Surprisal-based large language models reveal immunologic insights in lobular breast cancer is a paper published by the Allen Institute for AI on 27 August 2026, reporting the application of the institute's AutoDiscovery framework to the breast cancer cohort of The Cancer Genome Atlas (TCGA). The co-first authors are Bodhisattwa Prasad Majumder (Ai2) and Jessica Linak, with Peter Clark, Kelly Garneski Paulson and Sasha Stanton as co-senior authors; Paulson is the corresponding author. Contributing institutions are Ai2, the Providence Cancer Institute's Earle A. Chiles Research Institute, the Paul G. Allen Research Center at the Providence-Swedish Cancer Institute, the University of Massachusetts Amherst, MIT, and the Elson Floyd College of Medicine at Washington State University. The paper is hosted on allenai.org and carries no preprint-server identifier (Source: allenai.org).
The paper reports two findings that its authors keep separate: a negative result about unguided autonomous discovery, and a positive biological result obtained once an oncologist's annotations were fed back into the system.
The AutoDiscovery framework
AutoDiscovery combines large language models with Monte Carlo tree search to generate and evaluate hypotheses without predefined research questions, prioritising observations that are both surprising relative to prior expectations and reproducible. The system is organised as a six-role multi-agent pipeline: a Hypothesis Generator that formulates testable claims, an Experiment Generator that designs analysis plans, an Experiment Programmer that writes Python, a Code Executor that runs it, an Experiment Analyst that interprets results, and an Experiment Reviewer and Reviser that validates and repairs failed experiments. Generated code runs in a sandbox with a 600-second timeout and up to six debugging iterations.
Surprise is quantified as the difference between the means of prior and posterior belief distributions over whether a hypothesis is supported by the data. Beliefs are elicited by repeatedly sampling the model for a categorical judgment across five levels — Definitely True, Maybe True, Uncertain, Maybe False, Definitely False — scored 1, 0.75, 0.5, 0.25 and 0. Hypotheses are deduplicated by LLM-based hierarchical agglomerative clustering and ranked by whether they are positive and by the magnitude of the belief shift. The underlying method was published as Agarwal et al., "Autodiscovery: Open-ended scientific discovery via Bayesian surprise," NeurIPS 38 (2026).
The cold-start result
The unguided run was given only basic information about the structure of the TCGA breast cancer dataset — 1,097 patients with invasive ductal carcinoma (IDC) or invasive lobular carcinoma (ILC), with clinical variables, tumour characteristics, survival outcomes, somatic mutation data for 1,020 patients and gene expression data for 1,068 — and no hypotheses or clinical expert input.
It generated and analysed 100 hypotheses at a mean of 245 seconds each (median 231, range 126–630) and an approximate cost of $0.19 per hypothesis. Of these, 89 were unique and 27 were classified as positive, meaning the post-analysis result was more likely true than the pre-analysis expectation. The paper states that none was judged clinically useful or actionable, and the abstract records the outcome as "The system did not identify clinically meaningful novel findings without human input."
The warm-start result
A breast cancer physician-scientist with immunotherapy expertise annotated the 26 positive hypotheses from the first run (a 27th proved to be a reworded duplicate that automated deduplication had missed), giving brief feedback on interpretive limitations or clinical usefulness. Annotations ran from 1 to 47 words, with a median of 15. Review took roughly 10 minutes per hypothesis and approximately four hours in total.
The second run generated 500 hypotheses using both the dataset structure and the expert annotations. Of these, 15 were duplicates, 20 failed for technical reasons, 355 were negative in direction and 108 were classified as positive. Many positives recapitulated known biology — the highest-ranked implied enrichment of CDH1 alterations in ILC — which the authors treat as evidence the approach works rather than as new findings. Attention narrowed to the 13 hypotheses whose posterior probability rose more than 0.2 above the prior.
The third-ranked of those was that ILC tumours express the immune checkpoint genes PDCD1 and CD274 more highly than IDC tumours after controlling for hormone receptor and HER2 status. The reported belief shift was from a mean of 0.26 before analysis to 0.57 after. The authors describe the hypothesis as both surprising and clinically consequential because ILC has historically been treated as immunologically "cold" and excluded from immunotherapy trials on that basis.
Validation
The hypothesis was tested in METABRIC, an independent European breast cancer cohort in which fewer than 10% of the more than 2,000 cases are lobular. The analysis was restricted to the largest subtype — estrogen receptor-positive, HER2-negative primary tumours — comprising 696 IDC and 74 ILC cases. Median PDCD1 expression was 5.54 in ILC against 5.48 in IDC (p = 0.034), and median CD274 expression 5.51 against 5.45 (p = 0.032), by Mann–Whitney U test.
Tissue validation used multiplex immunofluorescence on 21 formalin-fixed specimens — 12 ILC (2 triple-negative, 4 HER2-positive, 6 hormone receptor-positive/HER2-negative) and 9 non-diseased breast samples — imaged with a Lunaphore COMET sequential immunofluorescence platform over 15 staining cycles and analysed in HALO AI digital pathology software. ILC tumours carried significantly greater numbers of PD-1-positive CD8-positive and CD4-positive T cells than normal breast tissue, a pattern that held within the hormone receptor-positive, HER2-negative subgroup.
By standard clinical immunohistochemistry with the SP263 antibody, only 3 of 12 ILC tumours (25%) were PD-L1 positive at a combined positive score above 1%, and 1 of 12 (8.3%) above 10%. The multiplex method found brisk PD-L1-positive T cells in a further 5 of the 9 tumours scored negative by the conventional assay. A separate system, DataVoyager, was run in user-guided mode against both TCGA and METABRIC to corroborate the signal.
Clinical argument
The authors argue the finding supports testing immunotherapy in lobular breast cancer, including early-stage ER+/HER2− disease, from which patients are currently excluded from large neoadjuvant immunotherapy trials. They note that a systematic review of 93 phase III/IV breast cancer trials found ILC excluded in 39, documented as included in 13, and analysed separately in 3. KEYNOTE-756 and CheckMate 7FL both made lobular pathology an exclusion criterion, despite an earlier response signal in KEYNOTE-028, where the objective response rate in ILC was 67% (2 of 3) against 8% (1 of 13) in IDC. The GELATO phase II study, which enrolled ILC alone, reported a 26% clinical benefit rate for checkpoint inhibition combined with low-dose carboplatin. SWOG2206 includes but does not stratify for IDC versus ILC.
Stated limitations
The authors state they lacked the resources to validate every surprising hypothesis the system generated, anticipate that some will be wrong, and hold that all findings require downstream validation. They also record that "AutoDiscovery did not operate in a fully unguided manner; performance was markedly improved by expert annotation rather than a true cold-start framework," and characterise the result as a hybrid in which human expertise and machine inference together outperform either alone. The abstract's closing claim is framed the same way: that domain experts working with LLMs "can derive more meaningful insights from complex data than either could achieve alone."
Provenance
Computational and tissue studies were approved under Providence IRB studies 2025000994 and 2024000051. Full code, analyses and results are published at autodiscovery.allen.ai. Funding came from the Paul G. Allen Research Center, the Swedish Foundation, the Kuni Foundation and Ai2. Several authors declare institutional funding and advisory income from pharmaceutical companies, listed in the paper's competing-interests statement.
Coverage of the paper, including Ai2's own announcement blog of 27 August 2026, renders the title as "...in breast cancer," dropping "lobular" (Source: allenai.org). The title as printed on the paper is used here.
Relationships
- supports: AI for Science — quantified cost, throughput and yield for LLM-driven hypothesis generation on an existing dataset.
- instance-of: AI for Science
- related: Allen Institute for AI (Ai2), Healthcare — AI Deployment, Agentic AI