Mila Tang Jian： Building ChatGPT in the field of life science and developing drugs with generative artificial intelligence.

"I have always been interested in biology. When I was in high school, my teacher told us that the 21st century was the century of biology. When I studied artificial intelligence, I believed that the 21st century was the century of artificial intelligence."

During the scientific lecture "Understanding the Future" hosted by the Future Forum, Tang Jian, an associate professor at the Research Center for Artificial Intelligence in Quebec, Canada (Mila) and a professor of artificial intelligence at the Canadian Institute of Advanced Studies (CIFAR), introduced the application of generative artificial intelligence in life sciences: how to build a "ChatGPT" in life sciences. He said, "Later, I felt that it might be the best choice if I could do something in the field where artificial intelligence and biology intersect."

Today, he is not only studying AI and biology at the same time, but also "left-handed scientific research and right-handed entrepreneurship". He said that he is a person who is willing to accept new things and challenges. From scientific research to entrepreneurship, many things are different, and it is difficult to carry out interdisciplinary cooperation, but this also reflects the value of doing these things.

Figure | Tang Jian (Tang Jian’s personal homepage)

According to reports, Tang Jian University studied geography at the beginning, transferred to the Department of Mathematics in his sophomore year, and successfully applied to the Department of Computer Science in his senior year. In his computer learning career, he first studied traditional machine learning, and in the last year of his doctoral degree, he turned to deep learning.

After graduating from the doctor’s degree, he officially embarked on the road of scientific research and made many achievements in the field of graphic representation learning. As one of the representative scholars in this field, his network representation learning method LINE (large-scale information network embedding) has been widely used, with more than 5,000 references.

"The double revolution of artificial intelligence and biotechnology broke out, and we are in the best era of doing research."

In 2018,Tang JianStart the research on the intersection of artificial intelligence and biology. At that time, he applied the graph representation learning technology to the research in the field of drug discovery, and did a lot of work in the prediction of three-dimensional structure of small molecules and the design of macromolecular proteins.

But a few years ago, only a few people devoted themselves to the research in this direction. In the past two years, more and more investment has been injected into this field. Tang Jian believes that the reason for this change is that artificial intelligence has brought great breakthroughs to the structural prediction and biological field in protein, and the spread of the COVID-19 epidemic has also aroused people’s focus on health and biomedicine.

"Now we are in the best era, because we are experiencing the dual technological revolution of artificial intelligence and biotechnology." Tang Jian said, "Therefore, we have seen great opportunities for generative artificial intelligence in the field of drug discovery, especially protein design."

Based on his accumulation in scientific research, he devoted himself to entrepreneurship and set up the artificial intelligence-driven biopharmaceutical company Baiao Geometry in 2022.

Build an AI macromolecular drug design platform to provide innovative drugs to save patients’ lives.

existTang JianIt seems that perhaps it is the entrepreneurial idea that has sprouted since childhood and the character of constantly seeking self-breakthrough that drives him to embark on the road of entrepreneurship.

The Baiao Geometry Company, founded by him, aims to develop a programmable protein by integrating the technologies of generative artificial intelligence and geometric deep learning, so as to overcome the problems of long cycle, high cost and low success rate of traditional macromolecular pharmacy and provide innovative drugs to save patients’ lives.

Tang Jian is mainly responsible for technical work in this company, and designs antibodies by building artificial intelligence models. As a new diffusion generation model, this model can model the amino acid sequence and three-dimensional structure of protein at the same time, and show the relationship between them, so as to generate protein with specific functions.

At present, the team has set up an AI macromolecular drug design platform, and is building a wet experiment verification platform for Qualcomm macromolecular drugs. At the same time, it also launched TorchProtein, an open source machine learning platform for macromolecules, in conjunction with companies such as NVIDIA, Intel and IBM.

"I hope that the technology I have made can really push several drugs into the clinic, solve more complicated diseases like cancer and save patients’ lives." Tang Jian said. In addition, he also said that with the development trend of intelligence, digitalization and automation in the biomedical field becoming inevitable, he expected to become a world-famous and international high-tech company in this field.

Based on generative artificial intelligence, build "ChatGPT" in life science field

In addition to starting a business, Tang Jian’s scientific research work at hand is also being carried out simultaneously. Based on generative artificial intelligence, he and his team have carried out research on three-dimensional structure prediction of small molecules and protein design of large molecules in the field of biomedicine.

In the prediction of the three-dimensional structure of small molecules, the team mainly modeled them based on the diffusion generation model. Specifically, a completely random structure is taken as the initial structure, and after several rounds of optimization, it finally converges to a stable protein structure. Each step of the optimization process is also called denoising. In addition, on the basis of expanding the research, the prediction of the complex structure was realized.

In the design of macromolecular proteins, the team proposed to design protein structure and sequence based on diffusion generation model, which not only realized the generation of antibody CDR Loop structure and sequence, but also designed α -helix transmembrane proteins with a specified number.

Not only that, the team also conducted research on antibody design and optimization. For example, it cooperated with the research team from Fudan University. Based on the antibody CR3022 obtained from newly crowned patients, it first designed the sequence with AI model, and then tested it on the wet experimental platform of Qualcomm, and finally successfully found the antibody molecule with strong affinity.

For generative artificial intelligence, its most critical ability is that it can generate new data. Then, if it is used in the biomedical field, it can help to generate a brand-new protein and help people find better drugs.

As a typical generative artificial intelligence model, ChatGPT has spread all over the world since it was introduced at the end of 2022. It can talk and communicate like a human being, and complete tasks such as writing papers and coding. "It is essentially a chat robot." Tang Jian said. As a large-scale pre-training language model, ChatGPT first pre-trains a large number of text and code data on the Internet, and then further optimizes it to enable it to be used for tasks such as dialogue.

The development of biotechnology, such as gene sequencing and gene synthesis, has brought a lot of data to the biomedical field. Therefore, Tang Jian believes that in the near future, people can also use these data to build ChatGPT in the field of life sciences.