Typing system and method for HR-positive / her2-negative breast cancer

By classifying HR-positive and HER2-negative breast cancer and developing personalized treatment plans, the lack of precise treatment guidance in existing technologies has been addressed, improving treatment outcomes and reducing the risk of recurrence and metastasis.

WO2024235337A1PCT designated stage expired Publication Date: 2024-11-21LUMITYPE MEDICAL (HUZHOU) CO LTD
View PDF 13 Cites 0 Cited by

Patent Information

Application Number
PCT/CN2024/094317
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2023-09-19
Filing Date
2024-05-20
Publication Date
2024-11-21

AI Technical Summary

Technical Problem

Current technologies lack precise treatment guidance for HR-positive HER2-negative breast cancer, resulting in limited efficacy of endocrine therapy and a high risk of distant recurrence and metastasis for patients.

Method used

Establish a classification system for HR-positive and HER2-negative breast cancer, classifying patients into four subtypes: classic luminal type, immunomodulatory type, proliferative type, and receptor tyrosine kinase-driven type. Based on the characteristics of each subtype, targeted treatment plans will be adopted, such as endocrine therapy, immunotherapy, CDK4/6 inhibitors, PARP inhibitors, or tyrosine kinase inhibitors.

Benefits of technology

It enables precise treatment of HR-positive and HER2-negative breast cancer patients, improves treatment outcomes, reduces the risk of distant recurrence and metastasis, and provides better clinical prognosis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN2024094317_21112024_PF_FP_ABST
    Figure CN2024094317_21112024_PF_FP_ABST
Patent Text Reader

Abstract

The present invention relates to a novel typing system and method for HR-positive / HER2-negative breast cancer. According to a multi-omics result of HR-positive / HER2-negative breast cancer, HR-positive / HER2-negative breast cancer has four subtypes, namely a classic luminal subtype, an immunomodulatory subtype, a proliferative subtype, and a receptor tyrosine kinase-driven subtype. Different treatment schemes are adopted for the four subtypes, so that accurate treatment for different subtypes of HR-positive / HER2-negative breast cancer can be achieved. The present invention further relates to a typing method for HR-positive / HER2-negative breast cancer based on AI digital pathology or a classifier and gene expression.
Need to check novelty before this filing date? Find Prior Art

Description

Classification system and method for HR-positive HER2-negative breast cancer

[0001] Cross-references

[0002] This application claims priority to Chinese application No. 202310567930.4 filed on May 18, 2023, Chinese application No. 202311210735.2 filed on September 19, 2023, Chinese application No. 202311210734.8 filed on September 19, 2023, and PCT application No. PCT / CN2023 / 119869 filed on September 19, 2023, and each priority application is incorporated by reference in its entirety for all purposes. Technical Field

[0003] The present invention relates to the field of cancer diagnosis and treatment, and in particular to the typing of HR-positive HER2-negative breast cancer. Background Art

[0004] Luminal breast cancer is mainly caused by breast cancer that is positive for estrogen (PR) or progesterone receptor (ER) (collectively referred to as hormone receptor HR) and negative for human epidermal growth factor receptor 2 (HER2) (HR + HER2 - ) and accounts for two-thirds of all breast cancers. Moreover, the vast majority of breast cancer recurrence and metastasis cases occur in patients with caval-facial breast cancer.

[0005] Luminal facial breast cancer is highly responsive to endocrine therapy, but factors such as tumor heterogeneity and endocrine resistance contribute to the continued risk of distant recurrence and metastasis. Despite the recent emergence of numerous immunotherapies and targeted therapies targeting luminal facial breast cancer, systematic guidance for the precise treatment of luminal facial breast cancer remains lacking. Therefore, further exploration of the heterogeneity of luminal facial breast cancer is needed to establish a more precise treatment system for this disease.

[0006] Summary of the Invention

[0007] The present invention provides a precision treatment regimen for HR-positive HER2-negative breast cancer, which is achieved by establishing a new classification system for HR-positive HER2-negative breast cancer. Specifically, the present invention provides a method for implementing precision treatment on patients with HR-positive HER2-negative breast cancer, the method comprising: determining the subtype of HR-positive HER2-negative breast cancer suffered by the patient, wherein the subtype is one of the following four types: classic luminal facial type (SNF1), immune regulatory type (SNF2), proliferative type (SNF3) and receptor tyrosine kinase driven type (SNF4); implementing precision treatment on the patient using the following regimen: endocrine therapy for patients with classic luminal facial type, immunotherapy for patients with immune regulatory type, administration of cyclin-dependent kinase (CDK) 4 / 6 inhibitors, or poly (ADP-ribose) polymerase (PARP) inhibitors, or a combination of CDK4 / 6 inhibitors and PARP inhibitors to patients with proliferative type, and administration of tyrosine kinase inhibitors such as platelet-derived growth factor receptor (PDGFR) signaling pathway inhibitors to patients with receptor tyrosine kinase driven type.

[0008] The present invention provides a HR-positive HER2-negative breast cancer classification system, which divides HR-positive HER2-negative breast cancer into the following four subtypes: classic luminal, immune-regulated, proliferative, and receptor tyrosine kinase-driven. The immunohistochemical scores of these four subtypes are:

[0009] Classic luminal surface type: α-SMA < 20%, CD8 < 3%, Ki67 < 25%, and all three conditions are met at the same time;

[0010] Immunomodulatory type: α-SMA < 20%, CD8 ≥ 3%, both of the above conditions are met;

[0011] Proliferative type: α-SMA < 20%, CD8 < 3%, Ki67 ≥ 25%, all three conditions are met;

[0012] Receptor tyrosine kinase-driven: α-SMA ≥ 20%.

[0013] The present invention also provides a method for classifying HR-positive HER2-negative breast cancer, which classifies HR-positive HER2-negative breast cancer into the following four subtypes: classic luminal subtype, immune-regulated subtype, proliferative subtype, and receptor tyrosine kinase-driven subtype. The method comprises: performing immunohistochemical scoring on the breast cancer and making the following determinations:

[0014] If α-SMA < 20%, CD8 < 3%, Ki67 < 25%, and the above three conditions are met at the same time, it is determined to be a classic luminal facial type; or

[0015] If α-SMA < 20%, CD8 ≥ 3%, and both conditions are met, the patient is classified as immune regulatory; or

[0016] If α-SMA < 20%, CD8 < 3%, Ki67 ≥ 25%, and the above three conditions are met at the same time, it is judged as proliferative type; or

[0017] If α-SMA ≥ 20%, the disease is determined to be receptor tyrosine kinase driven.

[0018] In some embodiments, the tumor sample is stained, for example, with hematoxylin-eosin, prior to immunohistochemistry.

[0019] The present invention also provides a system for typing HR-positive HER2-negative breast cancer, the system comprising:

[0020] 1) a full-scan digital pathology image data preprocessing module, which is used to cut the full-scan digital pathology image of the tumor tissue sample of the HR-positive HER2-negative breast cancer into a plurality of enlarged image blocks;

[0021] 2) a first neural network module, wherein the first neural network module is a tissue type segmentation neural network module, which is used to classify the image blocks into different tissue types and attach corresponding tissue labels;

[0022] 3) a second neural network module connected in series with the first neural network module, wherein the second neural network module is a typing prediction neural network module, and is used to perform typing on the sample based on the image block with the tissue label;

[0023] Among them, the classification refers to dividing HR-positive HER2-negative breast cancer into the following four subtypes: classic luminal type, immune regulatory type, proliferative type, and receptor tyrosine kinase driven type.

[0024] The present invention also provides a chip, which includes a processor for calling and running a computer program from a memory, so that a device equipped with the chip executes the typing method of the present invention.

[0025] The present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the typing method of the present invention is implemented.

[0026] The present invention provides a precise treatment plan for HR-positive HER2-negative breast cancer.

[0027] In one of the above-mentioned precision treatment plans of the present invention, the HR-positive HER2-negative breast cancer is determined to be of the classic cavitary type, and the patient is treated with endocrine therapy or given a PI3K inhibitor or a CDK4 / 6 inhibitor.

[0028] In one of the above-mentioned precision treatment schemes of the present invention, the HR-positive HER2-negative breast cancer is determined to be an immune-regulatory type, and the patient is subjected to immunotherapy.

[0029] In one of the above-mentioned precision treatment regimens of the present invention, the HR-positive HER2-negative breast cancer is determined to be a proliferative type, and the patient is given a cyclin-dependent kinase (CDK) 4 / 6 inhibitor, or a poly (ADP-ribose) polymerase (PARP) inhibitor, or a combination of a CDK4 / 6 inhibitor and a PARP inhibitor.

[0030] In one of the above-mentioned precision treatment schemes of the present invention, the HR-positive HER2-negative breast cancer is determined to be receptor tyrosine kinase-driven, and a tyrosine kinase inhibitor is administered to the patient.

[0031] In a preferred embodiment, the endocrine therapy comprises the administration of an estrogen receptor antagonist, preferably 4-hydroxytamoxifen and fulvestrant.

[0032] In a preferred embodiment, the immunotherapy comprises administering an immune checkpoint inhibitor, preferably a PD-1 antibody, a PD-L1 antibody, or a CTLA4 antibody.

[0033] In a preferred embodiment, the tyrosine kinase inhibitor is a VEGFR or PDGFR signaling pathway inhibitor, preferably famitinib, sorafenib, regorafenib, vandetanib, afatinib, lapatinib, pyrotinib, axitinib, sunitinib, foretinib, apatinib, or sitravatinib.

[0034] In a preferred embodiment, the CDK4 / 6 inhibitor is abemaciclib; and the PARP inhibitor is olaparib.

[0035] The present invention further provides the use of the substance in the preparation of a drug for the precision treatment of HR-positive HER2-negative breast cancer, specifically including the following applications:

[0036] (1) Use of an estrogen receptor antagonist in the preparation of a drug for treating classic luminal breast cancer. The estrogen receptor antagonist is preferably 4-hydroxytamoxifen and fulvestrant.

[0037] (2) Use of immune checkpoint inhibitors in the preparation of drugs for treating immunomodulatory breast cancer. The immune checkpoint inhibitors are preferably PD-1 antibodies, PD-L1 antibodies, or CTLA4 antibodies.

[0038] (3) Use of a CDK4 / 6 inhibitor, a PARP inhibitor, or a combination of a CDK4 / 6 inhibitor and a PARP inhibitor in the preparation of a medicament for treating proliferative breast cancer. The CDK4 / 6 inhibitor is preferably abemaciclib; the PARP inhibitor is preferably olaparib.

[0039] (4) Use of a tyrosine kinase inhibitor in the preparation of a drug for treating receptor tyrosine kinase-driven breast cancer. The tyrosine kinase inhibitor is preferably a VEGFR or PDGFR signaling pathway inhibitor, or preferably famitinib, sorafenib, regorafenib, vandetanib, afatinib, lapatinib, pyrotinib, axitinib, sunitinib, foretinib, apatinib, or sitravatinib.

[0040] Based on a new HR-positive HER2-negative breast cancer typing system, the present invention establishes a new HR-positive HER2-negative breast cancer typing method, and further establishes a precision treatment system for HR-positive HER2-negative breast cancer. The digital pathology and artificial intelligence-based HR-positive HER2-negative breast cancer typing system and method provided by the present invention can easily and efficiently perform the typing of HR-positive HER2-negative breast cancer patients in the clinic, thus providing the premise and means for achieving precise diagnosis and treatment of HR-positive HER2-negative breast cancer at the clinical level.

[0041] Based on the sensitivity of different subtypes of HR-positive HER2-negative breast cancer to different treatment options, the present invention can achieve better therapeutic effects than traditional breast cancer treatment options by implementing different precision treatment options, enabling it to achieve a more ideal prognosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] FIG1 is a summary of the four classifications of HR-positive HER2-negative breast cancer according to the present invention.

[0043] FIG2 is the distribution of PAM50 subtypes in the multi-omics cohort of the present invention.

[0044] FIG3 is a diagram showing the determination of the optimal number of clusters for the SNF cluster analysis of the present invention.

[0045] FIG4 is a silhouette plot of the SNF clustering of the present invention.

[0046] FIG5 shows the results of immunohistochemical analysis of SNF1-4 subtypes.

[0047] FIG6 shows the receiver operating characteristic curve analysis results of the three markers used in the immunohistochemical analysis.

[0048] Figure 7 is the result of comparing the predicted SNF typing proportion based on immunohistochemistry score of the present invention with the SNF typing proportion of the luminal-facial multi-omics cohort.

[0049] Figure 8 is the technical roadmap for digital pathology classification.

[0050] Figure 9 is the confusion matrix of the validation set classification results.

[0051] Figure 10 shows the comparison between the tissue segmentation results and the original pathological image.

[0052] FIG11 is a schematic diagram of the three-fold cross validation used in the modeling process of the present invention.

[0053] FIG12 is the ROC curve of each neural network model in the validation set for identifying the corresponding subtype.

[0054] Figure 13 shows typical pathological morphology of each subtype. Each subtype corresponds to four image blocks, of which the top two are the original WSI segmentation images, and the bottom two are the original images superimposed with the segmentation heatmap.

[0055] Figures 14-22 show the verification of key features of digital pathology.

[0056] FIG. 23 is a pathological image showing the survival of SNF1-4 subtypes under traditional endocrine therapy.

[0057] FIG. 24 shows the relative survival rates of SNF1-4 subtypes in response to traditional endocrine therapy.

[0058] FIG25 is a heat map showing the expression of the SNF1-4 subtype endocrine response module.

[0059] FIG26 is a comparison of flow cytometry results of SNF2 and non-SNF2 isoforms after immunotherapy, showing an increase in CD3+ total T activation markers.

[0060] Figure 27 is a comparison of flow cytometry results of SNF2 and non-SNF2 isoforms after immunotherapy, showing increased cytokines.

[0061] Figure 28 shows the relationship between SNF1-4 subtypes in early and late HR + / HER2 - Sensitivity to endocrine therapy in a real-world cohort of breast cancer.

[0062] Figure 29 is an analysis of PFS time after first-line endocrine therapy for SNF3 and non-SNF3 subtypes.

[0063] Figure 30 shows the sensitivity of SNF3 and non-SNF3 isoforms to endocrine therapy after long-term induction with 4-hydroxytamoxifen.

[0064] Figures 31 and 32 show that the PDO growth ability of SNF3 and non-SNF3 subtypes was significantly inhibited by abemaciclib and olaparib, respectively.

[0065] FIG33 shows that the combination of abemaciclib and olaparib significantly reduced the survival rate of SNF3-PDOs.

[0066] FIG34 shows the RNA-seq results (GSVA scores) of SNF1-4 subtype PDO.

[0067] FIG35 shows homologous recombination deficiency scores for SNF3 and non-SNF3 subtype tumors.

[0068] Figure 36 is the DRUG-seq results of SNF3 subtype tumors, showing the changes in signaling pathways after combined drug administration.

[0069] Figure 37 shows receptor expression in the SNF1-4 subtype receptor tyrosine kinase pathway.

[0070] Figure 38 shows the IC values ​​of different tyrosine kinase inhibitors against SNF4 tumors 50 value.

[0071] FIG39 is a pathological image showing the sensitivity of SNF4 tumors to different tyrosine kinase inhibitors.

[0072] FIG40 shows changes in PDGF-activated signaling pathways in SNF4 tumors after treatment with famitinib and sorafenib.

[0073] FIG41 shows the changes in SNF4 tumor cell viability after treatment with regorafenib and vandetanib.

[0074] Figure 42 shows that SNF4-expressing cancer-associated fibroblasts (CAFs) enhance tumor growth and are sensitive to sorafenib. (A) Heatmaps showing bulk RNA-seq (left) and scRNA-seq (center) expression of the RKT list in four SNF subtypes. The heatmap on the right shows the differences in phosphorylation abundance between SNF4 and other genes. These include kinase genes listed in "GOMF_TRANSMEMBRANE_RECEPTOR_PROTEIN_KINASE_ACTIVITY" (obtained from http: / / www.gsea-msigdb.org / gsea / msigdb) as well as previously published curated protein kinase gene lists. 58Protein kinase gene list. Phospho-sites enriched in SNF4 and detected at P < 0.05 are highlighted in purple. P values ​​are derived from a two-sided t-test. NA: No phospho-site detected.

[0075] (BC) Western blot images showing the expression of EGFR, PDGFRA, p-PDGFRA, and p-EGFR in SNF4 tumors or non-SNF4 tumors (B) and SNF4 tumor cells or SNF4 CAFs (C). All Western blot experiments were repeated three times with similar results.

[0076] (D-E) Tumor growth viability assay of SNF1 (n = 2) and SNF4 (n = 2) patient-derived cancer cells cultured in SNF4 tumor CAF-conditioned medium (CM). Scale bars: 100 μm. P(PDO#5) = 4.9e-05, P(PDO#4) = 4.1e-05, P(PDO#18) = 5.3e-05. P values ​​are derived from a two-sided Wilcoxon test. Box plots show the first and third quartiles as lower and upper limits, whiskers represent 1.5 times the IQR, and the center represents the median.

[0077] (FG) Efficacy of sorafenib (1 μM) against patient-derived cancer cells co-cultured with fibroblasts. Scale bar: 100 μm. Cancer cells were treated three times with sorafenib. Shown is one representative experiment. The experiment was repeated three times with similar results.

[0078] (H) Schematic showing how CAFs enhance the growth of RTK-driven (SNF4) subtype tumors.

[0079] Figure 43 shows the NG model building process (left) and the simplified model building process (right).

[0080] Figure 44 shows the overall real-world study design. (A) Molecular characterization of four HR+ / HER2- breast cancer subtypes. (B) Integration of data from multi-omics cohorts, real-world studies, drug testing platforms, and prospective clinical studies to validate precision medicine strategies for HR+ / HER2- breast cancer subtypes. CIN, chromosomal instability; RTK, receptor tyrosine kinase; CNA, copy number alteration; ER, estrogen receptor; HR, hormone receptor; HER2, human epidermal growth factor receptor 2.

[0081] Figure 45 shows the clinical characteristics and response to conventional therapy of a real-world cohort of HR+ / HER2- breast cancer patients, grouped by subtype. (A) Schematic diagram of the study design for the real-world cohort of HR+ / HER2- breast cancer. (B) Relapse-free survival of patients with early-stage HR+ / HER2- breast cancer receiving endocrine therapy, grouped by subtype. (C) Bar graph showing the distribution of primary and metastatic HR+ / HER2- breast cancer subtypes in the real-world cohort. P values ​​were calculated using the chi-square test. (D) Progression-free survival of patients with advanced HR+ / HER2- breast cancer receiving fulvestrant backbone therapy, grouped by subtype, in the FUSCC real-world cohort. (E) Efficacy of CDK4 / 6i in patients with advanced HR+ / HER2- breast cancer in the FUSCC real-world cohort, according to RECIST v1.1. P values ​​were calculated using the chi-square test. FUSCC, Fudan University Cancer Hospital; PR, primary lesion; MET, metastatic lesion; CDK4 / 6i, CDK4 / 6 inhibitor; PR, partial response; HR, hormone receptor; HER2, human epidermal growth factor receptor 2; RECIST, response evaluation criteria in solid tumors. Significance code: NS, not significant; *, P < 0.05; **, P < 0.01; ***, P < 0.001.

[0082] Figure 46 shows the validation of a subtype-directed precision therapy strategy in a prospective clinical cohort. (A) Schematic diagram of the Phase 1 analysis of the MULAN study. (B) Duration of immunotherapy in the fifth cohort of evaluable patients in the MULAN cohort. The dashed line indicates the data at 6 months of treatment. (C) Durability of CDK4 / 6i combined with PARPi in the second cohort of evaluable patients in the MULAN cohort. The dashed line indicates the data at 6 months of treatment. (D) Objective response rates were evaluated in the MULAN cohort for patients who received a subtype-directed precision therapy strategy (matched group) and those who did not receive matched treatment (unmatched group). (E) TKI persistence in a real-world clinical cohort of evaluable patients. The dashed line indicates the data at 6 months of treatment. (F) Progression-free survival of patients with advanced metastatic breast cancer who had received ≥3 prior lines of therapy using a matched and unmatched precision therapy strategy. (G) Baseline and representative images of tumor regression during matched precision therapy. (Top) Patients with the SNF2 subtype in the fifth cohort received immunotherapy; (Middle) Patients with the SNF3 subtype in the second cohort received CDK4 / 6i and PARPi; (Bottom) Patients with the SNF4 subtype in a real-world cohort received TKI treatment. (VIII) Study Summary. By integrating clinical cohort studies, omics analysis, and functional testing, we revealed the heterogeneity of treatment response in HR+ / HER2- breast cancer and proposed a subtype-guided precision medicine strategy. Patient subtypes were identified based on digital pathology using deep learning. The best overall treatment response was defined according to RECIST v1.1 criteria. CDK4 / 6i, CDK4 / 6 inhibitor; PARPi, PARP inhibitor; CR, complete response; PR, partial response; SD, stable disease; PD, progressive disease; TKI, tyrosine kinase inhibitor; mPFS, median progression-free survival; CI, confidence interval; RECIST, response evaluation criteria in solid tumors; HR, hormone receptor; HER2, human epidermal growth factor receptor 2. DETAILED DESCRIPTION

[0083] The systems, devices, and methods disclosed herein are described in detail by way of example and with reference to the accompanying drawings. The examples discussed herein are merely exemplary and are provided to help explain the apparatus, devices, systems, and methods described herein. No feature or component shown in the accompanying drawings or discussed below should be considered mandatory for any particular implementation of any of these devices, systems, or methods unless specifically designated as mandatory.

[0084] Furthermore, for any method described, whether or not the method is described in conjunction with a flowchart, it should be understood that, unless otherwise specified or required by the context, any explicit or implicit ordering of steps performed in the execution of the method does not imply that those steps must be performed in the order presented, but may instead be performed in a different order or in parallel.

[0085] As used herein, the term "exemplary" is used in the sense of "example," rather than "best" or "ideal." Additionally, the terms "a" and "an" herein do not denote a limitation of quantity, but rather denote the presence of one or more of the referenced items.

[0086] To explore the heterogeneity of HR-positive HER2-negative breast cancer, we optimized the traditional HR + / HER2 — Breast cancer treatment strategy, to build a precision treatment system for HR-positive HER2-negative breast cancer. The present invention uses multi-omics data such as genome, transcriptome, proteome and / or metabolome, and performs multi-omics fusion clustering through the similarity network fusion (SNF) algorithm to divide HR-positive HER2-negative breast cancer into four types: classic luminal (Canonical Luminal, SNF1), immunoregulatory (Immunogenic, SNF2), proliferative (Proliferative, SNF3) and receptor tyrosine kinases (RTK) driven (RTK-driven, SNF4). The characteristic genomic events of each subtype are revealed, and based on this, potential treatment options for each subtype are suggested. 1) Classic luminal: The prognosis is good, and the main characteristics are that it is basically composed of PAM50 LumA / B, high expression of hormone receptor-related pathways and high-frequency PIK3CA mutations. PI3K inhibitors can be considered for subsequent treatment. 2) Immune regulatory type: There is high expression of immune checkpoint molecules and high levels of tumor infiltrating lymphocytes, and there are a large number of PAM50 basal cell-like types and HER2 overexpression types mixed together, suggesting that immune checkpoint inhibitors may have a better effect on immune regulatory HR-positive HER2-negative breast cancer. 3) Proliferative type: Cell cycle pathways are activated and accompanied by CCND1, MDM2 amplification and ATM loss. E2F1, E2F2, and E2F-targeted genes are all highly expressed, suggesting that CDK4 / 6 inhibitors combined with PARP inhibitors can be used for this subtype. 4) Receptor tyrosine kinase-driven type: The prognosis is the worst, containing almost all PAM50 normal breast-like subtypes, with high EGFR expression and activation of receptor tyrosine kinase-related pathways, suggesting that it may benefit from receptor tyrosine kinase inhibitor treatment (see Figure 1).

[0087] The transcriptomic characteristics of each subtype can be summarized as follows:

[0088] SNF1: It is a mixed A / B isoform of the PAM50 luminal surface, enriched in hormone receptor-related pathways;

[0089] SNF2: This subtype encompasses nearly all PAM50 HER2-overexpressing and basal-like subtypes, with a relatively low proportion of PAM50 luminal A / B subtypes. It exhibits low activation of hormone receptor-related pathways and is enriched in numerous adaptive immune response pathways. Compared to other subtypes, SNF2 has lower estrogen receptor expression, higher expression of several immune checkpoint biomarkers, including the tumor inflammation index and programmed cell death protein-1 (PDCD1), and significantly elevated mRNA levels of several cytotoxic factors, including cluster of differentiation CD8A (CD8A), granzyme A (GZMA), and perforin 1 (PRF1).

[0090] SNF3: mainly composed of the PAM50 luminal B subtype, with more activated cell cycle pathways, high mRNA expression of CND1, CDK1, and CDK2, high mRNA expression of E2F1, E2F2, and E2F target genes, high mRNA expression of the G2 / M cell cycle activation gene MDM2, low mRNA expression of the G2 / M cell cycle inhibitory gene ATM, and enrichment of hormone receptor-related pathways;

[0091] SNF4: Contains a high proportion of PAM50 normal breast-like subtypes, low activation levels of hormone receptor-related pathways, enrichment of matrix component-related pathways, highly enriched expression features related to receptor protein kinases and fibroblasts, significantly enriched receptor tyrosine kinase pathways, and high expression of EGFR and several key receptor tyrosine kinases.

[0092] The metabolomic characteristics of each subtype can be summarized as follows:

[0093] SNF1: The metabolomic characteristics were not obvious, and no significantly upregulated or downregulated metabolic pathways were found;

[0094] SNF2: L-kynurenine and indoleamine 2,3-dioxygenase 1 (IDO1, the rate-limiting enzyme in the kynurenine metabolic pathway) levels are higher, and SNF2 patients with a worse prognosis have a higher abundance of phosphatidylserine;

[0095] SNF3: All metabolic pathways were more active than those of other subtypes. The abundance of amino acid and nucleotide-related metabolites was relatively high, while the abundance of citric acid cycle-related metabolites was relatively low. The abundance of glutamine and L-asparagine, as well as the expression of the amino acid transporter solute carrier 1A5 (SLC1A5), were increased.

[0096] SNF4: The overall metabolism is weaker than other subtypes, showing a metabolic "inert" state.

[0097] The proteomic characteristics of each subtype can be summarized as follows:

[0098] SNF1: The overall protein changes are similar to those in the transcriptome, see the relevant description in the transcriptome section;

[0099] SNF2: Protein changes are similar to those in the transcriptome (see transcriptome section for details). Typical events include significant increases in the protein levels of several cytotoxic factors, including CD8A, GZMA, and PRF1.

[0100] SNF3: The overall protein changes are similar to those of the transcriptome, see the relevant description in the transcriptome section, among which typical events include higher protein levels of CCND1, CDK1, and CDK2;

[0101] SNF4: The overall protein changes are similar to those of the transcriptome. See the transcriptome section for related descriptions.

[0102] The microenvironmental characteristics of each subtype can be summarized as follows:

[0103] SNF1: low expression of microenvironment-related molecules, tending to be cold tumor characteristics;

[0104] SNF2: Multiple immune cell subsets, including adaptive immune cells, increased significantly, with higher levels of iTILs and sTILs, and a higher diversity of T cell receptors (T cell receptor α and T cell receptor β), favoring hot tumor characteristics;

[0105] SNF3: low expression of microenvironment-related molecules, tending to be cold tumor characteristics;

[0106] SNF4: enriched in endothelial cells and fibroblasts, tends to have immune-blocking tumor characteristics.

[0107] The somatic mutation characteristics of each subtype can be summarized as follows:

[0108] SNF1: More than half of tumors have phosphatidylinositol 3-kinase CA (PIK3CA) mutations, while tumor protein P53 (TP53) mutations are less common;

[0109] SNF2: exhibits the highest TP53 mutation rate and has a lower copy number load;

[0110] SNF3: Cyclin D1 (CCND1) / fibroblast growth factor receptor 1 (FGFR1) / mouse double minute 2 (MDM2) copy number amplification, breast cancer 2 gene (BRCA2) copy number loss, G2 / M cell cycle activator gene MDM2 copy number amplification, G2 / M cell cycle inhibitor gene ataxia telangiectasia mutated gene (ATM) with higher levels of copy number loss, high chromosomal instability and whole genome duplication (WGD), chromosomal instability (CIN) score and homologous recombination deficiency (HRD) score were significantly increased. CCND1 amplification and ATM loss occurred more frequently in SNF3 recurrent and metastatic tumors.

[0111] SNF4: somatic mutation signature was not significant.

[0112] The morphological characteristics of each subtype (as shown in hematoxylin-eosin stained sections) are as follows:

[0113] SNF1: Partially retains the morphology of normal mammary glands;

[0114] SNF2: high immune cell infiltration;

[0115] SNF3: Abundant tumor cells with atypicality;

[0116] SNF4: enriched in fibroblasts surrounding tumor cells.

[0117] The clinicopathological characteristics of each subtype can be summarized as follows:

[0118] SNF1: best prognosis;

[0119] SNF2: no obvious features;

[0120] SNF3: higher Ki-67 proliferation index and histological grade;

[0121] SNF4: Worst prognosis.

[0122] In the HR-positive HER2-negative breast cancer classification system and method of the present invention, the immunohistochemical scores obtained for the four subtypes are:

[0123] Classic luminal surface type: α-SMA < 20%, CD8 < 3%, Ki67 < 25%, and all three conditions are met at the same time;

[0124] Immunomodulatory type: α-SMA < 20%, CD8 ≥ 3%, both of the above conditions are met;

[0125] Proliferative type: α-SMA < 20%, CD8 < 3%, Ki67 ≥ 25%, all three conditions are met;

[0126] Receptor tyrosine kinase-driven: α-SMA ≥ 20%.

[0127] That is, the subtype of a specific breast cancer can be determined based on the above-mentioned immunohistochemical characteristics. In actual operation, for convenience, the method of the present invention can be implemented in the following order:

[0128] If α-SMA ≥ 20%, it is determined to be receptor tyrosine kinase driven; or

[0129] If α-SMA < 20%, CD8 ≥ 3%, and both conditions are met, the patient is classified as immune regulatory; or

[0130] If α-SMA < 20%, CD8 < 3%, Ki67 ≥ 25%, and the above three conditions are met at the same time, it is judged as proliferative type; or

[0131] If α-SMA < 20%, CD8 < 3%, and Ki67 < 25%, and the above three conditions are met at the same time, it is determined to be a classic luminal facial type.

[0132] When implementing the above-mentioned HR-positive HER2-negative breast cancer typing method, the typing can be further confirmed based on other characteristics of each of the four subtypes according to actual needs or the judgment of the clinician. The other characteristics include one or more of the following characteristics: transcriptomic characteristics, metabolomic characteristics, proteomic characteristics, microenvironment characteristics, somatic mutation characteristics, morphological characteristics, and clinical pathological characteristics.

[0133] Although the classification of HR-positive, HER2-negative breast cancer can simultaneously reveal the biological characteristics of HR-positive, HER2-negative breast cancer and suggest subtype-specific treatment strategies, it is difficult to perform multi-omics sequencing on every HR-positive, HER2-negative breast cancer patient in actual clinical practice due to time and economic costs. Therefore, it is urgent to find a method that can clinically restore the classification of HR-positive, HER2-negative breast cancer patients in a simple, efficient, and relatively accurate manner.

[0134] The present invention provides a system for typing HR-positive HER2-negative breast cancer, the system comprising:

[0135] 1) a full-scan digital pathology image (WSI) data preprocessing module, which is used to segment the full-scan digital pathology image of the tumor tissue sample of the HR-positive HER2-negative breast cancer into a plurality of enlarged image blocks;

[0136] 2) a first neural network module, wherein the first neural network module is a tissue type segmentation neural network module, which is used to classify the image blocks into different tissue types and attach corresponding tissue labels;

[0137] 3) a second neural network module connected in series with the first neural network module, wherein the second neural network module is a typing prediction neural network module, and is used to perform typing on the sample based on the image block with the tissue label;

[0138] Among them, the classification refers to dividing HR-positive HER2-negative breast cancer into the following four subtypes: classic luminal type, immune regulatory type, proliferative type, and receptor tyrosine kinase driven type.

[0139] In some embodiments, the full-scan digital pathology image data preprocessing module is used to cut the full-scan digital pathology image into image blocks magnified 10-40 times, preferably into image blocks of 256*256 pixels, 512*512 pixels or 800*800 pixels at 20 times magnification.

[0140] The first neural network module is a tissue type segmentation neural network, which is used to classify the image blocks of the pre-processed full-scan digital pathology image into different tissue types and assign corresponding tissue labels. The tissue labels are preferably five tissue types: tumor, stroma, immune infiltration, normal glandular tissue, and necrosis or hemorrhage; and the tissue labels are preferably tumor tissue labels, stroma tissue labels, immune infiltration tissue labels, normal glandular tissue labels, and necrosis or hemorrhage tissue labels.

[0141] The second neural network module is a classification prediction neural network, which is used to determine the HR-positive HER2-negative breast cancer subtype of the sample based on image blocks with specific tissue labels. The specific tissue labels are preferably tumor tissue labels, stromal tissue labels, and / or immune infiltration tissue labels.

[0142] In a preferred embodiment of the present invention, the first neural network module and the second neural network module are constructed through artificial intelligence deep learning based on full-scan digital pathology images of tumor tissue samples of patients with known HR-positive HER2-negative breast cancer subtypes.

[0143] In some embodiments, the number of full-scan digital pathology images of the mass pathology sections of the known HR-positive HER2-negative breast cancer subtype patients is more than 100, preferably more than 200, more than 220, more than 230, more than 240, 200-300, 220-280, 230-270, 240-260, 242-250, 244; preferably, each full-scan digital pathology image of the mass pathology sections of the known HR-positive HER2-negative breast cancer subtype patient corresponds to one known HR-positive HER2-negative breast cancer subtype patient. Of course, this one-to-one correspondence is not the only way to complete the present invention. Those skilled in the art can make some equivalent substitutions based on the disclosure of the present invention, such as having some patients correspond to multiple full-scan digital pathology images, etc.

[0144] In one aspect, the present disclosure provides a method or system for classifying HR-positive HER2-negative breast cancer, which classifies HR-positive HER2-negative breast cancer into the following four subtypes: classic luminal subtype (SNF1), immune-regulated subtype (SNF2), proliferative subtype (SNF3), and receptor tyrosine kinase-driven subtype (SNF4), wherein:

[0145] The classic luminal surface type has one or more or all of the following characteristics: (1) in the tumor sample, at least 50% of the SNF1 highly expressed genes in Table 1 or 4 have enhanced expression relative to normal breast samples; (2) the tumor sample is negative for α-SMA expression, CD8 expression, and Ki67 expression; (3) the pathological image of the tumor sample has a ratio of normal ductal epithelial cells that is equal to or exceeds a predetermined threshold; (4) sensitive to endocrine therapy;

[0146] The immune regulatory type has one or more or all of the following characteristics: (1) in tumor samples, at least 50% of the SNF2 highly expressed genes in Table 1 or 4 have enhanced expression relative to normal breast samples; (2) the tumor sample is negative for α-SMA expression and positive for CD8 expression; (3) the pathological image of the tumor sample has a ratio of immune cells that is equal to or exceeds a predetermined threshold; (4) it is sensitive to immunotherapy;

[0147] The proliferative type has one or more or all of the following characteristics: (1) in the tumor sample, at least 50% of the SNF3 highly expressed genes in Table 1 or 4 have enhanced expression relative to normal breast samples; (2) the tumor sample is negative for α-SMA expression, negative for CD8 expression, and positive for Ki67 expression; (3) the pathological image of the tumor sample has a ratio of atypical tumor cells that is equal to or exceeds a predetermined threshold; (4) sensitive to cyclin-dependent kinase (CDK) 4 / 6 inhibitors, poly (ADP-ribose) polymerase (PARP) inhibitors, or a combination of CDK4 / 6 inhibitors and PARP inhibitors;

[0148] The receptor tyrosine kinase-driven type has one or more or all of the following characteristics: (1) in tumor samples, at least 50% of the SNF4 highly expressed genes in Table 1 or 4 have enhanced expression relative to normal breast samples; (2) the tumor sample is positive for α-SMA expression; (3) the pathological image of the tumor sample has a ratio of cancer-associated fibroblasts that is equal to or exceeds a predetermined threshold; (4) it is sensitive to tyrosine kinase inhibitors.

[0149] The present disclosure provides embodiments as described below, which may be practiced alone or in combination with embodiments described elsewhere herein.

[0150] Multi-omics typing

[0151] In one aspect, the present disclosure provides a multi-omics-based typing method for HR-positive HER2-negative breast cancer, comprising: obtaining target multi-omics data for each patient in a target cancer patient population, wherein the target multi-omics data includes expression profile data, somatic copy number variation data, and metabolomics data; calculating an omics similarity network corresponding to each omics in the target multi-omics data; and clustering the omics similarity network using a spectral clustering method to obtain the typing of the breast cancer.

[0152] In some embodiments, calculating the omics similarity network corresponding to each omics includes normalizing the data of each omics, calculating the squared Euclidean distance of each normalized data, and then constructing the omics similarity network based on the squared Euclidean distance.

[0153] In some further embodiments, the data of each group of studies are standardized by using the "standardNormalization" function in the R package "SNFtools". In some further embodiments, the squared Euclidean distance of each standardized data is calculated by the "dist2" function in the R package "SNFtools". In some further embodiments, the "affinityMatrix" and "SNF" functions in the R package "SNFtools" are used to construct a similarity network. In some further embodiments, when constructing the similarity network, the following parameter settings are used: K=15, sigma=0.5, t=20. In some further embodiments, the "Spectrum" function (NN=20) in the R package "Spectrum" is used to determine the optimal number of clusters.

[0154] In some further embodiments, spectral clustering is performed using Similarity Network Fusion Clustering (SNF) to obtain the breast cancer classification. In some further embodiments, spectral clustering is performed using the "spectralClustering" function in the R package "SNFtools" to obtain the breast cancer classification.

[0155] In some embodiments, the four SNF types are obtained by the multi-omics-based typing method disclosed herein, namely, the classic luminal-facial type, the immune-regulatory type, the proliferative type, and the receptor tyrosine kinase-driven type.

[0156] In one aspect, the present disclosure provides a HR-positive HER2-negative breast cancer typing, which is obtained by the multi-omics-based typing method disclosed herein.

[0157] In the present disclosure, the HR-positive HER2-negative breast cancer typing results obtained by the multi-omics-based typing method disclosed herein can be used as the gold standard for HR-positive HER2-negative breast cancer typing. In some embodiments, this gold standard is used to develop and validate other typing methods, such as for developing, training, and validating the digital pathology image-based or gene expression-based typing methods disclosed herein.

[0158] The multi-omics-based HR-positive HER2-negative breast cancer typing disclosed herein can simultaneously reveal the biological characteristics of HR-positive HER2-negative breast cancer and suggest subtype-specific treatment strategies, but considering the time cost and economic cost, it is difficult to perform multi-omics analysis on each HR-positive HER2-negative breast cancer patient in actual clinical work. Therefore, the present disclosure further provides a typing method that can be clinically simple, efficient and relatively accurate in restoring HR-positive HER2-negative breast cancer patients, including a typing method based on pathological images or morphological features, a typing method based on biomarkers, and a typing method based on transcriptome / gene expression. These typing methods are highly consistent with the multi-omics-based typing method.

[0159] Digital pathology classification

[0160] In one aspect, the present disclosure provides a system and method for classifying HR-positive and HER2-negative breast cancer based on digital pathology images.

[0161] In some technical solutions, the present disclosure provides a system for classifying HR-positive HER2-negative breast cancer based on digital pathology images, the system comprising:

[0162] 1) a digital pathology image (e.g., a full-scan digital pathology image) data preprocessing module, which is used to segment the digital pathology image of the tumor tissue sample of the HR-positive HER2-negative breast cancer into a plurality of image blocks, such as a plurality of enlarged image blocks;

[0163] 2) a tissue type segmentation module, which is used to identify the tissue type of the image block and add a corresponding tissue label;

[0164] 3) a typing prediction module connected in series with the tissue type segmentation module, which is used to perform the typing of the sample based on the image blocks with the tissue labels;

[0165] Among them, the classification refers to dividing HR-positive HER2-negative breast cancer into the following four subtypes: classic luminal type, immune regulatory type, proliferative type, and receptor tyrosine kinase driven type.

[0166] In some technical solutions, the present disclosure provides a system for typing HR-positive HER2-negative breast cancer, the system comprising:

[0167] 1) a digital pathology image (e.g., a full-scan digital pathology image) data preprocessing module, which is used to segment the digital pathology image of the tumor tissue sample of the HR-positive HER2-negative breast cancer into a plurality of image blocks, such as a plurality of enlarged image blocks;

[0168] 2) a tissue type segmentation module, which is used to classify the image blocks into different tissue types and assign corresponding tissue labels;

[0169] 3) a typing prediction module connected in series with the tissue type segmentation module, which is used to perform the typing of the sample based on the image blocks with the tissue labels;

[0170] Among them, the classification refers to dividing HR-positive HER2-negative breast cancer into the following four subtypes: classic luminal type, immune regulatory type, proliferative type, and receptor tyrosine kinase driven type.

[0171] In some embodiments, the present disclosure provides a system for typing HR-positive HER2-negative breast cancer, the system comprising:

[0172] 1) a digital pathology image (e.g., a full-scan digital pathology image) data preprocessing module, which is used to segment the digital pathology image of the tumor tissue sample of the HR-positive HER2-negative breast cancer into a plurality of image blocks, such as a plurality of enlarged image blocks;

[0173] 2) a first neural network module, wherein the first neural network module is a tissue type segmentation neural network module, which is used to classify the image block into different tissue types or identify the tissue type of the image block and add a corresponding tissue label;

[0174] 3) a second neural network module connected in series with the first neural network module, wherein the second neural network module is a typing prediction neural network module, and is used to perform typing on the sample based on the image block with the tissue label;

[0175] Among them, the classification refers to dividing HR-positive HER2-negative breast cancer into the following four subtypes: classic luminal type, immune regulatory type, proliferative type, and receptor tyrosine kinase driven type.

[0176] In one aspect, the present disclosure provides a computer-implemented method for classifying HR-positive HER2-negative breast cancer, the method comprising: receiving a digital pathology image of a tumor tissue sample corresponding to HR-positive HER2-negative breast cancer, and dividing the image image into a plurality of image blocks, such as a plurality of enlarged image blocks; applying a tissue type segmentation module to the image block to classify the image block into different tissue types or identify the tissue type of the image block, and attach a corresponding tissue label; applying a classification prediction module to the image block with the tissue label to perform the classification on the sample; wherein, the classification refers to classifying HR-positive HER2-negative breast cancer into the following four subtypes: classic luminal type, immune-regulated type, proliferative type, and receptor tyrosine kinase-driven type.

[0177] In some embodiments, the tissue type segmentation module and the typing prediction module are each independently a machine learning module, for example, a deep learning module, such as a CNN, a CNN trained using MIL, a recurrent neural network (RNN), a long-short-term memory RNN (LSTM), a gated recurrent unit RNN (GRU), a graph convolutional network, a support vector machine and / or a random forest module.

[0178] In some embodiments, the tissue type segmentation module or the first neural network module has been generated by processing a plurality of training digital pathology images (e.g., full scan digital pathology images) to identify an image block whose tissue type is selected from at least one of tumor, stroma, immune infiltration, normal gland, and necrosis or hemorrhage, and carries a corresponding tissue label. In some embodiments, the tissue type segmentation module or the first neural network module identifies an image block whose tissue type is a tumor. In some embodiments, the tissue type segmentation module or the first neural network module inputs the image block with the corresponding tissue label into the typing prediction module or the second neural network module, and the typing prediction module or the second neural network module performs the typing on the sample based on the image block with the tissue label. In some embodiments, the typing prediction module or the second neural network module has been generated by processing a plurality of training image blocks (e.g., image blocks with tissue labels) to perform typing on HR-positive HER2-negative breast cancer samples based on image blocks with tissue labels.

[0179] In some embodiments, the machine learning module is constructed by machine learning, such as artificial intelligence deep learning, based on digital pathology images (e.g., full-scan digital pathology images) of tumor tissue samples of patients with known HR-positive HER2-negative breast cancer subtypes. In some embodiments, the breast cancer subtype of the patient with known HR-positive HER2-negative breast cancer subtype is determined by the typing method described in the present disclosure, such as the multi-omics-based typing method of the present disclosure.

[0180] In some embodiments, the first neural network module and the second neural network module are constructed by machine learning, such as artificial intelligence deep learning, based on digital pathology images (e.g., full-scan digital pathology images) of tumor tissue samples of patients with known HR-positive HER2-negative breast cancer subtypes. In some embodiments, the breast cancer subtype of the patients with known HR-positive HER2-negative breast cancer subtypes is determined by the typing method described in the present disclosure, such as the multi-omics-based typing method of the present disclosure.

[0181] According to one or more embodiments, the machine learning system can be trained in different ways. For example, the training of the machine learning system can be performed by any one or any combination of supervised training, semi-supervised training, unsupervised training, classifier training, hybrid training, and / or uncertainty estimation. The type of training used can depend on the amount of data, the type of data, and / or the quality of the data.

[0182] The training techniques discussed in this disclosure can also be performed in stages, where images with better annotations are initially used for training, which can allow more effective later training using images with fewer annotations, less supervision, etc.

[0183] Training can begin with the most fully annotated images, relative to all possible training images. For example, training can begin using supervised learning. A first set of images can be received or determined along with associated annotations. Each image can have labeled and / or masked regions and can include information such as identifying regions of interest and tissue types. The first set of images can be provided to a training algorithm, such as a CNN, which can determine correlations between the first set of images and their associated annotations.

[0184] After training with the first set of images is complete, a second set of images with fewer annotations than the first set (e.g., with partial annotations) may be received or determined. In one embodiment, the annotations may only indicate that the image has a region of interest, but may not specify what cancer subtype it may belong to, etc. The second set of images may be trained using a different training algorithm (e.g., multiple instance learning) than the first set. The first set of training data can be used to partially train the system, and the second training round can be more effective in producing an accurate algorithm.

[0185] In this way, training can be performed in any number of stages using any number of algorithms based on the quality and type of the training images. These techniques can be used in situations where multiple training image sets are received, which can have different qualities, annotation levels, and / or annotation types.

[0186] In some embodiments, the number of digital pathology images (e.g., full-scan digital pathology images) of the mass pathology sections of the known HR-positive HER2-negative breast cancer subtype patients is more than 100, preferably more than 200, more than 220, more than 230, more than 240, 200-300, 220-280, 230-270, 240-260, 242-250, 244; preferably, each digital pathology image (e.g., full-scan digital pathology image) of the mass pathology sections of the known HR-positive HER2-negative breast cancer subtype patients corresponds to one known HR-positive HER2-negative breast cancer subtype patient. Of course, this one-to-one correspondence is not the only way to complete the present invention. Those skilled in the art can make some equivalent substitutions based on the disclosure of the present invention, such as having some patients correspond to multiple full-scan digital pathology images, etc.

[0187] In some embodiments, the images used for training can be from real sources (e.g., human patients) or can be from synthetic sources (e.g., graphics rendering engines, 3D models, etc.). Example sources of digital pathology images can include: (a) digitized slides stained with various stains (such as, but not limited to, H&E, etc.); and / or (b) digitized tissue samples from imaging devices. For example, a digital pathology image can include a digitized image of a microscope slide containing a specimen. Dye-based staining systems have been developed, including the periodic acid-Schiff reaction, Masson's trichrome, Nissl and methylene blue, and hematoxylin and eosin (H&E). For medical diagnosis, H&E is a widely used dye-based method in which hematoxylin stains cell nuclei blue, eosin stains the cytoplasm and extracellular matrix pink, and other tissue regions exhibit variations of these colors.

[0188] In some embodiments, the artificial intelligence deep learning uses image blocks with tissue labels and their corresponding HR-positive HER2-negative breast cancer subtypes to train and validate a machine learning model (e.g., a convolutional neural network (CNN) model), so that the machine learning model can identify or segment image blocks with specific tissue labels from digital pathology images (e.g., full-scan digital pathology images) of patients with HR-positive HER2-negative breast cancer and output image block prediction scores, and the prediction score for each patient is the average of the image block prediction scores of the image blocks with specific tissue labels included in the analysis of the patient; wherein the digital pathology classification of each patient is the HR-positive HER2-negative breast cancer subtype corresponding to the highest prediction score.

[0189] In some embodiments, the image blocks with specific tissue labels of the patient included in the analysis are randomly selected image blocks with specific tissue labels of the patient, and the number is more than 100, preferably more than 200, more than 220, more than 230, more than 240, 200-300, 220-280, 230-270, 240-260, 242-258, 250.

[0190] In some embodiments, a three-fold cross-validation algorithm is used to train and / or validate the machine learning model (e.g., a convolutional neural network (CNN) model). After three-fold cross-validation, the prediction scores of the three best models are standardized and averaged to generate a prediction score for each HR-positive HER2-negative breast cancer subtype for each patient, and the digital pathology classification of each patient is the HR-positive HER2-negative breast cancer subtype corresponding to the highest prediction score.

[0191] In some embodiments, the artificial intelligence deep learning uses an artificial intelligence support vector machine (SVM) nonlinear classifier and a one-to-many method to train and / or validate the convolutional neural network model using image blocks with tumor tissue labels and their corresponding HR-positive HER2-negative breast cancer subtype samples.

[0192] In some embodiments, a convolutional neural network is used to extract deep features, a support vector machine provides a binary classification result (e.g., SNF1-non-SNF1), and finally a one-to-many method (SNF1-SNF2-SNF3-SNF4) is used to determine which of the four subtypes the patient belongs to.

[0193] In some embodiments, the tissue type segmentation neural network model is obtained by performing neural network modeling using a ResNet-18 network structure.

[0194] In some embodiments, the tissue types are tumor, stroma, immune infiltration, normal glandular, and necrosis or hemorrhage. In some embodiments, the tissue labels are tumor, stroma, immune infiltration, normal glandular, and necrosis or hemorrhage.

[0195] In some embodiments, the tissue types are tumor, stroma, and immune infiltrate. In some embodiments, the tissue signatures are tumor, stroma, and immune infiltrate.

[0196] In some embodiments, the tissue types are tumor and stroma. In some embodiments, wherein the tissue labels are tumor and stroma. In some embodiments, the tissue type is tumor and the tissue label is tumor. In some embodiments, SNF1, SNF3, and SNF4 types use the "tumor" image block for classification prediction, while SNF2 type uses the "stroma" image block for classification prediction. In some embodiments, SNF1, SNF3, SNF2, and SNF4 types all use the "tumor" image block for classification prediction. In some embodiments, SNF1, SNF3, SNF2, and SNF4 types all use the "tumor" image block for classification prediction. In some embodiments, SNF1, SNF3, SNF2, and SNF4 types all use only the "tumor" image block for classification prediction.

[0197] In some embodiments, during the modeling stage, a judgment threshold is determined by ROC analysis of the model output score and the actual result using the Youden index as a standard, and this threshold can be applied to subsequent typing judgments.

[0198] In some embodiments, the enlarged image block is an image block enlarged 10-40 times. In some embodiments, the enlarged image block is an image block of 256*256 pixels, 512*512 pixels, or 800*800 pixels at 20 times magnification.

[0199] In some embodiments, the number of the pathology section images (eg, full scan pathology section images) is more than 100.

[0200] In some embodiments, the artificial intelligence deep learning includes training and validating a convolutional neural network model using image patches with tissue labels and their corresponding HR-positive HER2-negative breast cancer subtypes.

[0201] In some embodiments, the artificial intelligence deep learning includes using a three-fold cross-validation algorithm to train and validate the convolutional neural network model.

[0202] In some embodiments, the artificial intelligence deep learning includes training and validating the convolutional neural network model by using a nonlinear classifier of an artificial intelligence support vector machine and a one-to-many method.

[0203] In some embodiments, neural network modeling is performed using a ResNet-18 network structure to obtain a tissue type segmentation neural network model.

[0204] In one aspect, the typing system described in the present disclosure further comprises a scanning device, such as a scanner, which is configured to scan pathological slide images. In one aspect, the typing system described in the present disclosure further comprises a microscopic device, such as a microscope, which is configured to observe and / or acquire pathological slide images.

[0205] The present invention also provides a method for typing HR-positive HER2-negative breast cancer using the above-mentioned HR-positive HER2-negative breast cancer typing system.

[0206] In one aspect, the present disclosure further provides a chip, comprising a processor for calling and running a computer program from a memory, so that a device equipped with the chip executes the typing method of the present invention.

[0207] The present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the typing method of the present invention is implemented.

[0208] In some embodiments, the typing method of the present invention includes: receiving a digital pathology image of a tumor tissue sample corresponding to HR-positive HER2-negative breast cancer, and dividing it into multiple image blocks, such as multiple enlarged image blocks; applying a tissue type segmentation module to the image block to classify the image block into different tissue types or identify the tissue type of the image block, and attach a corresponding tissue label; applying a typing prediction module to the image block with the tissue label to perform the typing on the sample; wherein the typing refers to dividing HR-positive HER2-negative breast cancer into the following four subtypes: classic luminal type, immune regulation type, proliferative type, and receptor tyrosine kinase driven type.

[0209] Classification method based on morphological features of pathological images

[0210] In one aspect, the present disclosure provides an in vitro method for determining the subtype of HR-positive HER2-negative breast cancer based on morphological features, comprising:

[0211] (a) providing pathological images of tumor samples from patients with HR-positive HER2-negative breast cancer; and

[0212] (b) determining morphological features of the pathological image, comprising:

[0213] (b1) determining the proportion of normal ductal epithelial cells to all cells in the pathological image;

[0214] (b2) determining the proportion of immune cells to all cells in the pathological image;

[0215] (b3) determining the proportion of atypical tumor cells to all cells in the pathological image; and / or

[0216] (b4) determining the ratio of cancer-associated fibroblasts to all cells in the pathological image.

[0217] The higher the proportion of normal ductal epithelial cells to all cells, the more normal mammary ducts are retained. In some embodiments, if the proportion of normal ductal epithelial cells is equal to or greater than a predetermined threshold, the HR-positive HER2-negative breast cancer is determined to be a classic luminal type. In some embodiments, if the proportion of immune cells is equal to or greater than a predetermined threshold, the HR-positive HER2-negative breast cancer is determined to be an immune-regulated type. In some embodiments, if the proportion of atypical tumor cells is equal to or greater than a predetermined threshold, the HR-positive HER2-negative breast cancer is determined to be a proliferative type. In some embodiments, if the proportion of cancer-associated fibroblasts is equal to or greater than a predetermined threshold, the HR-positive HER2-negative breast cancer is determined to be a receptor tyrosine kinase-driven type.

[0218] Those skilled in the art will understand that the predetermined threshold values ​​described herein will vary depending on factors such as the assay methodology used (e.g., manual interpretation, artificial intelligence deep learning, etc.), detection reagents such as staining reagents, and detection instruments such as microscopes and scanners. Those skilled in the art will be able to determine a suitable predetermined threshold value based on the teachings of the present disclosure. As used herein, the term "predetermined threshold value" may refer to an average threshold value (abbreviated as: cut-off) calculated from a plurality of samples, wherein the plurality of samples are obtained from a plurality of subjects, in particular, subjects with known HR-positive HER2-negative breast cancer subtypes. In order to obtain a threshold value, the plurality of subjects may include subjects with different subtypes of tumors, such as subjects with the SNF tetrad breast cancer described in the present disclosure. In some embodiments, the predetermined threshold value may be an average threshold value obtained by calculating tumor samples of known HR-positive HER2-negative breast cancer subtypes.

[0219] In some embodiments, the optimal threshold point can be selected by the Youden index. The confusion matrix corresponding to the optimal threshold point on the ROC curve is the basis for calculating indicators such as sensitivity, specificity, and accuracy. The Youden index, also known as the correct index, refers to the sum of sensitivity and specificity minus 1: Youden index = Sensitivity + Specificity – 1. The Youden index ranges from 0 to 1, representing the overall ability of the classification model to detect true patients and non-patients. The larger the Youden index, the better the performance of the classification model.

[0220] In some embodiments, whether the ratio is equal to or exceeds a predetermined threshold is determined by manual interpretation or a typing prediction module described in the present disclosure, such as a neural network module (e.g., a typing prediction neural network module described in the present disclosure).

[0221] In some embodiments, the predetermined threshold for the proportion of normal ductal epithelial cells to total cells is 8%, 9%, 10%, 11%, 12%, 15% or more, such as 20%. In some embodiments, the predetermined threshold for the proportion of immune cells to total cells is 2%, 3% or more, such as 4% or 5%. In some embodiments, the predetermined threshold for the proportion of atypical tumor cells to total cells is 20%, 25% or more, such as 30%. In some embodiments, the predetermined threshold for the proportion of cancer-associated fibroblasts to total cells is 15%, 20% or more, such as 25%. In some embodiments, the predetermined threshold for the proportion of normal ductal epithelial cells to total cells is 10%, the predetermined threshold for the proportion of immune cells to total cells is 3%, the predetermined threshold for the proportion of atypical tumor cells to total cells is 25%, and the predetermined threshold for the proportion of cancer-associated fibroblasts to total cells is 20%. In some embodiments, the predetermined threshold for the proportion of normal ductal epithelial cells to all cells is 8%, the predetermined threshold for the proportion of immune cells to all cells is 4%, the predetermined threshold for the proportion of atypical tumor cells to all cells is 20%, and the predetermined threshold for the proportion of cancer-associated fibroblasts to all cells is 25%.

[0222] In some embodiments, the in vitro method of determining the subtype of HR-positive HER2-negative breast cancer based on morphological features disclosed herein can be performed using the typing system / method described elsewhere in this disclosure.

[0223] In some embodiments, the predetermined threshold or predetermined expression threshold described in the present disclosure can be determined based on the Youden Index.

[0224] Biomarker-based typing methods

[0225] In one aspect, the present disclosure provides a group of genes for HR-positive HER2-negative breast cancer typing, which include α-SMA gene, CD8 gene and Ki67 gene. In some embodiments, HR-positive HER2-negative breast cancer typing includes four subtypes: classic luminal type, immune regulatory type, proliferative type, receptor tyrosine kinase driven type. In some embodiments, HR-positive HER2-negative breast cancer typing is based on measuring the expression levels of the above genes in HR-positive HER2-negative breast cancer samples, such as protein expression levels or RNA transcript expression levels. In some embodiments, protein expression levels can be determined by immunohistochemical staining.

[0226] In some embodiments, if the expression level of α-SMA in the HR-positive HER2-negative breast cancer tumor sample reaches or exceeds the predetermined expression threshold value T of α-SMA, α-SMA, then the HR-positive HER2-negative breast cancer belongs to the receptor tyrosine kinase-driven type; if the expression level of α-SMA in the HR-positive HER2-negative breast cancer tumor sample is lower than the predetermined expression threshold T α-SMA The expression level of CD8 reaches or exceeds the predetermined expression threshold of CD8 T CD8 , then the HR-positive HER2-negative breast cancer belongs to the immune regulatory type; if the expression level of α-SMA in the HR-positive HER2-negative breast cancer tumor sample is lower than the predetermined expression threshold T α-SMA , the expression level of CD8 is lower than the predetermined expression threshold T CD8 , and the expression level of Ki67 reaches or exceeds the predetermined expression threshold of Ki67 T Ki67 , then the HR-positive HER2-negative breast cancer belongs to the proliferative type; if the expression level of α-SMA in the HR-positive HER2-negative breast cancer tumor sample is lower than the predetermined expression threshold T α-SMA , the expression level of CD8 is lower than the predetermined expression threshold T CD8 , and the expression level of Ki67 is lower than the predetermined expression threshold T Ki67 , then the HR-positive HER2-negative breast cancer belongs to the classic caval-facial type.

[0227] Those skilled in the art will understand that the predetermined expression threshold values ​​described in the present disclosure will vary depending on factors such as the detection methodology used (e.g., immunohistochemistry, PCR, chemiluminescence, etc.), detection reagents and / or instruments. Those skilled in the art will be able to determine a suitable predetermined expression threshold value based on the teachings of the present disclosure. As used herein, the term "predetermined expression threshold value" may refer to an average threshold value (abbreviated as: cut-off) calculated from a plurality of samples, wherein the plurality of samples are obtained from a plurality of subjects, in particular subjects suffering from cancer. In order to obtain a threshold value, the plurality of subjects may include subjects with different subtypes of tumors, such as subjects with the SNF tetrad breast cancer described in the present disclosure. The threshold value may represent the amount or concentration of RNA transcripts or the percentage of positive tumor cells in the total tumor cells by immunohistochemical staining. In one embodiment, the threshold value is given as a CT (cycle threshold) value (see below). In one embodiment, the threshold value is given as a percentage of positive tumor cells in the total tumor cells.

[0228] As used herein, the expression "reaches or is higher than a predetermined expression threshold" includes expression levels that are higher than or equal to the set expression threshold. An expression level that "reaches or is higher than a predetermined expression threshold" may also be referred to as "expression positive," while an expression level that is "lower than a predetermined expression threshold" may also be referred to as "expression negative." In some embodiments, the expression threshold for α-SMA is 15%, 20% or higher, such as 25%. In some embodiments, the expression threshold for CD8 is 2%, 3% or higher, such as 4% or 5%. In some embodiments, the expression threshold for Ki67 is 20%, 25% or higher, such as 30%. In some embodiments, the expression threshold for α-SMA is 20%, the expression threshold for CD8 is 3%, and the expression threshold for Ki67 is 25%.

[0229] In some embodiments, if α-SMA is positive, the HR-positive HER2-negative breast cancer is of the receptor tyrosine kinase-driven type; if α-SMA is negative and CD8 is positive, the HR-positive HER2-negative breast cancer is of the immune-regulated type; if α-SMA is negative, CD8 is negative, and Ki67 is positive, the HR-positive HER2-negative breast cancer is of the proliferative type; if α-SMA is negative, CD8 is negative, and Ki67 is positive, the HR-positive HER2-negative breast cancer is of the classic luminal type. Thus, in some embodiments, the receptor tyrosine kinase-driven HR-positive HER2-negative breast cancer is characterized by positive α-SMA expression; the immune-regulated HR-positive HER2-negative breast cancer is characterized by negative α-SMA expression and positive CD8 expression; the proliferative HR-positive HER2-negative breast cancer is characterized by negative α-SMA expression, negative CD8 expression, and positive Ki67 expression; and / or the classic luminal type HR-positive HER2-negative breast cancer is characterized by negative α-SMA expression, negative CD8 expression, and negative Ki67 expression.

[0230] In one aspect, the present disclosure provides an in vitro method for determining the subtype of HR-positive HER2-negative breast cancer tumors, comprising: (a) determining the expression level of α-SMA in a sample of the tumor; (b) determining the expression level of CD8 in a sample of the tumor; and / or (c) determining the expression level of Ki67 in a sample of the tumor. In some embodiments, steps (a), (b), and (c) may be performed in any order, or simultaneously. In some further embodiments, step (a) is performed before step (b) and / or step (c). In some further embodiments, step (b) is performed before step (c).

[0231] In some embodiments, the method further comprises determining whether the expression level of α-SMA, CD8 and / or Ki67 reaches, is above or is below the corresponding predetermined expression threshold. In some further embodiments, when the expression level of α-SMA in the tumor sample reaches or is above the predetermined expression threshold of α-SMA, α-SMA , the HR-positive HER2-negative breast cancer is determined to be receptor tyrosine kinase driven. In some further embodiments, when the expression level of α-SMA in the tumor sample is lower than the predetermined expression threshold T α-SMA The expression level of CD8 reaches or exceeds the predetermined expression threshold of CD8 T CD8 , the HR-positive HER2-negative breast cancer is determined to be immune-regulated. In some further embodiments, when the expression level of α-SMA in the tumor sample is lower than the predetermined expression threshold T α-SMA , the expression level of CD8 is lower than the predetermined expression threshold T CD8 , and the expression level of Ki67 reaches or exceeds the predetermined expression threshold of Ki67 T Ki67 In some further embodiments, when the expression level of α-SMA in the tumor sample is lower than the predetermined expression threshold T α-SMA , the expression level of CD8 is lower than the predetermined expression threshold T CD8 , and the expression level of Ki67 is lower than the predetermined expression threshold T Ki67 , then the HR-positive HER2-negative breast cancer is judged as the classic caval-facial type.

[0232] In some embodiments, the in vitro method of determining the subtype of a tumor in a patient with HR-positive HER2-negative breast cancer disclosed herein can be performed using the HR-positive HER2-negative breast cancer typing system / method described elsewhere in this disclosure.

[0233] In one aspect, the present disclosure provides a kit for genotyping HR-positive HER2-negative breast cancer, comprising reagents for detecting the expression levels of α-SMA, CD8, and / or Ki67. In some embodiments, the reagents for detecting the expression levels of α-SMA, CD8, and / or Ki67 can be packaged separately or in any combination.

[0234] In one aspect, the present disclosure provides use of a reagent for detecting the expression level of α-SMA, CD8, and / or Ki67 in the preparation of a diagnostic agent for categorizing HR-positive, HER2-negative breast cancer.

[0235] In one aspect, the present disclosure provides the use of α-SMA, CD8 and / or Ki67 as HR-positive HER2-negative breast cancer typing markers. In another aspect, the present disclosure provides the use of α-SMA, CD8 and / or Ki67 in preparing HR-positive HER2-negative breast cancer typing markers or in preparing reagents for detecting the expression level of HR-positive HER2-negative breast cancer typing markers.

[0236] In one aspect, the present disclosure provides a method for identifying genes that can be used for HR positive HER2 negative breast cancer typing, comprising the following steps: (a) performing mRNA expression profile difference analysis between subtypes of HR positive HER2 negative breast cancer to obtain candidate genes with statistically significant expression differences between subtypes. In some further embodiments, the method further comprises step (b) performing correlation analysis between mRNA expression levels and protein expression levels on the mRNA expression data and protein expression data of the mRNA expression profile to further screen candidate genes. In some further embodiments, the genes in the top 10, 9, 8, 7, 6, 5, 4, 3, 2 or 1 with the largest expression differences between subtypes are selected as candidate genes. In some further embodiments, the genes in the top 10, 9, 8, 7, 6, 5, 4, 3 or 2 with the largest expression differences between subtypes and the mRNA expression levels and protein expression levels are selected as candidate genes.

[0237] Classifier-based gene expression typing method

[0238] In one aspect, the present disclosure provides a method for classifying HR-positive HER2-negative breast cancer subtypes, comprising: obtaining a gene expression profile of a target test sample; and applying a machine learning classifier to the gene expression profile to classify the target test sample into one of the subtypes of classic luminal facial type (SNF1), immune regulatory type (SNF2), proliferative type (SNF3), and receptor tyrosine kinase driven type (SNF4).

[0239] In some embodiments, obtaining a gene expression profile of a target test sample includes detecting the expression level of a classifier gene. In some embodiments, the classifier gene includes or includes only at least one of the highly expressed genes listed in Tables 1 or 4, for example, includes or includes only at least one of the SNF1 highly expressed genes, at least one of the SNF2 highly expressed genes, at least one of the SNF3 highly expressed genes, and / or at least one of the SNF4 highly expressed genes listed in Tables 1 or 4. In some embodiments, the classifier gene includes or includes only at least two, at least 8, at least 16, at least 32, at least 48, at least 64, at least 100, at least 128, at least 150, at least 200, at least 240, at least 280, or at least 300 of the highly expressed genes listed in Tables 1 or 4, for example, includes or includes only at least two, at least 8, at least 16, at least 32, at least 48, at least 64, at least 100, at least 128, at least 150, at least 200, at least 240, at least 280, or at least 300 of the highly expressed genes listed in Tables 1 or 4. at least 8, at least 64, at least 88, or all of the highly expressed genes of SNF2, at least two, at least 8, at least 16, at least 32, at least 48, at least 64, at least 80, or all of the highly expressed genes of SNF3, and / or at least two, at least 8, at least 16, at least 32, at least 48, at least 64, at least 88, or all of the highly expressed genes of SNF4. In some embodiments, the classifier genes include or include only at least 10, at least 20, at least 25, at least 30, at least 35, or at least 35 of the highly expressed SNF1 genes, at least 10, at least 20, at least 25, at least 30, at least 35, or at least 35 of the highly expressed SNF2 genes, at least 10, at least 20, at least 25, at least 30, at least 35, or at least 35 of the highly expressed SNF3 genes, and / or at least 10, at least 20, at least 25, at least 30, at least 35, or at least 35 of the highly expressed SNF4 genes listed in Tables 1 or 4. In some embodiments, the classifier genes include or include only at least 25 of the highly expressed SNF1 genes, at least 25 of the highly expressed SNF2 genes, at least 25 of the highly expressed SNF3 genes, and / or at least 25 of the highly expressed SNF4 genes listed in Tables 1 or 4. In some embodiments, the classifier genes include or include only at least 30 of the SNF1 highly expressed genes, at least 30 of the SNF2 highly expressed genes, at least 30 of the SNF3 highly expressed genes, and / or at least 30 of the SNF4 highly expressed genes listed in Tables 1 or 4. In some embodiments, the classifier genes include or include only at least 30 of the SNF1 highly expressed genes, at least 25 of the SNF2 highly expressed genes, at least 30 of the SNF3 highly expressed genes, and / or at least 20 of the SNF4 highly expressed genes listed in Tables 1 or 4.In some embodiments, the classifier genes include or only include at least 30 of the SNF1 highly expressed genes, at least 25 of the SNF2 highly expressed genes, at least 30 of the SNF3 highly expressed genes, and / or at least 15 of the SNF4 highly expressed genes listed in Tables 1 or 4.

[0240] In some embodiments, the classifier genes include or are selected from the group consisting of SNF1 classifier genes, SNF2 classifier genes, SNF3 classifier genes, and SNF4 classifier genes. In some cases, the SNF1 classifier genes include or consist of at least two, at least eight, at least sixteen, at least thirty-two, at least forty-eight, at least sixteen, at least eighty-eight, or all of the highly expressed SNF1 genes listed in Tables 1 or 4. In some cases, the SNF2 classifier genes include or consist of at least two, at least eight, at least sixteen, at least thirty-two, at least forty-eight, at least sixteen, at least eighty-eight, or all of the highly expressed SNF2 genes listed in Tables 1 or 4. In some cases, the SNF3 classifier genes include or consist of at least two, at least eight, at least sixteen, at least thirty-two, at least forty-eight, at least sixteen, at least eighty-eight, or all of the highly expressed SNF3 genes listed in Tables 1 or 4. In some cases, the SNF4 classifier genes include or consist solely of at least two, at least 8, at least 16, at least 32, at least 48, or all of the highly expressed SNF4 genes listed in Tables 1 or 4.

[0241] In some embodiments, the gene expression profile is based on the expression of SNF1 highly expressed genes, SNF2 highly expressed genes, SNF3 highly expressed genes, and SNF4 highly expressed genes, such as the expression of some, substantially all, or all of the genes listed in Tables 1 or 4. In some embodiments, the gene expression profile is based on at least 5%, 10%, 15%, 20%, 25%, 30%, 35%, 40%, 45%, 50%, 60%, 70%, 80%, 90%, 95%, 96%, 97%, 98%, 99%, or 100% of the SNF1 highly expressed genes and at least 5%, 10%, 15%, 20%, 25%, 30%, 35%, 40%, 45%, 50%, 60%, 70%, 80%, 90%, 95%, 96%, 97%, 98%, 99%, or 100% of the SNF2 highly expressed genes in Tables 1 or 4. , 99% or 100% of the highly expressed genes in SNF3, and / or at least 5%, 10%, 15%, 20%, 25%, 30%, 35%, 40%, 45%, 50%, 60%, 70%, 80%, 90%, 95%, 96%, 97%, 98%, 99% or 100% of the highly expressed genes in SNF4.

[0242] In some embodiments, the classifier gene is a subtype-specific highly expressed gene, such as the SNF1 highly expressed gene, SNF2 highly expressed gene, SNF3 highly expressed gene, or SNF4 highly expressed gene listed in Table 1 or 4. As used herein, the term "highly expressed gene" refers to a gene whose expression in one SNF subtype of HR-positive HER2-negative breast cancer is statistically significantly increased compared to other SNF subtypes. In some cases, the expression is increased by at least 10%, at least 20%, at least 30%, at least 40%, 50%, at least 60%, at least 70%, at least 80%, 90%, at least 100%, at least 150%, at least 200%, at least 300%, at least 500% or more, such as 1000%.

[0243] In some embodiments, the machine learning classifier has been generated by processing a plurality of gene expression profile training sets of patients with known HR-positive HER2-negative breast cancer subtypes to predict the SNF subtype of HR-positive HER2-negative breast cancer. In some embodiments, the candidate SNF1 highly expressed gene set, SNF2 highly expressed gene set, SNF3 highly expressed gene set, and SNF4 highly expressed gene set that can be used to develop the classifier can be obtained by the following method: performing differential expression analysis between subtypes on the transcriptome data of patients with known HR-positive HER2-negative breast cancer subtypes (e.g., differential expression analysis by RNA-seq), obtaining the top m (e.g., top 100, top 80, or top 50) highly expressed genes for each subtype; optionally, reducing the candidate gene set for each subtype by collinearity elimination (e.g., removing genes with correlation coefficients higher than 0.75), thereby obtaining a candidate gene set for each subtype. In some embodiments, m is an integer less than or equal to 500, such as 400, 300, 200, 150, 100, 90, 80, 70, 60, 50, 40, 30, 20, or 10. In some embodiments, the candidate genes obtained by differential expression analysis between subtypes may be the same or different, such as the top 100 highly expressed genes of SNF1 and SNF2 subtypes and the top 80 highly expressed genes of SNF3 and SNF4 subtypes.

[0244] In some embodiments, the classifier is a random forest classifier or a support vector machine classifier. In some embodiments, the classifier can adopt a generalized linear model (GLM) or a decision tree; the decision tree can be, for example, a recursive partitioning and regression tree, gradient boosting (GBDT), XGBoost (eXtreme Gradient Boosting), LightGBM, etc. In some embodiments, ROC analysis is used to measure the accuracy of the developed classifier.

[0245] In some embodiments, the classifier genes used for SNF1 typing include or only include the top 5 to top 80 genes ranked by predicted relevance among the SNF1 genes in Table 4 or Table 5, such as the top 80, top 75, top 70, top 65, top 60, top 55, top 50, top 45, top 40, top 35, top 30, top 25, top 20, top 15, top 10, or top 5 genes.

[0246] In some embodiments, the classifier genes used for SNF2 typing include or only include the top 5 to top 80 genes ranked by predicted relevance among the SNF2 genes in Table 4 or Table 5, such as the top 80, top 75, top 70, top 65, top 60, top 55, top 50, top 45, top 40, top 35, top 30, top 25, top 20, top 15, top 10, or top 5 genes.

[0247] In some embodiments, the classifier genes used for SNF3 typing include or only include the top 5 to top 80 genes ranked by predicted relevance among the SNF3 genes in Table 4 or Table 5, such as the top 80, top 75, top 70, top 65, top 60, top 55, top 50, top 45, top 40, top 35, top 30, top 25, top 20, top 15, top 10, or top 5 genes.

[0248] In some embodiments, the classifier genes used for SNF4 typing include or only include the top 5 to top 50 genes ranked by predicted relevance among the SNF1 genes in Table 4 or Table 5, such as the top 45, top 40, top 35, top 30, top 25, top 20, top 15, top 10, or top 5 genes.

[0249] In some embodiments, the present disclosure provides a method for determining whether an HR-positive HER2-negative breast cancer tissue sample belongs to the SNF1 subtype, comprising: obtaining a gene expression profile of a SNF1 typing classifier gene of a target test sample; and applying a machine learning classifier to the gene expression profile to determine whether the HR-positive HER2-negative breast cancer tissue sample belongs to the SNF1 subtype. In some embodiments, the present disclosure provides a method for determining whether an HR-positive HER2-negative breast cancer tissue sample belongs to the SNF2 subtype, comprising: obtaining a gene expression profile of a SNF2 typing classifier gene of a target test sample; and applying a machine learning classifier to the gene expression profile to determine whether the HR-positive HER2-negative breast cancer tissue sample belongs to the SNF2 subtype. In some embodiments, the present disclosure provides a method for determining whether an HR-positive HER2-negative breast cancer tissue sample belongs to the SNF3 subtype, comprising: obtaining a gene expression profile of a SNF3 typing classifier gene of a target test sample; and applying a machine learning classifier to the gene expression profile to determine whether the HR-positive HER2-negative breast cancer tissue sample belongs to the SNF3 subtype. In some embodiments, the present disclosure provides a method for determining whether an HR-positive, HER2-negative breast cancer tissue sample belongs to the SNF4 subtype, comprising: obtaining a gene expression profile of an SNF4 typing classifier gene of a target test sample; and applying a machine learning classifier to the gene expression profile to determine whether the HR-positive, HER2-negative breast cancer tissue sample belongs to the SNF4 subtype.

[0250] In another aspect, the present disclosure provides a method for classifying subtypes of HR-positive, HER2-negative breast cancer in a test sample, comprising: a) obtaining a gene expression profile for each of classic luminal facial type (SNF1), immune regulatory type (SNF2), proliferative type (SNF3), and receptor tyrosine kinase driven type (SNF4), wherein the gene expression profile is based on the expression of some, substantially all, or all of the genes listed in Table 1 or 4; b) obtaining a gene expression profile of the test sample, wherein the gene expression profile of the test sample is based on the expression of the same genes in Table 1 or 4 used in step a); c) comparing the gene expression profile of the test sample with the gene expression profile in step a); d) assigning to the test sample one of the subtypes of classic luminal facial type (SNF1), immune regulatory type (SNF2), proliferative type (SNF3), and receptor tyrosine kinase driven type (SNF4) that has an expression profile most similar to the test sample.

[0251] In some embodiments, one, two, or three of steps a), c), and d) may be performed by a machine learning classifier described in the present disclosure.

[0252] In some embodiments, the comparison in step c) is performed by comparing the expression heat maps of steps a) and b).

[0253] In some embodiments, the gene expression profiles in steps a) and b) are based on at least 5%, 10%, 15%, 20%, 25%, 30%, 35%, 40%, 45%, 50%, 60%, 70%, 80%, 90%, 95%, 96%, 97%, 98%, 99% or 100% of the genes listed in Table 1 or 4. In some embodiments, in the gene expression profiles in steps a) and b), at least 5%, 10%, 15%, 20%, 25%, 30%, 35%, 40%, 45%, 50%, 60%, 70%, 80%, 90%, 95%, 96%, 97%, 98%, 99% or 100% of each of the SNF1 highly expressed genes, SNF2 highly expressed genes, SNF3 highly expressed genes, and SNF4 highly expressed genes in Table 1 or 4 are expressed. In some embodiments, the gene expression profiles in steps a) and b) are based on at least 80%, 90%, 95%, 96%, 97%, 98%, 99% or 100% of the highly expressed genes in SNF1, at least 80%, 90%, 95%, 96%, 97%, 98%, 99% or 100% of the highly expressed genes in SNF2, at least 80%, 90%, 95%, 96%, 97%, 98%, 99% or 100% of the highly expressed genes in SNF3, and at least 80%, 90%, 95%, 96%, 97%, 98%, 99% or 100% of the highly expressed genes in SNF4 in Tables 1 or 4. In some embodiments, the gene expression profiles in steps a) and b) are based on at least 90% of the highly expressed genes in SNF1, at least 90% of the highly expressed genes in SNF2, at least 90% of the highly expressed genes in SNF3, and at least 90% of the highly expressed genes in SNF4 in Table 1 or 4. In some embodiments, the gene expression profiles in steps a) and b) are based on at least 95% of the highly expressed genes in SNF1, at least 95% of the highly expressed genes in SNF2, at least 95% of the highly expressed genes in SNF3, and at least 95% of the highly expressed genes in SNF4 in Table 1 or 4. In some embodiments, the gene expression profiles in steps a) and b) are based on at least 99% of the highly expressed genes in SNF1, at least 99% of the highly expressed genes in SNF2, at least 99% of the highly expressed genes in SNF3, and at least 99% of the highly expressed genes in SNF4 in Table 1 or 4.

[0254] In yet another aspect, the present disclosure provides a method for classifying subtypes of HR-positive, HER2-negative breast cancer in a test sample, comprising: a) obtaining gene expression profiles of a plurality of training samples of HR-positive, HER2-negative breast cancer of known subtypes, wherein each of the classic luminal facial type (SNF1), immune regulatory type (SNF2), proliferative type (SNF3), and receptor tyrosine kinase driven type (SNF4) is present in the plurality of breast cancer samples, and wherein the gene expression profiles of the training samples are based on the expression of some, substantially all, or all of the genes listed in Table 1 or 4; b) using a supervised or unsupervised algorithm Constructing a centroid for each HR-positive HER2-negative breast cancer subtype in the training set; c) obtaining a gene expression profile for the test sample, wherein the gene expression profile of the test sample is based on the expression of the same genes in Table 1 or 4 used in step a); d) comparing the gene expression profile of the test sample with the centroid constructed in step b); e) calculating the distance from the gene expression profile of the test sample to each centroid; f) assigning one of the subtypes of classic luminal facial type (SNF1), immune regulatory type (SNF2), proliferative type (SNF3) and receptor tyrosine kinase driven type (SNF4) to the test sample based on the nearest centroid.

[0255] In some embodiments, the gene expression profiles in steps a) and c) are based on at least 5%, 10%, 15%, 20%, 25%, 30%, 35%, 40%, 45%, 50%, 60%, 70%, 80%, 90%, 95%, 96%, 97%, 98%, 99% or 100% of the genes listed in Table 1 or 4.

[0256] In some embodiments, the gene expression profiles in steps a) and c) are based on at least 5%, 10%, 15%, 20%, 25%, 30%, 35%, 40%, 45%, 50%, 60%, 70%, 80%, 90%, 95%, 96%, 97%, 98%, 99% or 100% of the genes listed in Table 1 or 4. In some embodiments, in the gene expression profiles in steps a) and c), at least 5%, 10%, 15%, 20%, 25%, 30%, 35%, 40%, 45%, 50%, 60%, 70%, 80%, 90%, 95%, 96%, 97%, 98%, 99% or 100% of the genes of each of the SNF1 highly expressed genes, SNF2 highly expressed genes, SNF3 highly expressed genes and SNF4 highly expressed genes in Table 1 or 4 are expressed. In some embodiments, the gene expression profiles in steps a) and c) are based on at least 5%, 10%, 15%, 20%, 25%, 30%, 35%, 40%, 45%, 50%, 60%, 70%, 80%, 90%, 95%, 96%, 97%, 98%, 99% or 100% of each of the SNF1 highly expressed genes, SNF2 highly expressed genes, SNF3 highly expressed genes, and SNF4 highly expressed genes in Table 1 or 4. In some embodiments, the gene expression profiles in steps a) and c) are based on at least 80%, 90%, 95%, 96%, 97%, 98%, 99% or 100% of the highly expressed genes in SNF1, at least 80%, 90%, 95%, 96%, 97%, 98%, 99% or 100% of the highly expressed genes in SNF2, at least 80%, 90%, 95%, 96%, 97%, 98%, 99% or 100% of the highly expressed genes in SNF3, and at least 80%, 90%, 95%, 96%, 97%, 98%, 99% or 100% of the highly expressed genes in SNF4 in Tables 1 or 4. In some embodiments, the gene expression profiles in steps a) and c) are based on at least 90% of the highly expressed genes in SNF1, at least 90% of the highly expressed genes in SNF2, at least 90% of the highly expressed genes in SNF3, and at least 90% of the highly expressed genes in SNF4 in Table 1 or 4. In some embodiments, the gene expression profiles in steps a) and c) are based on at least 95% of the highly expressed genes in SNF1, at least 95% of the highly expressed genes in SNF2, at least 95% of the highly expressed genes in SNF3, and at least 95% of the highly expressed genes in SNF4 in Table 1 or 4. In some embodiments, the gene expression profiles in steps a) and c) are based on at least 99% of the highly expressed genes in SNF1, at least 99% of the highly expressed genes in SNF2, at least 99% of the highly expressed genes in SNF3, and at least 99% of the highly expressed genes in SNF4 in Table 1 or 4.

[0257] In some embodiments, the SNF1 typing of the present disclosure has the following characteristics: at least 5%, 10%, 15%, 20%, 25%, 30%, 35%, 40%, 45%, 50%, 60%, 70%, 80%, 90%, 95%, 96%, 97%, 98%, 99% or 100% of the SNF1 highly expressed genes in Table 1 or 4 have enhanced expression in tumor samples relative to normal breast samples.

[0258] In some embodiments, the SNF2 typing of the present disclosure has the following characteristics: at least 5%, 10%, 15%, 20%, 25%, 30%, 35%, 40%, 45%, 50%, 60%, 70%, 80%, 90%, 95%, 96%, 97%, 98%, 99% or 100% of the SNF2 highly expressed genes in Tables 1 or 4 have enhanced expression in tumor samples relative to normal breast samples.

[0259] In some embodiments, the SNF3 typing of the present disclosure has the following characteristics: at least 5%, 10%, 15%, 20%, 25%, 30%, 35%, 40%, 45%, 50%, 60%, 70%, 80%, 90%, 95%, 96%, 97%, 98%, 99% or 100% of the SNF3 highly expressed genes in Tables 1 or 4 have enhanced expression in tumor samples relative to normal breast samples.

[0260] In some embodiments, the SNF4 typing of the present disclosure has the following characteristics: at least 5%, 10%, 15%, 20%, 25%, 30%, 35%, 40%, 45%, 50%, 60%, 70%, 80%, 90%, 95%, 96%, 97%, 98%, 99% or 100% of the SNF4 highly expressed genes in Tables 1 or 4 have enhanced expression in tumor samples relative to normal breast samples.

[0261] In some embodiments, the enhanced expression can be an increase in expression of at least 10%, at least 20%, at least 30%, at least 40%, 50%, at least 60%, at least 70%, at least 80%, 90%, at least 100%, at least 150%, at least 200%, at least 300%, at least 500% or more, e.g., 1000%, etc.

[0262] In another aspect, the present invention provides a method for determining the subtype of an HR-positive, HER2-negative breast cancer tissue sample obtained from a patient, the method comprising detecting the expression level of at least one classifier gene (e.g., at least one highly expressed gene of Table 1 or 4), wherein the detection of the expression level of the classifier gene specifically identifies the HR-positive, HER2-negative breast cancer tissue sample as a classic luminal type (SNF1), an immune regulatory type (SNF2), a proliferative type (SNF3), or a receptor tyrosine kinase driven type (SNF4). In some cases, the method further comprises comparing the detected expression level of the at least one classifier gene with the expression of the at least one classifier gene in at least one sample training set, wherein the at least one sample training set comprises expression data of the at least one classifier gene from a reference HR-positive HER2-negative breast cancer SNF1 sample, expression data of the at least one classifier gene from a reference HR-positive HER2-negative breast cancer SNF2 sample, expression data of the at least one classifier gene from a reference HR-positive HER2-negative breast cancer SNF3 sample, expression data of the at least one classifier gene from a reference HR-positive HER2-negative breast cancer SNF4 sample, or a combination thereof; and classifying the sample as SNF1, SNF2, SNF3, or SNF4 subtype based on the results of the comparing step. In some cases, the comparing step comprises applying a machine learning classifier or statistical algorithm, which comprises determining a correlation between the expression data obtained from the sample and the expression data from the at least one training set; and classifying the sample as SNF1, SNF2, SNF3, or SNF4 subtype based on the results of the machine learning classifier or statistical algorithm.

[0263] In some cases, the expression level of the classifier gene is detected at the nucleic acid level. In some cases, the nucleic acid level is RNA or cDNA. In some cases, the detection of expression level includes quantitative real-time reverse transcriptase polymerase chain reaction (qRT-PCR), RNAseq, microarray, gene chip, nCounter gene expression assay, gene expression serial analysis (SAGE), rapid analysis of gene expression (RAGE), nuclease protection assay, Northern blot or any other equivalent gene expression detection technology. In some cases, by carrying out RNA-seq detection table 1 or 4 expression level of the classifier gene. In some cases, by carrying out qRT-PCR detection table 1 or 4 expression level of the classifier gene. In some cases, the detection of expression level includes using at least one classifier gene of table 1 or 4 to have specific at least one pair of oligonucleotide primers. In some cases, sample is paraffin-embedded (FFPE) breast cancer tumor tissue sample fixed with formalin obtained from patient. In some cases, sample is fresh or frozen tissue sample. In some cases, this at least one classifier gene comprises multiple classifier genes. In some cases, the plurality of classifier genes comprises at least two classifier genes, at least 8 classifier genes, at least 16 classifier genes, at least 24 classifier genes, at least 32 classifier genes, at least 40 classifier genes, or at least 48 classifier genes, at least 60 classifier genes, at least 80 classifier genes, at least 100 classifier genes, at least 120 classifier genes, at least 150 classifier genes, at least 180 classifier genes, at least 200 classifier genes, at least 240 classifier genes, at least 270 classifier genes, or at least 300 classifier genes of Table 1 or 4. In some cases, the at least one classifier gene comprises all of the classifier genes of Table 1 or 4.

[0264] In the present disclosure, determining HR positive HER2 negative breast cancer subtypes can be improved by applying algorithms designed to normalize and / or improve the reliability of gene expression data. In some embodiments of the present invention, due to the large number of individual data points to be processed, data analysis utilizes a computer or other device, machine or apparatus to apply the various algorithms described herein. "Machine learning algorithm" refers to a computationally based prediction method, also referred to as a "classifier" by those skilled in the art, for characterizing one or more gene expression profiles, for example, in order to determine HR positive HER2 negative breast cancer subtypes. In one embodiment, the levels of biomarkers determined by, for example, microarray-based hybridization assays, sequencing assays, nanostring assays, etc., are subjected to an algorithm to classify the expression profiles. Supervised learning typically involves "training" a classifier to recognize the differences between SNF subtypes and then "testing" the accuracy of the classifier on an independent test set. Thus, for a new unknown sample, a classifier can be used to predict, for example, the category to which the sample belongs (e.g., each SNF1-4 subtype).

[0265] In one embodiment, the method of biomarker level data analysis includes the use of a pre-classifier algorithm. For example, the algorithm can use specific molecular fingerprints to pre-classify samples according to their composition and then apply correction / normalization factors. This data / information can then be input into the final classification algorithm, which will integrate this information to assist in the final typing.

[0266] In one embodiment, the method of biomarker level data analysis also includes the use of a classifier algorithm as provided herein. In one embodiment of the invention, diagonal linear discriminant analysis, k-nearest neighbor algorithm, support vector machine (SVM) algorithm, linear support vector machine, random forest algorithm or a method based on a probability model or a combination thereof is provided for the classification of microarray data. In some embodiments, based on the statistical significance of the differences in biomarker levels between the categories of interest, predictive markers are selected to distinguish sample subtypes.

[0267] Table 1. SNF tetrad highly expressed genes

[0268] Treatment

[0269] In one aspect, the present disclosure provides a method for treating classic luminal breast cancer, comprising administering endocrine therapy, such as an estrogen receptor antagonist, to a patient. In some embodiments, classic luminal breast cancer is characterized by expression of α-SMA in a tumor sample below a predetermined expression threshold T. α-SMA, the expression level of CD8 is lower than the predetermined expression threshold T CD8 , and the expression level of Ki67 is lower than the predetermined expression threshold T Ki67 In other embodiments, the classic luminal breast cancer is characterized in that the pathological image of the tumor sample has a partially preserved normal mammary duct morphology (i.e., a proportion of normal mammary ducts equal to or exceeding a predetermined threshold is preserved, or the pathological image has a proportion of normal ductal epithelial cells equal to or exceeding a predetermined threshold). In other embodiments, the classic luminal breast cancer is characterized in that

[0270] In one aspect, the present disclosure provides a method for treating immune-regulated breast cancer, comprising administering immunotherapy, such as an immune checkpoint inhibitor, to a patient. In some embodiments, immune-regulated breast cancer is characterized by expression of α-SMA in a tumor sample below a predetermined expression threshold T. α-SMA And the expression level of CD8 reaches or exceeds the predetermined expression threshold T CD8 In other embodiments, the immune-regulated breast cancer is characterized by a proportion of immune cell infiltration exceeding a predetermined threshold in a pathological image of a tumor sample.

[0271] In one aspect, the present disclosure provides a method for treating proliferative breast cancer, comprising administering to a patient a CDK4 / 6 inhibitor, a PARP inhibitor, or a combination of a CDK4 / 6 inhibitor and a PARP inhibitor. In some embodiments, the proliferative breast cancer is characterized by an expression level of α-SMA in a tumor sample below a predetermined expression threshold T. α- SMA , the expression level of CD8 is lower than the predetermined expression threshold T CD8 , and the expression level of Ki67 reaches or exceeds the predetermined expression threshold of Ki67 T Ki67 In other embodiments, the proliferative breast cancer is characterized by having an abundance of atypical tumor cells exceeding a predetermined threshold in a pathology image of the tumor sample.

[0272] In one aspect, the present disclosure provides a method for treating receptor tyrosine kinase-driven breast cancer, comprising administering a tyrosine kinase inhibitor to a patient. In some embodiments, the receptor tyrosine kinase-driven breast cancer is characterized by expression of α-SMA at or above a predetermined expression threshold T in a tumor sample. α-SMA In other embodiments, the receptor tyrosine kinase-driven breast cancer is characterized by a proportion of cancer-associated fibroblasts exceeding a predetermined threshold in a pathology image of the tumor sample.

[0273] In one aspect, the present disclosure provides a method for treating HR-positive HER2-negative breast cancer, comprising:

[0274] (a) determining the subtype of HR-positive HER2-negative breast cancer by the typing system and / or method disclosed herein; and

[0275] (b) If the HR-positive HER2-negative breast cancer is determined to be of the classic luminal type, the patient is given endocrine therapy; if the HR-positive HER2-negative breast cancer is determined to be of the immune-regulatory type, the patient is given immunotherapy; if the HR-positive HER2-negative breast cancer is determined to be of the proliferative type, the patient is given a CDK4 / 6 inhibitor, or a PARP inhibitor, or a combination of a CDK4 / 6 inhibitor and a PARP inhibitor; or if the HR-positive HER2-negative breast cancer is determined to be of the receptor tyrosine kinase-driven type, the patient is given a tyrosine kinase inhibitor.

[0276] In some embodiments, immune checkpoint inhibitors are inhibitors of immune checkpoints selected from the following: PD-1, PD-L1, CTLA-4, TIGIT, NKG2A, B7-H3, VISTA, LAG3, TIM-3, 2B4, BTLA, LAIR1, CD160, and / or TGFRβ. In some embodiments, inhibition can be performed at the DNA, RNA, or protein level. In some embodiments, inhibitory nucleic acids (e.g., dsRNA, siRNA, or shRNA) can be used to inhibit the expression of inhibitory molecules. In other embodiments, the inhibitor of an inhibitory signal is a polypeptide, e.g., a soluble ligand (e.g., PD-1-Ig or CTLA-4Ig) or an antibody or antigen-binding fragment thereof that binds to an inhibitory molecule; e.g., an antibody or fragment thereof (also referred to herein as an "antibody molecule") that binds to PD-1, PD-L1, CTLA-4, TIGIT, NKG2A, B7-H3, VISTA, LAG3, TIM-3, 2B4, BTLA, LAIR1, CD160, and / or TGFRβ, or a combination thereof.

[0277] In some embodiments, the tyrosine kinase inhibitor is a receptor tyrosine kinase inhibitor. In some further embodiments, the tyrosine kinase inhibitor is selected from one or more of the tyrosine kinase inhibitors targeting the following kinases: ALK, ACVR1C, NTRK3, KIT, EGFR, PDGFR (especially PRGFRA or PDGFRB), TGFBR2, EPHB1, MET, NTRK2, TIE1, DDR2, TEK, ACVRL1, EPHA5, ROS1, FLT4, EPHB6, TYRO3, EPHA3, EPHA4, LTK, EPHA2, MERTK, CSF1R, ACVR2A, INSRR, KDR, AXL, FLT3, FGFR1, BMPR1A, ACVR1, NTRK1 and FLT1. In some embodiments, the tyrosine kinase inhibitor is selected from the tyrosine kinase inhibitors targeting the kinases shown in the table below. In some embodiments, the tyrosine kinase inhibitor is selected from the drugs shown in the table below.

[0278] In some further embodiments, the tyrosine kinase inhibitor is selected from the group consisting of an EGFR inhibitor, a Her3 inhibitor, an IGFR inhibitor, and a Met inhibitor. For example, tyrosine kinase inhibitors include, but are not limited to, erlotinib hydrochloride; linifanib (N-[4-(3-amino-1H-indazol-4-yl)phenyl]-N'-(2-fluoro-5-methylphenyl)urea, also known as ABT869, available from Genentech); sunitinib malate; bosutinib (4-[(2,4-dichloro-5-methoxyphenyl)amino]-6-methoxy-7-[3-(4-methylpiperazin-1-yl)propoxy]quinoline-3-carbonitrile, also known as SKI-606, and described in U.S. Pat. No. 6,780,996); dasatinib, pazopanib, sorafenib, vandetanib (ZD6474); and imatinib or imatinib mesylate.

[0279] In some further embodiments, epidermal growth factor receptor (EGFR) inhibitors include but are not limited to erlotinib hydrochloride; gefitinib (N-[4-[(3-chloro-4-fluorophenyl)amino]-7-[[(3"S")-tetrahydro-3-furyl]oxy]-6-quinazolinyl]-4(dimethylamino)-2-butenamide); vandetanib; lapatinib (3R,4R)- 4-Amino-1-((4-((3-methoxyphenyl)amino)pyrrolo[2,1-f][1,2,4]triazin-5-yl)methyl)piperidin-3-ol (BMS690514); canertinib dihydrochloride (CI-1033); 6-[4-[(4-ethyl-1-piperazinyl)methyl]phenyl]-N-[(1R)-1-phenylethyl]-7H-pyrrolo[2,3-d]pyrimidin-4-amine (AEE788, CAS 497839-62-0); Mubritinib (TAK165); Peltinib (EKB569); Afatinib, Neratinib (HKI-272); N-[4-[[1-[(3-fluorophenyl)methyl]-1H-indazol-5-yl]amino]-5-methylpyrrolo[2,1-f][1,2,4]triazin-6-yl]-carbamic acid, (3S)-3-morpholinomethyl ester (B MS599626); N-(3,4-dichloro-2-fluorophenyl)-6-methoxy-7-[[(3aα,5β,6aα)-octahydro-2-methylcyclopenta[c]pyrrol-5-yl]methoxy]-4-quinazolinamine (XL647, CAS781613-23-8); and 4-[4-[[(1R)-1-phenylethyl]amino]-7H-pyrrolo[2,3-d]pyrimidin-6-yl]-phenol (PKI166, CAS187724-61-4).

[0280] In some further embodiments, EGFR antibodies include but are not limited to Cetuximab; Panitumumab; Tocilizumab (EMD-72000); Nimotuzumab (hR3); Zalutumumab; TheraCIM h-R3; MDX0447 (CAS 339151-96-1); and ch806 (mAb-806, CAS 946414-09-1).

[0281] Other inhibitors include, but are not limited to, Neratinib (HKI-272, (2E)-N-[4-[[3-chloro-4-[(pyridin-2-yl)methoxy]phenyl]amino]-3-cyano-7-ethoxyquinolin-6-yl]-4-(dimethylamino)but-2-enamide) and are described in PCT Publication No. WO 05 / 028443); lapatinib or lapatinib ditosylate (3R,4R)-4-amino-1-((4-((3-methoxyphenyl)amino)pyrrolo[2,1-f][1,2,4]triazin-5-yl)methyl)piperidin-3-ol (BMS690514); (2E)-N-[4-[(3-chloro-4-fluorophenyl)amino]-7-[[(3S)-tetrahydro-3-furyl]oxy]-6-quinazolinyl]-4-(dimethylamino)-2-butenamide (BIBW-2992, CAS 850140-72-6); N-[4-[[1-[(3-fluorophenyl)methyl]-1H-indazol-5-yl]amino]-5-methylpyrrolo[2,1-f][1,2,4]triazin-6-yl]-carbamic acid, (3S)-3-morpholinylmethyl ester (BMS 599626, CAS 714971-09-2); canertinib dihydrochloride (PD183805 or CI-1033); and N-(3,4-dichloro-2-fluorophenyl)-6-methoxy-7-[[(3aα,5β,6aα)-octahydro-2-methylcyclopenta[c]pyrrol-5-yl]methoxy]-4-quinazolinamine (XL647, CAS 781613-23-8).

[0282] In some further embodiments, HER3 inhibitors include but are not limited to LJM716, MM-121, AMG-888, RG7116, REGN-1400, AV-203, MP-RM-1, MM-111, and MEHD-7945A.

[0283] In some further embodiments, MET inhibitors include but are not limited to Cabozantinib (XL184, CAS 849217-68-1); Foretinib (GSK1363089, formerly known as XL880, CAS 849217-64-7); Tivantinib (ARQ197, CAS1000873-98-2); 1-(2-hydroxy-2-methylpropyl)-N-(5-(7-methoxyquinolin-4-yloxy)pyridin-2-yl)-5-methyl-3-oxo-2-phenyl-2,3-dihydro-1H-pyrazole-4-carboxamide (AMG458); Crizotinib (PF-02341066); (3Z)-5-(2,3-dihydro-1H-indol-1-ylsulfonyl)-3-({3,5-dimethyl-4-[(4-methylpiperazin-1-yl)carbonyl]-1H-pyrrol-2-yl}methylene)-1,3-dihydro-2H-indol-2-one (S U11271); (3Z)-N-(3-chlorophenyl)-3-({3,5-dimethyl-4-[(4-methylpiperazin-1-yl)carbonyl]-1H-pyrrol-2-yl}methylene)-N-methyl-2-oxoindoline-5-sulfonamide (SU11274); (3Z)-N-(3-chlorophenyl)-3-{[3,5-dimethyl-4-(3-morpholin-4-ylpropyl)-1H-pyrrol-2-yl]methylene}-N-methyl-2-oxoindoline-5-sulfonamide (SU11606); 6-[difluoro[6-(1-methyl-1H-pyrazol-4-yl)-1,2,4-triazolo[4,3-b]pyridazin-3-yl]methyl]-quinoline (JNJ38877605, CAS 943540-75-8); 2-[4-[1-(quinolin-6-ylmethyl)-1H-[1,2,3]triazolo[4,5-b]pyrazin-6-yl]-1H-pyrazol-1-yl]ethanol (PF04217903, CAS 956905-27-4); N-((2R)-1,4-dioxan-2-ylmethyl)-N-methyl-N'-[3-(1-methyl-1H-pyrazol-4-yl)-5-oxo-5H-benzo[4,5]cyclohepta[1,2-b]pyridin-7-yl]sulfonamide (MK2461, CAS 917879-39-1); 6-[[6-(1-methyl-1H-pyrazol-4-yl)-1,2,4-triazolo[4,3-b]pyridazin-3-yl]thio]-quinoline (SGX523, CAS 1022150-57-7);and (3Z)-5-[[(2,6-dichlorophenyl)methyl]sulfonyl]-3-[[3,5-dimethyl-4-[[(2R)-2-(1-pyrrolidinylmethyl)-1-pyrrolidinyl]carbonyl]-1H-pyrrol-2-yl]methylene]-1,3-dihydro-2H-indol-2-one (PHA665752, CAS 477575-56-7). ;

[0284] In some further embodiments, IGFR inhibitors include but are not limited to BMS-754807, XL-228, OSI-906, GSK0904529A, A-928605, AXL1717, KW-2450, MK0646, AMG479, IMCA12, MEDI-573, and BI836845. See, for example, the review by Yee, JNCI, 104; 975 (2012).

[0285] In some embodiments, CDK4 / 6 inhibitors include but are not limited to palbociclib (CAS Registry Number: 571190-30-2, also known as PD-0332991, or 6-acetyl-8-cyclopentyl-5-methyl-2-{[5-(1-piperazinyl)-2-pyridinyl]amino}pyrido[2,3-d]pyrimidin-7(8H)-one), ribociclib (CAS Registry Number: 1211441-98-3). (CAS Registry Number: 1211441-98-3, also known as LEE011 or 7-cyclopentyl-N,N-dimethyl-2-((5-(piperazin-1-yl)pyridin-2-yl)amino)-7H-pyrrolo[2,3-d]pyrimidine-6-carboxamide), abemaciclib (CAS Registry Number: 1231929-97-7), also known as LY835219 or N-[5-[(4-ethyl-1-piperazinyl)methyl]-2-pyridinyl]-5-fluoro-4-[4-fluoro-2-methyl-1-(1-methylethyl)-1H-benzimidazol-6-yl]-2-pyrimidinamine), dalcil Trilacizide or compounds disclosed in International Application Publication No. WO 2010 / 020675, US Patent Nos. 8,415,355 and 8,685,980 (incorporated herein by reference in their entireties).

[0286] In some embodiments, the PARP inhibitor can be a PARP1, PARP2, PARP3, PARP4, PARP5, or PARP6 inhibitor. In some embodiments, the PARP inhibitor can be, for example, olaparib, rucaparib, niraparib, tazoparib, fluzoparib, pamiparib, and / or veliparib.

[0287] Reagent test kit

[0288] Kits for practicing the methods of the present invention may further be provided. A "kit" may encompass any article (e.g., packaging or container) comprising at least one reagent (e.g., an antibody, nucleic acid probe, or primer, etc.) for specifically detecting the expression of a biomarker or classifier gene of the present invention. The kit may be promoted, distributed, or sold as a unit for performing the methods of the present invention. In addition, the kit may include a package insert describing the kit and its use.

[0289] In one embodiment, a test kit for practicing the inventive method is provided. Such test kits are compatible with manual and automatic immunocytochemistry techniques (e.g., cell staining). These test kits comprise at least one antibody for a biomarker of interest or a sorter gene, chemicals for detecting the antibody combined with the biomarker, counterstains, and optionally blueing agents (to promote the identification of positively stained cells). Any chemical for detecting antigen-antibody binding can be used in the practice of the present invention.

[0290] Example

[0291] The following experiments illustrate the specific embodiments of the present invention.

[0292] (1) Cancer HR + / HER2 - Breast cancer patient samples, clinical data, and multi-platform comprehensive analysis

[0293] A Chinese HR study including 579 patients was established + / HER2 - A large-scale multi-omics cohort of breast cancer (CBCGA cohort) was included. Among them, 512 patients had whole-exome sequencing (WES) data of primary tumor tissue and paired blood samples, 527 patients had somatic copy number variation (SCNA) data, 569 patients had RNA sequencing data, 182 patients had TMT (Tandem Mass Tag)-based mass spectrometry quantitative protein data, 384 patients had metabolomics data, 421 patients had digital pathology data, and 9 patients had single-cell RNA-Seq data based on 10X Genomics.

[0294] In order to study the intrinsic biology of this disease from different dimensions, multi-omics clustering of somatic copy number variation, mRNA and metabolite abundance was performed using similarity network fusion (SNF) in 351 patients with overlapping data sets (referred to herein as "CBCGA cohort 351 samples" or "multi-omics gold standard cohort"). The distribution of PAM50 subtypes is consistent with that reported in the literature (Cancer Genome Atlas, N. Comprehensive molecular portraits of human breast tumours. Nature 490, 61-70 (2012)) (Figure 2), indicating the accuracy and reliability of multi-omics data. The eigenvalues ​​(y-axis) of the eigenvectors (x-axis) are plotted from the Laplacian matrix of the data graph. The largest eigengap is between the third and fifth eigenvectors, indicating that the optimal number of clusters for SNF cluster analysis is 4, i.e., SNF1-4 (Figure 3, Figure 4). In this way, the SNF classifier combining RNA expression, somatic copy number variation and metabolite abundance data can achieve HR + / HER2 - Statistical robustness and complete delineation of molecular heterogeneity in breast cancer.

[0295] This study investigated the distinct characteristics of each SNF subtype. Regarding the association between SNF subtypes and PAM50 subtypes, SNF1 is a mixed PAM50 luminal A / B subtype, while SNF3 primarily includes PAM50 luminal B subtypes. SNF2 encompasses nearly all PAM50 HER2-overexpressing subtypes and basal-like subtypes, and SNF4 contains a high proportion of PAM50 normal breast-like subtypes.

[0296] SNF1 and SNF3 tumors were almost exclusively composed of the PAM50 luminal A / B subtype. More than half of SNF1 tumors harbored PIK3CA mutations (approximately 51%), while fewer tumors harbored TP53 mutations (approximately 11%). SNF3 tumors had more CCND1 / FGFR1 / MDM2 copy number gains and more cell cycle pathway activation. High chromosomal instability and homologous recombination deficiency scores were also observed in SNF3 tumors, indicating high genomic instability. In addition, SNF1 and SNF3 tumors were enriched for hormone receptor pathways.

[0297] SNF2 and SNF4 tumors contained fewer PAM50 luminal A / B isoforms and less hormone receptor pathway activation. SNF2 tumors were enriched in adaptive immune response pathways. Expression markers associated with receptor protein kinases and extracellular matrix structural components were highly enriched in SNF4 tumors.

[0298] In this disclosure, the typing results obtained based on the multi-omics data of 351 samples from the CBCGA cohort are used as the gold standard for SNF typing in this article, which can be used as the gold standard for SNF typing in known HR + / HER2 - Breast cancer typing, used to train and validate the AI-based digital pathology typing method and the classifier gene expression-based typing method described in this disclosure.

[0299] (2) HR + / HER2 - The metabolomic landscape of breast cancer

[0300] Next, we explored differences in metabolic reprogramming signatures among the four SNF subtypes. We first discovered that expression distances were significantly greater than those in normal samples. We applied Euclidean distance to investigate the overall differences in metabolic gene expression between SNF subtypes and their corresponding normal tissues. We found that the SNF3 subtype had the highest level of metabolic dysregulation, while the SNF4 subtype had the lowest metabolic dysregulation. We then performed network analysis to explore subtype-specific metabolic genes, polar metabolites, and lipids, and to detail the metabolic signatures between subtypes. We found that the majority of dysregulated pathways were upregulated in the SNF3 subtype but downregulated in the SNF4 subtype. In addition, SNF1 and SNF2 isoforms were characterized by phospholipid metabolism, and high expression of genes involved in phospholipid synthesis such as PIP5K1A (East, MP, Laitinen, T. & Asquith, CRMPIP5K1A: a potential target for cancers with KRAS or TP53 mutations. Nat Rev Drug Discov 19, 436 (2020).) and PIK3CG (Semba, S. et al. Down-regulation of PIK3CG, a catalytic subunit of phosphatidylinositol 3-OH kinase, by CpG hypermethylation in human colorectal carcinoma. Clin Cancer Res 8, 3824-31 (2002)) was identified.

[0301] Furthermore, a global analysis was performed to investigate the association between polar metabolites and lipids and genomic events. This was to analyze the association between genomic variation and metabolic reprogramming. The relationship between metabolic genes, genomic mutations, and copy number variations and the abundance of polar metabolites and lipid subclasses in each subtype of the SNF cluster was investigated, with a focus on HR + / HER2 -Nineteen known cancer-related genes with high frequency mutations in breast cancer were selected, and linear regression models (controlling for confounding factors) were applied to examine the relationship between somatic mutations and metabolite abundance. TP53 mutations were found to be positively correlated with the abundance of nucleotides such as deoxyinosine and oxidized prostaglandins. Regarding copy number variation, the copy number values ​​of chromosomal regions (8q23.3 or 1q32.1) containing specific breast cancer oncogenes (such as TRPS1 and ADORA1) were found to be positively or negatively correlated with variant metabolites (nucleotides or amino acids). In addition, the association between the expression of cell cycle-related genes and DNA synthesis-related metabolites was also observed. For example, AURKA mRNA expression was observed to be positively correlated with deoxyinosine abundance, and CCND3 mRNA expression was negatively correlated with dUMP abundance.

[0302] (III) Different clinical characteristics of the four SNF subtypes

[0303] Immunohistochemistry (IHC) staining analysis revealed that SNF3 tumors had a significantly higher Ki67 index and tumor grade, while SNF2 tumors had lower ER expression. There was no significant difference in HER2 status between the various SNF subtypes. Interestingly, the distant metastasis-free survival and recurrence-free survival rates of SNF4 were significantly lower than those of the other three subtypes. In particular, for patients with PAM50 luminal surface A, the prognosis of SNF4 was far worse than that of the other subtypes. Univariate and multivariate Cox analyses showed that SNF4 subtype was an independent prognostic indicator.

[0304] (IV) Upregulation of cell cycle signaling pathways in SNF3 isoforms

[0305] CDK4 / 6 inhibitors can significantly improve HR + / HER2 -Patient survival rate. However, there are still many non-responders to this indicator, especially in Asian populations (Mayer, ELet al.Palbociclib with adjuvant endocrine therapy in early breast cancer (PALLAS): interim analysis of a multicentre, open-label, randomised, phase 3 study. Lancet Oncol 22, 212-222 (2021); Johnston, SRDet al.Abemaciclib Combined With Endocrine Therapy for the Adjuvant Treatment of HR+, HER2-, Node-Positive, High-Risk, Early Breast Cancer (monarchE). J Clin Oncol 38, 3987-3998 (2020).). Genomic analysis found that the copy number variation frequency of cell cycle genes in SNF3 samples increased overall. The relationship between these copy number variations and RNA or protein expression was then analyzed. Significant increases in RNA and protein expression of oncogenes such as CCND1, while decreased expression of tumor suppressor genes such as RB1, were observed, suggesting cis-activation of cell cycle signaling driven by genomic variants in SNF3. Specifically, a cis-effect between copy number variation and RNA expression of CCND1 was observed in both SNF3 and other tumors, while a cis-effect between copy number variation and protein levels was observed only in SNF3 tumors. Similarly, gene set enrichment analysis (GSEA) indicated that cell cycle pathways were activated in SNF3, with increased mRNA and protein abundance of other cell cycle genes, such as phosphorylated retinoblastoma protein (phospho-Rb), which was significantly elevated.

[0306] We also observed cis-regulation of cell cycle signaling in SNF3 by two regulators of G2 / M, the G2 / M activator MDM2 and the G2 / M inhibitor ATM. Furthermore, SNF3 tumors had significantly higher homologous recombination deficiency (HRD) and chromosomal instability (CIN) scores.

[0307] (5) HR + / HER2- The microenvironmental landscape of breast cancer

[0308] The tumor microenvironment (TME) is involved in tumor development and influences treatment options (Maman, S. & Witz, IPA history of exploring cancer in context. Nat Rev Cancer 18, 359-376 (2018)). Therefore, a comprehensive analysis of the microenvironmental characteristics of the four SNF subtypes was conducted.

[0309] The tumor microenvironment characteristics of SNF2 and SNF4 tumors were further analyzed using single-cell RNA sequencing (scRNA-Seq). A total of 59,479 cells from nine tumors were analyzed, and 11 distinct cell clusters were identified, including B cells, cancer-associated fibroblasts (CAFs), CD4+ T cells, CD8+ T cells, ISG (interferon-stimulated) T cells, NK cells, tumor cells, endothelial cells, myeloid cells, non-tumor epithelial cells, and proliferating cells. A higher proportion of CD8+ T cells was observed in SNF2 samples, and a higher proportion of CAFs was observed in SNF4 samples, validating the deconvolution results of bulk RNA-Seq data. Furthermore, single-cell sequencing was used to investigate the functional differences in immune cells between patient subtypes. Compared with non-SNF2 patients, CD8+ T cells from SNF2 patients were enriched in interferon response pathways, indicating an activated immune response. The cytotoxicity score, GZMK, and GNLY levels of CD8+ T cells from SNF2 patients were higher, indicating that CD8+ T lymphocytes in these patients were more cytotoxic.

[0310] Furthermore, immunohistochemistry experiments were performed to verify the tumor microenvironment characteristics of the four SNF subtypes. The score of the stromal cell marker α-smooth muscle actin (α-SMA) was significantly higher in SNF4; the immune infiltration marker CD8 was significantly higher in SNF2; and the proliferative cell-associated antigen marker Ki67 was higher in SNF3 than in SNF1 (Figure 5A, Figure 5B, Figure 5C).

[0311] The immunohistochemistry assay was performed as follows:

[0312] Paraffin-embedded tissue sections were deparaffinized at 60°C for 20 minutes, cleared in xylene, and treated with a graded alcohol series. For hematoxylin-eosin staining, slides were stained with Mayer's hematoxylin and 0.1% sodium bicarbonate and counterstained with eosin Y solution. For immunohistochemical analysis, slides were heated in saline-sodium citrate buffer at 95–100°C. After cooling, slides were blocked with blocking solution (2% goat serum, 2% bovine serum albumin, and 0.05% Tween in PBS) at room temperature and incubated with primary antibodies diluted in blocking solution at 4°C. Endogenous peroxidase activity was quenched with 0.3% H₂O₂. Slides were incubated with horseradish peroxidase-conjugated secondary antibodies at room temperature, developed with 3,3′-diaminobenzidine substrate, counterstained with hematoxylin, and dehydrated with a graded alcohol series. Positive static density was measured by a computerized imaging system consisting of a charge-coupled device camera connected to a microscope. The percentage of positively stained cells (0-100%) was assessed independently by two pathologists.

[0313] Based on the results of immunohistochemistry experiments, a new SNF classification system for HR-positive HER2-negative breast cancer can be summarized. The immunohistochemical scores of the four subtypes are as follows:

[0314] SNF1: α-SMA < 20%, CD8 < 3%, Ki67 < 25%, all three conditions are met;

[0315] SNF2: α-SMA < 20%, CD8 ≥ 3%, both of the above conditions are met;

[0316] SNF3: α-SMA < 20%, CD8 < 3%, Ki67 ≥ 25%, all three conditions met;

[0317] SNF4: α-SMA ≥ 20%.

[0318] In order to verify the discriminatory efficacy of the above-mentioned markers, receiver operating characteristic (ROC) curve analysis was performed on these three markers, and the results are shown in Figures 6A, 6B, and 6C. According to the area under the ROC curve (AUC), the above typing method can be determined as an appropriate model. In addition, the above-mentioned predicted SNF typing ratio based on immunohistochemistry score was compared with the SNF typing ratio of the luminal-facial multi-omics cohort of the present invention, and the results are shown in Figure 7. It can be seen from the results that the above-mentioned predicted SNF typing based on immunohistochemistry score can reproduce the SNF typing of the HR-positive HER2-negative breast cancer multi-omics cohort.

[0319] (VI) SNF4 subtype tumor-derived cancer-associated fibroblasts (CAFs) can promote tumor growth

[0320] Since patients with SNF4 subtype have HR + HER2- breast cancer patients have the worst prognosis, and it is necessary to explore the underlying mechanism. GSEA was performed using RNA-Seq data and found that the receptor tyrosine kinase pathway, including several important receptor tyrosine kinases, was significantly enriched in SNF4 patients. The receptor tyrosine kinase pathway is a key factor in promoting HR. + / HER2 - One of the downstream pathways of breast cancer progression (Burstein, HJ Systemic Therapy for Estrogen Receptor-Positive, HER2-Negative Breast Cancer. N Engl J Med 383, 2557-2570 (2020)). Receptor tyrosine kinases can be highly expressed in tumor or stromal cells (Du, Z. & Lovly, CM Mechanisms of receptor tyrosine kinase activation in cancer. Mol Cancer 17, 58 (2018); A. PDGF receptors in tumor stroma: Biological effects and associations with prognosis and response to treatment. Adv Drug Deliv Rev 121, 117-123 (2017).). Given that SNF4 tumors have a higher abundance of CAFs, receptor tyrosine kinases may exert their functions through CAFs. Through a comprehensive analysis of RNA-Seq, scRNA-Seq, and phosphorylation proteomics (CPTAC (Clinical Proteome Tumor Analysis Consortium) breast cancer cohort), it was observed that two receptor tyrosine kinases highly expressed in SNF4 (i.e., EGFR and PDGFRA) were primarily expressed and highly phosphorylated in CAFs. By Western blotting, it was verified that PDGFRA and EGFR were highly expressed and highly phosphorylated in SNF4 tumor samples, especially in SNF4 tumor-derived CAFs.

[0321] CAF can promote tumor growth (Gui, Y. et al. Metastatic Breast Carcinoma-Associated Fibroblasts Have Enhanced Protumorigenic Properties Related to Increased IGF2Expression. Clin Cancer Res 25, 7229-7242 (2019); Bertero, T. et al. Tumor-Stroma Mechanics Coordinate Amino Acid Availability to Sustain Tumor Growth and Malignancy. Cell Metab 29,124-140.e10(2019); Jungwirth, U.et al.Impairment of a distinct cancer-associated fibroblast population limits tumor growth and metastasis. Nat Commun 12,3516(2021).). To determine the effect of CAF on HR + / HER2 - To investigate the effects of cytotoxicity on breast cancer tumor growth, cancer cells were isolated from SNF1 (less aggressive isoform) and SNF4 (more aggressive isoform) tumors and cultured with conditioned medium from CAFs isolated from SNF4 tumors. Furthermore, conditioned medium from SNF4-derived CAFs was observed to promote the proliferation of tumor cells derived from both SNF1 and SNF4 tumors, suggesting that CAF abundance is a major mechanism of aggressiveness in the SNF4 isoform.

[0322] It has been reported that phosphorylated EGFR and PDGFRA can activate the downstream MAPK signaling pathway (Perrone, F. et al. PDGFRA, PDGFRB, EGFR, and downstream signaling activation in malignant peripheral nerve sheath tumor. Neuro Oncol 11, 725-36 (2009)). Accordingly, higher phosphorylation levels of key molecules of the MAPK pathway, such as RAF (RAF1) and ERK1 / 2 (MAPK1 / 3), were observed in SNF4 samples in the CPTAC cohort.

[0323] (VII) Establishment of a precise classification system for HR-positive and HER2-negative breast cancer based on digital pathology and artificial intelligence

[0324] 1) Digital pathology image collection, quality control and preprocessing (WSI image data preprocessing module)

[0325] Pathological hematoxylin and eosin (HE)-stained sections of tumor masses from patients with HR-positive, HER2-negative breast cancer treated at the Fudan University Cancer Center (FUSCC) between 2009 and 2016 were retrieved from the Department of Pathology. The slides were scanned using a Nanozoomer S210 digital slide scanner, and the resulting WSIs were quality-controlled. Exclusion criteria included the following: a) contaminated, broken, or severely wrinkled sections; b) samples of carcinoma in situ with microinvasion; and c) samples that had received neoadjuvant therapy before surgery. A total of 244 quality-controlled sections were included for subsequent analysis. Each section corresponded to one WSI, resulting in 244 WSIs for each of the 244 quality-controlled sections.

[0326] The ImageScope tool was used to delineate the region of interest (ROI) containing the majority of invasive cancer components. Using the OpenSlide library in MATLAB, each WSI was segmented into 256-by-256-pixel square blocks at 20x magnification using a sliding window, excluding background blocks. Each WSI had a different pixel size, resulting in a different total number of blocks. However, during subsequent modeling, an equal number of blocks were extracted from each WSI to ensure that all WSIs were evenly represented in the modeling.

[0327] 2) Constructing a neural network model for tissue type segmentation (the first neural network module)

[0328] Under the guidance of a pathologist, we manually annotated 20 WSIs with tissue types, depicting five tissue types: tumor, stroma, lymphatic infiltration (also known as immune infiltration), normal mammary ducts (also known as normal glands), and necrosis or hemorrhage. The depicted areas were then segmented into image blocks with classification labels, as shown in Figure 8. These 20 WSIs were not derived from the 244 quality control samples described in step 1, but from 20 additional breast cancer patients, with one WSI for each patient.

[0329] Fifteen randomly selected image patches from the 20 WSIs were used to build a neural network model for tissue segmentation. These 15 images generated a total of 126,967 patches, which were randomly divided into a training set and a validation set. The training set contained 100,134 patches, and the validation set contained 26,833 patches.

[0330] A ResNet-18 network architecture was selected for neural network modeling to obtain a tissue segmentation neural network model. Hyperparameter settings were as follows: loss function, cross entropy loss function; batch size, 256; learning rate, 0.001; and optimizer, ADAM. This modeling process can be performed as follows: manually annotating N full-scan digital pathology images of breast cancer patients, that is, drawing five tissue labels on them, namely, tumor tissue label, stromal tissue label, immune infiltration tissue label, normal glandular tissue label, and necrotic or hemorrhagic tissue label; cutting the N full-scan digital pathology images of the breast cancer patients into M image blocks at 20 times magnification of 256*256 pixels, 512*512 pixels or 800*800 pixels to obtain image blocks with tissue labels for training and validating the tissue type segmentation neural network model, wherein N>10, preferably N>14, 14<N<21, N=15, N=20; M>100000, preferably M>120000, M>126000, M>126900, M>126960, 120000<M<130000, M=126967.

[0331] In the validation set, the confusion matrix was used to evaluate the identification effect of the tissue type segmentation neural network model on each tissue type (Figure 9).

[0332] By integrating the prediction results of the tissue type segmentation neural network model for all image blocks of the same WSI, a tissue type segmentation heat map of the WSI can be generated. By comparing it with the original WSI image, the effectiveness of the tissue type segmentation neural network model is intuitively demonstrated (Figure 10).

[0333] The tissue segmentation neural network model was then applied to all image blocks in subsequent experiments, and image blocks with tumor tissue labels, stromal tissue labels, and immune infiltration tissue labels were selected for subsequent HR-positive and HER2-negative breast cancer typing analysis. The specific tissue labels in this invention refer to tumor tissue labels, stromal tissue labels, and / or immune infiltration tissue labels.

[0334] 3) Classification based on tissue type segmentation neural network model (second neural network module)

[0335] Using image patches with different tissue labels (e.g., tumor tissue, stromal tissue, and immune-infiltrated tissue) for each patient, a CNN model was trained and validated using a three-fold cross-validation algorithm to ultimately identify each patient's SNF classification. This three-fold cross-validation minimizes the impact of the training / validation set split on the model's predictive performance.

[0336] Specifically, the WSIs that had completed preprocessing and tissue type segmentation in the early stage (i.e., the 244 WSIs described in parts 1 and 2) were divided into three data sets, and the distribution of SNF subtypes was ensured to be consistent across the data sets. In each round, image blocks from two data sets were selected for training the CNN model, and image blocks from the remaining data set (i.e., the validation set) were used to validate the CNN model, with a total of three rounds of training (Figure 11). 500 image blocks with specific labels were randomly selected from each patient's WSI and annotated with SNF typing. In the training set, image blocks with specific labels and SNF annotations were input for deep learning of the CNN model. After deep learning of a large number of image blocks with specific labels and SNF annotations, the CNN model was able to use a one-to-many strategy to develop a CNN model for identifying each SNF subtype.

[0337] During training, each image patch was assigned a prediction score generated by the CNN model (range 0 to 1; the modeling process was replicated by setting a random seed (see Artificial intelligence for multimodal data integration in oncology, Cancer Cell 40, October 10, 2022, 1095–1110). The prediction score for each patient was defined as the mean of the scores for all image patches for that patient, and receiver operating characteristic (ROC) curve analysis was performed. After 50 cumulative CNN training rounds, the model with the largest area under the ROC curve (AUC) in the validation set was identified as the optimal model. Other hyperparameters were set as follows: loss function, cross-entropy loss function; batch size, 256; learning rate, 0.001; and optimizer, ADAM.

[0338] Consistent with the previous study, a three-fold cross-validation approach divided patients equally into three sets. Each time, one dataset was selected as the validation set to verify the CNN model's performance. Three rounds of training and validation generated three optimal models, and the prediction scores of the three optimal models were standardized and averaged. CNN models trained on image patches with different tissue labels were compared, and the accuracy of the prediction scores was evaluated through receiver operating characteristic (ROC) analysis. The CNN model with the highest area under the curve (AUC) was selected for SNF subtype discrimination (Figure 12). Finally, a prediction score for each SNF subtype was generated for each patient (Table 2). The digital pathology classification for each patient was the SNF subtype corresponding to the highest prediction score.

[0339] Pathological evaluation of the pathological features of each SNF subtype predicted by the CNN model (Figure 13). For each SNF subtype, the representative image blocks of the patient with the highest prediction score were visually reviewed. The image blocks of the SNF1 subtype partially retain the morphology of the normal breast. The characteristic image blocks of the SNF2 subtype have high immune cell infiltration. The representative image blocks of the SNF3 subtype show densely distributed tumor cells. The image blocks of the SNF4 subtype reflect tumor cell invasion into the stroma.

[0340] The computer hardware configuration used to construct the classification model through deep learning algorithm is: CPU model is I9-10900K, memory is 32G, hard disk is 4TB, and graphics card is RTX-3090-24G.

[0341] The software used includes Matlab R2021a (Mathworks Inc., USA), Python, and the Pycharm integrated development environment. The required libraries are as follows: joblib == 1.0.0, numpy == 1.19.1, opencv-python == 4.3.0.36, openslide-python == 1.1.2, pytorch == 1.9.1, sklearn == 0.0, tensorboard == 2.4.1, tensorboardx == 2.1, and torchvision == 0.10.1. Their general functions are: Matlab R2021a WSI generates image blocks; Pycharm Python code editor and integrated development environment; Joblib improves execution speed; Numpy Python basic library; Opencv-python openslide-python WSI processing and image block calling; Pytorch deep learning modeling; Tensorboard monitors and reproduces the modeling process; and Sklearn generates ROC curves.

[0342] Table 2 shows the pathological prediction scores of the patients' SNF1-4 subtypes.

[0343] Table 2. Pathological prediction scores of SNF1-4 subtypes in 244 patients

[0344] 4) Perform digital pathology typing in the validation cohort

[0345] In a validation cohort of 166 patients with HR-positive and HER2-negative breast cancer from the Affiliated Cancer Hospital of Fudan University, it was demonstrated that this digital pathology classification model can reproduce the key features of multi-omics classification.

[0346] Paraffin-embedded sections were obtained from 166 patients with HR-positive, HER2-negative breast cancer who were treated at the Fudan University Cancer Hospital between 2009 and 2016. Genomic data from all 166 patients with HR-positive, HER2-negative breast cancer were obtained using high-throughput sequencing.

[0347] The ImageScope tool was used to outline the region of interest (ROI) containing the majority of invasive cancer components. Each WSI was segmented into 256*256 pixel square blocks using the OpenSlide library of Matlab software in a sliding window format, excluding background blocks.

[0348] A prediction score for each subtype was calculated for these 166 new cases using a three-fold cross-validation method. These prediction scores reflect the likelihood that a sample is of the corresponding SNF subtype. Each patient received four final prediction scores for each of the four SNF subtypes, with the SNF subtype corresponding to the highest score being the patient's presumed SNF subtype.

[0349] Fresh tumor tissue was collected postoperatively from 244 patients with HR-positive, HER2-negative breast cancer. RNA was extracted and then subjected to next-generation sequencing. This yielded transcriptome data (reflecting the expression of individual genes). Pathway enrichment analysis was then performed using this transcriptome data, ultimately yielding enrichment scores for each pathway within a single sample. The resulting digital pathology classification was then validated for biological properties at the transcriptome level, as shown in Figure 14. Each row on the left side of the heatmap in Figure 14 represents a pathway, arranged in the order of the SNF1 / SNF2 / SNF3 / SNF4 signature pathways. Each column represents the enrichment analysis results for each pathway within a sample, with each cell's color varying depending on the enrichment score. The results suggest that the digital pathology classification can replicate key transcriptomic features across the four categorizations of HR-positive, HER2-negative breast cancer. For example, patients with the SNF1 subtype (blue) exhibit elevated expression of estrogen-related pathways. Patients with the SNF2 subtype (green) exhibit highly activated immune pathways, particularly adaptive immunity. Patients with the SNF3 subtype (yellow) exhibit elevated expression of cell cycle pathways. Receptor tyrosine kinase pathways are more highly upregulated in patients with SNF4 than in other subtypes.

[0350] The correspondence between digital pathology SNF classification and PAM50 classification is consistent with the results of multi-omics sequencing-based classification studies, that is, SNF1 type is mainly PAM50 luminal surface A / B, most PAM50 basal cell breast cancers and HER2-overexpressing breast cancers belong to SNF2, and almost all normal breast types belong to SNF4 (Figure 15).

[0351] A typing study based on multi-omics sequencing found that the tumor immune microenvironment of SNF2 patients was highly activated and accompanied by high expression of immune checkpoint molecules. Figure 16 analyzed the microenvironment of digital pathology SNF2 patients through transcriptome analysis and found that immune activation such as T cells in these patients was activated. Figure 13 shows a typical pathological section of a digital pathology SNF2 patient, whose tumor contains infiltrating lymphocytes such as CD8-positive T cells and activated NK cells (cells with smaller nuclei (dark blue) in the SNF2 image block in Figure 13). In Figure 17, immune checkpoint molecules such as PD-1 and IDO1 are highly expressed in digital pathology SNF2, and CD8A and GZMA are also highly expressed in SNF2, which is consistent with the previously identified SNF2 characteristics.

[0352] Using the Oncoscan kit, previously collected tumor tissue was analyzed for gene copy number variation, thereby detecting gene amplification and deletion. As shown in Figure 18, ATM deletion and MDM2 amplification were significantly more common in SNF3 than in other subtypes, consistent with previously identified SNF3 characteristics.

[0353] Multi-omics sequencing-based phenotyping studies confirmed that upregulation of cell cycle-related pathways, such as G0 / G1 and G2 / M, is a hallmark of SNF3 patients. As shown in Figure 20, SNF3 patients have high enrichment scores (NES>1.8) for cell cycle-related pathways, such as G0 / G1 and G2 / M, suggesting a high enrichment of these pathways, consistent with previously identified SNF3 characteristics.

[0354] Previous studies have shown that SNF3 patients characteristically overexpress E2F1, E2F2, and genes targeted by E2F. As shown in Figure 19, genes classically targeted by E2F1, E2F2, and E2F (such as AURKA and MYBL2) are also highly expressed in SNF3 patients, consistent with previously recognized SNF3 characteristics.

[0355] It was previously believed that SNF4 patients characteristically overexpress EGFR and RTK pathway-related genes. As shown in Figure 21, SNF4 patients overexpress RTK pathway genes such as EGFR, ALK, MET, and PDGFRA. As shown in Figure 22, SNF4 patients have high RTK pathway enrichment scores (NES>1.8), suggesting RTK pathway activation, which is consistent with the previously identified SNF4 characteristics.

[0356] (8) SNF subtype method based on transcriptomics data

[0357] 1) Development of SNF subtyping methods based on transcriptomics data

[0358] Referring to the method of Migliozzi et al. (Migliozzi, S. et al. Integrative multi-omics networks identify PKCδand DNA-PK as master kinases of glioblastoma subtypes and guide targeted cancer therapy. Nat Cancer 4, 181-202 (2023)), we developed a random forest classifier to classify HR+ / HER2- tumors into SNF subtypes using the transcriptomic data of 351 samples from the CBCGA cohort described above. First, we performed differential expression analysis using the R software package DESeq2 (v1.34) and retrieved the top 100 highly expressed genes for each subtype based on the log2 (fold change) value. Secondly, we performed collinearity removal to reduce the candidate gene set. We deleted all genes with a Spearman correlation coefficient higher than 0.75. Only the genes with the highest correlation with the corresponding SNF subtype were retained, namely the SNF1-SNF4 highly expressed genes listed in Table 1. Finally, based on the expression of the remaining genes, we developed a random forest classifier using the multi-omics typing results of 351 samples from the CBCGA cohort as the gold standard and measured its accuracy using ROC analysis. Based on this cohort of 351 samples, a transcriptome molecular typing prediction model (herein referred to as the "NG model") was constructed using the random forest algorithm. The model contains 309 predictive genes (i.e., classifier genes) with an accuracy of 84% for predicting molecular typing, and the AUC for four SNF subtypes is greater than 0.9. The NG modeling process is also described in the left panel of Figure 43.

[0359] 2) Simplification of classification model

[0360] To further reduce the number of classifier genes and increase their clinical application value, we simplified the 309 genes of the NG model based on the following process (Figure 43 right):

[0361] 1. 319 candidate gene sets were identified, including 309 NG model predicted genes and 10 manually added SNF multi-omics subtype characteristic functional genes (Table 3).

[0362] Table 3: SNF subtype characteristic genes

[0363] 2. Model Construction

[0364] The samples used for model construction are shown in the following table: the number of subtypes in CBCGA (multi-omics gold standard cohort) and PCMC (precision sequencing cohort).

[0365] 2.1 The most relevant eigengenes were selected from the 319 candidate gene sets using the maximal relevance and minimal redundancy (mRMR) algorithm. The genes were ranked from highest to lowest based on their correlation with the predictions, as shown in Table 4 below.

[0366] Excluding the 10 manually added functional genes from Table 4, the correlation of the 309 predicted genes in Table 1 is ranked from large to small, as shown in Table 5 below.

[0367] Table 5

[0368] 2.2 Using the SVM algorithm and leave-one-out cross-validation, the TOP N genes (the top N genes with correlation), Bottom N genes (the bottom N genes with correlation), and Random N genes (random N genes) were selected from the 319 candidate genes for modeling and accuracy calculation. The results are shown in Table 6 below.

[0369] Table 6

[0370] 2.3 Performance Evaluation at Different Gene Counts: The 319 candidate genes were ranked by relevance, with intervals of 30 genes each. The number of genes with the highest accuracy within each interval was selected. SVM models were then constructed using these genes on the entire cohort of 351 samples (the "Multi-Omics Gold Standard Cohort"), and the accuracy and AUC for each subtype were calculated. The results are shown in Table 7 below.

[0371] Table 7

[0372] 3. Balanced gene distribution

[0373] 3.1 Random selection of TOP genes may lead to uneven distribution of predicted genes for each subtype. To ensure that a considerable number of genes for each subtype are included in the modeling, the TOP5-TOP50 genes associated with each subtype were selected for modeling.

[0374] 3.3 Performance evaluation: The model was evaluated for its accuracy in predicting CBCGA, the mean square error of the proportion of each subtype in the test set (PCMC) and training set (CBCGA), and the AUC for predicting SNF subtypes.

[0375] The results are shown in Table 8 below.

[0376] Table 8

[0377] As an example, the top 25 genes associated with each subtype are shown in Table 9:

[0378] Table 9

[0379] In the tables above, Train.ACC is the SNF typing accuracy of the model predicted by the CBCGA cohort for all CBCGA samples, SNFx(AUC) is the AUC of each subtype predicted by the model, and test.RMSE is the mean squared error of the proportion of each subtype in the test set (PCMC) and the training set (CBCGA). CBCGA is the gold standard typing result, and PCMC is the typing result using the NG model.

[0380] The above results show that multiple simplified gene combination models of 319 candidate genes can achieve the SNF typing of the present invention with high accuracy and AUC.

[0381] Furthermore, based on the ranking of the 319-gene modeling, the 10 manually added functional genes all ranked low in relevance, with the highest ranking being 124, corresponding to the gene E2F1. This gene had little impact on the candidate gene modeling, so re-modeling was not necessary to validate the simplified model of the 309 genes. The conclusions regarding the simplified gene combination confirmed by the 319-gene modeling also apply to the simplified model of the 309 genes.

[0382] (IX) Establishment of drug response testing platform

[0383] 1) Preparation and culture of patient-derived organoids (PDOs)

[0384] A biobank for organoid storage was established using the recognized methods reported in previous literature (Gong Y, Ji P, Yang YS, Xie S, Yu TJ, Xiao Y, et al. Metabolic-Pathway-Based Subtyping of Triple-Negative Breast Cancer Reveals Potential Therapeutic Targets. CELL METAB. 2021; 33 (1): 51-64 doi 10.1016 / j.cmet.2020.10.012). Briefly, fresh breast cancer tissue was cut into small pieces with a sterile scalpel. The tissue was digested and resuspended in 10 mL of TAC buffer, incubated for 3 minutes to remove red blood cells, and passed through a 100 mm cell strainer (Corning). For passage, 5 ml of harvest solution (Trevigen, 3700-100-01) was used to digest the matrix gel (BME) and incubated on ice for 1 hour. Subsequently, the organoids were centrifuged at 350 g for 5 minutes, washed in digestion buffer, and centrifuged. Next, 3 mL of TrypLE Express (Invitrogen) was added, and the organoids were incubated at room temperature for 3 minutes before being mechanically dissociated into small cell clusters using a pipette. Organoids were passaged every 2-3 weeks at a dilution of 1:2-3.

[0385] 2) Drug response test of PDO

[0386] The drug response test of triple-negative breast cancer (TNBC) organoids was performed according to the method of the published paper (Xiao Y, Ma D, Yang YS, Yang F, Ding JH, Gong Y, et al. Comprehensive metabolomics expands precision medicine for triple-negative breast cancer. CELL RES. 2022; 32(5): 477-90 doi 10.1038 / s41422-022-00614-0.). Organoids were not treated with drugs. Organoids in good condition were harvested and digested into single cells. 25 μl of organoid suspension was added to a black cell exclusion surface (Greiner 781976-SIN) in a clear bottom 384-well plate, and 1×10 3 ~3×10 3Cells were added and cultured for an additional 5-6 days before drug treatment. Organoids were cultured with drugs for 1 week and then tested for viability. Organoid cell viability was assessed using the CellTiter-Glo 3D Cell Viability Assay (Promega, G9683) according to the manufacturer's instructions.

[0387] 3) Evaluation of the efficacy mechanism of endocrine therapy

[0388] SET Index Analysis

[0389] The SET index is a marker indicating the patient's sensitivity to endocrine therapy. Its definition can be found in the literature (Speers CW, Symmans WF, Barlow WE, Trevarton A, The S, Du L, et al. Evaluation of the Sensitivity to Endocrine Therapy Index and 21-Gene Breast Recurrence Score in the SWOG S8814 Trial. J CLIN ONCOL. 2023: O2201499 doi 10.1200 / JCO.22.01499; Du L, Yau C, Brown-Swigart L, Gould R, Krings G, Hirst GL, et al. Predicted sensitivity to endocrine therapy for stage II-III hormone receptor-positive and HER2-negative (HR+ / HER2-) breast cancer before chemo-endocrine therapy. ANN ONCOL. 2021; 32(5): 642-51 doi 10.1016 / j.annonc.2021.02.011.; Sinn BV, Fu C, Lau R, Litton J, Tsai TH, Murthy R, et al. SET (ER / PR): a robust 18-gene predictor for sensitivity to endocrine therapy for metastatic breast cancer. NPJ Breast Cancer. 2019; 5:16 doi 10.1038 / s41523-019-0111-0.).

[0390] Single-sample gene set enrichment analysis (ssGSEA)

[0391] For single-sample gene set enrichment analysis, the gene sets listed in Supplementary Table S2 were used to calculate the enrichment of the tamoxifen (TAM) / SERD response module and endocrine therapy resistance pathway using the “GSVA” function in R.

[0392] 4) Patient-derived tumor fragment (PDTF) platform

[0393] PDTF cultures

[0394] As previously reported, the collected tissue materials suitable for PDTF culture were cut on ice into 1-2 mm 3 Large small tumor fragments (Voabil P, de Bruijn M, Roelofsen LM, Hendriks SH, Brokamp S, van den Braber M, et al. An ex vivo tumor fragment platform to dissect response to PD-1 blockade in cancer. NAT MED. 2021; 27(7): 1250-61 doi 10.1038 / s41591-021-01398-3.). Single PDTFs from different regions within the tumor were mixed to ensure uniform representation of the entire tumor. Single PDTFs were embedded in an artificial extracellular matrix prepared as follows: tumor culture medium supplemented with 1.1% sodium bicarbonate, 1 mM sodium pyruvate, 1x MEM non-essential amino acids, 2 mM L-glutamine, 10% fetal bovine serum, 1% penicillin-streptomycin and collagen (final concentration of 1 mg / ml) and mixed with ice-cold Matrigel (final concentration 4 mg / ml). Take a 96-well plate, apply 40 μl of matrix per well as the bottom layer, and expose to 37 ° C for 20-30 minutes to solidify. Place a single PDTF in each well on top of the pre-solidified matrix, and then place a second layer of 40 μl matrix. The plate is then placed in an incubator at 37 ° C for 20-30 minutes to wait for solidification. Next, 140 μl of tumor culture medium is added to the top, which is supplemented with nivolumab (final concentration 20 μg / ml) or human anti-β-Gal-hIgG4 as a control. 8 PDTFs were used for each condition, and the PDTF culture was kept at 37 ° C for 48 hours before flow cytometric analysis.

[0395] PDTF flow cytometric analysis

[0396] PDTF was analyzed by flow cytometry to assess T cell activation after incubation with nivolumab or IgG to determine the tumor immune environment. T cell activation was detected using antibodies from BioLegend as described (Voabil P, de Bruijn M, Roelofsen LM, Hendriks SH, Brokamp S, van den Braber M, et al. An ex vivo tumor fragment platform to dissect response to PD-1 blockade in cancer. NAT MED. 2021; 27(7): 1250-61 doi 10.1038 / s41591-021-01398-3.). PDTFs for flow cytometric analysis were manually retrieved from the matrix and processed into single-cell suspensions for each experimental condition by enzymatic digestion in a digestion mixture (DMEM / F12 medium supplemented with 1% penicillin-streptomycin, 10% bovine serum albumin, 0.5 μg / ml cortisol, 5 μM Y-27623, 1X insulin, 1 mg / ml type IV collagenase [Worthington], hyaluronidase [Sigma], and dornase alfa [Sigma]) with gentle rotation at 37°C for 1–2 hours. The digested samples were filtered twice through a 70 μM filter, washed in PBS, resuspended in 100 μl of PBS, and incubated with Fc receptor blocker (BioLegend) and Zombie NIR (BioLegends) for 20 minutes at 4°C. The cells were washed, resuspended in 50 μl of staining buffer (BioLegend) containing the above-mentioned antibodies, and incubated at 4°C for 20 minutes. After washing twice, cells were transferred to 100 μl of cell staining buffer and subjected to flow cytometry analysis.

[0397] Analysis of secreted mediators

[0398] For analysis of cytokines, chemokines, and cytotoxicity mediators, supernatants from PDTF cultures were collected after 48 hours of culture. The supernatants were immediately frozen and stored at -80°C. The supernatants were thawed on ice and, except for single-fragment analysis, the supernatants from each condition were combined and assayed for the presence of the aforementioned cytokines and chemokines using LEGENDplex human Th cytokines, human proinflammatory chemokines (BioLegend), or IFN / TNF human ELISA kits (BioLegend) according to the manufacturer's instructions.

[0399] 5) Quantification and statistical analysis

[0400] Frequency tables and summary statistics were used to characterize data distribution. Wilcoxon and Kruskal-Wallis tests were used to compare continuous and categorical variables, and Pearson chi-square tests and Fisher's exact tests were used to compare unordered categorical variables. Survival curves were constructed using the Kaplan-Meier product limit method and compared with the log-rank test to obtain RFS and PFS. RFS was defined as the time from diagnosis to the first recurrence or diagnosis of contralateral breast cancer. PFS was defined as the time interval from the start of treatment to disease progression or death due to any cause (whichever occurred first), or to the last PFS assessment for patients with progression-free survival. Univariate and multivariate Cox proportional hazards regression models were constructed to estimate hazard ratios (HRs) and corresponding 95% confidence intervals (CIs). All analyses were performed using the R software package version 3.6.1 (https: / / cran.r-project.org / ).

[0401] (X) Sensitivity of SNF1 subtypes to endocrine therapy

[0402] The sensitivity of three traditional endocrine therapies (estrogen deprivation, 4-hydroxytamoxifen (4OHtam), and fulvestrant) to four subtypes of PDO (n=88) was compared on the PDO drug testing platform. It was observed that SNF1-PDO had the lowest survival rate among the three traditional endocrine therapies, while SNF4-PDO had a relatively high survival rate (Figures 23 and 24). To investigate the potential mechanism by which SNF1 tumors are sensitive to endocrine therapy, endocrine therapy-related pathways in the multi-omics data of the Affiliated Cancer Hospital of Fudan University were analyzed. It was observed that the SET index was highly enriched in SNF1 tumors, and the endocrine therapy response module was also highly expressed in SNF1 tumors (Figure 25). In addition, the expression of endocrine therapy resistance pathways in the four subtypes was also compared. The heat map showed that the expression of endocrine therapy resistance pathways (cell cycle pathway, RTK pathway, WNT pathway, EMT pathway, and PI3K pathway) was reduced in SNF1 subtype tumors (Figure 25). These results demonstrate that patients with the SNF1 subtype are sensitive to endocrine therapy, with relatively high activation of the response module and low expression of various resistance modules.

[0403] (XI) Sensitivity of SNF2 subtypes to immunotherapy

[0404] As described in the literature (Xi Jin, Yi-Fan Zhou, Ding Ma, Shen Zhao, Cai-Jin Lin, Yi Xiao, et al. Integrative Molecular Classification of HR+ / HER2-Breast Cancer Guides Precision Treatment. Nature Genetics under minor revision. 2023.02.26.), SNF2 subtype tumors are rich in immune cells and may therefore be sensitive to immunotherapy, especially immune checkpoint inhibitors (such as PD-1 antibodies, PD-L1 antibodies, CTLA4 antibodies, etc.). To verify this hypothesis, we used the PDTF platform to analyze the immune response of human tumor tissues to ex vivo PD-1 blockade. Using cytokines and T cell activation markers as readouts, we analyzed the effects of PD-1 blockade on 49 PDTFs after 48 hours of incubation with anti-PD-1 antibodies. Flow cytometry results showed that CD3+ total T activation markers (ICOS, CD137) increased more significantly from baseline in the SNF2 subtype after immunotherapy compared with other subtypes (Figures 26A and 26B). Importantly, in the SNF2 subtype, PD-1 blockade not only significantly increased those parameters that directly reflect T cell activation, but also significantly increased cytokines (TNF-α and IFN-γ) (Figures 27A and 27B).

[0405] (XII) Susceptibility of SNF3 subtypes to treatment with CDK4 / 6 inhibitors and PARP inhibitors

[0406] SNF3 isoforms in early and late HR + / HER2 -Breast cancer real cohorts show different sensitivities to endocrine therapy (Figure 28A, Figure 28B), so it is necessary to explore its underlying mechanisms and find new treatment strategies. An analysis of the PFS time of first-line endocrine therapy for stage IV patients in a real-world cohort revealed that most SNF3 subtype patients progressed after 6 months of endocrine therapy (Figure 29), suggesting that SNF3 subtype patients may have developed secondary resistance to endocrine therapy according to the international consensus guidelines for advanced breast cancer. In order to gain a deeper understanding of the biological characteristics of the SNF3 subtype, low-concentration 4-hydroxytamoxifen was used to induce PDO for a long time to establish a model for studying acquired resistance to endocrine therapy (Xi Jin, Yi-Fan Zhou, Ding Ma, Shen Zhao, Cai-Jin Lin, Yi Xiao, et al. Integrative Molecular Classification of HR+ / HER2-Breast Cancer Guides Precision Treatment. Nature Genetics under minor revision. 2023.02.26.). Surprisingly, unlike other subtypes, SNF3-PDO became less sensitive to endocrine therapy after long-term induction with 4-hydroxytamoxifen (Figure 30A, Figure 30B). These results suggest that endocrine therapy alone may not be sufficient for the SNF3 subtype and requires intensive treatment.

[0407] Given that SNF3 tumors are highly enriched for homologous recombination deficiency scores, chromosomal instability, and cell cycle activity, we tested the response of SNF3 PDOs to CDK4 / 6 and PARP inhibitors. The growth ability of SNF3-PDOs was significantly inhibited by the CDK4 / 6 inhibitor abemaciclib (Figure 31) and the PARP inhibitor olaparib (Figure 32). Interestingly, the combination of abemaciclib and olaparib significantly reduced the survival of SNF3-PDOs compared to either drug alone, and this effect was even more pronounced after induction with 4-hydroxytamoxifen (Figure 33). RNA-seq of PDOs revealed that cell cycle signaling pathways were more active in SNF3-PDOs than in other subtypes (Figure 34). Furthermore, SNF3 tumors had significantly higher homologous recombination deficiency scores, regardless of BRCA mutation status (Figure 35). These results indicate that the SNF3 subtype is susceptible to treatment with CDK4 / 6 inhibitors and PARP inhibitors, so CDK4 / 6 inhibitors and PARP inhibitors or their combination may be potential precision treatment strategies for SNF3 tumors.

[0408] To further investigate the mechanisms behind the drug combination, high-throughput transcriptome analysis was performed on a drug testing platform called DRUG-seq (small-scale high-throughput transcriptome sequencing) (Ye C, Ho DJ, Neri M, Yang C, Kulkarni T, Randhawa R, et al. DRUG-seq for miniaturized high-throughput transcriptome profiling in drug discovery. NAT COMMUN. 2018; 9(1):4307 doi 10.1038 / s41467-018-06500-x.). DRUG-seq showed that after combined drug therapy, oncogene signaling pathways related to cell cycle and DNA damage repair in SNF3 tumors were significantly downregulated (Figure 36). Unlike the changes in signaling pathways after monotherapy, some new signaling pathways underwent significant changes after combined drug therapy, suggesting a potential mechanism for combined drug therapy in SNF3 tumors.

[0409] (XIII) Inhibition of receptor tyrosine kinase pathways can inhibit the progression of SNF4 subtypes

[0410] Given that receptor tyrosine kinase signaling pathways are enriched in the SNF4 subtype, the responses of the four subtypes to tyrosine kinase inhibitors were evaluated. According to the literature method (Lemmon MA, Schlessinger J. Cell signaling by receptor tyrosine kinases. CELL. 2010; 141(7): 1117-34 doi 10.1016 / j.cell.2010.06.011.), 20 receptor tyrosine kinase subfamilies and their major receptors were identified. To screen for key therapeutic targets, the expression of receptors in the receptor tyrosine kinase pathway was first compared. It was found that most receptors in some receptor tyrosine kinase pathways (such as PDGFR and VEGFR pathways) were significantly overexpressed in the SNF4 subtype (Figure 37).

[0411] Next, based on the identified key receptors, 12 tyrosine kinase inhibitors were selected to construct a "tyrosine kinase inhibitor panel." These included tyrosine kinase inhibitors targeting the ErbB family (afatinib, lapatinib, pyrotinib), the PDGFR family (axitinib, sunitinib), the VEGFR family (foretinib, apatinib), and TAM receptors (sitravatinib), as well as several multi-target tyrosine kinase inhibitors (sorafenib, vandetanib, regorafenib, and famitinib). This panel was used in PDO drug sensitivity testing to identify the tyrosine kinase inhibitor most sensitive to SNF4 tumors (Figure 38). The results showed that in SNF4 tumors, the IC50 values ​​of most tyrosine kinase inhibitors, particularly famitinib, sorafenib, regorafenib, and vandetanib, were lower than those of other subtypes (Figures 38 and 39). Following treatment with famitinib (Figure 40A) and sorafenib (Figure 40B), SNF4 tumors showed significant downregulation of PDGF-activated signaling pathways. Following treatment with regorafenib (Figure 41A) and vandetanib (Figure 41B), SNF4 tumor cell viability was significantly decreased (Wilcoxon test, *, p < 0.05; **, p < 0.01; ***, p < 0.001). Combined with the drug sensitivity results for the 12 tyrosine kinase inhibitors and the targets of the four drugs mentioned above, VEGFR and PDGFR signaling were preliminarily identified as potential therapeutic susceptibilities for SNF4 tumors.

[0412] 14. Real-world research

[0413] We established a multi-omics cohort (n=351), a multicenter real-world clinical cohort (n=643), a prospective clinical cohort (MULAN trial), and a drug testing platform that included organ tissues (n=126) and patient-derived tumor fragments (n=49) derived from HR+ / HER2- breast cancer patients. By integrating "Bench" and "Bedside" data, we conducted comprehensive preclinical and clinical translational studies of precision strategies for HR+ / HER2- breast cancer.

[0414] We established a comprehensive classification system for HR+ / HER2- breast cancer, encompassing four distinct subtypes. Through clinical cohort studies, omics analyses, and functional assays, our study demonstrated the efficacy and underlying mechanisms of precision medicine approaches tailored to each subtype: endocrine therapy alone for the classic luminal-endoluminal subtype; the addition of a CDK4 / 6 inhibitor and a PARP inhibitor for the proliferative subtype; immunotherapy for the immune-regulatory subtype; and tyrosine kinase inhibitors for the receptor tyrosine kinase-driven subtype. Using this clinically applicable subtyping approach, we validated in a real-world cohort that a matched precision medicine strategy outperformed an unmatched approach, nearly doubling the median progression-free survival (9.83 months [95% CI, 5.74-13.93] vs 4.77 months [95% CI, 3.35-6.18]; hazard ratio, 0.64 [95% CI, 0.41-0.99]) for patients with refractory advanced HR+ / HER2- breast cancer. Importantly, a phase I analysis of the MULAN Phase II umbrella clinical trial (NCT04355858) confirmed a higher objective response rate (88.9%, 95% CI: 51.7%-99.7%) with subtype-directed precision medicine.

[0415] Our study highlights the superiority of subtype-directed precision treatment strategies for HR+ / HER2- breast cancer, improves the traditional "one-size-fits-all" therapy, and promotes the translation of precision treatment strategies from laboratory to clinic.

[0416] method

[0417] Clinical cohort

[0418] Multi-omics cohort

[0419] Using the CBCGA cohort described above, a total of 579 patients with HR+ / HER2- breast cancer were enrolled. Of these, 512 had whole-exome sequencing (WES) data, 527 had somatic CNA (SCNA) data, and 569 had RNA sequencing data. Multi-omics SNF subtype classification was available for 351 patients. Follow-up for these patients ended on June 30, 2021, with a median follow-up of 83.0 months.

[0420] Real-world retrospective clinical cohort

[0421] This study involved two real-world retrospective cohorts.

[0422] A real-world cohort of patients with early-stage HR+ / HER2- breast cancer

[0423] We conducted a retrospective review of patients who underwent surgical resection at FUSCC between January 2014 and December 2021. Patient selection criteria were as follows: 1) histopathologically confirmed HR+ / HER2- unilateral invasive carcinoma; 2) availability of digital pathology images of the tumor; 3) no evidence of distant metastasis; and 4) a complete medical history available for further analysis. Their clinicopathological characteristics, including tumor size, lymph node status, histological grade, and ER, PR, and Ki67 expression, were obtained from original medical records.

[0424] A real-world cohort of patients with advanced HR+ / HER2- breast cancer

[0425] To gain a deeper understanding of the landscape of metastatic HR+ / HER2- breast cancer, we constructed a real-world cohort of patients with advanced HR+ / HER2- breast cancer from the FUSCC and invited three other centers to participate in our real-world study. Ultimately, 462 patients were enrolled based on the following criteria: 1) treatment at the four centers between January 2014 and December 2021; 2) biopsy or surgical resection tumor specimens confirmed as sites of HR+ / HER2- breast cancer metastasis by histopathology; 3) availability of digital pathology images; and 4) a complete medical history to facilitate further analysis of their treatment response. Patients in this cohort received either fulvestrant backbone therapy or targeted therapy including CDK4 / 6 inhibitors and TKIs. Their tumor progression was assessed according to RECIST v1.1.

[0426] Prospective clinical cohort

[0427] A prospective clinical cohort of HR+ / HER2- breast cancer was derived from the MULAN clinical trial (ClinicalTrials.gov Identifier: NCT04355858). Patients must meet all of the following inclusion criteria to be included in this study: 1) treated at FUSCC between April 2020 and May 2023; 2) histologically confirmed HR+ / HER2- invasive breast cancer; 3) locally advanced breast cancer (incapable of receiving treatment); 3) locally advanced breast cancer (ineligible for local radical treatment) or recurrent metastatic breast cancer; 4) patients with HR+ / HER2- advanced breast cancer with a history of endocrine resistance (previously treated with CDK4 / 6 inhibitors) and currently in progressive disease are eligible if they meet one of the following conditions: (1) patients who meet the criteria for primary endocrine resistance (recurrence or metastasis within two years of starting adjuvant endocrine therapy, or disease progression within 6 months of late endocrine therapy) and have received at least one cycle of chemotherapy in the late stage; (2) patients with liver metastases who have received at least one cycle of endocrine therapy and one cycle of chemotherapy in the late stage; (3) patients without liver metastases who have received at least two cycles of endocrine therapy and one cycle of chemotherapy in the late stage.

[0428] result

[0429] Multipronged Exploration of Precision Medicine: Study Cohorts and Research Designs

[0430] To optimize the traditional "one-size-fits-all" treatment strategy for HR+ / HER2- breast cancer, we established molecular subtypes based on a multi-omics cohort (n=351) and successfully divided the HR+ / HER2- population into four SNF subtypes (Figure 44A).

[0431] This study adopted a multi-pronged precision therapy exploration approach (Figure 44B). We established a multicenter real-world retrospective clinical cohort (n = 643) to further explore the clinical characteristics, prognosis, and treatment sensitivity of the four subtypes of HR+ / HER2- breast cancer. In addition, as described above, we also constructed a PDO model (n = 126) and a PDTF model (n = 49) for drug testing and proposed a subtype-oriented precision therapy strategy. Finally, we validated the efficacy of the precision therapy strategy in a prospective breast cancer clinical cohort. According to the deep learning-based digital pathology classification pipeline described in this article, we used hematoxylin and eosin (H&E)-stained sections to classify patients into SNF subtypes in our study. In summary, we conducted a comprehensive and in-depth study of the subtype-directed precision therapy strategy for HR+ / HER2- breast cancer, achieving a transformation from multi-omics analysis to functional testing and from bench to bedside.

[0432] Clinical characteristics of the four subtypes in a real-world cohort

[0433] We analyzed the clinical characteristics of the four SNF subtypes (determined by digital pathology) in a multicenter real-world cohort that included patients with HR+ / HER2- breast cancer (Figure 45A). In the real-world cohort of early-stage HR+ / HER2- breast cancer (early cohort, n=216), the pathological characteristics of the four SNF subtypes were similar to those in the multi-omics cohort, as described herein. The morphological characteristics of each SNF subtype are as follows: the SNF1 subtype partially retains the morphology of the normal breast; the SNF2 subtype has high immune cell infiltration; the SNF3 subtype is rich in atypical tumor cells; and the SNF4 subtype is rich in tumor cell clusters with surrounding fibroblasts. In the real-world HR+ / HER2- breast cancer cohort, the SNF2 subtype has low ER expression. In addition, the Ki-67 index of SNF3 tumors was significantly higher, as measured by immunohistochemistry (IHC) staining. More importantly, the proportions of each subtype in the real-world early cohort were also consistent with the proportions in the multi-omics cohort of the same early-stage patients. These results were consistent with the subtype characteristics of the multi-omics classification, further demonstrating the feasibility of differentiating the four subtypes based on digital pathology.

[0434] Response to conventional therapy in patients by subtype in a real-world cohort

[0435] In a real-world cohort, we also analyzed the response of each subtype to traditional clinical treatment. In the early-stage cohort, among patients with HR+ / HER2- breast cancer who received postoperative endocrine therapy, patients with the SNF4 subtype had worse recurrence-free survival (RFS), while patients with SNF1 had better RFS (log-rank p < 0.01, Figure 45B). We also explored whether tumor size and lymph node status were potential confounders in our analysis and found no significant imbalance in these variables among the four SNF types. Compared with the early-stage cohort, the proportion of patients with the SNF4 subtype in a real-world cohort of patients with advanced HR+ / HER2- breast cancer (late-stage cohort) was significantly increased, indicating that SNF4 had a worse prognosis (P < 0.01, Figure 45C). In the late-stage cohort from the Affiliated Cancer Hospital of Fudan University (FUSCC, n = 277), SNF3 and SNF4 subtypes were associated with worse progression-free survival (PFS) with fulvestrant backbone therapy (p = 0.033, Figure 45D). Surprisingly, we found that patients with the SNF3 subtype were more likely to achieve a partial response in patients treated with CDK4 / 6i (p < 0.05, Figure 45E). This significant difference in treatment response suggests the limitations of the traditional "one-size-fits-all" treatment strategy and the need for precision therapy guided by the four molecular subtypes.

[0436] Prospective clinical cohort validation of subtype-directed precision therapy strategies

[0437] To validate this precision medicine strategy across four subtypes, we conducted efficacy analyses in prospective clinical cohorts of HR+ / HER2- breast cancer patients stratified by SNF subtype within the Phase II umbrella trial MULAN (NCT04355858) (Figure 46A). At the May 2023 cutoff, we found that sufficient patients had been enrolled in Arm II (CDK4 / 6i plus PARPi therapy) and Arm V (anti-PD-1 backbone therapy) to meet the protocol-specified prespecified boundaries for the first-stage analysis. In the first stage of Arm II and Arm V, objective response rates (ORRs) reached 85.7% (95% CI: 42.1%-99.6%) and 57.1% (95% CI: 18.4%-90.1%), respectively (Figures 46B-C). Due to the strong efficacy, Arms II and V of the MULAN study will proceed to the next stage, as per the Simon two-stage design. Safety data were consistent with the known safety profile of the drugs. No grade 5 treatment-related adverse events (AEs) were reported.

[0438] In our exploratory analysis of the MULAN trial, we found that compared with other subtypes, patients with the SNF2 subtype in the V arm (identified by deep learning-based digital pathology) had a longer duration of response to immunotherapy, and more patients achieved a partial response (PR) according to RECIST v1.1 (Figure 7B). In addition, patients with advanced SNF2 breast cancer achieved PR and maintained long-term stability after receiving immunotherapy even after progression after multiple lines (≥5 lines) of metastatic treatment. In addition, in the patients in the II arm who received a combination of CDK4 / 6i and PARPi, all four patients with advanced breast cancer of the SNF3 subtype achieved PR, one of whom had a response duration of more than 12 months and another had a tumor regression of more than 70% of diffuse metastases in the liver (Figure 7C). Notably, in the MULAN clinical study, patients who received a precision therapy strategy guided by SNF subtype (matched group) were more likely to achieve an objective response compared to the unmatched group (88.9% [95% CI: 51.7%-99.7%] vs 40.0% [95% CI: 5.3%-85.3%], p = 0.095, Figure 7D). Furthermore, in a real-world clinical cohort of advanced HR+ / HER2- breast cancer, patients with the SNF4 subtype who were resistant to endocrine therapy were relatively sensitive to TKI treatment (Figure 46E). These findings suggest that adopting a subtype-directed precision therapy strategy for specific subtypes can bring the benefits of matched therapy.

[0439] To assess the overall benefit of a subtype-directed precision therapy strategy in a real-world clinical cohort of patients with advanced HR+ / HER2- breast cancer, we included patients who had received three or more lines of therapy in the metastatic setting (Supplementary Table S10). Patients who received precision therapy based on their SNF subtype were assigned to the matched group, while those who did not receive such therapy were assigned to the unmatched group. Supplementary Table S11 provides detailed criteria for determining the group into which patients were assigned according to the SNF subtype-guided precision therapy strategy. Notably, patients with specific SNF subtypes who received matched therapy based on the subtype-directed precision therapy strategy had significantly better progression-free survival than those who received unmatched physician-choice therapy (9.83 months [95% CI, 5.74-13.93] vs 4.77 months [95% CI, 3.35-6.18]; hazard ratio 0.64 [95% CI, 0.41-0.99], p = 0.045; Figure 46F). Figure 7G shows examples of patients who achieved an objective response.

[0440] In conclusion, our clinical cohort study highlights the superiority of a subtype-specific precision therapy strategy for HR+ / HER2- breast cancer. These findings demonstrate the benefits of a personalized medicine approach and underscore the importance of using subtype-specific precision therapy strategies for the treatment of HR+ / HER2- breast cancer.

Claims

1. A system for classifying HR-positive HER2-negative breast cancer, the system comprising: 1) a full scan digital pathology image data preprocessing module, which is used to cut the full scan digital pathology image of the tumor tissue sample of the HR-positive HER2-negative breast cancer into a plurality of enlarged image blocks; 2) a first neural network module, wherein the first neural network module is a tissue type segmentation neural network module, which is used to classify the image blocks into different tissue types and add corresponding tissue labels; 3) a second neural network module connected in series with the first neural network module, wherein the second neural network module is a typing prediction neural network module, which is used to perform the typing of the sample based on the image block with the tissue label; in, The classification refers to dividing HR-positive HER2-negative breast cancer into the following four subtypes: classic luminal type, immune regulation type, proliferative type, and receptor tyrosine kinase driven type.

2. The system of claim 1, wherein the first neural network module and the second neural network module are constructed through artificial intelligence deep learning based on full-scan digital pathology images of tumor tissue samples of patients with known HR-positive HER2-negative breast cancer subtypes.

3. The system of claim 1 or 2, wherein the tissue types are tumor, stroma, immune infiltration, normal gland, and necrosis or hemorrhage, and the tissue labels are tumor, stroma, immune infiltration, normal gland, and necrosis or hemorrhage.

4. The system of claim 1 or 2, wherein the tissue signature is tumor, stroma and / or immune infiltration.

5. The system of claim 1 or 2, wherein the artificial intelligence deep learning comprises training and validating a convolutional neural network model using image patches with tissue labels and their corresponding HR-positive HER2-negative breast cancer subtypes.

6. The system of claim 5, wherein the artificial intelligence deep learning includes training and validating the convolutional neural network model using a three-fold cross-validation algorithm.

7. The system of claim 5, wherein the artificial intelligence deep learning includes training and validating the convolutional neural network model by using a nonlinear classifier of an artificial intelligence support vector machine and a one-to-many method.

8. The system of claim 5, wherein the neural network modeling is performed using a ResNet-18 network structure to obtain a tissue type segmentation neural network model.

Citation Information

Patent Citations

  • Process for the preparation of 7-substituted-3 quinolinecarbonitriles

    US6780996B2

  • Pyrrolopyrimidine compounds and their uses

    US8415355B2

  • Pyrrolopyrimidine compounds and their uses

    US8685980B2

  • Protein tyrosine kinase enzyme inhibitors

    WO2005028443A2

  • Pyrrolopyrimidine compounds as CDK inhibitors

    WO2010020675A1