Nick Edwards, PhD, CEO of Potato, has a “burning desire” to speed up scientific discovery. In an interview with GEN, he explained how Potato’s AI “scientist,” named Tater, recently replicated a main discovery from his PhD research investigating circuit specificity of the brain’s reward center.1 As a graduate student, Edwards recalls that experiments could span five to six hours, while troubleshooting flaws in protocols would take weeks. In contrast, Tater was able to take unpublished raw data, complete a statistical analysis, generate code to reproduce results, and create figures to summarize findings within a few hours.
“Tater found some relevant papers and told me to try injecting the virus at different locations to ensure the right spot. That’s work that took me about a week to do,” said Edwards. He also emphasized that an AI scientist should not just be able to come up with new ideas, but also test hypotheses in the real world.
Another example involves the SARS-CoV-2 main protease (Mpro), a key COVID-19 therapeutic target due to its role in viral replication. Tater identified single-point mutations in Mpro that could lead to drug resistance. The AI scientist then applied computational analysis to thousands of DNA variants to identify mutations that would occur close to the binding site for testing in the lab.
In recent years, the release of general-purpose AI models that interface with human prompts, such as ChatGPT from OpenAI, Gemini from Google, and Claude from Anthropic, has correspondingly led to autonomous systems across disciplines, including the scientific research lab.
Potato executes scientific protocols from end-to-end by automating tasks, from summarizing literature and conducting web searches to generating code and operating lab tools. These tasks are based on an array of scientific input, including repositories of peer-reviewed publications, experimental protocols, lab notebooks, and experimental raw data. Founded in 2023 by Edwards and Ryan Kosai, former CTO of Pioneer Square Labs and principal engineer at ExtraHop Networks, Potato closed a $4.5 million seed round in April. Potato has since established a partnership with Wiley to integrate repositories of peer-reviewed scientific publications into the company’s AI scientist capabilities.
More scientists
Rory Kelleher, senior director, global head of business development, healthcare and life sciences at NVIDIA, highlights that amplifying the number of scientists is one of the powerful advantages of AI agents.
“This is a tool you can give to a recent PhD graduate and have them conduct six months of work in just a few days, or allow a lab of 15 researchers to operate like a team of 100,” said Kelleher. “Just like in other areas of health care where there’s a shortage of nurses and doctors, the world needs more scientists—research and scientific discoveries are what drive breakthroughs in economic development, human health, and more.”
As an “AI company that works with every AI company,” NVIDIA has partnered with a wealth of life science players applying foundation models, AI agents, and robotics to scientific discovery. These partners include FutureHouse, an AI scientist non-profit backed by former Google CEO Eric Schmidt and co-founded by Sam Rodriques, PhD, previously group leader at the Francis Crick Institute and current CEO, and Andrew White, PhD, head of science at FutureHouse.
As AI scientists have traditionally been limited by their ability to synthesize information at large scale, FutureHouse’s latest AI scientist, Kosmos, can read 1,500 papers and execute 42,000 lines of analysis code in one run. In November, FutureHouse launched Edison Scientific, a new commercial spinout that will focus on further developing and deploying the AI scientist for commercial applications.
Domain-specific
Other companies are developing autonomous systems to fit domain-specific tasks. Dyno Therapeutics, a Boston-based genetic medicines company that has traditionally focused on capsid engineering for delivery, has recently developed AI agents to streamline workflows that leverage both structure-based and language models for protein design.
Dyno’s Structure agent, p0, reasons around protein structures to facilitate therapeutic design. [Dyno Therapeutics]In November, Dyno announced Parser, an autonomous agent that participates in decision flow from raw lab experiments to product discovery, and Knowledge agent, which generates executive summaries to allow users to oversee the history of gene therapy products at Dyno. Additionally, the Structure agent, p0, was introduced to reason across molecular structures and facilitate payload engineering.
Eric Kelsic, PhD, CEO of Dyno says p0 serves as a bridge toward more partnerships for therapeutic design.
“In protein design, there’s a lot of routine work, such as finding the right file, loading it into structure view, and manipulating it for visualizations,” explained Kelsic. “With a simple language prompt, Dyno p0 agent automates those steps to understand information in the context of the research topic of interest.”
Co-scientists
Le Cong, PhD, assistant professor of pathology and genetics at Stanford University, states that much of the lab “know-how” is currently locked in the hands of a few experts. By having agents watch, document, and generalize those practices, the expertise can be democratized across scientists and institutions.
“Can we turn these AI agents into ‘co-scientists’ that understand both the digital world of code and papers, and the physical world of cells, operations, and equipment in the lab?” he posed.
In 2023, Cong set out to build an AI agent for gene editing, named CRISPR-GPT, after using ChatGPT to design guide RNAs and CRISPR experiments “failed completely.” CRISPR-GPT leverages the reasoning capabilities of large language models (LLMs) for complex task decomposition and decision making, including selecting CRISPR systems, experiment planning, designing guide RNAs, choosing delivery methods, drafting protocols, designing assays, and analyzing data.2 The agent illustrates the importance of incorporating domain-specific knowledge to tackle biological design problems.
Using CRISPR-GPT to guide biological insight, researchers were able to achieve greater than 80% editing across all loci for Cas12 editing. In addition, the team discovered and validated a new natural killer (NK) cell immunotherapy target (CEACAM6) and identified intersectin 1 (ITSN1) as a regulator of cell–cell fusion.
“We thought, maybe we should take the lead now to build an AI agent for CRISPR to scale and automate our work. Now, when a new student or collaborator comes into our lab, we let them use CRISPR-GPT instead of spending hours to teach them,” said Cong.
LabOS enables AI to see what scientists see through smart glasses and supports real-time experimentation. [Cong Lab]
In the physical lab, Cong’s team went a step further and announced LabOS as a preprint on bioRxiv in October.3 LabOS is an AI co-scientist that connects multi-model AI agents, smart glasses, and human-AI collaboration to allow AI to see what scientists see, understand experimental context, and assist in real-time execution.
As much of the “cost” in biology is iteration over protocols, LabOS can reduce those cycles by catching mistakes early and standardizing execution. Additionally, the system auto-records steps, flags likely errors, and provides stepwise guidance to junior scientists via XR glasses, which reduces failed experiments and the need for mentor supervision. According to the preprint, a new student working in cell engineering was able to achieve 85% efficiency compared to a senior postdoc within one week.
At NVIDIA GTC in Washington, DC in October, a collaboration was announced between NVIDIA, VITURE, Nebius, and more partners to bring XR smart glasses directly to research labs. NVIDIA XR AI is a platform that connects XR devices to an organization’s full computational power, enabling agents to operate across cloud, data center, and workstations.
NVIDIA’s BioNeMo is an open-source machine learning framework for building and training deep learning models for biopharma. AI model builders doing biomolecular research with DNA, RNA, and protein data can access tools to scale their research.
All told, the AI scientist is poised to manage information flow and streamline physical tasks, freeing human researchers to focus on creativity, troubleshooting, and new directions. How dramatically this shift will accelerate scientific discovery will unfold on the horizon.
References
1. Edwards NJ, Tejeda HA, Pignatelli M, et al. Circuit specificity in the inhibitory architecture of the VTA regulates cocaine-induced behavior. Nat Neurosci 2017;20(3):438–448; doi: 10.1038/nn.4482.
2. Qu Y, Huang K, Yin M, et al. CRISPR-GPT for agentic automation of gene-editing experiments. Nat Biomed Eng 2025; doi: 10.1038/s41551-025-01463-z.
3. Cong L, Zhang Z, Wang X, et al. LabOS: The AI-XR Co-Scientist That Sees and Works With Humans. bioRxiv 2025;2025.10.16.679418; doi: 10.1101/2025.10.16.679418.
The post Virtual Scientists Poised to Accelerate Discovery appeared first on GEN – Genetic Engineering and Biotechnology News.
