Google’s AI Co-Scientist Racks Up Two Wins

Google! What if, instead of determining reminders or bringing restaurant reviews, have you helped break the secrets of biology?

This fun question alludes to a radical vision that is now tested in laboratories. Artificial intelligence systems are not reformulated as digital secretaries, but as scientific partners-a pilot was designed to dream of bold, testable ideas.

The stadium looks revolutionary. But it also makes many Bristle scientists. How much is the true grandmother that the device can evoke? Is it not likely to reconstitute the past instead of detecting something really new?

For several months, immersed the controversy over “artificial intelligence scientists”: noise for hope, and the parrot for discovery. But two new studies offer some of the strongest evidence so far that LLMS models can really generate new scientific ideas, jumping into unclear visions that may require many years of painstaking laboratory work. Both studies are presented by a Google Scientific Researcher, known as the participating teacher of Amnesty International.

Dylan Brasad, a neurosurgery researcher at Northwestern University and an external observer written about The possibility to generate hypotheses participating in artificial intelligence. “You have artificial intelligence agents produce a scientific discovery! It’s very exciting.”

Artificial intelligence takes the drug replacement

In one of these demonstrations is to prove the concept, it leads a team led Gary PigzStanford, AI’s assistant, AI assistant to find medicines already in the market that can be filled for liver fibrosis, is a condition on the scale of organs with a few effective treatments.

It has paid the tool to search for medications directed to the Laginian organizers – proteins that control how genes are operating or stopping them without changing the essential DNA – and AI, after extracting biological medical literature, with three reasonable suggestions. Peltz added candidates on its own, and put all the five medications through a battery of tests on the liver tissue planted in the laboratory.

Two of artificial intelligence choices – but nothing from Peltz – showed fibrosis and even showed signs of enhancing liver regeneration in laboratory tests. Bilz, who published the results on September 14 in the magazine Advanced scienceHopefully, the results will pave the way for a clinical experience of one prominent candidate, cancerous drug fibroostat, in liver fibrosis.

Bacterial ambiguity solution

In the second verification study of health, a team led by microbiologists leads Jose Benadis and Thiago Costa In Import College London, stabbed artificial intelligence with a thorny question about bacterial development. In 2023, the researchers showed that parasitic DNA scraps can spread within the bacterial population by riding horse on the tails of infected viruses. But this mechanism seemed confined to one host type. How, then, did it identical from the DNA surface in completely different types of bacteria?

So they assigned artificial intelligence to solve mystery. They fed their data, background papers, and a pointed question about the hidden mechanism that may explain the jump. Artificial intelligence, after “thinking” and treatment for two days, suggested a handful of solutions – the leading that DNA fragments can extract viral tails not only from its host cell but also from the adjacent bacteria to completing their journey.

It was true incorrect.

What the system cannot know is that Penadés and Costa have already unpublished data that exactly alludes to this mechanism. In fact, Amnesty International has jumped to the same conclusion that it had taken years of work to work on deduction, a rapprochement that amazed the imperial team and a credible loan to the tool.

“I really shocked,” says Penadés, who initially believed that artificial intelligence had penetrated his computer and access to additional data to reach the right result. I reassured him that he did not do so. “Even for those who were not true, thinking was very good,” says Benadis.

Amnesty International Scientific Method

This vocal logic of the imperial team has pushed to explore one of the ideas of runner-up in artificial intelligence-logic that might transmit bacteria the DNA fragments directly to each other. By working with microbial genetics in France, the group is now investigating this possibility, with promising early results. It seems that our initial data indicates that hypothesis [also] “It is correct,” says Costa.

He and Penadés published both the predictions of artificial intelligence And its experimental results in the magazine cell Earlier this month.

It is worth noting that the imperial researchers have also tried many LLMS, not specifically designed for scientific thinking. These included systems from Openai, Anthropic, Deepseek, and the GEMINI 2.0 general purposes of Google. None of the Jack Claude models have preceded the hypotheses that have proven experimental correction.

Google DeepMind, who helped develop the co -coexist, believes that he knows what explains that edge. It refers to the design of the multi -agent system, which appoints the various roles of Amnesty International to create, criticize, refine and rank in the repetitive episodes, all supervised by the “supervisor” who manages goals and resources. Unlike LLM year, it puts ideas in external tools and literature, and strategically prepares a strategic point of view for deeper thinking, and veterinarians’ assumptions through automated championships.

According to Natarajan, academic institutions around the world are now experimenting with the system, with plans to expand access scope – although the company’s “trusted laboratory program” is currently and new applications are not accepted. He says: “It is clear that we see a lot of potential.” “We imagine that every time you will try to solve a new problem, you will use the participating teacher to come on the trip with you.”

A group of participating scholars

Google is not alone in chasing this vision. In July, the computer world Kyle Swanson Description of his colleagues at Stanford University Virtual LaboratoryLLM system that combines thinking steps through biology data collections to suggest new experiences.

Competitors move quickly as well: biomniAnother system led by Stanford, helps to carry out a wide range of research tasks independently in life sciences, while non -profit organizations the future It is building a similar platform. Each of them competes to show that its approach can turn language models into real engines for discovery.

Many spectators have admired, noting that studies provide some of the clearest evidence so far that LLMS can generate ideas worth testing on the bench. “This will make our jobs much easier,” said Rodrigo Ebara Chavez, a microbiologist at Copenhagen University in Denmark, who is studying a kind of bacterial genetic stop that the imperial team explored.

But critics warn that excessive dependence on the hypotheses generated from artificial intelligence in scientific risks risks the creation of a closed episode that recycles old information rather than producing new discoveries.

“We need tools that increase our creativity and critical thinking, not to fill out current information using an alternative language,” Kriti Gaur wrote from Elucidata Science Analysis Company in white paper evaluating Google platform. “Until this” co -scientists of artificial intelligence “can explain original visions, which can be verified, meaningful to scientific scrutiny, it remains a strong assistant, but it is definitely not a co -expert.”

The blue section of the figure shows an experimental search pipeline that has discovered the transmission of the DNA between bacterial species. The orange section explains how artificial intelligence quickly reached the same conclusions.José R. Penadés, Juraj Gottweis, et al.

Logic, not only remember

Supporters contradict that the last generation of models shows a glimmer of what scientists can call it “intelligence” reasonably. Not only do regimes such as Google experts from Google and the synthesis of vast libraries, but also a cause through competing capabilities, ignoring weakest ideas, and improving strong ideas in ways that can feel a striking person.

“I find it very active,” says Bilz. “It is like a conversation with someone who knows more than you.”

However, magic does not happen automatically. Swanson notes that the extraction of valuable hypotheses requires accurate occurrence and repetitive reactions and the desire to participate in a type of dialogue with artificial intelligence. It is similar to pressing a button to answer and more similar to directing a novice colleague – when taking the right questions, responding to shallow thinking, and pushing the system towards more clear visions.

“For the moment, you still need to be a little expert to get the most useful use of these systems,” said Swanson. “But if you ask a good designed question, you can get really good answers.”