Artificial intelligence is beginning to play a more concrete role in medical research—not as a substitute for scientists, but as a tool for finding signals that might otherwise go unnoticed. Ai2 and the Providence Swedish Cancer Institute have expanded their collaboration to apply AutoDiscovery to cancer research using protected clinical data.
The first major result came from a study of invasive lobular breast cancer, a subtype that accounts for approximately 15% of breast cancers diagnosed each year in the United States. The analysis suggests that these tumors may have a stronger immune response than previously thought.
AutoDiscovery looks for hypotheses, not automatic answers
Biomedical research is going through a period of data abundance. Public repositories such as The Cancer Genome Atlas, known as TCGA, bring together genomic, molecular, clinical, and imaging information from millions of patients. Research centers also store large volumes of protected data.
The problem is clear: having more data does not mean we fully understand it. How many relevant signals remain hidden because no team can explore every possible combination?
AutoDiscovery aims to complement the traditional scientific method. Instead of starting only with a hypothesis defined by researchers, the system uses language models to generate and evaluate surprising hypotheses from complex datasets.
Its operation is based on two main criteria:
- Surprise: identifies observations that differ from what would be expected based on prior knowledge.
- Reproducibility: prioritizes signals that appear again when the analyses are repeated or when they are examined in other datasets.
This does not turn a correlation into a medical discovery. The tool helps determine which questions deserve deeper investigation, while scientists interpret the results, design validations, and decide whether the evidence is solid.
The idea is not for AI to replace scientific expertise, but to help systematically explore a data landscape that is too vast to analyze manually.
An unexpected signal in lobular breast cancer
A team led by Dr. Kelly Paulson of the Paul G. Allen Research Center and Ai2 researcher Bodhisattwa Majumder applied AutoDiscovery to the TCGA dataset.
Among its findings was an unexpected observation about invasive lobular carcinoma, or ILC. This breast cancer subtype had historically been considered “immunologically cold,” a way of describing tumors with little immune activity and a potentially limited response to immunotherapy.
However, AutoDiscovery detected a stronger immune signature than previous research had recognized. An immune signature is a set of biological signals that suggests the presence or activity of immune system cells and processes within a tumor.
The team did not stop at the initial result. They validated the observation in an independent patient dataset and then turned to laboratory analyses of tumor samples. That combination is essential: an AI-generated finding needs to withstand external testing before it can become a useful foundation for further research.
The results, published in the paper “Surprisal-based large language models reveal immunologic insights in breast cancer,” suggest that invasive lobular breast cancer deserves broader evaluation in future immunotherapy studies.
This does not mean that a new treatment already exists or that all patients with ILC will respond to immunotherapy. It means that a category once considered less promising could contain signals that justify additional trials and analyses.
From public datasets to protected clinical data
The expanded partnership marks an important change for AutoDiscovery. Until now, the platform had worked with public datasets. The next step will be to deploy it within Providence’s cloud environment, where it can operate on research data and protected clinical data.
The information will remain within Providence’s infrastructure. The PARC research computing team will install and run the platform, as well as support the researchers who use it.
Why local deployment matters
In medicine, moving sensitive data to external services can create privacy, regulatory compliance, and institutional control risks. Local deployment reduces the need to transfer that information outside the organization’s environment.
It also allows the center to maintain greater control over technical aspects such as:
- Which data the system uses.
- Who can access the results.
- How the analyses are recorded.
- Which model versions and configurations produce each finding.
- How experiments are repeated to verify their reproducibility.
For AI to be accepted in biomedical science, accuracy alone is not enough. Researchers need traceability, transparency, independent validation, and a clear way to understand how each conclusion was reached.
A collaboration model for scientists and AI
AutoDiscovery was designed to work alongside researchers. Experts can guide the system toward promising areas, incorporate clinical knowledge, and decide which signals are valuable enough to be tested in the laboratory.
That point is especially important in cancer. A statistical relationship may be caused by factors such as patient composition, sample quality, or the way the data was collected. Scientific expertise helps distinguish an interesting lead from a misleading result.
The collaboration between Ai2 and the Providence Swedish Cancer Institute presents a practical model: AI explores vast spaces of hypotheses, while scientists provide context, judgment, and validation methods.
Is this the future of research? It probably is not a matter of choosing between humans and machines. The most useful path appears to be combining AI’s ability to review enormous volumes of information with the human responsibility to interpret, verify, and decide.
Deploying the system on protected data will be an important test. If the platform can generate reproducible signals without compromising privacy and truly becomes part of scientific workflows, it could become a tool for accelerating discoveries across different cancer research programs.
For now, the lobular breast cancer case shows something more valuable than a futuristic promise: an AI-generated hypothesis, validated with independent data and laboratory analyses, that opens a new medical question. In science, that is often the beginning that matters.