stressed that the final evaluation must return to the clinic, ultimately extending patient survival and improving quality of life.

Zhang Shuming, chief strategy officer of Deep Principle, introduced a materials R&D path combining deep learning, first-principles calculations and automated experiments. He said industrial customers are shifting their focus from model performance and automation of local tasks to actual delivery: what value is created and whether their problems are solved. Deep Principle's materials R&D agent platform, Mira, adapts to different computational software, experimental equipment and data systems, bringing scattered R&D resources into unified scheduling. Zhang cited a data center coolant project: starting from limited chemical elements, the team used generative models to build about 2.3 million molecular structures, then screened them according to physical property requirements, synthesis steps and raw material costs, ultimately recommending nearly 200 structures for experimental exploration. The project achieved laboratory-stage results in about four and a half months; the related formula has been patented and is undergoing material compatibility testing with downstream server customers. He also shared a collaboration with L'Oreal on formula stability, which expanded from stability optimization to biological activity improvement and new molecule design.

Yang Mengyang, co-founder of Kaiwuji, described the company's layout across the full materials R&D process. He said Kaiwuji hopes to connect data, foundation models, laboratories and factories so AI can participate from material design and formula validation to scaled production. At the model level, the team built a material property prediction model and a generative model that generates candidate structures from target performance. Around these capabilities, the team combined models, patents, data and tools into an agent called System 0 to help researchers access computational capabilities and explore connecting it further to laboratories and production. Yang said the team released the materials dataset Elementa and initially opened 40 million DFT calculation data points to support materials foundation model research. For commercialization, he said China's manufacturing and supply chain conditions provide important support for AI materials R&D. Whether AI for Materials succeeds ultimately depends on whether materials can be synthesized in the physical world, commercialized, and meet customers' actual needs.

Yang Ling, an assistant professor at Peking University, presented the team's research route from code generation and agents to long-horizon tasks. Recursive self-improvement centers on letting a system use feedback to improve itself, and the availability of feedback in different tasks determines the difficulty of the process. Using code generation as an example, Yang said code execution and tests can provide relatively direct and timely feedback. The team explored a framework in which code and unit tests improve together, using tests to check code and then using code execution performance to check test quality. For more complex agent tasks, the team tried to let the task environment and model evolve together and introduced a process reward model to supervise intermediate steps. As task chains lengthen, problems such as goal loss, incomplete memory retrieval and chaotic tool calls become more prominent. To address these difficulties, the team maintained working memory to continuously record task status, final goals and required tools. On this basis, Yang pointed to 'discovery intelligence' as the next stage: AI needs to judge what problems are worth solving and how to evaluate new scientific conjectures. He said scientific discovery is open-ended, often has no fixed answer, and may involve hundreds or even thousands of steps, making it hard to evaluate the whole process by final results alone. Autonomous systems for scientific discovery therefore need to address open-ended evaluation, long-horizon task execution and real experimental validation. Computational proposals must be tested by wet experiments.