BioMap's Chief AI Scientist Li Ziqing on How AI Agents Are Driving a Paradigm Shift in Scientific Research: AI as Both Tool and Partner

20The 2024 Nobel Prizes in Physics and Chemistry Have Been Awarded to the Fields of Artificial Intelligence and AI Life Sciences, this milestone event announces to the world: we are in the midst of a scientific research paradigm revolution led by AI.

As of now,The exploration of life sciences has fully entered the era of large models.: Relying on massive data and enormous computing power for training and optimization, large models fully demonstrate their advantages in accuracy, efficiency, transferability, and emergence, pushing the boundaries of human understanding of the complexity of life systems in an unprecedented way.

The revolution of large models in scientific research practice goes far beyond the improvement of algorithm performance; more importantly, it has given rise to a new generation of infrastructure and platform systems, driving scientific discovery from single-model breakthroughs towards full-process intelligent closed loops, enablingHighly complex and large-scale scientific research tasks can achieve autonomous decision-making, dynamic optimization, and continuous evolution.。

As a pioneer in global life science large models, BioMapFull-Modal Biomass Large Model to be Released in October 2024xTrimo V3, to210 billion parameters refresh the record for the world's largest foundational AI model in the life sciences. Driven by large models, BioMap has constructedCoverageInformation Collection-Bio Insights - Intelligent ExperimentTheEntire ProcessAI Generative Discovery System, helping life science customers and partners improve R&D efficiency and accelerate business closure.

It is reported that,BioMap willApril 25th Meeting"Intelligent Evolution, Discover the Future" Generative Discovery System Launch Event, which can intelligently invoke self-developed core tools and external resources. Users can drive it without complex operations.The full process of "Design-Build-Test-Learn," and through the sharing and co-construction of knowledge and models, form a dynamic, open, and win-win intelligent scientific and technological innovation ecosystem to accelerate breakthroughs in the entire life sciences field.

Register Now | BioMap Life Science Generative Discovery System Launch Event

Recently, Zhiyao Bureau interviewedChief Scientist of BioMap (AI Large Model) Professor Ziqing Li, as a world-renownedAI Scholar, Professor Li guides and leads the research, development, and application of multiple large model projects within the company, and participates in the planning and execution of the company's overall technology strategy.

In this interview, we willIn-depth exchanges were held on cutting-edge innovations and application implementations of AI for Life Science. Standing at the historical juncture where AI is profoundly transforming life science discoveries, the exploration trajectory, vision, and mission of one individual and one company have been clearly presented.

Chief Scientist of BioMap (AI Large Model) Professor Ziqing Li

Li Ziqing (Professor Stan Z. Li is a world-renowned AI scholar, IEEE Fellow, and IPAR Fellow. He has published over 500 papers, with more than 76,000 citations, and an H-Index of 153.And inIn the 2024 World Scientist and University Rankings, he ranked first globally in the "AI for Science" field. He has served as an associate editor of the top artificial intelligence journal, IEEE Transactions on Pattern Analysis and Machine Intelligence (IEEE T-PAMI), among other important academic positions. He has long been active on the front lines of top international AI academia and enjoys widespread recognition in both the global academic and industrial communities.

Professor LiSince 1991, he has served at Nanyang Technological University until becoming a tenured associate professor, and in 2000 joined Microsoft Research Asia as a Lead Researcher, accumulating extensive experience in academic research and industry. He invented the world's first real-time facial recognition system. Since 2004, he has served as a senior researcher at the National Laboratory of Pattern Recognition, Chinese Academy of Sciences, leading the development of more than ten major national projects. Since 2019, he joined Westlake University as a Chair Professor of Artificial Intelligence, leading the work of the AI Research and Innovation Lab at Westlake University, and has devoted significant effort to the AI + life sciences field, achieving a series of groundbreaking academic results. Professor Li, as the principal investigator and chief scientist, has led two major projects under the national "New Generation Artificial Intelligence" initiative (AI + protein computation, drug design), and one key project from the National Natural Science Foundation (AI + multi-omics analysis), becoming a frontier explorer in the field of AI + life sciences.

Q: What was the opportunity that led you to shift from computer vision (such as facial recognition) to the AI + life sciences field? What are the commonalities and differences in core methodologies between the two fields?

Li Ziqing:I have been working on computer vision (Computer Vision), especially in the research and development of facial recognition technology. During his time at Microsoft Research, he successfully developed the world's first real-time facial recognition system, Eye-CU, which was personally demonstrated and recommended by Mr. Bill Gates in a CNN interview.

During his tenure at the Chinese Academy of Sciences, he led a team to apply multimodal face recognition systems and intelligent video surveillance solutions to multiple national security projects, includingIn 2005, I designed and built the Shenzhen Luohu-Hong Kong automatic customs clearance system, and in 2008, the security system for the Beijing Olympics and the 2010 Shanghai World Expo, among other pioneering innovative applications. Later, with the successful application of deep learning and the vigorous development of related AI enterprises in China, facial recognition became a mature industry. I realized that my mission in the field of facial recognition had been accomplished.

In 2019, I joined Westlake University as a Chair Professor of Artificial Intelligence. Life science is one of the advantageous disciplines at Westlake University, providing me with an opportunity for transformation. Through collaboration with Principal Investigators (PIs) in life sciences, I began to venture into proteomics research.Although the research field has shifted from face recognition to life sciences, the underlying methods remain mathematics, pattern recognition, and machine learning.Based on the feature space mapping model constructed using deep neural networks, we have successfully developed a deep manifold transformation from high-dimensional data space to representation space. This core technology has been applied to multiple cutting-edge fields such as early cancer diagnosis, protein modeling, and single-cell analysis.

Q: Regarding the research on AI for Life Science, what are the main areas you are currently focusing on? How does BioMap's "Life Science Foundation Large Model" align with your research vision?

Li Ziqing:In the pastOver the past five years, I have led the team from scratch, starting with applied research in proteomics, gradually extending to protein structure and function design, and then expanding to modeling of the central dogma of biology and target drug development.TherebyBuilt a relatively complete research system from DNA, RNA, to proteins and drug design., which belong to the AI + molecular biology level.This research path, from molecular mechanisms to application implementation, is highly aligned with BioMap's strategic direction.

In the next few years, I will expand my research to the AI + cell biology level., leveraging AI and big data, to buildA large model of cell base that characterizes cellular operating mechanisms, cell differentiation, and cell fate regulation, empowering research on cellular mechanisms and their applications in life sciences, healthcare, and synthetic biology.

BioMap is committed toThe research and industry of AI large models in life sciences also include two levels: AI molecular biology and AI cell biology. From model research to application implementation, this strategic direction aligns closely with my research interests.

At the implementation level, each has its own focus. My laboratory focuses on exploring cutting-edge methods, while BioMap is committed to the scaled, engineering validation and industrial application of large model methods.

Q: What are the key challenges currently faced by AI for Life Science? How will you and BioMap further address these issues?

Li Ziqing：The deep integration of AI and life sciences across disciplines is one of the key points for achieving breakthroughs.

InAlphafold 2The breakthrough of AlphaFold 2, for example, is backed by DeepMind's interdisciplinary team, including experts in molecular dynamics, biologists, chemists, AI scientists, and engineers. The close collaboration and exchange of ideas within such a cross-disciplinary team led to the remarkable achievements of AlphaFold 2 and also initiated a new paradigm for AI for Science research.

Another example is the collaboration between Stanford University, Arc Institute, Nvidia, and other institutions.Evo 2To develop a capable large model in the life sciences, it is necessary to deeply integrate AI with life sciences and embed the intrinsic biological rules at all levels into the model, rather than simply applying an AI framework.

Another key point is the availability of biological big data,Behind Alphafold is the PDB (Protein Data Bank) as its data foundation, which at the time contained around 200,000 protein structures. Without such data on protein sequences and structures, there would be no Alphafold.The reason why current AI cannot well solve many problems in life sciences is that the biotechnology field has not yet developed appropriate and sufficient testing technologies to support effective AI modeling.

In-depth interdisciplinary collaboration, sufficient data, and of course, computing power, areThe Necessary Conditions for AI for Science to Achieve Breakthroughs.

Q: In the field of protein research, you and your team have successively launched models such as PiFold and FoldToken series, which have demonstrated more efficient advantages compared to similar models. What is the secret behind this?

Li Ziqing:The team in my lab is a very young and creative group, mostly from computer science backgrounds, but also including talents from fundamental disciplines like mathematics and physics. They are highly intelligent, capable, and courageous in pursuing progress. New students, under the guidance of their seniors, learn by working alongside them.SOTA accumulates fundamental capabilities, and then carries out paradigm-shifting work in cutting-edge fields.

InIn PiFold, we conducted a comprehensive analysis of model design at various levels, ultimately simplifying the approach to introduce the first non-autoregressive sequence design graph model, achieving dual breakthroughs in efficiency and accuracy; in FoldToken, we performed a detailed analysis and improvement of basic vector quantization methods, proposing the first tokenization-based protein sequence-structure modeling method.

We believe that innovation in fundamental approaches is the most important. Only by making breakthroughs in fundamental methods can progress be achieved in various fields. We also hope to promote advances in protein research through this work.

As the architect of the laboratory,Will intentionally avoid crowded tracks, but to seek more novel and morePromising directions to explore, which also fits the characteristics of Westlake University's "high starting point, small and precise, research-oriented"办学特点. Cutting-edge achievements can be scaled up through BioMap to become a part of large models.

Q: You previously proposed the idea that "all biomolecules can be tokenized." How should this be understood? Compared to large language models, what are the unique aspects of data construction and training paradigms for large life science models? What efforts has BioMap made?

Li Ziqing:Sequences in biological data (such as protein sequences) are naturally suitable forToken-based representation, rather than sequential data (such as protein structures, images), can be converted into discrete tokens through vector quantization.

The mathematical and physical principles behind this process are：Compared with natural language, life science data is a higher-dimensional data.And continuous space (such asThe representation of n-dimensional vectors) contains a large amount of information redundancy, while tokenization compresses information through discretization, retaining only key patterns. This may align with the discrete nature of physical quantization and also suppress noise in the data.

Another reason is that biological moleculesAfter tokenization, it can be adapted to widely used general architectures like Transformer, making modeling convenient. Of course, as mentioned earlier, it is necessary to cleverly utilize the biological rule constraints of the data.

In the pastOver the more than four years, BioMap has been dedicated to the construction of data graphs based on raw data, algorithmic innovation for biological language and different modalities, the establishment of high-throughput experimental systems, and the accumulation of a large amount of self-produced data. The results have ultimately been validated across various application scenarios such as drug design, target discovery, and bio-manufacturing.

BioMap and I are recently working on a project, which is toMolecules such as DNA, RNA, and proteins can be deeply integrated through the principles of the central dogma and embedded into the modeling process., we believe this can improve the quality of large models and generate significant industry value.

Q: In your opinion, in which fields will large multimodal biological models covering proteins, DNA, RNA, etc. be first implemented and truly change our lives?

Li Ziqing:The advantage of large language models lies in their ability to extend multi-dimensional downstream tasks.In the pharmaceutical field,The xTrimo platform has reached SOTA level in over 200 task models across application scenarios such as AI target discovery, protein design and generation, life science tools, and disease mechanism research.Have supported customers in achievingMore than 10 validated antibodies, over 10 innovative target authorizations, and other breakthrough achievements, serving more than 400 users globally, generating significant value at the industry level.

In the field of biomanufacturing,xTrimo can empower processes such as strain modification, enzyme design, and process fermentation.。The industrialization projects we are actually advancing mainly focus on industrial application scenarios, covering fields such as chemical raw material production, feed processing, and environmental protection. From the perspective of commercialization paths, the initial breakthroughs will prioritize high value-added pharmaceutical intermediates and basic chemical raw materials, which have clear market demand and high technical feasibility.

Q: You recently gave a presentation on virtual cells, a technology that Demis Hassabis, Nobel Prize winner in Chemistry, has described as "revolutionizing biological research entirely." What work have you and BioMap done in this area?

Li Ziqing:Currently, I am working onAI Cell Research Has Two Major Directions: One is Related to Life Sciences, and the Other is Synthetic Biology. Both Areas Have Significant Social Implications.

We are building a name called"5M" Multi-Dimensional Research Framework，5M, or 5-Multi, includesMulti-omics, Multi-modal, Multi-perturbation, Multi-scale, Multi-task，Aiming atBased on the "5M" data, a model of cell states and their changes over time and space is constructed. As I mentioned earlier, the development of AI life sciences relies on breakthroughs in biochemical detection technologies, particularly the coordinated development of sequencing and imaging technologies. The path for AI virtual cells is still long.

Compared with life sciences, synthetic biology is easier to implement. I focus on the industrial end of synthetic biology with BioMap.Hope to analyze the principles of single-cell microorganisms and apply them to biomanufacturing., including how to design, transform, and optimize microorganism-enabled biomanufacturing, and how to optimize the process to significantly improve cell production efficiency.

Q: Not long ago, BioMap announced the open-source release of xTrimoPGLM, the world's first large protein model with 100 billion parameters. What do you think is the impact on industry development?

Li Ziqing:We hope to achieve through open sourcexTrimoPGLM, Promote the EntireThe Development of AI + Life Science. Currently, xTrimoPGLM has achieved comprehensive processing capabilities for tasks such as protein structure prediction, functional analysis, and sequence generation. It has reached an internationally leading level in the fields of antibody sequence generation and complex structure prediction. In April 2025, xTrimoPGLM will also be featured in a top journal.《Nature Methods》。

Secondly, it lowers the industry threshold and accelerates innovation transformation. After the model is open-sourced,Researchers can directly fine-tune for vertical tasks such as enzyme stability prediction and affinity analysis, significantly reducing the data and computational power costs required for training large models from scratch, providing a low-barrier R&D path for small and medium-sized enterprises.We also hope to promote industry standardization by opening up the ecosystem.Taking the Model Hub as an example, the platform has integrated dozens of vertical models, and is expected to attract more industry contributors to jointly build an open ecosystem in the future.

In simple terms, this open source not only provides world-leading proteinAI tools, through platform construction, are reconstructing the R&D ecosystem, driving life sciences from single-point breakthroughs towards systematic innovation, and are expected to bring exponential improvements in R&D efficiency and fundamental optimization of cost structures to the industry.

Q: In the AI field, Agent (intelligent agent) has recently become a hot topic, and the industry has also designated 2025 as the "Year One of Intelligent Agents." You and BioMap have done quite a lot of work in this area. Could you share some related progress?

Li Ziqing:AI Agent technology is reshaping the global industrial landscape in a disruptive manner.BioMap is set to release its generative discovery system at the end of April, reconstructing the technological foundation through a multi-agent system.Based onThe xTrimo multimodal large model with 210 billion parameters enables agents to autonomously invoke self-developed core tools and external resources, achieving "deep research" capabilities that surpass traditional automation.

This technological breakthrough has enabledAI Upgraded from a Single Execution Tool to an "Intelligent Research Partner" with Proactive Reasoning Abilities, completing complex task collaboration in target discovery, molecule generation, and other aspects, marking a significant advancement in the life sciences field.The paradigm leap of AI applications from auxiliary tools to research subjects.

In terms of interactive scenario innovation, BioMap reconstructs the scientific research workflow through natural language dialogue. By building"Dialogue as Research" Agent Interaction Interface,Researchers can drive the full-process experimental closed loop covering design, build, test, and learn without mastering professional programming skills.This low-threshold human-machine collaboration model liberates researchers from repetitive tasks, allowing them to focus more on scientific hypotheses and innovative breakthroughs, reshaping the human-machine collaboration ecosystem in life science R&D. 

At the ecosystem construction level, enterprises are committed to building an open network ecology for intelligent agents. By sharing models, data, and knowledge bases, they connect multi-party resources from industry, academia, and research to form a dynamic collaboration network. This open architecture not only accelerates technological iteration and knowledge accumulation but also creates scalable value for niche fields such as drug discovery and synthetic biology through the collaborative computing capabilities of intelligent agents, driving the entire life sciences industry towards intelligence and platformization.

Q: Some studies at home and abroad have shown that generative AI can effectively help accelerate the process of scientific discovery, giving rise to the concept of "AI scientists." What is your take on this? In what ways will AI bring about a transformation in scientific discovery?

Li Ziqing:GenerativeAI is driving a revolution in scientific research paradigms,“"AI Scientist" enhances the efficiency of scientific exploration to an unprecedented height by integrating full-process capabilities such as literature analysis, hypothesis generation, experimental design, data validation, and paper writing., while also sparking in-depth academic reflections on the technical potential and ethical risks.

Currently"AI Scientists" Still Face Multiple Bottlenecks: First, insufficient multimodal capabilities, particularly relying on human intervention in visual information processing and experimental operation stages; second, limited logical reasoning ability, such as large language models often making errors in numerical comparisons; third, the evaluation system is not yet mature, with the explainability and transparency of AI-generated conclusions needing urgent improvement. Ethical risks cannot be ignored either—automated paper production may exacerbate academic bubbles, while the misuse of technologies in areas like biosafety requires global regulatory collaboration.

An article titled "Empowering biomedical discovery with AI agents" in Cell delves into how AI agents accelerate breakthroughs in biomedical research and play a key role in collaborative efforts with researchers. The article outlines the development of AI agents across four levels:

The First Layer——AI is only used as a tool, such as AlphaFold for predicting the three-dimensional structure of proteins;

The Second Layer——AI agents complete specific tasks under the guidance of researchers, such as performing bioinformatics analysis in genome-wide association studies (GWAS);

The Third Layer——AI agents emerge as "partners" for researchers, capable of participating in hypothesis generation and experimental planning. For instance, AI agents can automatically propose hypotheses about certain genes being related to specific diseases based on existing genetic data and design experiments to validate these hypotheses. At this point, AI agents not only execute the instructions of researchers but also provide suggestions for improving experimental protocols, continuously adjusting the research direction based on experimental results, becoming an important partner in scientific research;

The Fourth Layer——AI agents are envisioned as "AI scientists" with the ability to make independent scientific discoveries. They can autonomously propose new scientific hypotheses based on existing knowledge and independently complete experimental verification. These AI agents are not merely tools or assistants but research partners that can work alongside researchers. At this point, AI agents need to possess a high level of learning and reasoning capabilities to make reasonable judgments when facing complexity and uncertainty.

The academic community generally believes that,AI Will Propel Scientific Research into the "Fifth Paradigm". AI not only accelerates data processing but also, through knowledge graph construction and interdisciplinary associations, gives rise to entirely new scientific hypotheses. Open scientific resources will become crucial for innovation, while the core role of human scientists will shift towards strategic planning and creativity stimulation.

In this scientific research revolution of human-machine collaboration,"AI Scientist" is both a tool and a partner.It cannot replace human intuition and inspiration, but through supercomputing power and pattern recognition, it can free scientists from repetitive tasks, allowing them to focus on more fundamental explorations. With the evolution of multimodal models and systems, a new era of more autonomous and creative scientific discovery is accelerating.

Q: In which areas is BioMap currently focusing its research and investment in large-scale life science models? What is the vision for the iteration of these large models in the next 1-3 years?

Li Ziqing:As a life scienceAs a pioneer in large AI models, BioMap has been continuously deepening the layout and innovation of large life science models in recent years. At the technical core, BioMap has built xTrimo V3, the world's first large life science model covering seven modalities including proteins, DNA, RNA, cells, and small molecules.

In the next three years, BioMap plans to further expand the model parameters and add new modalities such as metabolomics and microbiome, achieving full-chain modeling from molecules to ecosystems.The cross-scale modeling technology we are developing, such as combining cell interaction models with clinical data to predict drug side effects, may redefine the paradigm of drug research and development.  

Application scenarios will expand in depth into fields such as synthetic biology and cell gene therapy.The company is developing a foundation model based on single-cell transcriptomics.scFoundation's "Cell-Level Life Simulator"，Previous achievements have been selected."Top Ten Advances in Chinese Bioinformatics in 2024." In the field of biomanufacturing, plans are in place to optimize industrial strain modification and improve the efficiency of pilot-scale fermentation processes through AI.

In addition, the company will continue to implement its open source strategy.Following the open sourceAfter xTrimoPGLM, the company plans to release larger-scale model capabilities, enabling small and medium-sized institutions to use AI tools at a low cost.At the same time, we are accelerating the construction of the global developer community, and we hope to become the core provider of bio-computing infrastructure. 

BioMap“Basic Large Model+Vertical Scenarios +Open Ecosystem"Three-dimensional strategy, striving to lead China to occupy the commanding heights in the global bio-computing competition.

Note:This interview has been edited and approved by the interviewee. We also welcome readers to interact through comments and share your thoughts on this interview. For more information about BioMap's generative discovery system, please stay tuned this month.The press conference on the 25th will be live-streamed by Zhi Medicine Bureau.

Recommended Reading