How researchers are using AutoDiscovery

Today we released AutoDiscovery as an experimental feature in AstaLabs. AutoDiscovery is an AI-powered tool that explores structured datasets autonomously—generating hypotheses, designing and running statistical experiments, and surfacing findings that researchers might never have thought to look for.

Most AI tools for science are goal-driven: they wait for you to provide a research question, then help you answer it. AutoDiscovery works differently. Give it a dataset and let it explore overnight. When it finishes, you'll have a complete list of novel research directions—each one with reproducible code, statistical results, and a clear path to follow-up investigation.

AutoDiscovery started as a research project we published last year with open-source code. During development, we worked with researchers across disciplines to apply it to real scientific problems. Now we're making it broadly available in AstaLabs.

In oncology

Dr. Kelly Paulson, a medical oncologist and Head of the Center for Immuno-Oncology at Providence Swedish Cancer Institute, has used AutoDiscovery to analyze large clinical, genomic, and gene expression datasets from breast cancer and melanoma research programs. These cancers remain challenging because current treatments work well for some patients but not others, and Paulson's team aims to understand the factors associated with improved outcomes.

"Traditional approaches to large cancer datasets, including many current AI tools, often rely on a framework in which researchers begin with a specific hypothesis and test that idea against the data," Paulson says. "This limits discovery to what we already suspect. AutoDiscovery takes a different approach by exploring the full dataset without requiring predefined assumptions."

The system has already provided value by confirming several expected findings – such as the importance of immune activity in melanoma and the relevance of the PI3K pathway in breast cancer – which increased the team's confidence in its reliability. But even early runs pointed to new hypotheses of clinical and biological relevance: in melanoma, potential associations involving very strong immune responses, and in breast cancer, relationships that may be linked to the risk of lymph node spread. These findings weren't part of the team's initial hypotheses, and they're now testing and validating them in follow-up studies.

"AutoDiscovery's ability to reveal discoveries that may be hiding in plain sight is especially valuable in cancer research, where substantial amounts of clinical and genomic data remain only partially explored," Paulson says.

Read Paulson’s report here.

In marine ecology

Researchers at the Scripps Institution of Oceanography have used AutoDiscovery to explore more than 20 years of rocky reef monitoring data from the Gulf of California. They already understood that recent marine heatwaves were negatively affecting fish populations, but AutoDiscovery helped them move beyond broad patterns by framing more original, structured, and mechanistic hypotheses. In particular, the system surfaced relationships between productivity across trophic levels that would have required extensive manual iterations to discover. 

"The ability to generate multiple hypotheses that can then be thoroughly evaluated by the user is extremely powerful," Fabio Favoretto, a marine ecologist at Scripps, says. "Even when some hypotheses are less valid, there is always something to learn from the process." Favoretto emphasized that AutoDiscovery's transparency – the ability to inspect reasoning and analytical steps in depth – builds the trust that scientists need when working with AI tools. 

Read a detailed technical report by Favoretto here.

In social science

Sanchaita Hazra, an economist in the College of Social and Behavioral Science at the University of Utah, has used AutoDiscovery to explore social and economic datasets, surfacing relationships between variables that might otherwise require extensive manual hypothesis testing to uncover. Her findings were published in a peer-reviewed paper last November.

"The biggest competitive advantage of AutoDiscovery is its systematic search over the possible hypothesis space spanned by the data," Hazra says. "While chat-based AI interfaces allow for local exploration of ideas, even with data, they never go beyond a local search, often repeating themselves. AutoDiscovery is almost like deep research with data, but on steroids."

In her paper, Hazra examined how academic authors edit AI-generated writing. Her original focus was comparing behavior under human versus AI conditions. But AutoDiscovery helped surface something she hadn't been looking for: a significant effect of education level on editing patterns. Authors with doctoral degrees edited AI-generated abstracts substantially more than those with undergraduate or master's degrees, suggesting that more experienced researchers engage more critically with AI-generated text. 

Later, Hazra independently confirmed the results.

"This is a clear example of where my original research focus would have missed this finding or needed several regressions to get to this interesting result," she says. "Surfacing results that are non-trivial, relevant, but novel to the existing literature is a full-time job for a PhD student for a couple of weeks. AutoDiscovery did it in a few hours."

Looking back, Hazra notes that AutoDiscovery was able to replicate many major findings from experimental data she collected over years of doctoral research. "Had I had access to this tool, it would have powercharged my research, potentially seeding ideas for novel extensions," she says. 

Our collaborations with experts have shaped AutoDiscovery's development—helping us understand where open-ended discovery is most valuable and how it fits into real research workflows. Science has always had more data than time to explore it. AutoDiscovery won't replace the intuition and expertise that drive great research, but it can help surface the questions hiding in your data so you can focus your effort where it matters most.

Try AutoDiscovery in AstaLabs today, and let us know what you find—and learn more in our technical blog post.

Sign up for Asta Preview to gain early access to features like AutoDiscovery. Learn more here.