An 'artificial intelligence (AI) laboratory' led by AI scientists, instead of humans, is drawing attention in the scientific community for its achievements in developing a treatment for COVID-19. The research process underwent rigorous steps typically seen in actual laboratories, yet the speed was overwhelmingly fast, completing treatment design in just a few days.
James Zou, a professor at Stanford University in the United States, and John Pak, a researcher at the Chan Zuckerberg Biohub, noted on the 30th, 'We have successfully developed an AI-based virtual laboratory that quickly solves complex scientific problems such as the design of COVID-19 treatments.' The research results were published in the international journal 'Nature' on the same day.
The virtual laboratory was centered around an AI serving as the chief researcher, with various AI scientists collaborating in different areas of expertise. AI acting as immunologists, machine learning specialists, and computational biologists formed teams, along with AI serving as overseers to critique ideas or warn of mistakes.
The AIs conducted brainstorming sessions like humans. The difference is that meetings can last mere seconds if brief or a few minutes at most, and the same AIs could participate in multiple meetings simultaneously. Professor Zou explained, 'Good research comes from collaboration among researchers from different fields,' and added, 'AI transcends barriers that obstruct collaboration, allowing for faster and more creative research.'
AI is not just a chatbot that answers questions. It requests necessary tools, conducts data analysis independently, and adjusts experimental designs on its own. Some AIs have also requested specific software they needed.
In actual experiments, the task of developing new treatments capable of responding to variants of the novel coronavirus (COVID-19) was proposed. However, ideas that required excessive expenses, making verification in real laboratories challenging, were excluded.
The AI laboratory proposed designs for 92 new nanobodies in just a few days, based on smaller nanobodies compared to existing antibodies. Nanobodies are small antibodies found in camelids such as llamas, measuring only a quarter the length of human antibodies and weighing merely one-tenth. Their small size and simple structure make design and computer experiments easy and precise.
Antibodies attach to the spike protein, which is the part of the virus that binds to human cells, blocking human infection. These are known as neutralizing antibodies. Viruses bound by neutralizing antibodies are attacked and destroyed by other immune cells.
Vaccines work on the principle of inducing the human body to produce neutralizing antibodies by exposing it to weakened forms of the virus or parts of it. Antibody therapies artificially introduce neutralizing antibodies into the body to prevent mild COVID-19 patients from deteriorating into severe cases. The designs of these nanobodies leveraged AI systems including Meta's ESM, Google DeepMind's AlphaFold Multimer, and Rosé from Washington University, which analyze and predict protein structures.
As a result of synthesizing the nanobodies proposed by the AI in their laboratory, two were found to strongly bind to the coronavirus. The binding strength of an antibody to a virus indicates how tightly the antibody adheres to the virus and is a key factor determining neutralizing efficacy.
In contrast, the two nanobodies showed little reaction to other substances, indicating low side effects. They reacted strongly not only to recently emerged variant viruses but also to the early COVID-19 virus identified five years ago, confirming the potential for universal treatment.
The research team is currently relaying the experimental data back to the AI virtual laboratory to refine molecular designs with greater precision. Professor Zou stated, 'The datasets collected in the fields of biology and medicine are incredibly complex,' adding, 'I find it truly fascinating that AI can often produce results that surpass what researchers have published.'
References
Nature (2025), DOI: https://doi.org/10.1038/s41586-025-09442-9