Ewan K.A. Millar

The analysis of histopathology images with artificial intelligence aims to enable clinical decision support systems and precision medicine. The success of such applications depends on the ability to model the diverse patterns observed in pathology images. To this end, we present Virchow, the largest foundation model for computational pathology to date. In addition to the evaluation of biomarker prediction and cell identification, we demonstrate that a large foundation model enables pan-cancer detection, achieving 0.95 specimen-level area under the (receiver operating characteristic) curve across nine common and seven rare cancers. Furthermore, we show that with less training data, the pan-cancer detector built on Virchow can achieve similar performance to tissue-specific clinical-grade models in production and outperform them on some rare variants of cancer. Virchow's performance gains highlight the value of a foundation model and open possibilities for many high-impact applications with limited amounts of labeled training data.

PURPOSE: To determine the clinical utility of intrinsic molecular phenotype after breast-conserving therapy (BCT) with lumpectomy and whole-breast irradiation with or without a cavity boost. PATIENTS AND METHODS: Four hundred ninety-eight patients with invasive breast cancer were enrolled into a randomized trial of BCT with or without a tumor bed radiation boost. Tumors were classified by intrinsic molecular phenotype as luminal A or B, HER-2, basal-like, or unclassified using a five-biomarker panel: estrogen receptor, progesterone receptor, HER-2, CK5/6, and epidermal growth factor receptor. Kaplan-Meier and Cox proportional hazards methodology were used to ascertain relationships to ipsilateral breast tumor recurrence (IBTR), locoregional recurrence (LRR), distant disease-free survival (DDFS), and death from breast cancer. RESULTS: Median follow-up was 84 months. Three hundred ninety-four patients were classified as luminal A, 23 were luminal B, 52 were basal, 13 were HER-2, and 16 were unclassified. There were 24 IBTR (4.8%), 35 LRR (7%), 47 distant metastases (9.4%), and 37 breast cancer deaths (7.4%). The overall 5-year disease-free rates for the whole cohort were: IBTR 97.4%, LRR 95.6%, DDFS 92.9%, and breast cancer-specific death 96.3%. A significant difference was observed for survival between subtypes for LRR (P = .012), DDFS (P = .0035), and breast cancer-specific death (P = .0482), but not for IBTR (P = .346). CONCLUSION: The 5-year and 10-year survival rates varied according to molecular subtype. Although this approach provides additional information to predict time to IBTR, LRR, DDFS, and death from breast cancer, its predictive power is less than that of traditional pathologic indices. This information may be useful in discussing outcomes and planning management with patients after BCT.

Breast cancer is a common malignancy with current biological therapies tailored to steroid hormone (ER, PR) and HER2 receptor status. Understanding the biological basis of resistance to current targeted therapies and the identification of new potential therapeutic targets is an ongoing challenge. The PI3K pathway is altered in a high proportion of breast cancers and may contribute to therapeutic resistance. We undertook an integrative study of mutational, copy number and expression analyses of key regulators of the PI3K pathway in a cohort of 292 invasive breast cancer patients with known treatment outcomes. The alterations identified in this cohort included PIK3CA mutations (12/168, i.e. 7%), PIK3CA copy number gain (28/209, i.e. 14%), PTEN loss (73/258, i.e. 28%) and AKT activation (62/258, i.e. 24%). Overall at least 1 parameter was altered in 72% (139/193) of primary breast cancers. PI3K pathway activation was significantly associated with ER negative (p = 0.0008) and PR negative (p = 0.006) status, high tumor grade (p = 0.032) and a "basal-like" phenotype (p = 0.01), where 92% (25/27) of tumors had an altered pathway. In univariate analysis, PI3K pathway aberrations were associated with death from breast cancer; however, this relationship was not maintained in multivariate analysis. No association was identified between an activated pathway and outcome in tamoxifen- or chemotherapy-treated patients. We concluded that >70% of breast cancers have an alteration in at least 1 component of the PI3K pathway and this might be exploited to therapeutic advantage especially in "basal-like" cancers.

Inositol polyphosphate 4-phosphatase-II (INPP4B) is a regulator of the phosphoinositide 3-kinase (PI3K) signaling pathway and is implicated as a tumor suppressor in epithelial carcinomas. INPP4B loss of heterozygosity (LOH) is detected in some human breast cancers; however, the expression of INPP4B protein in breast cancer subtypes and the normal breast is unknown. We report here that INPP4B is expressed in nonproliferative estrogen receptor (ER)-positive cells in the normal breast, and in ER-positive, but not negative, breast cancer cell lines. INPP4B knockdown in ER-positive breast cancer cells increased Akt activation, cell proliferation, and xenograft tumor growth. Conversely, reconstitution of INPP4B expression in ER-negative, INPP4B-null human breast cancer cells reduced Akt activation and anchorage-independent growth. INPP4B protein expression was frequently lost in primary human breast carcinomas, associated with high clinical grade and tumor size and loss of hormone receptors and was lost most commonly in aggressive basal-like breast carcinomas. INPP4B protein loss was also frequently observed in phosphatase and tensin homolog (PTEN)-null tumors. These studies provide evidence that INPP4B functions as a tumor suppressor by negatively regulating normal and malignant mammary epithelial cell proliferation through regulation of the PI3K/Akt signaling pathway, and that loss of INPP4B protein is a marker of aggressive basal-like breast carcinomas.