You’re reading an issue of “The AI Economy,” my newsletter exploring the forces shaping the AI era—tracking how AI is rewriting business, work, technology, and culture. Subscribe to get expert insights and curated updates delivered straight to your inbox.
Last February, the Allen Institute for AI (Ai2) introduced AutoDiscovery, an open-source experimental AI tool that analyzes massive scientific datasets to uncover new lines of inquiry. The non-profit lab applied it in a partnership with the Paul G. Allen Research Center (PARC) at Providence Swedish Cancer Institute, using AutoDiscovery to analyze The Cancer Genome Atlas, one of the most comprehensive cancer datasets.
Today, the two organizations announced a surprise discovery: invasive lobular carcinoma (ILC), a breast cancer subtype affecting roughly 15 percent of U.S. patients each year and long considered unresponsive to immunotherapy, appears to have a stronger immune signature than previously known. Tumors like these are classified as “immune cold”—the immune system isn’t engaging with them, so the drugs that work by unleashing that response have nothing to unleash. The finding was validated across an independent patient dataset and through a lab analysis of tumor samples.
According to Dr. Kelly Paulson, PARC’s lead for the Center for Immuno-Oncology: This is good news. “There’s immune therapies that are already available today that we could think about testing on this type of cancer, and we can also use this insight to help develop and apply new forms of immunotherapy treatments,” she told The AI Economy in an email, adding that it was “exciting” to witness the discovery of something that could be immediately translatable into a new cancer treatment.
Details of the discovery are documented in a paper titled “Surprisal-based large language models reveal immunologic insights in breast cancer.” The team said that the manuscript has been submitted to MedRxiv, the most relevant preprint server for the work—it hasn’t been peer-reviewed yet—and there are plans to provide it to leading medical and science journals. Researchers suggest that ILC “may warrant broader investigation” in future immunotherapy research.
“Cancer researchers have access to extraordinary datasets, but the challenge is no longer collecting data; it’s understanding everything those datasets have to tell us,” Paulson said in a statement. “AutoDiscovery helped us identify a promising signal that we may not have otherwise investigated, and from there we were able to validate the finding through additional datasets and laboratory research.”
Two Findings, 65,000 Hypotheses
The ILC revelation is the second publicly disclosed hypothesis coming from AutoDiscovery. Back in February when the tool was introduced, Ai2 revealed that it had started a partnership with PARC. The team discovered an unexpected signal, documented in a paper it submitted to the Conference on Neural Information Processing Systems (NeurIPS): among patients with a PIK3CA gene mutation, mutations in another gene, TP53, occurred less often than chance would predict. This means that because they rarely occur together, they might serve the same biological function—either that, or cancer cells carrying both couldn’t survive.
Paulson shared that before now, AutoDiscovery was validated by Ai2 on multiple different data sets, but wasn’t applied specifically to cancer. There was concern about how it would perform in real-world biomedical settings, despite promising early data. “When we get an AutoDiscovery run back we get information back on surprising hypotheses, but the system also tests hypotheses that are very likely to be true,” she said. “Most of those true hypotheses validate as real, and the data and code are provided. These ‘true positives’ help to provide support.”
AutoDiscovery has already generated 65,000 hypotheses by scientists working in domains like oncology, neuroscience, and social science, according to Bodhisattwa Majumder, Ai2’s senior research scientist in an email to The AI Economy. “Local deployments like the one Providence Swedish Cancer Institute is doing now, will unlock even more discoveries as they begin to apply our tools to their private clinical data.”
There’s an enormous amount of data sitting in public databases—genomic, molecular, clinical, and imaging—accumulated over decades from millions of patients, not to mention the privately protected data from cancer centers and research institutes that are largely unexamined. That scale makes it difficult to start with a goal, so AutoDiscovery approaches the problem from the opposite direction.
How AutoDiscovery Works
Built from Ai2’s Asta framework, AutoDiscovery autonomously explores complex datasets to generate and evaluate surprising hypotheses with large language models. It’s designed to work collaboratively with scientists, allowing them to guide promising directions, apply domain expertise, and assess which findings should be probed further.
“We are taking away the goal,” Majumder said in February. “We’re saying that it is open-ended. It doesn’t wait for the user to give a goal. It looks at the data and starts figuring it out, [coming] up with its own goal and exploring autonomously.”
When asked how this latest finding validates AutoDiscovery’s purpose, he said that Ai2 was “very proud” of the collaboration, adding that Ai2 is actively working with domain scientists at other organizations to develop more Asta agents, though he declined to provide specifics.
AutoDiscovery Expands to Active Cancer Research
Since its launch, AutoDiscovery has been limited to public research datasets. However, its partnership with the Providence Swedish Cancer Institute marks the start of a new chapter in the tool’s use—it’s now being used in active cancer research programs at a leading oncology research center.
The Providence Swedish Cancer Institute said it’s bringing AutoDiscovery into its cloud environment, which will give the AI tool access to the institute’s private research and clinical data. Doing so not only keeps the data secure, but ensures that PARC’s computational research team is doing all the work: installing AutoDiscovery, running it, and providing support to internal researchers using it.
“Our early research gave us confidence that AutoDiscovery could complement the way our scientists already work by helping surface hypotheses that thus far have stood up to rigorous validation,” PARC’s Lead Data Scientist, Zachary Reitz, said in a statement.
Beyond local deployment, Providence Swedish Cancer Institute plans to eventually import Ai2’s agents locally, enabling them to analyze private clinical data in the hopes of empowering new drug and treatment discoveries.
“Scientific discovery depends on trust, and trust is earned through close collaboration with researchers and rigorous validation of every finding,” Peter Clark, Ai2’s interim chief executive, said. “We believe partnerships like this will help define how AI is used to accelerate discovery across medicine.”
