AI News

Zhejiang Lab and BGI Release Genos, a 100-Billion-Parameter Human Genome Foundation Model

People's DailyUpdated Nov 13 05:51

AI summary

Genos, the world's first 100-billion-parameter human genome foundation model jointly developed by Zhejiang Lab and BGI Life Science Research Institute, was released in October 2025. Over the past year, it has collaborated with nearly 40 hospitals, universities and other institutions, attempting to use AI to interpret genetic variants and assist in disease diagnosis research and drug target discovery. The model's training data brings together 636 "telomere-to-telomere"-level human genomes, and through a mixture-of-experts architecture it achieves precise modeling of ultra-long context sequences of one million bases as well as single bases. The team has about 130 members with an average age of 30, and subsequently also developed the OneGenome model, which integrates genome sequences with biomedical knowledge.

Body · AI translation

General Secretary Xi Jinping has emphasized that young talent are the new force in basic research, and that many major achievements have been made by young people.

In the opening year of the 15th Five-Year Plan, basic research achievements have emerged in China, and behind this is the innovation, enterprise and pursuit of excellence of a vast number of young talents, who contribute their youthful strength to advancing Chinese modernization.

How can young talent take the lead and play the main role on the front line of basic research, and how can conditions be created for them to fully display their talents? Starting today, the Science and Technology section launches the series "We Have Confidence in Taking the Lead and Playing the Main Role," offering in-depth discussion from different angles.

— The Editors

In October 2025, Genos, the world's first hundred-billion-parameter human genome foundation model jointly developed by Zhejiang Lab and BGI Life Science Research Institute, was officially released. Over the past year, the team has carried out cooperation with nearly 40 hospitals, universities and other institutions, attempting to use AI to interpret genetic variants and find new clues for disease diagnosis research and drug target discovery.

Behind this frontier exploration is a young research team. The joint R&D team has about 130 members, with an average age of only 30. Coming from different professional backgrounds such as computer science and life sciences, they are trying to enable AI to learn from massive genomic data and search for the codes of life that humanity has not yet understood.

Exploring the unknown within billions of base pairs

The human genome contains billions of base pairs and records a person's genetic information. In 2003, the Human Genome Project was declared complete, and scientists for the first time had a high-quality reference sequence of the human genome. Since then, with continuous technological progress, massive genomic data have accumulated at an accelerating pace. However, how to understand the genetic information in genome sequences remains a major challenge facing life science research.

Liu Shiping, a chief scientist in the field at BGI Life Science Research Institute, said that so far scientists still have many gaps in their understanding of the human genome. For example, if a certain position in the genome changes, will it affect health? How do different genes influence one another? These mysteries remain to be discovered. "At present, only more than 20,000 protein-coding genes are known, and even for these known genes, their specific roles and relationships with diseases have far from been thoroughly studied," Liu Shiping said.

The rapid development of AI has brought new ideas: since AI models can learn language patterns from massive amounts of text, can they also learn from large numbers of genome sequences, allowing raw scientific data to directly enter foundation models and helping scientists find connections that were difficult to discover in the past?

At that time, Zhejiang Lab was exploring ways to bring scientific data into foundation models, and life sciences was one of the key areas; BGI Life Science Research Institute had accumulated data and experience in genome research, and the two sides hit it off immediately. Liu Shiping recalled that after the two sides met for the first time and exchanged views for about an hour, they decided to form a joint research team to train a foundation model to tackle the "hard nut" of the human genome and look for scientific clues related to disease and health.

Liu Junchen, born in 1995 and a young core researcher at the Life Science Computing Research Center of Zhejiang Lab, was one of the earliest team members to join. In 2023, after receiving his doctorate in electronic science and technology from Beijing University of Posts and Telecommunications, he joined Zhejiang Lab. Liu Junchen, who was originally unfamiliar with biology, had the simplest aspiration when participating in the Genos project: "After learning about many cases of rare diseases related to genes, I thought that if I could use my knowledge of computer science to help solve biological problems, it might spare some families suffering," Liu Junchen said.

Fighting alongside Liu Junchen was Deng Kaiwen, a young core researcher at BGI Life Science Research Institute, also born in 1995. During his school years, he experienced several years of rapid development of large language models. Seeing the continuous progress in AI's ability to process human language, he wondered: since human genome sequences and natural language are both one-dimensional information, could AI also learn to "interpret" human genome sequences? "At the time in the laboratory, I paid great attention to this matter," Deng Kaiwen said. As foreign teams successively carried out related explorations, he increasingly wanted to try it himself, but limited by computing resources, it was difficult to organize large-scale foundation model training.

After graduating in July 2025, Deng Kaiwen joined BGI Life Science Research Institute. Hearing that the Genos project was being carried out at Zhejiang Lab, having been on the job for less than a week, he hurried to Hangzhou, Zhejiang, to participate in the development of the initial model. "I had always wanted to try training a genome foundation model myself and see what effect it could achieve," Deng Kaiwen said.

Liu Junchen brought AI technology into life sciences, while Deng Kaiwen brought biological questions in search of new methods. In this way, a group of young researchers from different fields such as computer science and life sciences came to the same research topic, trying to enable AI to understand the massive amounts of human genetic sequence information.

Getting the "languages" of two disciplines to speak together

The first challenge in the joint research was how to communicate using the languages of different disciplines.

Liu Junchen still remembers a discussion about "sequence length." Facing the same term, colleagues with a computer science background were talking about the "token" length processed by the model, while colleagues with a life science background were thinking about the base length of DNA (deoxyribonucleic acid). Only after the discussion did everyone realize that they had not been talking about the same thing at all.

Similar "little episodes" are not uncommon in this interdisciplinary team: computer researchers are familiar with models and algorithms, but they need to learn the laws of life in the genome; life science researchers understand the functions and roles of genes, but they need to learn how to train large models. In the first month of the joint effort, young people from different specialties sat together to fill in their basic knowledge, and Ma Jing, deputy director of the Life Science Computing Research Center at Zhejiang Lab, was responsible for leading the team in evaluating the effects of different approaches.

Once the communication barrier was crossed, greater scientific challenges followed. Previously, some genome models were trained mainly on reference genomes, and limited reference data made it difficult to fully represent genetic differences among different populations. To make AI truly understand the human genome, it must be exposed to more diverse genetic information.

Just as the team was pondering data sources, the Human Pangenome Reference Consortium released a batch of high-quality genome data covering different regions and populations. "This was a crucial moment. This open scientific data made the team firm up its research direction, take human genetic diversity as an important consideration, and let the model learn the genomic information of different individuals, providing a more reliable foundation for studying the links between genetic variation and disease," Ma Jing recalled.

With data in hand, how to train the model better remained a challenge. The human genome sequence appears to consist of only four bases represented by letters, but the internal relationships are extremely complex. A change in one base may have an important impact on life activities, and two sequences far apart may also jointly affect the expression of a gene. How to enable the model to analyze longer sequences without missing subtle changes became a difficult problem in development.

At the time, Deng Kaiwen participated in model evaluation, comparing the performance of different versions of the model in specific research tasks from a biological perspective. The team found that training data covering more individual genomes helped improve the model's performance in identifying human-related variants, providing a basis for the team to subsequently expand the data scale.

At Zhejiang Lab, scientific discussions like this can happen at any time. At 9 a.m., 10 team members gathered around a whiteboard to discuss optimization paths for a new version of the genome foundation model. After several rounds of exchange, they had solutions for the programming problems that had emerged the night before.

"Being able to discuss problems face to face at any time is very convenient," Liu Junchen said. He and Deng Kaiwen work in the same office, and young people with different academic backgrounds have gradually learned to think about problems from each other's perspective.

This close cooperation also benefits from innovation in the way research is organized. Chen Gaoxiang, deputy director of the Life Science Computing Research Center at Zhejiang Lab, introduced that the lab's "seed class" mechanism allows external institutions and relevant lab personnel to work together in one place, achieving "zero-distance" communication in physical space and allowing them to discuss problems and tackle challenges together at any time.

After months of concentrated effort, the team successfully overcame the above challenges. The Genos model, for the first time, brought together 636 human genomes at the "telomere-to-telomere" level from around the world as training data, representing human genetic diversity more comprehensively; at the same time, through a mixture-of-experts architecture, it achieved precise modeling of ultra-long context sequences of one million bases and single bases, like reading a monumental work: it can remember details without missing a word while also grasping the grand narrative of the whole.

Starting from the model, continue to pursue new scientific questions

After the initial version of Genos was released, the team still devoted most of its energy to upgrading the model. Liu Shiping admitted that the current model's understanding of the human genome is not yet deep enough, and there is still a long way to go before discovering more credible laws of life. Deng Kaiwen and Liu Junchen now each lead research on one module in the new model, and their work happens to correspond to two types of challenges in the genome.

Deng Kaiwen is working on "spatial relationships." For example, DNA on a computer screen is a string of bases arranged in sequence, but in human cells, DNA exists in a complex three-dimensional structure. Two regions that are far apart in sequence may be close to each other in space and jointly affect the activity of a gene. How can the model identify these complex connections from one-dimensional sequences? He is trying to adjust the model's attention mechanism so that it can learn the connections between distant regions.

Liu Junchen focuses his research on the model's "memory." Existing models cannot yet process a person's complete genome in one go. He hopes to first select the most representative information so that when the model analyzes local DNA sequences, it can also make judgments by combining the individual's overall genetic characteristics.

The two's work often intersects. Whether a new research idea can be implemented computationally and whether it has biological significance requires them to find answers together.

AI's entry into basic research has changed the traditional scientific research paradigm, and how to find truly worthwhile scientific questions among many ideas has become a new test. "AI has now lowered the threshold for us to try new ideas, but which idea can truly promote scientific discovery requires cultivating an 'aesthetic' for such judgment," Deng Kaiwen said.

Liu Junchen feels this deeply. AI is making the pace of research faster and faster, with new papers and new results constantly emerging. He sometimes worries that if his new ideas are not produced quickly, they may soon be achieved first by other teams. But he also constantly reminds himself not to lose the rigor that research should have. "We must use AI to improve efficiency, but we must not, for the sake of convenience, leave loopholes in the research content," Liu Junchen said.

Ma Jing introduced that the team has now further developed the OneGenome model, integrating genome sequences with biomedical knowledge so that more people can conveniently use AI to conduct life science research.

From proposing ideas and training models to pursuing new scientific questions, the team's young people have deeply integrated AI with basic research and continuously expanded their scientific horizons. Among the billions of base pairs in the human genome, there is still too much unknown. They are still persisting in exploration, using AI to read out more raw information and bring scientific research one step closer to the mysteries of life.

Source

People's Daily