Agentic AI scientists and the rise of virtual laboratories

To the Editor, Recently, artificial intelligence (AI) has made impressive progress in medical research and diagnostics. The progress of AI is progressing very quickly. AI models, such as AlphaFold and large language models, have been developed over time and have helped advance the medical field. The AI-driven AlphaFold model can predict protein structure from amino acid sequences. Now, it is able to develop these protein structures more accurately. This AI model helps researchers identify drug targets and understand how proteins interact with other molecules. It helps in structure-based drug design, target identification and validation, and therapeutic discovery. Thereby, the model advances medical science and biomedical research[1]. Similarly, another AI-driven model, called large language models (LLMs), has emerged recently. One significant example of this model is ChatGPT. It is transforming the medical field very quickly. It helps in clinical decision-making, medical documentation, literature review, and patient communication. These AI models can rapidly analyze large amounts of biomedical information. They are now assisting clinicians and researchers. It helps in knowledge discovery, medical education, and evidence-based practice. Recent bibliometric analyses highlight the growing global research interest and the potential of LLMs to enhance health care and medical research[2]. Several AI-enabled LLMs in drug discovery support drug discovery and development through structure-based drug molecule design and de novo drug design. The models also help in drug target identification and validation[3]. All these AI-enabled models help advance medical research, education, and clinical practice, accelerating the field’s progress. Now, researchers have wanted to create AI systems that can make big discoveries on their own. During discoveries, the AI system learns without human help. This idea has been called “AI scientist.” Recent advances in AI technology are making it increasingly possible. Now, with the development of new agentic AI, researchers can create AI scientists. The agentic AI scientists can think, learn, and work together using LLMs, generative AI (GenAI), machine learning (ML) tools, and experimental platforms to solve problems[4,5]. It is possible through the human–AI collaboration. With the agentic AI scientists, AI is transforming scientific research and discovery. AI is redefining how scientists approach future research by integrating intelligent systems. Modern AI can do more than analyze data. Modern AI agents are conducting scientific discovery[6]. They operate autonomously, generating hypotheses and designing experiments. All within virtual laboratory settings, and they run simulations, interpret results, and refine methods[7]. Agentic AI represents a major shift in science and technology by enabling systems to work independently toward broad goals rather than single decisions[6]. These systems undertake complex tasks requiring planning and reflection, building on GenAI’s earlier capabilities to make discoveries[8–10]. Agents interact with tools and environments, engage in deep multi-step reasoning, plan, reflect, set complex goals, operate in dynamic environments, and manage resources independently. Using Agentic AI scientists, researchers are now trying to perform their research within virtual laboratory settings. Merriam-Webster defines a laboratory, which illustrates the laboratory as a place for experimental study and testing, as well as for observation, in any field. Traditionally, it is a physical space. However, laboratories are increasingly adopting digital tools and automation. However, the core concept remains same. A virtual laboratory is a digital environment that offers all the elements of a classical laboratory, enabling research tasks in a modern, digitally driven context[11]. Virtual laboratories expand upon tools that scientists already use. Computational methods such as simulations, digital twins of instruments, robotic measurement devices, and experimental design approaches are now standard in modern science. Data analysis and statistical estimation methods are also integral to this toolkit[11]. In this way, virtual laboratories represent a natural progression for research environments. Medical-to-biomedical research is changing due to advances in computational intelligence. Right now, AI only helps with simple, low-risk tasks in the virtual laboratory. However, scientists must review all AI results generated by the virtual laboratory with agentic AI scientists. The proposed agent-based AI systems can learn and reason autonomously by combining different parts. It may use LLM-based systems and other ML tools. They include experimental platforms and, if necessary, human researchers’ intervention. Sometimes, they combine all of these elements. Humans and AI need to interact continuously. However, researchers need to build safe sandboxes where AI agents can fail and learn from mistakes. This is one way to achieve trustworthy AI[4]. Recently, scientists have been conducting experiments using Agentic AI Scientists to create a virtual laboratory. Using this approach, Swanson et al. have developed nanobodies against the S-protein of current SARS-CoV-2 variants and ancestral viral variants. Here, they first understood the performance of virtual laboratory and agentic AI scientists, and their impact on science. Therefore, they have created the virtual laboratory using agentic AI. The virtual laboratory is a collaboration between AI and humans. This AI agent leads the project in collaboration with human scientists. The virtual laboratory also has several LLM scientist agents. Each scientist agent gets assigned research tasks. They have attended meetings and contributed ideas. The human researcher provides guidance and high-level feedback. The team has done several works, from designing the virtual laboratory to designing a nanobody. These nanobodies target recent SARS-CoV-2 variants. The virtual laboratory developed a pipeline of computer-based design pipeline using different software such as AlphaFold, Rosetta, and ESM. They have designed 92 new nanobodies. The new nanobodies were tested through the experiments. The results indicated that they bind well throughout the variants SARS-CoV-2. Two nanobodies performed especially well. However, two new nanobodies bind better to the recent KP.3 or JN.1 variants. Both still bind strongly to the original virus[12]. JN.1 and KP.3 are now circulating globally[13,14]. These nanobodies can help fight these recent circulating variants. While researchers will perform the AI experiments, they need to develop a proper benchmarks for its application. Recently, Agha and his colleagues explained transparency standards for AI-related experiment of report. They created a standard guideline called TITAN[15]. This guideline explains how to use AI properly. It might serve as a standard document for researchers, with benchmarks. It will help researchers use AI correctly in their studies. It is necessary to develop agentic AI scientists capable of handling multiple tasks simultaneously in virtual laboratories to address next-generation medical-to-biomedical science problems. To accomplish this, agentic AI scientists must plan discovery workflows, use ML feedback loops, assess and revise their own work, and identify knowledge gaps. All of these should be performed with support from AI–human collaboration. While their potential to support research is impressive, current usage remains limited. In the future, research will focus on answering proper scientific questions instead of broad open-ended exploration.