Ai2 finds immune signal in breast cancer once thought immune-resistant

The numbers give this weight. This subtype makes up about 15% of breast cancers diagnosed each year in the United States. That means a patient population effectively excluded from immunotherapy consideration until now could be reconsidered.

Invasive lobular carcinoma is a particularly tricky subtype of breast cancer. Rather than forming a solid mass, its cancer cells spread out in single-file lines, which makes them hard to catch on imaging, and the way they interact with immune cells within tissue differs from other subtypes too. That's why it has long been left out of immunotherapy discussions.

Verified twice

What stands out in this announcement isn't the discovery itself but the verification process. First, researchers checked whether the same signal reappeared in a separate breast cancer dataset. Then they examined tumor samples using immunofluorescence imaging. Alongside hormone receptor markers, they could actually see T cells surrounding the tumor.

Immunofluorescence imaging attaches fluorescent tags to specific proteins so researchers can visually pinpoint where certain cells sit within tissue. It serves as the final checkpoint, confirming whether a statistically detected signal actually shows up in the same place in real tissue.

Form a hypothesis, reproduce it in different data, confirm it visually — only after filling in all three boxes did the two organizations announce the expanded partnership.

The paper carries researcher names from both sides. Dr. Kelly Paulson took part from Providence, and Dr. Sasha Stanton from the Earle A. Chiles Research Institute.

Ai2 said that for a discovery to feed into future research, it needs to be transparent, reproducible, and independently verified. That's why the announcement spends more time explaining the process than the discovery itself.

Keeping only what's both surprising and reproducible

AutoDiscovery is a tool that generates hypotheses using a large language model and then evaluates them itself. It filters candidates by two criteria: whether a finding is surprising relative to prior expectations, and whether it holds up across multiple analyses.

The key is requiring both conditions at once. Surprising alone is just noise, and reproducible alone is just something already known. Only by combining the two do you end up with candidates worth a researcher's attention.

The dataset analyzed was The Cancer Genome Atlas, considered one of the most comprehensive cancer datasets ever assembled.

A warehouse analogy makes this easier to picture: everything is already in stock, but nobody knows exactly where anything is.

Kelly Paulson, director of the Immuno-Oncology Center at Providence Swedish Cancer Institute, put it precisely: "Cancer researchers have access to remarkable datasets. The problem is no longer collecting data — it's understanding everything those datasets are telling us."

The tool goes where the data lives

The expanded partnership covers four areas: Providence researchers and Ai2 scientists co-authoring papers, deploying AutoDiscovery inside Providence's own cloud environment, giving the tool access to protected research and clinical data within that environment, and having the cancer institute's computational research team handle installation and support.

What matters here is where the tool is deployed. Rather than pulling data out to reach the tool, the tool goes to where the data already sits. When working with institutions that handle patient data, negotiations don't even get off the ground unless it works this way.

Bodhisattwa Majumder, a senior research scientist at Ai2, summed up the tool's role this way: "AutoDiscovery wasn't built to replace scientific expertise. It was built to amplify it."

The design intent points the same way. AutoDiscovery wasn't built to stand in for researchers but to work alongside them as a complementary tool. It narrows down candidates, and researchers decide which ones to pursue.

Ai2 is a non-profit research institute. It is currently operating under interim CEO Peter Clark, and it has staked out a different position from commercial labs by openly releasing its models and data.

Editor's take

The most valuable part of this announcement isn't the discovery — it's the verification process. Over the past year, dozens of stories have surfaced about AI generating hypotheses. Very few of them were announced only after being reproduced in a separate dataset and visually confirmed with tissue sample imaging.

The "immunologically cold tumor" classification has sat in textbooks for a long time. It's hard for a human researcher to be the first to doubt a signal that contradicts it — questioning something already considered settled rarely attracts funding or time. What the AI did here wasn't provide an answer. It opened a drawer nobody had bothered to reopen.

This case lines up precisely with the five-walls document on AI-assisted scientific research that Ai2 published yesterday. The concerns laid out there show up here as an actual working process: require both surprise and reproducibility, verify twice, and leave the final call to a human.

What hospitals and research institutes in Korea should take away from this isn't the model itself — it's the deployment approach. Medical data here is difficult to move outside institutional walls, which is usually where AI-adoption discussions stall, at the negotiation over exporting data.

Providence flipped that order. It brought the tool into its own cloud and had its own computing team handle installation and operation. The data never moves an inch.

The 15% figure also needs context. It's not a measure of treatment efficacy — it's the size of the population now worth reconsidering. Whether immunotherapy actually works for this group is a question only clinical trials can answer, and that answer is years away. Even so, the fact that a candidate population now exists where none existed before is itself a starting point for trial design.

Claims that AI is transforming science have become commonplace. What sets this case apart is what got changed — not a paper summary or a bit of code, but patient classification. AI has stepped into the process of redrawing the line that decides which patients count as candidates for which treatments.