Artificial intelligence (AI) research laboratories Google DeepMind and Futurehouse have recently introduced AI research assistants capable of generating scientific hypotheses, crafting experiments, and analyzing data. The researchers associated with these innovations believe that they could expedite scientific advancement, for instance, by discovering clinically approved medications for unrelated conditions or illnesses.
AI tools have thus far aided scientists in processing vast amounts of data, predicting protein structures, or even creating small molecule drugs that are beginning to undergo clinical trials. ‘[But] how can we instruct AI systems to think like scientists do?’ queries Vivek Natarajan from Google DeepMind. Current interactive AI tools – like Chat GPT – frequently provide rapid responses, which ‘is not representative of how science operates’, he states. ‘[Scientific thinking] is more methodical, significantly more rigorous, structured, and it’s a process that often spans extensive durations.’
To address this challenge, Google DeepMind created Co-Scientist – an AI assistant specifically designed to collaborate with researchers. DeepMind made the program available in a preprint last February. Non-profit Futurehouse presented its counterpart, Robin, several months afterward, and both organizations have now shared their results in Nature.
Co-Scientist builds upon Google’s AI assistant Gemini and employs multiple AI programs or ‘agents’ to react to a research objective proposed by a user. The agents subsequently generate preliminary ideas, sift through existing scientific literature to assess hypotheses, and repeat this process numerous times to enhance ideas. This operation is akin to how Google’s AlphaGo determines which move to make next in the game of Go, as noted by Natarajan. Researchers can then conduct the proposed experiments and utilize the findings to prompt Co-Scientist to produce more refined hypotheses.
What is AI?
Artificial intelligence (AI) is a broad term often misused to describe a range of interconnected yet simpler processes.
AI refers to the capacity of computers and software to perform tasks that are typically within the human domain, such as reasoning, responding to feedback, and making decisions.
Generative AI represents a newer form of AI that analyzes and identifies patterns within training datasets to create original text, images, and video in response to user requests. ChatGPT, Microsoft Copilot, Google Gemini, and more recently, X’s Grok are all chatbots utilizing generative AI.
Neural networks consist of a linked array of artificial neurons, similar to biological brains, that detect, analyze, and learn from statistical patterns in data.
Machine learning is a branch of AI that empowers machines to learn from datasets and make predictions based on new information, without explicit commands from programmers. Machine learning models enhance their performance as they accumulate more data.
Deep learning is an advanced type of machine learning that applies neural networks with numerous layers to examine complex data from very large datasets. Uses of deep learning include speech recognition, image creation, and translation.
Large language models or LLMs are a specific kind of deep learning trained on extensive datasets to comprehend and generate language. LLMs grasp patterns in text by anticipating the subsequent word in a sequence, and these models now possess the ability to write narratives, analyze text from the internet, and engage in dialogues with users.
Natarajan indicates that the team has collaborated with hundreds of scientists around the world to explore how the research community might utilize the tool. ‘Most of the scientists [we converse with] are quite astonished by the state of AI, what it is producing [and] how quickly it is responding,’ he remarks.
For example, the team partnered with microbiologists at Imperial College London, UK, who were studying how certain genes transfer between bacteria, enabling them to develop antimicrobial resistance. ‘When we operated the [Co-Scientist] system for a couple of days, it effectively retraced their entire research trajectory and made the exact same predictions,’ says Natarajan.
Tiago Dias da Costa, who led the study at Imperial, states that Co-Scientist had ‘independently formulated a hypothesis that closely corresponded to the mechanism we discovered through years of experimental work, but ‘it did not supplant the experimental discovery process’.
‘The significance of Co-Scientist was not in simply “providing us with the answer” but in illustrating how AI can assist in generating, prioritizing, and fine-tuning biologically significant hypotheses.’ He emphasizes that ‘AI-generated hypotheses only evolve into discoveries when they undergo rigorous laboratory testing’.
Co-Scientist will serve as the hypothesis-generating tool within Google’s broader Gemini for Science initiative, which is expected to be accessible to researchers in ‘the coming weeks and months,’ he adds. However, Natarajan believes that ‘a few major advancements are required before we achieve a system capable of achieving what some of the remarkable scientists of the past [have accomplished] – like formulating a true original breakthrough [or] a paradigm-shifting theory’.
Repurposing existing drugs to explore new approaches to treat diseases.