Ataraxis Breast

Genomic assays such as Oncotype DX read prognosis out of gene expression, and the price is physical: tissue is consumed, results take ten to thirty days, and most patients who take one land in an intermediate band that does not resolve a treatment decision — 526 of the 858 assayed patients in this study's comparison cohorts. Meanwhile the H&E slide the pathologist has already read carries prognostic morphology nobody has to specify in advance: tumor architecture, stromal texture, the placement of the immune infiltrate.

Ataraxis Breast is the test built on that observation. Developed by Ataraxis AI with NYU Grossman School of Medicine and a dozen cancer centers across seven countries, it passes a digitized H&E slide from a core needle biopsy or resection through Kestrel — the company's pan-cancer pathology foundation model, a vision transformer self-supervised with DINOv2 on 400 million tissue patches — and feeds the resulting patch embeddings to supervised time-to-event models. A second score comes from six routine clinical variables, and the two are averaged into a continuous risk value between 0 and 1.

Architecturally this is a thin supervised head on a frozen general-purpose backbone, and Kestrel belongs to the same generation of DINOv2-pretrained pathology encoders as UNI and Prov-GigaPath. Where CHIEF spreads one backbone across many prediction tasks, this work aims a single head at one endpoint and spends its effort evaluating it across independent patient populations.

#Key Features

• Frozen pan-cancer feature extractor: Kestrel is a 303M-parameter ViT-L trained with DINOv2 on 400 million 256x256 patches from 45,000 histopathology slides. Each input slide yields 8,000 to 16,000 foreground patches at 20x magnification.
• Two-branch ensemble: ten pathology models — Cox proportional-hazards and discrete-time heads varying in pooling, loss and feature extractor — are averaged, as are ten clinical models, and the two modality scores are averaged again.
• Six routine clinical inputs: T and N stage, patient age, ER, PR and HER2 status, and ductal versus lobular histology, all of which already appear on a standard pathology report.
• One checkpoint across every evaluation cohort: fit once on 4,659 patients in 10 cohorts, then applied unchanged to 3,502 patients in five cohorts excluded from training and hyperparameter selection.
• Coverage beyond HR+/HER2−: prognostic in triple-negative and HER2-positive disease, subtypes for which NCCN guidelines recommend no prognostic assay.

#Technical Details

For the primary endpoint, disease-free interval, the pooled concordance index across the five evaluation cohorts was 0.71 [0.68–0.75], with a hazard ratio of 3.63 [3.02–4.37] per 0.2-unit increase in score. Per cohort: 0.74 in Providence (n=1,733), 0.70 in TCGA-BRCA (n=911), 0.70 in Basel (n=269), 0.67 in UChicago (n=421) and 0.62 in Karmanos (n=168). Pooled time-dependent AUC was 0.76, 0.71 and 0.73 at three, five and seven years, and the score stayed prognostic for distant recurrence-free interval (0.70) and overall survival (0.65).

Across the 858 patients also assayed with Oncotype DX, the pooled C-index was 0.67 [0.61–0.74] against 0.61 [0.49–0.73]; the journal version presents this as numerically higher discrimination, with overlapping confidence intervals. In a multivariate Cox model adjusting for Oncotype DX score, Nottingham grade, race and dataset, the AI score carried an adjusted hazard ratio of 2.95 [1.82–4.79, p < 0.001] while Oncotype DX's was not significant (1.43 [0.91–2.27, p = 0.12]). The high/low cutoff is the 80th percentile of scores in the HR+/HER2− subpopulation of the NYU training cohort, a threshold the authors call preliminary.

#Applications

Because the input is a slide already cut at diagnosis, the workflow adds no wet-lab step: accessioning, scanning and inference take under an hour against ten to thirty days for a genomic assay, and the tissue block survives for later sequencing if the disease progresses. The paper cites a Medicare price of $706 for the first digital pathology procedure codes against roughly $3,800 for genomic assays. In the Oncotype-tested cohorts the model moved 666 of 858 patients into a different risk category and gave all 526 intermediate-risk patients a low or high call. Ataraxis AI sells the model commercially as Ataraxis Breast.

#Impact

The work appeared as a preprint in October 2024 and in Nature Communications in 2026, and its evaluation is among the largest assembled for a breast cancer prognostic model: 8,161 women across 15 cohorts in seven countries, with more than 40% held out for external testing. The evidence is retrospective and observational, and the test is prognostic rather than predictive — it was never trained to model treatment effect, so a high score does not identify patients who will benefit from chemotherapy. The study population is women with non-metastatic invasive breast cancer, the Oncotype DX comparison is confined to HR+/HER2− patients in three of the five evaluation cohorts, and the codebase and weights are proprietary and unreleased, leaving external groups unable to compute the score independently.