Real-world evidence of diagnostic testing and treatment patterns among US breast cancer patients

Abstract

Techniques are provided for analyzing gene expression data contained in real-world data and real-world evidence to evaluate biological pathways for identifying molecular subtypes. The systems and methods include determining pathway scores using gene expression data for multiple biological pathways and determining summary scores for the multiple biological pathways. The summary scores may be compared to one or more enrichment scores associated with each predetermined molecular subtype. The molecular subtype is determined based on the comparison. Various heuristics may be applied to filter the pathways prior to summary scoring. Furthermore, the techniques diagnose HER2 status for a patient by identifying discordant HER2 status results between immunohistochemistry (IHC) and fluorescence in situ hybridization (FISH) and diagnosing the HER2 status based on the gene expression data.

Description

[Technical field]

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS This application claims the benefit of priority under 35 U.S.C. §119(e)(1) to, and claims the benefit of the filing date of, U.S. Provisional Patent Application No. 62 / 947,431, filed December 12, 2019, the entire disclosure of which is incorporated herein by reference.

[0002] The present disclosure relates to techniques for the analysis of real-world data and / or real-word evidence, and more specifically, to techniques for the analysis of gene expression data contained in real-world data and real-word evidence to evaluate biological pathways to identify molecular subtypes.

[0003] Technical Field The present disclosure relates to technologies for analyzing real-world data and / or real-word evidence, and more specifically, to technologies for replacing imaging assays using real-world data and real-word evidence RNA sequencing analysis to evaluate biological pathways to identify molecular subtypes. [Background technology]

[0004] The description of the background art provided herein is for the purpose of generally presenting the context of the present disclosure. The work of the presently named inventors to the extent described in this Background section, and aspects of the description that may not otherwise be considered prior art at the time of filing, are not admitted, expressly or impliedly, as prior art to the present disclosure.

[0005] Increasing studies are exploring real-world data (RWD) and subsequent real-world evidence (RWE) to accelerate treatment for cancer patients. RWD relates to patient information acquired during usual care, whereas RWE is clinical evidence derived from RWD. The feasibility of this approach is improving with technological advances and regulatory support to continuously capture and integrate medical data sources. Various studies have demonstrated the ability of RWE to guide clinical development strategies, expand product labeling, and address knowledge gaps by examining clinical aspects not captured in clinical trials. Although limitations of dissemination and room for improvement still exist, RWE has the power to impact patient care.

[0006] A critical step to strengthen the validity of RWE is to demonstrate consistency between population statistics derived from observational RWD and those from controlled experimental data. Despite overwhelming support for the utility of RWE in oncology, technical barriers must be addressed before RWD / RWE can reach its full clinical potential. Incorporating administrative data, ancillary data, and unstructured clinical text from various institutions to generate RWE is a complex task. For example, no standardization exists for abstracting and structuring highly heterogeneous data sources, and many natural language processing algorithms cannot account for these inconsistencies. As a result, clinical endpoints may not be accurately captured, and extraneous variables in raw RWD may introduce confounding biases even when data are appropriately abstracted and prepared for analysis. Similarly, the integration of omics data and RWD requires a controlled approach for large-scale data analysis.

[0007] RWE and integrated omics data have the power to impact patient care. Although various studies show the added value of molecular tumor profiling using RWD for clinically relevant breast cancer insights, further advances in this field require the integration of genetic and clinical data from various institutions along with omics-focused functional and data analytics. One potential avenue to enhance the value of breast cancer RWD is transcriptomics, as RNA-based gene expression analysis has shown prognostic, predictive, and treatment-oriented value beyond DNA sequence insights. Whole-transcriptome RNA sequencing (RNA-seq) can help classify cancer types and breast cancer biomarkers, overcoming the inconclusive pathological assessments, insufficient tissue amounts, and interobserver variability of immunohistochemical or in situ hybridization assays.

[0008] There is a need for techniques to analyze real-world evidence of diagnostic testing and treatment pattern analysis in cancer patients, particularly evidence related to treatment biomarkers, where such analysis is derived from analysis of RNA sequencing data and / or imaging data indicative of pathway data. Summary of the Invention

[0009] This application describes a system and method that addresses the complexities of structuring and analyzing real-world data (RWD). The technology herein demonstrates the feasibility of retrospective RWD analysis and shows that clinical study results can be replicated using longitudinal RWD from a large, representative breast cancer cohort. Applying the technology using clinical information such as patient demographics, clinical characteristics, molecular markers, treatment patterns, and overall survival (OS) outcomes, we show that they can reveal discrepancies in real-world test records, such as HER2 test records. The technology provides integration of RWD and transcriptomic profiling to gain clinically relevant insights through analysis of RWD and molecular data. The technology can enhance the value of real-world evidence (RWE) by tailoring molecular subtypes, revealing pathway-driven insights, and identifying patients who may benefit from RNA sequencing analysis.

[0010] According to an embodiment, a computer-implemented method for determining a molecular subtype of a cancer specimen includes: determining a pathway score for each of a plurality of predefined biological pathways using gene expression data for a plurality of nucleic acids associated with the specimen; creating a summary score for the plurality of biological pathways based on the pathway score for each biological pathway; comparing the summary score to one or more enrichment scores each associated with a predefined molecular subtype; and returning a determined molecular subtype based on the comparison of the summary score to the one or more enrichment scores.

[0011] In some examples, the method includes receiving gene expression data corresponding to a plurality of available biological pathways and applying a pathway heuristic filter to identify a subset of the available biological pathways, the subset being a predetermined biological pathway.

[0012] In some embodiments, the path heuristic filter is a path overlap filter.

[0013] In some embodiments, the method includes filtering the available biological pathways by applying a pathway overlap filter to identify and filter pathways that have 90% or more genes in common with the pathways retained in the subset.

[0014] In some embodiments, the method includes filtering the available biological pathways by applying a pathway overlap filter to identify and filter pathways that have 80% or more genes in common with the pathways retained in the subset.

[0015] In some embodiments, the method includes filtering the available biological pathways by applying a pathway overlap filter to identify and filter pathways that have 50% or more genes in common with the pathways retained in the subset.

[0016] In some embodiments, the pathheuristic filter is a gene expression data filter.

[0017] In some embodiments, the pathheuristic filter is a molecular subtype filter.

[0018] In some embodiments, the pathway score for each biological pathway is a z-score.

[0019] In some examples, the summary score is the average of the z-scores for the biological pathway.

[0020] In some embodiments, the method includes scaling one or more of the z-scores before creating the average of the z-scores.

[0021] In some examples, scaling one or more of the z-scores includes reversing the sign of one or more of the z-scores.

[0022] In some examples, scaling one or more of the z-scores includes reversing the sign of one or more of the z-scores.

[0023] In some embodiments, the methods include reversing the sign of the z-scores to have an average negative z-score in the group of positive gene expression samples and an average positive z-score in the group of negative gene expression samples.

[0024] In some examples, the methods include reversing the sign of z-scores having negative scores below a negative threshold or positive scores above a positive threshold.

[0025] In some examples, the method includes scaling the pathway scores for each of the plurality of biological pathways prior to determining the summary score.

[0026] In some embodiments, the gene expression data is RNA-seq gene expression data.

[0027] In some embodiments, the predetermined biological pathway is one or more of the hallmark pathways.

[0028] In some embodiments, the predetermined biological pathway is one or more of all of the hallmark pathways.

[0029] In some embodiments, the predetermined biological pathway is one or more pathways associated with estrogen signaling.

[0030] In some embodiments, the predetermined biological pathway is one or more of a downstream pathway of human epidermal growth factor receptor 2 (HER2), a downstream pathway of RAS, or a downstream pathway of mTOR.

[0031] In some embodiments, the predetermined biological pathway is one or more immune-related pathways.

[0032] In some embodiments, the predetermined biological pathway is one or more immune-related hallmark pathways.

[0033] In some embodiments, one or more enrichment scores in are determined by UMAP analysis.

[0034] In some embodiments, the determined molecular subtype is an HR+ subtype, an HR+ / HER2+ subtype, an HR- / HER2+ subtype, or a HER2- subtype.

[0035] In some embodiments, the determined molecular subtype is a triple negative subtype.

[0036] In some embodiments, the specimen is from a patient diagnosed with breast cancer.

[0037] In some embodiments, the specimen is a breast cancer specimen.

[0038] According to another embodiment, a system having a memory and a processor, the memory storing instructions that, when executed, cause the processor to perform any of the methods described above.

[0039] According to another embodiment, a system having a memory and a processor, the memory storing instructions that, when executed, cause the processor to: determine a pathway score for each of a plurality of predefined biological pathways using gene expression data for a plurality of nucleic acids associated with a specimen; create a summary score for the plurality of biological pathways based on the pathway score for each biological pathway; compare the summary score to one or more enrichment scores each associated with a predefined molecular subtype; and return a determined molecular subtype based on the comparison of the summary score to the one or more enrichment scores.

[0040] According to another embodiment, there is provided a computer-implemented method for diagnosing HER2 status for a patient, the method including obtaining a human epidermal growth factor receptor 2 (HER2) status for the specimen from analysis of an immunohistochemistry (IHC) image from a first sample from the patient, obtaining a HER2 status for the specimen from analysis of a fluorescent in situ hybridization (FISH) image from a second sample from the patient, identifying a discordant HER2 status result between the HER2 status from the IHC and the HER2 status from the FISH, and in response to identifying the discordant HER2 status, diagnosing the HER2 status based on gene expression data from at least a third sample from the patient.

[0041] In some embodiments, the method further comprises generating a HER2 mismatch status report indicating biological pathways in the gene expression data.

[0042] In some embodiments, the method further includes generating a HER2 mismatch status report that includes a representation of the model of molecular subtype gene expression used to distinguish mismatched HER2.

[0043] In some embodiments, the model of molecular subtype gene expression comprises a linear gene expression model.

[0044] In some embodiments, the model of molecular subtype gene expression comprises a pathway gene expression model.

[0045] In some embodiments, the method further includes generating a HER2 mismatch status report that includes a list of pathways identified by the pathway gene expression model.

[0046] In some embodiments, the model of molecular subtype gene expression comprises a multi-gene linear regression gene expression model.

[0047] In some embodiments, the method further comprises generating a HER2 mismatch status report comprising a list of genes identified by the multigene linear regression gene expression model.

[0048] In some embodiments, the method further comprises generating a HER2 mismatch status report indicating the molecular subtype determined from the gene expression data.

[0049] In some embodiments, the molecular subtype is an HR+ subtype, an HR+ / HER2+ subtype, an HR- / HER2+ subtype, or a HER2- subtype.

[0050] In some embodiments, the molecular subtype is ER+, ER-, PR+, or PR-.

[0051] In some embodiments, the molecular subtype is a triple negative subtype.

[0052] In some examples, the methods further include adjusting a therapeutic treatment protocol based on the pattern in the gene expression data.

[0053] In some examples, diagnosing the HER2 status based on the gene expression data from at least a third sample of the patient includes determining a pathway score for each of a plurality of biological pathways in the gene expression data, creating a summary score for the plurality of biological pathways based on the pathway score for each biological pathway, and comparing the summary score to one or more enrichment scores each associated with a predetermined molecular subtype, and diagnosing the HER2 status based on the gene expression data from at least a third sample of the patient includes determining a molecular subtype of the gene expression data corresponding to the HER2 status based on a comparison of the summary score to the one or more enrichment scores.

[0054] In some embodiments, obtaining the HER2 status from analysis of the immunohistochemistry (IHC) image includes applying the IHC image to a trained IHC classification model trained on histopathological slide image data to classify the HER2 status, and obtaining the HER2 status from analysis of the FISH image includes applying the IHC image to a trained FISH classification model trained on the histopathological slide image data to classify the HER2 status.

[0055] In some embodiments, the trained IHC classification model and the trained FISH classification model are convolutional neural networks.

[0056] In some examples, the method further includes providing the discordant HER2 status result and the diagnosed HER2 status based on at least the gene expression data to a hybrid classification model in response to identifying a discordant HER2 status result between the HER2 status from IHC and the HER2 status from FISH and diagnosing the HER2 status based on at least the gene expression data from a third sample of the patient.

[0057] In some embodiments, the hybrid classification model includes a convolutional neural network.

[0058] In some embodiments, the first sample, the second sample, and the third sample are from a single biopsy block.

[0059] In some embodiments, the first sample, the second sample, and the third sample are different slices from a biopsy block.

[0060] According to another embodiment, a system is provided having a memory and a processor, the memory storing instructions that, when executed, cause the processor to obtain a human epidermal growth factor receptor 2 (HER2) status for the specimen from analysis of an immunohistochemistry (IHC) image from a first sample from the patient, obtain a HER2 status for the specimen from analysis of a fluorescent in situ hybridization (FISH) image from a second sample from the patient, identify a discordant HER2 status result between the HER2 status from the IHC and the HER2 status from the FISH, and in response to identifying the discordant HER2 status, diagnose the HER2 status based on gene expression data from at least a third sample from the patient.

[0061] In some embodiments, the memory stores instructions that, when executed, cause the processor to generate a HER2 mismatch status report indicating biological pathways in the gene expression data.

[0062] In some embodiments, the memory stores instructions that, when executed, cause the processor to generate a HER2 mismatch status report including a representation of the model of molecular subtype gene expression used to identify mismatched HER2.

[0063] According to another embodiment, a system having a memory and a processor, the memory storing instructions that, when executed, cause the processor to perform any of the methods described above. [Brief description of the drawings]

[0064] The drawings described below depict various aspects of the systems and methods disclosed herein. It should be understood that each figure depicts an example of an aspect of the systems and methods. [Figure 1] FIG. 1 is a schematic diagram of an exemplary computer processing system for performing biological pathway analysis on real-world data and / or real-world evidence, according to an embodiment. [Diagram 2] FIG. 2 is a block diagram of an exemplary process for performing pathway analysis and molecular subtype identification from gene expression data, as performed by the processing system of FIG. 1, according to an embodiment. [Figure 3A] FIG. 2 is a block diagram of an exemplary process for performing image assay displacement analysis and molecular subtype identification from RNA-seq data, as performed by the processing system of FIG. 1, according to an embodiment. [Figure 3B] FIG. 3C is a block diagram of an example classification model training process that may be used to train the classification process of the example process of FIG. 3B, according to an embodiment. [Figure 4] 1 is a plot of patients grouped by year of initial diagnosis showing the distribution of patients by year of initial diagnosis across the clinical abstraction cohort, according to an embodiment. [Figure 5A] 5A and 5B illustrate plots of breast cancer molecular biomarkers and subtypes in a clinical abstraction cohort according to an embodiment. Figure 5A is a plot of the number of patients with positive, negative, or equivocal IHC or FISH test results for ER, PR, HR, and HER2 status at initial diagnosis. [Figure 5B] 5A-5C illustrate plots of breast cancer molecular biomarkers and subtypes in a clinical abstraction cohort, according to an embodiment. FIG. 5B is a plot of the distribution of breast cancer molecular subtypes as determined by abstracted ER, PR, and HER2 test results at the time of initial diagnosis. [Figure 5C] 5A-5C are plots of breast cancer molecular biomarkers and subtypes in a clinical abstraction cohort, according to an embodiment. FIG. 5C is a plot of the distribution of ER and PR status combinations across the entire cohort. [Figure 5D] 5A-5D are plots of breast cancer molecular biomarkers and subtypes in a clinical abstraction cohort, according to an embodiment. FIG. 5D is a plot of the number of patients with high, intermediate, low, indeterminate, or equivocal Ki67 IHC test results or physician notes indicating status at initial diagnosis, separated by molecular subtype. [Figure 6A]6A-6B are plots showing anti-HER2 treatment by HER2 status in a clinical abstraction cohort, according to an embodiment. Anti-HER2 treatment initiation patterns among HER2+ (FIG. 6A) and HER2- (FIG. 6B) patients who received anti-HER2 treatment at some point in clinical care. M, month; Y, year. [Figure 6B] 6A-6B are plots showing anti-HER2 treatment by HER2 status in a clinical abstraction cohort, according to an embodiment. Anti-HER2 treatment initiation patterns among HER2+ (FIG. 6A) and HER2- (FIG. 6B) patients who received anti-HER2 treatment at some point in clinical care. M, month; Y, year. [Figure 7] 7A-7F are plots of overall survival from date of primary diagnosis in clinical abstraction cohorts, according to an embodiment; 10-year survival rates for stage I-IV patients stratified by stage in FIG. 7A and subtype in FIG. 7B; 5-year survival rates stratified by HER2 status for all patients in FIG. 7C and stage IV patients in FIG. 7D, ER status for all patients in FIG. 7E and stage IV patients in FIG. 7F, and TNBC status for all patients in FIG. 7G and stage IV patients in FIG. 7H. [Figure 8A] FIG. 8 illustrates the molecular characteristics of a molecularly sequenced cohort, according to an embodiment. FIG. 8A is a plot of the distribution of patients with mutations in the most frequently reported genes across the cohort. The number of patients with mutations in each gene is indicated above the bar. [Figure 8B] FIG. 8B illustrates the molecular characteristics of a molecularly sequenced cohort, according to an embodiment. FIG. 8B is a plot of the number of variants classified as alterations, amplifications, or deletions within each of the most frequently reported genes in the cohort. [Figure 8C] FIGURE 8C illustrates the molecular characteristics of a molecularly sequenced cohort, according to an embodiment. FIGURE 8C is a plot of the distribution of patients with pathogenic germline alterations in NCCN-designated familial high-risk genes. [Figure 8D]

[0023] Figure 8D illustrates the molecular characteristics of a molecularly sequenced cohort, according to an embodiment. Figure 8D is a plot of tumor mutation burden (TMB) across the cohort. [Figure 9A]9A and 9B are plots related to RNA-based receptor status prediction analysis of molecularly sequenced cohorts, according to an embodiment. FIG. 9A is a plot of UMAP transcriptome clustering of 19,147 genes in the cohort, color-coded by molecular subtype. Circles correspond to samples with available IHC or FISH test results for all proteins, and X symbols correspond to patients with predicted status for at least one protein. [Figure 9B] FIG. 9B is a plot related to RNA-based receptor status prediction analysis of a molecular sequenced cohort, according to an embodiment. FIG. 9B plots the relationship between ER, PR, and HER2 receptor status and log10 transformed and normalized gene expression of ESR1, PGR, and ERBB2. The left panel represents samples with available receptor status from abstracted test results, and the right panel represents transcriptome-based receptor status prediction. HER2 predictions for samples reported as unclear are plotted as white dots. [Figure 10A] 10A-10C are plots relating to single-gene logistic model performance for ER, PR, and HER2 status prediction in molecularly sequenced cohorts, according to an embodiment. [Figure 10B] Figure 10B is a plot related to the performance of single gene logistic models for ER, PR and HER2 status prediction in molecular sequenced cohorts according to an embodiment. Figure 10B illustrates the precision and recall of transcriptome-based receptor status prediction evaluated on a test set consisting of cohort RNA sequenced samples with abstracted receptor status results in the database. [Figure 10C] Figure 10C illustrates a confusion matrix showing transcriptome-based ER, PR, and HER2 status prediction performance. [Figure 11A]Figure 11 illustrates breast cancer pathway analysis from RNA-seq data of molecular sequenced cohorts according to MSigDB and Hallmark pathways by example. Figure 11A illustrates the Pearson correlation between ERBB2 expression and enrichment score (GSVA) ​​for each HER2-related pathway in MSigDB among cohorts. Figure 11B illustrates the correlation between ESR1 expression and enrichment score for each ER-related pathway in MSigDB among cohorts. [Figure 11B] FIG. 11B illustrates breast cancer pathway analysis from RNA-seq data of molecularly sequenced cohorts according to MSigDB and hallmark pathways, by way of example. FIG. 11B illustrates the correlation between ESR1 expression and enrichment scores for each ER-related pathway in MSigDB among cohorts. [Figure 12A] 12A and 12B are plots related to RNA-seq breast cancer pathway analysis of molecularly sequenced cohorts, according to an embodiment. HER2 (FIG. 12A) and ER (FIG. 12B) pathway metascores for patients with abstracted HER2 IHC or FISH test results are shown. FIG. 12C is a UMAP of the 50 hallmark enrichment scores. Patients with molecular subtypes based on at least one abstracted receptor status are indicated by circles, and patients with molecular subtypes determined solely from RNA-predicted status are indicated by X symbols. Distributions of enrichment Z-scores for HR- / HER2+ (FIG. 12D) and TNBC (FIG. 12E) related pathways are shown. [Figure 12B] 12A and 12B are plots related to RNA-seq breast cancer pathway analysis of molecularly sequenced cohorts, according to an embodiment. HER2 (FIG. 12A) and ER (FIG. 12B) pathway metascores for patients with abstracted HER2 IHC or FISH test results are shown. FIG. 12C is a UMAP of the 50 hallmark enrichment scores. Patients with molecular subtypes based on at least one abstracted receptor status are indicated by circles, and patients with molecular subtypes determined solely from RNA-predicted status are indicated by X symbols. Distributions of enrichment Z-scores for HR- / HER2+ (FIG. 12D) and TNBC (FIG. 12E) related pathways are shown. [Figure 12C]12A and 12B are plots related to RNA-seq breast cancer pathway analysis of molecularly sequenced cohorts, according to an embodiment. HER2 (FIG. 12A) and ER (FIG. 12B) pathway metascores for patients with abstracted HER2 IHC or FISH test results are shown. FIG. 12C is a UMAP of the 50 hallmark enrichment scores. Patients with molecular subtypes based on at least one abstracted receptor status are indicated by circles, and patients with molecular subtypes determined solely from RNA-predicted status are indicated by X symbols. Distributions of enrichment Z-scores for HR- / HER2+ (FIG. 12D) and TNBC (FIG. 12E) related pathways are shown. [Figure 12D] 12A and 12B are plots related to RNA-seq breast cancer pathway analysis of molecularly sequenced cohorts, according to an embodiment. HER2 (FIG. 12A) and ER (FIG. 12B) pathway metascores for patients with abstracted HER2 IHC or FISH test results are shown. FIG. 12C is a UMAP of the 50 hallmark enrichment scores. Patients with molecular subtypes based on at least one abstracted receptor status are indicated by circles, and patients with molecular subtypes determined solely from RNA-predicted status are indicated by X symbols. Distributions of enrichment Z-scores for HR- / HER2+ (FIG. 12D) and TNBC (FIG. 12E) related pathways are shown. [Figure 12E] 12A and 12B are plots related to RNA-seq breast cancer pathway analysis of molecularly sequenced cohorts, according to an embodiment. HER2 (FIG. 12A) and ER (FIG. 12B) pathway metascores for patients with abstracted HER2 IHC or FISH test results are shown. FIG. 12C is a UMAP of the 50 hallmark enrichment scores. Patients with molecular subtypes based on at least one abstracted receptor status are indicated by circles, and patients with molecular subtypes determined solely from RNA-predicted status are indicated by X symbols. Distributions of enrichment Z-scores for HR- / HER2+ (FIG. 12D) and TNBC (FIG. 12E) related pathways are shown. [Figure 13]Illustrates plots of the distribution of enrichment z-scores for each HER2-related pathway in MSigDB among patients in the molecular sequenced cohort. Shown are abstracted or predicted patients with negative (blue), equivocal (orange), or positive (green) HER2 test results. P-values ​​listed for each pathway represent the results of a Kruskal-Wallis test for the difference between enrichment scores from HER2-, HER2-equivocal, and HER2+ patients. [Figure 14] Figure 1 illustrates a plot of the distribution of enrichment z-scores for each ER-related pathway in MSigDB among patients in the molecular sequenced cohort. Abstracted or predicted patients with negative (blue) or positive (green) ER test results are displayed. P-values ​​listed for each pathway represent the results of a Wilcoxon rank sum test for the difference between enrichment z-scores from ER+ and ER- patients. [Figure 15A] FIG. 13 is a plot illustrating that for each Hallmark pathway, the significance of differential enrichment between molecular subtypes was determined by Kruskal-Wallis test of enrichment scores. The vertical line indicates P=0.001, and values ​​to the right of the line were considered significant. [Figure 15B] Illustrated are the distributions of z-scores among HR+ / HER2- (blue), HR+ / HER2+ (green), HR- / HER2+ (orange), and TNBC (gray) patients for the two electrogen response hallmark pathways with the most significant differential enrichment among molecular subtypes. [Figure 16] FIG. 2 is a block diagram of an exemplary process for performing hybrid pathway analysis and molecular subtype identification from gene expression and image data, as performed by the processing system of FIG. 1, according to an embodiment. [Figure 17] 1 illustrates an exemplary computing device for implementing the system of FIG. 1 and the processes of FIGS. 2, 3A, 3B, and 16, according to an embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0065] The expanding utility of RWE is evident from the growing number of related studies and regulatory considerations. However, compared to randomized controlled trials, RWD analysis is complicated by the lack of standardization between records, and the introduction of external factors such as natural language processing errors and uncontrolled confounding variables. In various embodiments, the present application addresses these concerns by providing systems and methods to 1) increase the statistical power of analysis with relatively large cohort sizes, 2) incorporate various data sources beyond electronic health records to benefit downstream analysis, and 3) demonstrate consistency between the characteristics of real-world cohorts and the results of previous clinical studies.

[0066] Using only a portion of breast cancer patient records from an extensive clinicogenomics database, we were able to perform a retrospective analysis using the techniques provided herein that provide further evidence of the feasibility and value of generating clinically relevant RWE. We show that longitudinal RWD can capture important information about patient history, treatment course, and outcomes. RWD analysis, using the techniques described in the examples herein, replicated previously published clinical results and generated valid RWE that was generally consistent with established databases, demonstrating feasibility. While the majority of cohort characteristics were consistent with previous clinical studies, our analysis also highlighted the complexity of breast cancer RWD. For example, the proportion of premenopausal and postmenopausal patients was similar to previous clinical trial data, but menopausal status was confidently abstracted in only approximately 51% of the cohort. Upon further review, many RWD breast cancer studies applied simplified definitions of menopause, such as age restrictions, reported missing status in electronic records, or did not include menopausal status at all. Advantageously, in various embodiments, the present technology enables the integration of simplification rules for abstraction that fill such gaps in RWD, such as in the definition of real-world progression-free survival, but may also affect the validity of the conclusions.

[0067] In various embodiments, rules were established and applied to perform relevant analyses and derive statistics from sample cohorts to enhance the validity of the RWE presented herein. In various embodiments, the techniques herein include applying biological pathway analysis techniques that facilitate the definition of molecular subtypes from multiple abstracted test results. For example, applying the techniques herein to HER2 test result analysis confirmed existing inconsistencies in the interpretation of standard tests, a problem highlighted by recent American Society of Clinical Oncology (ASCO) guidelines, previous clinical studies, and meta-analyses. Specifically, findings of intra-test discrepancies in IHC exemplify the subjectivity of IHC testing, prompting improvements in standard tests and biomarker discovery.

[0068] After observing discrepancies in abstracted HER2 test results, another cohort containing complete biopsy data was selected to test the effectiveness of the whole transcriptome model in predicting molecular subtypes. By combining clinical and molecular data with the approach herein, we show that transcriptome profiling can complement RWD and reveal fundamental biological differences between patients. RNA sequencing can complement the interpretation of standard tests by providing clinically relevant insights when biopsy test data is inconclusive, as exemplified by the resolution of molecular subtypes in patients with ambiguous status. In various embodiments, ambiguous status can be avoided and appropriate molecular subtypes can be identified to inform better treatment decisions, for example, better breast cancer treatment decisions.

[0069] Moreover, applying the present technology in various embodiments, as further discussed in the Examples below, we have uncovered potential pathway-related therapeutic targets, such as oncogenic signaling through the mTOR pathway, for subtypes such as TNBC, where previous signaling pathway investigations have limited available drug therapies. Furthermore, the present pathway analysis can also elucidate treatment-related tumor characteristics not captured by standard diagnostic and prognostic tests, such as additional biomarkers or amplifications that can be targeted in HER2+ breast cancer patients. Expression-based immune signatures can also predict response to neoadjuvant treatment with several experimental agents / combinations added to standard chemotherapy, including the addition of pembrolizumab in early-stage TNBC. Biomarker selection for immune therapy in early-stage TNBC becomes essential for therapeutic strategies, given its substantial toxicity.

[0070] Thus, in various embodiments, the present technology provides a data pipeline that integrates longitudinal RWD and comprehensive molecular sequencing data into a structured clinical genetic database that can generate real-time actionable clinical evidence. Although RWD is inherently complex, cancer cohort selection and data insights are feasible using structured data sources and rigorously defined analytical criteria. As we have discovered, integrating RNA-seq data with RWD can improve clinically actionable evidence related to clinical markers, potential therapeutic targets, and optimal treatment selection in breast cancer.

[0071] FIG. 1 illustrates a system 100 for performing biological pathway analysis and molecular subtyping from analysis of (i) image data, such as fluorescent in situ hybridization (FISH), hematoxylin & eosin staining (H&E), or immunohistochemistry (IHC) image data, and image assay results obtained from such images, (ii) gene expression data, such as RNA-seq data, or (iii) a combination thereof. In the illustrated embodiment, the system 100 includes a pathway analysis computing device 102 communicatively coupled to a plurality of datasets and data sources via a communications network 104. For example, the computing device 102 may be configured to receive gene expression data from a number of different sources via the network 104. The computing device 102 may be coupled, for example, to a network-accessible RNA-seq database (or dataset) 106, where such gene expression data may or may not be pathway-marked. The computing device 102 may be coupled to a network-accessible RNA-seq database 108 that includes pathways that have already been marked. In some examples, either of these databases 106 and 108 may contain gene expression data sets that have been pre-normalized or formatted according to a predetermined normalization protocol. In some examples, either of these databases 106 and 108 is not normalized prior to being received at the computing device 102. As used herein, the term biological pathway is also referred to simply as a "pathway" and includes genetic pathways, cellular pathways, signaling pathways, and metabolic pathways.

[0072] The computing device 102 may be communicatively coupled to other sources of gene expression data, such as an RWD / RWE server 110 that makes available gene expression data stored in the form of RNA sequencing datasets, including gene expression data and other data collected in clinical settings such as may be included in research institution computing systems, lab computing systems, hospital computing systems, physician group computing systems, electronic health records, patient-generated health data mobile devices, and wearables and mobile apps, observational study systems, patient / condition registries, etc.

[0073] The computing device 102 may be communicatively coupled to a histopathological image data repository 112. Any number of histopathological image data sources may be accessible using the computing device 102. The histopathological images may be images captured by any suitable optical histopathological slide scanner, including any dedicated digital medical image scanner, for example, 10x and / or 40x resolution magnification scanners. In yet other examples, the images may be received from genomic sequencing systems, for example, TCGA and NCI Genomic Data Commons, or other sources of gene expression data. Additionally, the histopathological image data repository 112 may be from an organoid modeling lab. These image sources may communicate image data, image assay result data, genomic data, patient data, treatment data, historical data, or other RWD / RWE data according to the techniques and processes described herein. Each of the image sources may represent multiple image sources. Additionally, each of these image sources may be considered a different data source, which may generate and provide different image data than other providers, hospitals, etc. Imaging data between different sources can potentially differ in one or more respects, resulting in biases inherent to the various data sources, such as different dyes, fixation, embedding, and staining protocols for biological samples, as well as different pathology imaging equipment and settings.

[0074] Although server 110 is specifically labeled as an RWD and / or RWE server, it will be understood that any of data sources 106, 108, 110, and 112 may be considered to include RWD and / or RWE datasets.

[0075] The functionality of the computing device 102 may be implemented across distributed computing devices connected together via communication links. The functionality of the system 100 may be distributed across any number of devices, including the illustrated portable personal computers, smartphones, electronic documents, tablets, and desktop personal computing devices. In some embodiments, some or all of the functionality of the computing device 102 may be executed on a network-accessible server 114 coupled to a network 104. The network 104 may be a public network, such as the Internet, a private network, such as a research or corporate network, or any combination thereof. The network may include a local area network (LAN), a wide area network (WAN), a cellular, satellite, or other network infrastructure, whether wireless or wired. The network may utilize communication protocols, including packet-based and / or datagram-based protocols, such as the Internet Protocol (IP), Transmission Control Protocol (TCP), User Datagram Protocol (UDP), or other types of protocols. Additionally, the network may include multiple devices that facilitate network communication and / or form the hardware infrastructure of the network, such as switches, routers, gateways, access points (such as wireless access points as illustrated), firewalls, base stations, repeaters, backbone devices, etc.

[0076] In the illustrated example, the computing device 102 is configured with a pipeline configuration illustrating three different pathway analysis pipelines: An initial pre-processing 120 is provided for pre-processing of gene expression data as described in various embodiments herein.

[0077] The first pipeline 130 includes an image assay permutation analysis module 132 that receives histopathological image data (e.g., including image assay result data) from the pre-processing layer 120. The image-based module 132 includes an image processing stage 134 and a trained molecular subtype model 136, which is a model trained to classify the presence of one or more molecular subtypes from RWD / RWE determined based on the analysis of the image assay results, i.e., based on histopathological image data. The subtype discrepancy resolver 138 receives the molecular subtype data determined from different images and determines whether there is a match or a discrepancy between them. For example, as discussed further below, the trained molecular subtype model 136 can be trained to identify proteins associated with breast cancer in both IHC image data and FISH image data. However, as discussed further below, the molecular subtype determination of IHC image data often does not match the molecular subtype determination of FISH image data. The subtype mismatch resolver can compare the resulting molecular subtypes to determine when there is a mismatch, e.g., when the data is ambiguous, and determine whether the mismatch needs to be resolved using another pipeline process, such as pipeline 140 or pipeline 150.

[0078] In contrast to the image assay permutation determination of pipeline 130, a second pipeline 140 performs pathway analysis on gene expression data only. In the illustrated example, pipeline 140 includes an RNA-seq pathway analysis module 142 having an RNA normalization stage 143, a pathway generator stage 144, a pathway heuristic-based filter stage 146, a pathway scorer stage 148, and a pathway aggregation evaluator 149.

[0079] The third pipeline 150 forms a hybrid process configured to perform pathway analysis on a combination of histopathological image data (including image assay result data) and gene expression data. In the illustrated embodiment, the third pipeline 150 includes a hybrid pathway analysis module 152 that includes an image processing stage 154, RNA-seq based pathway aggregation data 156 obtained, for example, from module 142, a trained molecular subtype neural network 158, and a pathway evaluator stage 159.

[0080] In the illustrated embodiment, each of the pipelines 130, 140, and 150 coupled to transmit their respective pathway analysis data to the molecular subtype processing layer 160 is further configured as a report generator layer. The molecular subtype processing layer 160 includes a molecular subtype identifier 162 that compares pathway summary scores received from one or more of the pipelines 130, 140, and 150 with enrichment scores of candidate groups of molecular subtypes stored in or accessible to the layer 160. From this comparison, the identifier 162 determines a predicted molecular subtype and provides the subtype to a report template generator 164. In some embodiments, the identifier 162 may access available treatment / therapy options 166 corresponding to the determined molecular subtype and provide additional information thereof to the report generator 164. A user preferences module 168 is further provided for storing user-settable preferences, e.g., preferences entered into the computing device 102 via the display and graphical user interface. Examples of user preferences include report templates, rankings of available treatments, user-predetermined molecular subtype identifier rules, and the like.

[0081] 2 illustrates an exemplary process 200 of pathway analysis of RWD data, particularly gene expression data, that may be performed by computing device 102, and in particular by RNA-seq-based pathway analysis module 142. At block 202, RWD or RWE data in the form of gene expression data, e.g., RNA-seq data, is received into the pathway analysis system. In the example of FIG. 1, such data may be received, for example, from RNA-seq datasets 106 and / or 108, or server 110.

[0082] In various embodiments, block 204 performs a normalization process on the provided and received RNA-seq data. Thus, as illustrated in FIG. 1, pipeline 140 may include an RNA-seq data normalizer, such as that disclosed in U.S. Patent Application No. 16 / 581,706, entitled "Methods of Normalizing and Correcting RNA Expression Data," filed September 24, 2019, which is incorporated by reference in its entirety for all purposes. In some examples, block 204 performs a transcriptomics data deconvolution process, such as that disclosed in U.S. Provisional Patent Application No. 62 / 786,756, entitled "Transcriptome Deconvolution of Metastatic Tissue Samples," filed December 31, 2018, and U.S. Provisional Patent Application No. 62 / 924,054, entitled "Calculating Cell-type RNA Profiles for Diagnosis and Treatment," filed October 21, 2019, which are incorporated by reference in their entireties for all purposes. The deconvoluted information can then be passed to other aspects of the platform, such as variant calling, RNA expression calling, or insight engines.

[0083] In various embodiments, block 204 may normalize the newly obtained gene expression (e.g., RNA-seq) dataset to eliminate biases caused by, for example, gene content (GC), gene length, and sequencing depth. The conversion factor may be generated by comparing the obtained gene expression dataset to a standard gene expression dataset using a statistical mapping model. Examples of statistical mapping models include, but are not limited to, standard linear models, generalized linear models (e.g., using a gamma distribution for count data), or non-parametric techniques such as data transformation to ranks.

[0084] For example, the gene expression dataset from block 202 may contain RNA sequence data. A gene information table containing information such as gene name, start and end points (to determine gene length), gene content ("GC"), etc. may be accessed by block 202, and the resulting information may be used to determine sample regions for analyzing the gene expression dataset. From there, block 204 may perform additional normalization. For example, normalization of GC content may be performed using a quantile normalization process such as a first full quantile normalization process, e.g., the R packages EDASeq and DESeq normalization processes (available from Bioconductor, Roswell Park Comprehensive Cancer Center, Buffalo, NY, https: / / bioconductor.org / packages / release / bioc / html / DESeq.html). The GC content of the sampled data may then be normalized to the gene expression dataset. A second full quantile normalization may then be performed to the gene length of the sample data. To correct for sequencing depth, a third normalization process can be used that can correct for global differences in sequencing depth between samples without being overly influenced by outlier gene expression values ​​within any given sample. For example, a global reference can be determined by calculating the geometric mean of expression for each gene across all samples. A size factor can be used to adjust the samples to match the global reference. The expression values ​​of the samples can be compared to the global reference geometric mean to create a set of expression ratios (i.e., ratios of sample expression to global reference expression) for each gene. A size factor is determined as the median of these calculated ratios. The samples are then adjusted to match the global reference with a single size factor correction, for example, by dividing the gene expression value for each gene by the size factor of the sample. The GC-wide normalized, gene length normalized, and sequencing depth corrected RNA-seq data can be stored as normalized RNA-seq data.A correction process can then be performed on the normalized RNA-seq data by sampling the RNA-seq data multiple times and performing statistical mapping or applying a statistical transformation model, such as a linear transformation model, to each gene. Corresponding intercept and beta values ​​can be determined from the linear transformation model and used as correction factors for the RNA-seq data.

[0085] In some examples, to incorporate multiple datasets, block 204 performs a gene expression batch normalization process that adjusts for known biases in the datasets, including but not limited to GC content, gene length, and sequencing depth.

[0086] In various embodiments, process 204 may receive the normalized gene expression data and additionally perform a deconvolution process to rectify the data using a clustering process to optimize the number of clusters, K, such that one or more gene expression clusters associated with one or more cell types of interest are detected. Subsequent analysis of the gene expression clusters may determine cancer-specific cluster types within such data.

[0087] The deconvoluted gene expression data can be used in downstream gene expression data analysis to obtain more accurate results than analyzing mixed sample gene expression data. For example, analysis of gene expression data of a mixed sample may return results that reflect the background tissues of the mixed sample, not the cancer tissues. Such deconvolution can be beneficial for downstream gene expression data analysis, including pathway analysis to determine which genes are over- or under-expressed along different pathways, to determine consensus molecular subtypes, predict the type of cancer present in the sample (especially for tumors of unknown origin), detect infiltrating lymphocytes, determine which cellular activity pathways are dysregulated, discover biomarkers, match therapeutics or clinical trials based on the results of these downstream analyses, and design clinical trials or organoid experiments based on the results of these downstream analyses.

[0088] Additional normalization processes at block 204 include normalizing the RNA-seq data to account for technology differences between datasets, such as for the assay used (polyA / exome-capture), for the probe set used for capture, for the sequencer used, and / or for the flow cell. Further examples and descriptions of normalization processes that may be performed by block 204 include those described in U.S. Patent Application No. 17 / 112,877, entitled "Systems And Methods For Automating RNA Expression Calls In A Cancer Prediction Pipeline," filed December 4, 2020, and U.S. Patent Application No. 2019 / 0033934, entitled "Methods Of Normalizing And Correcting RNA Expression Data," both of which are incorporated herein by reference in their entireties for all purposes.

[0089] One or more of these processes in block 204 may be performed in the RNA normalizer 143 or preprocessing layer 120 of the pipeline 140.

[0090] In some implementations, the received gene expression data includes pathway data, such as that contained in an RNA-seq dataset 108. In some implementations, the gene expression data does not contain pathway data, such as the RNA-seq dataset 106. Thus, block 206 is provided to identify pathway data in the received normalized gene expression data. The number of identified pathways may be a small number, 1-10, 10-100, 100-1000, 1000-10000, or more than 100,000.

[0091] If no pathway data is provided, then block 206 may apply one or more pathway models to the gene expression data to generate pathways. In some examples, block 206 may identify the presence of existing pathway data, determine whether the existing pathway data meets predefined criteria, and if the data does not, then apply a new pre-approved pathway model to even reconstruct the pathway data from the dataset. Block 206, which may be performed by pathway generator 146 in pipeline 140, may apply an appropriate pathway model, such as a hallmark pathway model, an immune-related hallmark pathway model, a related estrogen signaling model, a downstream HER2 model, a downstream RAS model, downstream of the mTOR model, and / or an immune-related pathway model. Block 206 may be configured to apply a model selected by a user, for example, stored in user preferences module 168. In various embodiments, block 206 is configured to generate a sufficiently large number of diverse pathways to imply aggregation in the following process.

[0092] Pathways may be determined (or previously determined) using gene set enrichment analysis GSEA, single sample gene set enrichment analysis (ssGSEA), or another analysis. Pathways may differ in genes, gene rankings, gene expression, and nodes depending on pathway activity. Pathways may correspond to different cancers or different cancer subtypes. In various embodiments, pathways may correspond to signaling pathway activity (MAPK, RAS, NOTCH), immune function (inflammation, cytokine signaling), tumor microenvironment (hypoxia, angiogenesis), overexpression of genes (leveraging their downstream effects on the expression of other genes), and the like.

[0093] With the pathways identified from process 204, process block 208 optionally performs heuristic-based filtering on the pathways to generate a reduced pathway set for scoring and aggregate analysis. In some embodiments, process 208 compares the pathways to determine whether any of the pathways overlap (e.g., contain genes in common) by a threshold amount or percentage. If two or more pathways overlap by a sufficient amount, process 208 determines which pathways should remain in the reduced pathway set and then eliminates the other pathways. For example, the process may retain the pathways with the greatest pathway length or the pathways with the greatest amount of biological expression (gene expression, protein expression, cell expression). In some examples, process 208 may filter the pathways to minimize overlap in gene sets or to ensure that there is no overlap. Process 208 may filter the pathways for analysis based on other heuristic processes, such as identifying only upregulated pathways, i.e., pathways that exhibit biological expression, or identifying only downregulated pathways, i.e., pathways with suppressed biological expression. Other examples of heuristic processes include filtering pathways based on the amount of upregulation or the amount of downregulation. For example, pathways with an amount of downregulation above a threshold may be removed. Other examples of heuristic processes include filtering pathways based on pre-assigned molecular subtypes, such as filtering to include pathways previously determined to be associated with HER2+ or HER2-. Other examples of heuristic processes include filtering pathways that include upregulation and downregulation in the same pathway. In some examples, filtering may be based on requiring pathways with scores that are significantly different (e.g., p<.05 by Wilcoxon) between known positives (e.g., by HER2 IHC / FISH) or negatives. In some examples, pathways may be required to have an average fold difference (e.g., 1.5) between known positives and negatives. In the example of FIG. 1, process 208 may be performed by pathway heuristic filter stage 148.

[0094] In the illustrated example of FIG. 2, once a set of pathways for analysis has been determined, process 210 performs scoring of each pathway according to a scoring rule. In some examples, a pathway score is used that ranks the genes in the pathway. In some examples, scoring is performed by determining a z-score for each pathway. The z-score is the difference between the average of the gene expression values ​​that form the known pathway (reference gene expression) and the average of the gene expression values ​​in the RNA-seq dataset, for example after normalization. In various embodiments, the reference gene expression for a pathway is represented by one or more enrichment scores of the pathway, which indicate upregulation or downregulation of gene expression of the pathway. In this way, in various embodiments, the enrichment scores are associated with a given molecular subtype. The reference pathways may be obtained from a known database of pathways and their gene sets. Generally speaking, the z-score for a pathway is an example of a scoring metric that can take into account the directional effect of one or more molecules on a process and the direction of change of those molecules in the dataset (e.g., upregulation or downregulation of those molecules). The z-score metric can predict activation or inhibition of regulators based on their relationship to the dataset genes and the direction of change of the dataset genes. For example, a positive z-score indicates that the genes in that pathway are upregulated relative to the average, and a negative z-score indicates that the genes in that pathway are downregulated relative to the average. By way of example only, a positive z-score pathway may, in some examples, indicate a pathway associated with gene expression indicative of a HER2+ molecular subtype, and a negative z-score may indicate a pathway associated with gene expression inhibition indicative of a HER2- molecular subtype. Although various examples are described with reference to z-scores, process 210 may apply any number of pathway scoring rules that exhibit similar properties to z-scores, such as, for example, quantile normalization or machine learning algorithms (MLA).

[0095] In various embodiments, process 210 may further apply scaling to enable summary aggregation of pathway scores without canceling positive and negative scores across average gene expression levels. The scaling applied by process 210 can be pathway-level scaling or gene-expression-level scoring. For pathway-level scaling, for example, process 210 can identify all pathways having a negative z-score (negative control) in the majority of patients / specimens of training data with a HER2 status, or the average negative z-score in negative subjects, and can invert the sign of the pathway z-score. For each pathway, there can be a range of z-scores that are negative and a range that are positive for a series of patients. In some examples, scaling can include inverting z-scores having an average negative z-score in positive samples and inverting z-scores having an average positive z-score in negative samples. In some examples, process 210 can identify those pathways having negative scores below a threshold and either invert the sign only for those pathways or change the score values of those pathways to the threshold. For example, process 210 can be configured to assign a z-score greater than zero, 0, or a z-score within a range such as -2 < z-score < +2. In another example, the signs of all pathway names including HER2 and dn (for downregulation) are inverted. In other examples, scaling can be performed to reduce the amount of skew in the distribution of pathway scores. In gene-expression-level scaling, variable gene-expression levels can be scaled based on reference gene-expression levels for the training data. Such scaling can be based on z-scores, quantile normalization, or other scoring modalities.

[0096] In various embodiments, process 210 may further perform an aggregation of the pathway scores to determine an aggregated score across the reduced set of pathways. For example, process 210 may determine an average of the z-scores across all pathways after a scaling process. In some examples, aggregation is performed on the z-scores without performing a scaling process. Features of process 210 may be performed by pathway scorer 148 and aggregate evaluator 149 of module 142 of FIG. 1. The aggregate score is an example of a summary score of the analysis pathways established for the subject.

[0097] At block 212, the summary score from process 210 is compared to the enrichment score of a given molecular subtype to determine if a match exists, which may be determined using matching rules. For example, the aggregate score from process 210 may be compared to an enrichment score in the form of a given HER2+ or HER2 pathway gene expression score. If the aggregate score is statistically sufficiently associated with the enrichment score, process 212 identifies a corresponding candidate molecular subtype at block 214, and a report including the identified molecular subtypes is generated for display to a user, for storage by a computing device, and / or for remote communication to a user via a communication network. In some implementations, process 212 may identify multiple candidates for a molecular subtype, and process 214 may generate a report including each of them, and in some examples, the molecular subtypes are ranked according to a similarity score. In various embodiments, the report may include a report element indicating one or more treatments that are likely to work for a given molecular subtype based on the stored association between the two. In various embodiments, the report may be ranked by the best match for the determined molecular subtype.

[0098] Optionally, in some examples, the molecular subtype is provided to block 216, which determines a recommended set of treatments / therapies from a larger population of available treatments / therapies corresponding to the replacement molecular subtype. For example, available treatments generally associated with molecular subtypes may be stored in a ranked manner on computing device 102. Thus, block 214 may generate a report that includes a report element indicating one or more treatments / therapies that are likely to work for a given molecular subtype based on a stored association between the two, or a known association (e.g., published research studies and / or clinical trials, or NCCN or FDA guidelines). In various embodiments, the report may be ranked by best match for the determined molecular subtype.

[0099] In contrast to pipeline 140, pipeline 130 of FIG. 1 is an image displacement analysis pipeline. An exemplary implementation of the pipeline is shown in FIG. 3 and process 300. At block 302, process 300 receives image assay result data, e.g., IHC image results and / or FISH image results with molecular subtype conclusions. For example, these image assay results may result from pathologist-labeled IHC stained images and FISH stained images labeled with HER2 status, HER2+ or HER2-. In some examples, as discussed herein, these image assay results may result from a trained image-based molecular subtype classification model, such as model 136. At block 304, the image assay result data is compared and examined for discrepancies. For example, in block 304, a comparison of the molecular subtypes resulting from the different imaging assay results is made to determine whether there is a discrepancy between the results, including a discrepancy between molecular subtypes, for example, one image set (e.g., IHC) results in the identification of the presence of HER2+ while another image set (e.g., FISH) results in a negative presence of HER2+ or an equivocal determination of HER2.

[0100] In response to determining a discordance between the resulting molecular subtypes at block 306, gene expression data (e.g., RNA-seq data) corresponding to the specimen associated with the image assay results is analyzed to determine the molecular subtype. In various embodiments, for example, the associated RNA-seq data is analyzed by block 306 configured with a linear regression model trained on gene expression data training data, such as RNA expression training data. For example, the linear regression model may be trained to look at HER2+ and HER2- outcomes with the gene expression data. Block 306 may analyze the RNA-seq data with a logistic regression model of HER2 gene expression levels. For other molecular subtypes of the IHC target, the logistic regression model will be the gene expression levels of that molecular subtype or the expression levels of genes associated with the IHC target.

[0101] In various other examples, block 306 is configured to use pathway analysis, e.g., implementing pipeline 140, to examine the RNA-seq data, determine which pathways best match the data using a trained model, and identify a molecular subtype based on the pathway analysis. For example, in response to identifying a discrepancy in molecular subtype classification between imaging assay results, block 306 may provide the molecular subtype data and patient data to a gene expression-based pipeline for analysis, such as using process 200 of FIG. 2.

[0102] These various processes of blocks 306 and 308 may be performed by the subtype mismatch resolver 138 according to the pipeline 140 and other features of such processes.

[0103] In still other various embodiments, block 306 implements a multi-gene lasso model that examines the expression levels of multiple genes in the RNA-seq data, and uses the trained classification model to apply a lasso linear regression process to examine whether the RNA-seq data is closer to one gene expression group (e.g., gene expression levels associated with one molecular subtype) compared to another gene expression group (e.g., gene expression levels associated with another molecular subtype). From the comparison, a determination of the molecular subtype, e.g., HER2+ or HER2-, is made. For example, process 306 may also compare gene expression data with genes in predictive models for different molecular subtypes, estrogen receptor (ER) for ER+ breast cancer, progesterone receptor (PR) / PGR genes for PR-positive cancer, and HER2 for HER2+ breast cancer, resulting in a corresponding gene list as follows: For each gene, an ensembl_gene_id is provided along with a coefficient value, which is a weighting coefficient corresponding to the particular molecular subtype, and an hgnc_symbol. For example, the lasso regression model may be applied to the following gene list: [Table 1] [Table 2] [Table 3]

[0104] At block 308, a molecular subtype determination based on gene expression is made, for example, from a linear regression analysis, and assigned as a surrogate molecular subtype for the discordant imaging assay results. For example, if the IHC imaging assay result indicates HER2+ and the FISH indicates HER2-, and the processes of blocks 306 and 308 determine HER2-, the IHC imaging assay result may be substituted in the RWD / RWE data for the sample with a substitution determination of HER2-.

[0105] Advantageously, in various embodiments herein, discrepancies between imaging assay results may result in processes 306 and 308 identifying completely different molecular subtypes from either of the imaging assay results. For example, processes 306 and 308 may supplement the HER2 call with an ER call (ER+ / -) or a PR call (PR+ / -). In fact, in various embodiments, processes 306 and 308 may be performed even in the absence of discrepancies identified in block 304. For example, processes 306 and 308 may be performed before receiving imaging assay results. In fact, in some examples, processes 306 and 308 may be performed first before imaging analysis or if imaging analysis is not available. Processes 306 and 308 may be performed, and only if a particular result, such as HER2+, is determined, the computing device will send a request for an image-based analysis. In fact, in these methods, process 300 may be used to predict other biomarkers in IHC results by examining gene expression levels in associated RNA-seq data. That is, process 300 may be used to identify ER, PR, or Ki67 (e.g., to determine luminal A vs. luminal B) and then determine a molecular subtype, such as HER2+ / -, ER+ / -, PR+ / -, or triple negative. Further, if the molecular subtype is ER+ or PR+, it determines whether the subtype is luminal A / B. An illustrative example of biomarkers that may be identified by these imaging assay replacement techniques is described in U.S. Patent Application No. 16 / 888,357, entitled "A pan-cancer model to predict the pd-l1 status of a cancer cell sample using RNA expression data and other patient data," filed May 29, 2020, and incorporated herein by reference in its entirety for all purposes.

[0106] Optionally, in some examples, the replacement molecular subtype is provided to block 310, which determines a recommended set of treatments / therapies from a larger population of available treatments / therapies corresponding to the replacement molecular subtype. For example, the available treatments generally associated with the molecular subtypes may be stored in a ranked manner in the computing device 102. At block 312, a report is generated indicating the determined molecular subtypes, including resolved subtypes from the analysis, and, in some examples, image-based discordant subtypes. In various embodiments, the report may include a report element indicating one or more treatments / therapies that are likely to work for a given molecular subtype, based on a stored association between the two, or a known association (e.g., published research studies and / or clinical trials, or NCCN or FDA guidelines). In various embodiments, the report may be ranked by best match for the determined molecular subtype.

[0107] FIG. 3B illustrates another example of performing a training process 350 that may be executed by computing device 102 using resolver 138 to train molecular subtype model 136, for example, for use in performing the process of block 306. At block 352, a target biomarker and / or IHC target classification, e.g., a biomarker / target for predicting cancer type and / or cancer molecular subtype, is selected. For example, for breast cancer, it is known that IHC results for HER2, ER, PR / PGR, and Ki67 biomarkers are used. At block 354, training data is obtained using a plurality of patient specimens, each associated with RNA-seq data and a positive or negative status for that biomarker / target. At optional block 356, a pathway score (z-score) is calculated for each specimen based on the RNA-seq data (the RNA-seq data may be transcriptome-wide or a target panel). At block 358, the pathway scores for each specimen group (positive vs. negative) are analyzed to select a z-score threshold that separates positive values ​​from negative values. The result is a trained pathway analysis model 360 that may be stored as a model of molecular subtype gene expression levels 362 of process 306. Alternatively, in training a logistic regression model, in block 364, the gene expression levels for the IHC targets after splitting the specimens into positive and negative values ​​are used to train a logistic regression model. The resulting linear regression model 366 is stored as model 362. Alternatively, in training a multi-gene model, in block 368, to train a Lasso model, the gene expression levels of the RNA-seq data are fed to the Lasso model as a table where each row is a specimen, one column is an IHC status, and each other column is a gene with gene expression level for that specimen. The resulting trained Lasso model 370 is stored as model 362. As will be appreciated, although each of these training pathways is shown separately, depending on the sufficiency of the training data, model 362 may collectively include a classification model trained by all of these training pathways.

[0108] An exemplary implementation of a gene expression pathway analysis pipeline according to processes 200 and 300 is described below. EXAMPLES

[0109] In an exemplary implementation, cohort selection was performed to identify patient data for analysis using the techniques herein, including the process of FIG. 2. For cohort selection, two retrospective breast cancer cohorts were randomly selected from a clinical genetics database after applying clinically relevant selection criteria. All data were anonymized in accordance with the Health Insurance Portability and Accountability Act (HIPAA). The dates used for the analysis were relative to the date of breast cancer primary diagnosis (PDX), with the year of PDX randomly offset. PDX within the cohorts from 1990 to 2018.

[0110] The first data group was a clinical abstraction (CA) cohort of 4,000 breast cancer patients randomly selected as a representative sample of structured RWD in oncology databases. Cohort inclusion required data on breast cancer PDX, date of PDX, age, race, sex, stage, histological subtype, and estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2) status. Recorded stage and histological subtype had to be within 30 days relative to the PDX date, whereas receptor status could have been recorded within 30 or 50 days depending on testing methodology.

[0111] The second cohort was a randomly selected molecular sequencing (MLC) cohort including 400 primary breast cancer patients with a PDX date and tumor biopsy who underwent RNA-seq and targeted DNA sequencing (DNA-seq) with either a whole exome panel or two targeted sequencing panels between 2017 and 2019. Only patients with reported mutations in the cohort were included, although less than 1% of all breast cancer cases in the database have no reported DNA mutations.

[0112] As part of the image-based pipeline, abstracted molecular markers were determined from histopathological imaging data using the first cohort. Protein expression from immunohistochemistry (IHC) results for ER and PR, and IHC and fluorescent in situ hybridization (FISH) results for HER2 were curated during clinical data abstraction. Receptor results contained abstracted equivocal, positive, or negative status. Hormone receptor (HR) status was classified by a combination of ER and PR status. When available, Ki-67 index was determined from categorical interpretation of expression levels reported from pathology reports and / or progress notes in routine clinical practice. Normalized Ki67 results contained indeterminate, low, equivocal, intermediate, or high status. Chi-squared tests assessed the significance of differences in distribution of Ki67 test results. Fisher's exact test was performed for post-hoc analysis, and P values ​​were adjusted for multiple hypothesis testing using Bonferroni correction.

[0113] Molecular subtype determination was performed as follows. Molecular subtype of each CA patient was classified as HR+ / HER2-, HR+ / HER2+, HR- / HER2+, or triple-negative breast cancer (TNBC) based on receptor status at diagnosis. HR status was determined from the most recent IHC result or physician's note recorded within 30 days of the date of PDX. HR+ status included ER+ / PR+, ER+ / PR-, and ER- / PR+. HER2 status was determined from the most recent FISH result recorded within 50 days of the date of PDX. In the absence of HER2 FISH data, the most recent IHC result or physician's note within 30 days of the date of PDX was utilized. References to results under "initial diagnosis" refer to these 30-day and 50-day time frames. Molecular subtype of the MLC cohort was determined from IHC or FISH results linked to the patient's pathology report.

[0114] RWE or clinical data abstraction was performed as follows: Clinical data were extracted from a real-world oncology database of longitudinal structured and unstructured data from geographically diverse oncology practices, including integrated delivery networks, academic institutions, and community practices. Many of the records included in this study were acquired collaboratively through ASCO CancerLinQ. Structured data from electronic health record systems were integrated with unstructured data collected from patient records via technically possible chart abstraction and corresponding molecular data, when applicable. Data were harmonized and normalized to standard terminology from MedDRA, NCBI, NCIt, NCIm, RxNorm, and SNOMED.

[0115] In the examples, a determination of menopausal status was made. Menopausal status was determined using relevant abstracted text fields when available. Patients were considered premenopausal if a single, undated, menopausal negative (perimenopausal, premenopausal, or menstrual) status was recorded on or before the PDX date and no menopausal positive (menopausal or postmenopausal) status was indicated prior to diagnosis. Patients were considered premenopausal at PDX if a menopausal event was recorded 1 year after the PDX date.

[0116] Similarly, patients with undated positive menopausal status and those with menopausal or postmenopausal status documented before the date of PDX were considered postmenopausal. Patients were also considered postmenopausal if no menopausal information was available before the date of PDX, but menopausal or postmenopausal status was indicated within 1 year.

[0117] Menopausal status outside these criteria was indicated as "unknown."

[0118] Upon ingestion and analysis of the data, an overall survival analysis was performed. Overall survival (OS) was calculated for all stage I-IV CA cohort patients with invasive breast cancer (n=3,952). Patients with unknown relative death dates were right-censored at the date of their most recent relative clinical interaction. Survival curves were generated in R using the survival (v2.43-4) and survminer (v0.4.3) packages with P values ​​calculated by the log-rank test. Results show the proportion of patients surviving per year, stratified based on stage and HER2, ER, and triple-negative status.

[0119] For the genomic testing pipeline, a second cohort of data was used. In particular, the reported variants in the MLC cohort were generated from formalin-fixed paraffin-embedded (FFPE) slides of primary breast tumor biopsies and, when possible, targeted DNA sequencing of matched saliva or blood samples. Whole-transcriptome RNA sequencing was performed on samples from the same tissue blocks. Most samples were sequenced with one of two targeted DNA sequencing assays that detect oncological targets in solid tumors and hematological malignancies as previously described. Two patient samples were sequenced with an updated and improved version of the xT panel targeting clinically relevant exons of 596 genes, and their reported variants were merged for analysis. Additionally, one sample in the MLC cohort was sequenced with a whole-exome panel targeting 19,396 genes across a genomic region of 39 megabases (Mb).

[0120] Because each assay targeted a different set of genes, the MLC cohort variant analysis included only genes that were tested across all 400 samples. Variants were classified and reported according to previously established clinical guidelines. Reported variants were classified as amplifications or deletions, alterations, fusions, or copy number variations. Alterations included variants of unknown significance (VUS), biologically relevant or potentially actionable alterations, and both germline VUS and pathogenic variants.

[0121] Tumor mutation burden (TMB): TMB was calculated by dividing the number of nonsynonymous mutations by the adjusted panel size of each assay (2.4 Mb, 5.86 Mb, and 36 Mb, respectively). All nonsilent somatic coding variants with >100x coverage and >5% allelic fragmentation, including missense, indel, and stop-loss variants, were counted as nonsynonymous variants.

[0122] RNA-based prediction of molecular subtypes: Transcriptomic models were used to predict receptor status in an MLC cohort that included patients with missing IHC or FISH data. In this example, single-gene logistic models were trained on an independent set of RNA-seq'd breast cancer samples according to normalized gene expression of ESR1, PGR, or ERBB2 using the R glm package v2.0-16. In contrast to the MLC cohort, this independent training set included samples from both primary and metastatic breast cancers. Model performance was evaluated separately for primary samples, metastatic samples, and the combined set using 10-fold cross-validation (Table 1). Performance was evaluated on a test set consisting of RNA-seq'd samples from the MLC cohort, where IHC or FISH results from the database (ER n=308, PR n=306, HER2 n=261) were abstracted and removed from the training set. Abstracted results from clinical IHC or FISH testing were derived from the same tissue as the subsequent RNA-seq data. Positivity thresholds for the IHC prediction models were selected using Youden's J statistic to optimize sensitivity and specificity. Normalized values ​​of ≥2.36 for ESR1, ≥1.20 for PGR, and ≥3.59 for ERBB2 expression were considered positive for ER, PR, and HER2 receptor status, respectively.

[0123] Gene expression collection, processing, and normalization: Gene expression was generated through RNA-seq of FFPE tumor samples using an exome capture-based protocol. Quantification of transcript levels to GRCh37 was performed using Kallisto0.44. Transcript counts were then corrected for GC content and length using quantile normalization and adjusted for sequencing depth via the size factor method. Normalized counts across protein-coding transcripts covered by the exome panel were then summed to obtain gene-level counts. Subsequent expression analysis was performed using log 10 This was performed using converted counts.

[0124] RNA-seq pathway analysis: Gene sets were downloaded from the MSigDB website (http: / / software.broadinstitute.org / gsea / msigdb / index.jsp) and pathway enrichment scores were calculated from normalized gene expression using the ssGSEA function in the Gene Set Variation Analysis (GSVA) ​​R Bioconductor package v1.0.6. ER- and HER2-related pathways were identified as pathways containing the terms "ESR1" or "estrogen" and "ERBB2" or "HER2", respectively. Z-scores were calculated for each set of enrichment scores and the sign was reversed for all pathways containing "DN" (down) or "inhibited". For select analyses, the average of z-scores across pathways was calculated to generate pathway metascores for patients, as an example of summary scores. Except for metascore analyses of the HER2 and ER signaling pathways, receptor status was derived from both abstracted and predicted protein expression. Significance was determined by Wilcoxon test for comparisons between two groups and Kruskal-Wallis test for comparisons between three or more groups, with P<0.05 considered significant. A separate gene set analysis was performed to examine differences in enrichment between the four molecular subtypes compared to 50 hallmark pathways, a highly curated list from the MSigDB database. To determine how patients were clustered by pathway score, a second UMAP analysis was performed featuring the enrichment scores of each hallmark pathway.

[0125] The results of these examples are as follows:

[0126] First, we present results on real-world evidence from the Clinical Abstraction (CA) breast cancer cohort, the first cohort.

[0127] Patient demographics and clinical characteristics in the CA cohort: We first determined whether key demographic and clinical characteristics captured in the RWD replicated clinical studies and found that the de-identified data were consistent with previous large breast cancer cohort studies (Table 2). The cohort is comprised primarily of women (99.3%, n=3,970) with a median age at diagnosis of 61.0 years. Years of diagnosis in the cohort ranged from 1990 to 2018 (Figure 4). Self-reported race was 83.3% White (n=3,332), 13.1% Black or African American (n=523), and 3.6% Asian or Pacific Islander (n=145). Of the 2,042 women with menopausal data, 87.4% (n=1,784) were postmenopausal. Abstracted stage at initial diagnosis consisted primarily of stages I (49.6%, n=1,986) and II (33.3%, n=1,333), followed by III (10.5%, n=420), IV (5.5%, n=219), and 0 (1.1%, 42). Most tumors had a histologic classification of invasive ductal carcinoma (77.4%, n=3,095), and 9.5% (n=378) had an invasive ductal component or were NOS. Several rare cancer types were also represented.

[0128] Molecular subtype determination in the CA cohort: We evaluated the extent to which RWD captures molecular marker information from clinical test results. The distribution of all abstracted receptor test results at initial diagnosis was shown in Figure 5A. Consistent with previous US breast cancer statistics, the most common molecular subtype was HR+ / HER2- (71.5%, n=2,859), followed by TNBC (12.3%, n=491) (Figure 5B). Among HR+ patients with unclear status, most were ER+ / PR+ (71.0%, 2,839 out of 3,996), followed by ER+ / PR- (10.4%, n=417) and ER- / PR+ (1.4%, n=57) (Figure 5C). Finally, abstracted Ki67 IHC test results were consistent with Ki67 expression levels, which usually indicate a specific breast cancer subtype (Figure 5D). The distribution of Ki67 results was significantly greater among molecular subtypes (chi-square, P = 1.75x10 -9), especially HR+ / HER2- vs HR+ / HER2+ patients (P=0.015) and TNBC vs HR+ / HER2- patients (P=6.38x10 -9 The highest proportion of high Ki67 IHC results were in TNBC (82.0%, n=50 of 61) and HR- / HER2+ patients (75.0%, n=15 of 20), whereas the lowest Ki67 results were in HR+ / HER2- patients (44.0%, n=140 of 318).

[0129] Analysis of anti-HER2 therapy in the CA cohort: We next examined anti-HER2 therapy treatment patterns from the longitudinal RWD. Selected anti-HER2 therapies included trastuzumab, ado-trastuzumab emtansine, neratinib, lapatinib, and pertuzumab. Among CA cohort patients, 13.7% (n=546) were HER2+ at initial diagnosis, of which 74.2% (n=405) received anti-HER2 therapy at some point during clinical care. Approximately 70.0% of patients who received anti-HER2 therapy did so within 3 months of a positive test result, and the majority (73.5%) had early-stage cancer (Figure 6A). These results are consistent with previous breast cancer cohort studies. Furthermore, a small proportion of HER2- patients showed evidence of receiving anti-HER2 therapy (1.1%, 36 of 3,352 HER2 patients). Among these patients, 27.8% received anti-HER2 therapy within 3 months of initial diagnosis (n=10), 30.6% 3–12 months after initial diagnosis (n=11), and 33.3% >1 year after initial diagnosis (n=12), while 8.3% did not have a start date for anti-HER2 therapy recorded (n=3) (Figure 6B). Additionally, 33.3% (n=12) had evidence of discordant results at the time of initial diagnosis, 44.4% (n=16) had HER2 results only, and 22.2% (n=8) had HER2 equivocal or positive results recorded since initial diagnosis. A small proportion of patients (n=37) did not have a HER2 treatment time window assigned due to date quality issues.

[0130] Analysis of HER2 test results in the CA cohort: To assess inter- and intra-test concordance, we compared HER2 IHC and FISH results among patients with both tests performed close to initial diagnosis (17.7%, n=709). Among patients with HER2+ IHC results and subsequent FISH testing, 62.2% (n=51 of 82) had inter-test concordance (Table 3); however, 31.7% of HER2+ IHC were HER2- by FISH (n=26 of 82). This discordance is greater than in a previously reported meta-analysis of IHC and FISH HER2 testing worldwide. 54 Four of these 26 patients had received anti-HER2 therapy during the clinical timeline.Among patients with HER2-IHC results, 3.9% (n=7 of 182) were HER2+ by FISH, similar to previous reports. 54 The majority of these patients (n=6 of 7) received anti-HER2 therapy. Equivocal IHC results for HER2 (HER2 IHC2+) were observed in 62.8% (n=445 of 709) of the cohort. Among these patients with equivocal results, 7.8% (n=35 of 445) were subsequently confirmed as equivocal by FISH testing. However, 80.7% (n=359 of 445) had subsequent HER2- and 11.5% (n=51 of 445) HER2+ FISH results.

[0131] Furthermore, intratest discordance was analyzed in patients with multiple HER2 results at initial diagnosis. Among patients with multiple HER2 IHC results at diagnosis (7.1%, n=253 of 3,561 with HER2 IHC), 18.6% (n=47) showed intratest discordance. Among patients with multiple HER2 FISH results (4.5%, n=52 of 1,157), 21.2% (n=11) showed intratest discordance.

[0132] Overall survival in the CA cohort: OS analysis from longitudinal RWD revealed that overall 5- and 10-year survival rates (92.2% and 85.7%, respectively) were relatively consistent with the mean US percentages. Survival was high as expected and varied as expected by stage (P<0.0001) (Figure 7A) and receptor status (P=0.016) (Figure 7B). Five-year survival rates were 93.5% for stage I-IV HER2+ patients and 92.0% for HER2- patients (P=0.45), and 74.3% and 57.1% for stage IV patients, respectively (P=0.098) (Figure 7C, Figure 7D). The 5-year survival rates were 92.7% among ER+ patients and 89.8% among ER- patients with stages I-IV (P = 0.052), and 63.7% and 55.5% among stage IV patients, respectively (P = 0.12) (Figure 7E, Figure 7F). TNBC patients had significantly worse OS compared with other subtypes, with the 5-year survival rate of stage IV TNBC patients being 36.3% and 65.1% among stage IV non-TNBC patients (P = 0.0024) (Figure 7G, Figure 7H).

[0133] Second, we present results on genomic testing insights from the molecularly sequenced cohort, the second cohort.

[0134] Patient demographics and clinical characteristics in the MLC cohort: Abstracted clinical characteristics and patient demographics from the 400 MLC cohort patients were evaluated (Table 2) and found to be relatively consistent with the CA cohort and other large breast cancer cohort studies. The cohort had a slightly younger median age at diagnosis of 55.8 years (range 45.2-66.4) and a higher proportion of Black or African American (14.6%, n=35) and Asian or Pacific Islander patients (5.4%, n=13) than the CA cohort. Patients with known stage information were mostly stage II (38.4%, n=83) at diagnosis, followed by stages IV (26.4%, n=57), III (21.3%, n=46), and I (13.9%, n=30), indicating an overall higher-risk population compared to the CA cohort. A total of 75.0% (n=267) of the tumors were invasive ductal carcinoma, with several rare cancer types also represented in the cohort.

[0135] DNA sequence analysis of the MLC cohort: The top three genes with reported alterations were found to be TP53, PIK3CA, and GATA3, in 55.0% (n=220), 29.0% (n=116), and 13.8% (n=55) of the MLC cohort, respectively (Figure 8A). These findings are consistent with previous analysis of breast cancer data from The Cancer Genome Atlas. Figure 8B shows the distribution of variants in the 20 most frequently reported genes. Evaluation of patients with tumor / normal matched DNA sequences (n=356) identified 18 patients (5.1%) with pathogenic germline mutations in 12 NCCN-designated familial high-risk genes (Figure 8C). This subpopulation may have been underrepresented because exon-level duplications or deletions were not included. Among the 18 patients harboring a pathogenic germline mutation in one of these 12 genes, most included variants in BRCA1 (n=6), BRCA2 (n=6), CHEK2 (n=6), ATM (n=2), and / or PALB2 (n=2). In this embodiment, as TMB and MSI status are integrated biomarker measurements, we observed a wide range of TMB across the cohort with a median of 1.7 mutations / Mb (Figure 8D). Consistent with previous studies, the majority of patients (84.7%, n=339) had stable MSI, while only 0.3% (n=1) had MSI-high and 0.5% (n=2) had MSI-low.

[0136] RNA-based prediction of receptor status of molecular subtypes: We developed a whole-transcriptome model based on 19,147 genes to predict IHC receptor status and resolve molecular subtypes in the MLC cohort. The predicted RNA-based subtypes are largely consistent with the abstracted IHC-based subtypes (Figure 9A). Similar to the literature, transcriptomic signatures differ between molecular subtypes using TNBC clustering individually. Seventeen samples clustered with TNBC but were predicted or abstracted as another subtype, and samples clustered outside the group may benefit from further testing or analysis. ESR1, PGR, and ERBB2 gene expression correlated with the respective abstracted and predicted receptor status (Figure 9B).

[0137] RNA-based receptor status prediction was highly accurate for ER (95.5%, AUROC 98.1%) and HER2 (94.6%, AUROC 93.8%) compared to abstracted status, whereas PR status was predicted with slightly lower accuracy (87.9%, AUROC 95.2%) (Figure 10). Prediction accuracy for all receptors was 92.7%. A detailed overview of validation data and model performance is available in Table 1. Patients with incompletely abstracted molecular subtypes (n=150) were classified by receptor status predicted from the transcriptomics model. Importantly, patients with equivocal HER2 status abstracted from IHC and / or FISH results (n=36) were predicted by the model as HER2+ (n=7) or HER2- (n=29).

[0138] RNA-based HER2 and ER pathway analysis: To further evaluate the potential of RNA-seq to enrich clinical data in breast cancer, gene set enrichment analysis was performed using the MSigDB database. First, we assessed whether ambiguous or unclear IHC and FISH test results could be resolved by measuring the activity of signaling pathways. Multigene sets that putatively measure such pathway activity were identified by searching for "ERBB2", "HER2", "ESR1", or "estrogen" in the MSigDB database (Figure 11A and Figure 11B). The results of the pathway analysis were expressed as metascores to avoid biases introduced when selecting a single pathway. HER2 IHC-positive and FISH-positive samples were enriched for HER2 activity metascores, as expected, but the HER2 signaling results contained considerable variability in pathway activity (Figure 12A). Notably, GO_ERBB2_SIGNALING_PATHWAY, which directly measures HER2 activity, showed a strong correlation with HER2 expression (r=0.453) (Figure 11A) and significantly different enrichment between HER2 status (P=0.00031) (Figure 13). ER enrichment scores were more distinct between IHC-positive and IHC-negative patients, consistent with the relatively high reliability of ER IHC compared to HER2 testing, and variability was also observed in ER signaling results (Figure 12B, Figure 14).

[0139] Next, RNA-seq data was analyzed in the context of a highly curated set of hallmark pathway genes to determine differential activation of biological pathways between breast cancer subtypes. Most hallmark pathways (32 out of 50) showed significantly different enrichment scores between molecular subtypes (Figure 15A). UMAP using only scores from these 50 pathways recapitulated the TNBC clustering observed in the full-transcriptome UMAP (Figure 12C). As expected, HR+ samples, but not HR- or TNBC samples, were highly enriched for two pathways related to estrogen signaling (Figure 15B). Among HR- / HER2+ cancers, we observed enrichment for pathways known to be downstream of HER2, RAS, and mTOR (Figure 12D). HER2-driven tumors also showed enrichment for all immune-related hallmark pathways, a finding consistent with the literature. Many oncogenic signaling pathways, including Wnt, mTOR, PI3K, Hedgehog, and Notch, were enriched in TNBC (Figure S12E), consistent with TNBC tumors' reliance on ER-, PR-, and HER2-independent pathways. TNBC samples were also enriched for pathways associated with mitotic index, as expected, due to their relatively high growth rate, consistent with a high Warburg effect, and potentially targetable glycolysis and cancer / testis antigens.

[0140] A third type of pipeline in the computing device 102 is a hybrid pipeline 150, which uses pathway data obtained from gene expression data and molecular subtype data obtained from histopathological analysis to inform each other to generate an aggregated pathway score that can be used for more accurate assessment of molecular subtypes. An exemplary implementation of the pipeline 150 is shown in FIG. 16 and process 400. At block 402, histopathological images, such as IHC and / or FISH images, are acquired in a pathway analysis system. In some embodiments, image processing is performed similar to that of process 302 in FIG. 3. At block 404, the images are provided to a trained image molecular subtype classification model, which may be implemented by the trained molecular subtype neural network 158. However, in process 400, instead of examining discrepancies between different image sets, the classification model of process 404 may be trained to classify one or more types of image sets. For example, the classification model may be trained to classify only IHC images, or may be trained to classify only FISH images. In this manner, an existing image classification model may be used. In some embodiments, the classification models of process 404 contain classification models for multiple different histopathological image set types.

[0141] For example, in various embodiments, different sets of histopathological images may be received, with each set determined using a different image generation process, e.g., a first set of images are IHC stained images and a second set of images are FISH images. Other image set types include hematoxylin and eosin (H&E) stained images. At block 404, initial image processing may be performed on these received images, e.g., by image processing stage 154. For example, the images may be analyzed on a whole slide basis or on a tile basis. For the latter, in some examples, the image is segmented into multiple tile images by applying a tiling mask to the digital image, with each tile image containing a portion of the digital image. These tile masks may generate tiles of the same or different sizes, and the tiles may be rectangular shaped, square shaped, etc.

[0142] At block 404, the image, whole slide or tile may be provided to a trained image-based molecular subtype classification model, such as model 158. The classification model may be a neural network, such as a convolutional neural network. In some examples, the classification model is a tile-resolution fully convolutional network (FCN) classification model. In some examples, the classification model is a tissue classification model trained using a set of training images with annotations identifying various tissue types, the training images including histopathological images that are fed to a deep learning framework that trains the tissue classification model using a convolutional neural network. In some examples, the classification model is a cell segmentation classification model trained using a set of training images with annotations identifying cell boundaries, cell interiors, and cell exteriors, the training images being histopathological images that are fed to a deep learning framework that trains the cell segmentation classification model using a convolutional neural network. In some examples, the classification model is a UNet classification model.

[0143] At block 406, the resulting molecular subtypes from the classification process 404 are compared to the aggregate pathway analysis data and molecular subtype calls from gene expression, as may be performed, for example, by process 200 of FIG. 2. The comparison at block 406 allows for a generally less accurate call, such as a HER2+ or HER2 call made from an IHC classification model, to be compared to a highly accurate result based on aggregate pathway analysis using a gene expression-based pipeline. The comparison also applies to FISH analysis of HER2, although FISH-based classification models are generally more accurate than IHC-based classification models. At block 406, the comparison may identify discrepancies between the resulting molecular subtype data and pass the different calls to a report generator block 408. The resulting report may include data from both the calls obtained from the histopathological images and the calls obtained from the gene expression. In some examples, block 406 passes the pathway aggregate scoring, molecular subtype, and gene expression data for one or more of the set of pathways to block 408 for report generation.

[0144] In some embodiments, block 410 performs a quantitative comparison between the determinations to determine whether one of the determinations, the more likely gene expression derived one, has a higher level of confidence relative to the others. In such embodiments, block 406 may determine the molecular subtype as the one having the highest confidence level value, for example, by using a softmax function to generate a probability for each pathway prediction. This molecular subtype value is transmitted to process 408 for inclusion in the generated report.

[0145] In various embodiments, the pathways, pathway scores, and gene expression values ​​from the gene expression pipeline are provided by block 406 to train a hybrid molecular subtype classification model in block 4. The hybrid classification model may be trained with gene expression and pathway scores from histopathological images (e.g., IHC and / or FISH images) and gene expression data (e.g., RNA-seq data) corresponding to the same cohort samples. The hybrid classification model may, for example, perform multi-gene classification on the subsequently received histopathological images, allowing multiple different subtypes to be classified. The hybrid classification model may be a neural network, such as a convolutional neural network, configured as a whole image classifier or a tile-based classifier, like the classification model of block 304.

[0146] FIG. 17 illustrates an exemplary computing device 500 for implementing the pathway analysis computing device 100 of FIG. 1. As illustrated, the computing device 100 can be implemented on one or more processing units 510, which may represent the computing device 500, particularly a central processing unit (CPU), and / or on one or more graphic processing units (GPUs) 511, which may include a cluster of CPUs and / or GPUs, and / or on one or more tensor processing units (TPUs) (also labeled 511), any of which may be cloud-based. Features and functions described for the computing device 100 can be stored and implemented from one or more non-transitory computer-readable media 512 of the computing device 500. The computer-readable media 512 can include, for example, an operating system 514 and a pathway analysis system (e.g., pipeline, etc.) 516 having corresponding elements of the pipelines 130, 140, and 150, as well as the pre-processing layer 120, and the molecular subtype layer 160. More generally, the computer readable medium 512 may store trained classification models, executable code, etc. used to implement the techniques herein. The computer readable medium 512 and processing unit 510 and TPU / GPU 511 may store histopathological images, gene expression data, pathway expression data, pathway scores, aggregate pathway scores, enrichment scores, molecular subtype classification data, and other data in one or more databases 513 herein.

[0147] The computing device 500 includes a network interface 524 communicatively coupled to a network 550 for communicating to and / or from a portable personal computer, a smart phone, an electronic document, a tablet, and / or a desktop personal computer, or other computing device. The computing device further includes an I / O interface 526 connected to devices such as a digital display 528, a user input device 530, etc. In some embodiments, the computing device 500 generates the molecular subtype / pathway data as an electronic document 515 that can be accessed and / or shared over the network 550, as described herein.

[0148] In the illustrated embodiment, the system 100 is implemented on a single server 500. However, the functionality of the system 100 may be implemented across distributed devices 500, 502, 504, etc., connected to each other via communication links. In other embodiments, the functionality of the system 100 may be distributed across any number of devices, including the illustrated portable personal computers, smartphones, electronic documents, tablets, and desktop personal computing devices. In other embodiments, the functionality of the computing device 100 may be cloud-based, such as, for example, one or more connected cloud TPUs customized to perform machine learning processes. The network 550 may be a public network, such as the Internet, a private network, such as a research institute or corporate private network, or any combination thereof. The network may include a local area network (LAN), a wide area network (WAN), cellular, satellite, or other network infrastructure, whether wireless or wired. The network may utilize communication protocols, including packet-based and / or datagram-based protocols, such as the Internet Protocol (IP), Transmission Control Protocol (TCP), User Datagram Protocol (UDP), or other types of protocols. Additionally, the network may include a number of devices that facilitate network communications and / or form the hardware infrastructure of the network, such as switches, routers, gateways, access points (such as wireless access points as shown), firewalls, base stations, repeaters, backbone devices, etc.

[0149] The computer-readable medium may include executable computer-readable code stored on a computer (including, for example, a processor and a GPU) for programming a computer to the techniques of the present specification. Examples of such computer-readable storage media include hard disks, CD-ROMs, digital versatile disks (DVDs), optical storage devices, magnetic storage devices, ROMs (read-only memories), PROMs (programmable read-only memories), EPROMs (erasable programmable read-only memories), EEPROMs (electrically erasable programmable read-only memories), and flash memories. More generally, the processing unit of the computing device 1300 may represent a CPU-type processing unit, a GPU-type processing unit, a TPU-type processing unit, a field programmable gate array (FPGA), another class of digital signal processor (DSP), or other hardware logic components that can be driven by a CPU.

[0150] While the exemplary classification models and neural networks herein have been described as being configured with exemplary machine learning architectures (FCN and UNET configurations), it should be noted that any number of suitable convolutional neural network architectures can be used. Generally speaking, the classification models herein may implement any suitable statistical model (e.g., neural networks or other models implemented through machine learning processes) that is applied to each of the received images. As discussed herein, the statistical model may be implemented in a wide variety of ways. In some examples, a classifier is developed that uses machine learning to evaluate training images and / or corresponding gene expression data and correlates predetermined image features to specific categories or molecular subtypes of gene expression. In some examples, image features can be identified using a learning algorithm, such as a neural network, a support vector machine (SVM), or other machine learning process, as a training classifier. Once the classifier in the statistical model is properly trained with a set of training images, the statistical model can be used in real time to analyze subsequent images that are provided as input to the statistical model to predict the status of the biomarker. In some examples, when the statistical model is implemented using a neural network, the neural network can be configured in a wide variety of ways. In some examples, the neural network may be a deep neural network and / or a convolutional neural network. In some examples, the neural network may be a distributed and scalable neural network. The neural network may be customized in a wide variety of ways, such as providing a specific top layer, such as a logistic regression top layer. A convolutional neural network may be viewed as a neural network that includes a set of nodes with associated parameters. A deep convolutional neural network may be viewed as having a structure in which multiple layers are stacked. Neural networks or other machine learning processes may include a variety of sizes, numbers of layers, and levels of connectivity.Some layers can correspond to stacked convolutional layers (optionally followed by contrast normalization and max pooling) followed by one or more fully connected layers. For neural networks trained with large datasets, the number of layers and layer size can be increased by using dropout to address the potential problem of overfitting. In some cases, neural networks can be designed to forego the use of fully connected upper layers at the top of the network. By forcing the network to perform dimensionality reduction in the intermediate layers, very deep neural network models can be designed while dramatically reducing the number of learned parameters.

[0151] A system for performing the methods described herein may include a computing device, and more specifically may be implemented in one or more processing units, e.g., a central processing unit (CPU), and / or one or more graphics processing units (GPUs), including a cluster of CPUs and / or GPUs. The features and functions described may be stored on and implemented from one or more non-transitory computer-readable media of the computing device. The computer-readable media may include, for example, an operating system and software modules, or "engines," that implement the methods described herein. More generally, the computer-readable media may store batch normalization process instructions of the engine for implementing the techniques herein. The computing device may be a distributed computing system, such as an Amazon Web Services cloud computing solution.

[0152] The computing device includes a network interface communicatively coupled to a network for communicating to and / or from a portable personal computer, a smart phone, an electronic document, a tablet, and / or a desktop personal computer, or other computing device. The computing device further includes an I / O interface connected to devices such as a digital display, a user input device, etc.

[0153] The engine's functionality may be implemented in distributed computing devices, etc., interconnected via communication links. In other examples, the system's functionality may be distributed across any number of devices, including the illustrated portable personal computers, smartphones, electronic documents, tablets, and desktop personal computing devices. The computing devices may be communicatively coupled to a network and another network. The network may be a public network, such as the Internet, a private network, such as a research or corporate network, or any combination thereof. The network may include a local area network (LAN), a wide area network (WAN), a cellular, satellite, or other network infrastructure, whether wireless or wired. The network may utilize communication protocols, including packet-based and / or datagram-based protocols, such as the Internet Protocol (IP), Transmission Control Protocol (TCP), User Datagram Protocol (UDP), or other types of protocols. Additionally, the network may include multiple devices, such as switches, routers, gateways, access points (such as wireless access points as illustrated), firewalls, base stations, repeaters, backbone devices, etc., that facilitate network communication and / or form the hardware infrastructure of the network.

[0154] The computer-readable medium may include executable computer-readable code stored on a computer (including, for example, a processor and a GPU) for programming a computer to the technology herein. Examples of such computer-readable storage media include hard disks, CD-ROMs, digital versatile disks (DVDs), optical storage devices, magnetic storage devices, ROMs (read-only memories), PROMs (programmable read-only memories), EPROMs (erasable programmable read-only memories), EEPROMs (electrically erasable programmable read-only memories), and flash memories. More generally, the processing unit of a computing device may represent a CPU-type processing unit, a GPU-type processing unit, a field programmable gate array (FPGA), another class of digital signal processor (DSP), or other hardware logic components that can be driven by a CPU.

[0155] The above methods and systems can be utilized in conjunction with or as part of a digital and laboratory healthcare platform generally directed to medical care and research. It should be understood that many uses of the above methods and systems are possible in conjunction with such platforms. One example of such a platform is described in U.S. Patent Application No. 16 / 657,804, entitled "Data Based Cancer Research and Treatment Systems and Methods," filed October 18, 2019, and incorporated herein by reference in its entirety for all purposes.

[0156] In one embodiment, when the platform includes a genetic analysis system, the genetic analysis system may include a targeted panel and / or a sequencing probe. One example of a targeted panel is disclosed, for example, in U.S. Provisional Patent Application No. 62 / 902,950, entitled "System and Method for Expanding Clinical Options for Cancer Patients using Integrated Genomic Profiling," filed September 19, 2019, which is incorporated by reference in its entirety for all purposes. An example of a next generation sequencing probe design is disclosed, for example, in U.S. Provisional Patent Application No. 62 / 924,073, entitled "Systems and Methods for Next Generation Sequencing Uniform Probe Design," filed October 21, 2019, which is incorporated by reference in its entirety for all purposes.

[0157] In one example, if the platform includes a bioinformatics pipeline, the above methods and systems may be utilized after completion or substantial completion of the systems and methods utilized in the bioinformatics pipeline. As an example, the bioinformatics pipeline may return a set of binary files, such as one or more BAM files, reflecting RNA read counts aligned to a reference genome. The above methods and systems may be utilized, for example, to incorporate RNA read counts and generate cellular pathway activation and / or predicted protein expression information as a result. Other inputs, such as DNA read counts, may also be used as described herein.

[0158] As described above, the pipeline may include an RNA data normalizer, such as disclosed in U.S. Patent Application No. 16 / 581,706, entitled "Methods of Normalizing and Correcting RNA Expression Data," filed September 24, 2019, which is incorporated herein by reference in its entirety for all purposes. The pipeline may include a genetic data deconvoluter, such as disclosed in U.S. Provisional Patent Application No. 62 / 786,756, entitled "Transcriptome Deconvolution of Metastatic Tissue Samples," filed December 31, 2018, and U.S. Provisional Patent Application No. 62 / 924,054, entitled "Calculating Cell-type RNA Profiles for Diagnosis and Treatment," filed October 21, 2019, which is incorporated herein by reference in its entirety for all purposes. The deconvoluted information may then be passed to other aspects of the platform, such as variant calling, RNA expression calling, or insight engines.

[0159] The pipeline may include an automated RNA expression caller. One example of an automated RNA expression caller is disclosed in U.S. Provisional Patent Application No. 62 / 943,712, entitled "Systems and Methods for Automating RNA Expression Calls in a Cancer Prediction Pipeline," filed December 4, 2019, and is incorporated herein by reference in its entirety for all purposes.

[0160] In another example, the methods and systems disclosed herein may be implemented in one or more microservices running on a platform. In another example, one or more of such microservices may be part of an order management system in the platform that orchestrates the sequence of events required to perform the disclosed methods at the appropriate time and in the appropriate order of events required to perform gene sequencing, such as sequencing a patient's tumor or normal tissue for precision medicine deliverables to cancer patients. In another example, a bioinformatics microservice may include one or more sub-microservices for provisioning and executing various stages of a bioinformatics pipeline. One such stage of a bioinformatics pipeline includes the methods and systems described herein. A microservice-based order management system is disclosed, for example, in U.S. Provisional Patent Application No. 62 / 873,693, entitled "Adaptive Order Fulfillment and Tracking Methods and Systems," filed July 12, 2019, which is incorporated herein by reference in its entirety for all purposes.

[0161] In another embodiment where the platform includes a report generation engine, the above methods and systems may be utilized to generate a summary report of information for presentation to a physician. For example, the report may provide the physician with information regarding the activation state of cellular pathways and / or predicted protein expression levels. The report may include treatments and / or clinical trials matched based on some or all of the information. For example, the treatments may be matched according to the systems and methods disclosed in U.S. Provisional Patent Application No. 62 / 804,724, entitled "Therapeutic Suggestion Improvements Gained Through Genomic Biomarker Matching Plus Clinical History," filed February 12, 2019, which is incorporated herein by reference in its entirety for all purposes. For example, clinical trials may be matched according to the systems and methods disclosed in U.S. Provisional Patent Application No. 62 / 855,913, entitled "Systems and Methods of Clinical Trial Evaluation," filed May 31, 2019, which is incorporated herein by reference in its entirety for all purposes. The report may include a comparison of the results to a database of results for many specimens. One example of a method and system for comparing results to a database of outcomes is disclosed in U.S. Provisional Patent Application No. 62 / 786,739, entitled "A Method and Process for Predicting and Analyzing Patient Cohort Response, Progression and Survival," filed December 31, 2018, and incorporated by reference in its entirety for all purposes. The information can be further used, in some cases, in combination with similar information from additional specimens and / or clinical response information to discover biomarkers or design clinical trials.

[0162] In a third example, the above method and system may be applied to organoids developed in association with the platform. In this example, the method and system may be used to analyze gene sequencing data derived from the organoids to provide information about the activation state and / or predicted protein expression levels of cellular pathways associated with the organoids. The report may include treatments collated based on some or all of the deconvoluted information. These treatments may be tested on the organoids, derivatives of the organoids, and / or similar organoids to determine the sensitivity of the organoids to those treatments. For example, organoids can be cultured and examined according to the systems and methods disclosed in U.S. Provisional Patent Application No. 16 / 693,117, entitled "Tumor Organoid Culture Compositions, Systems, and Methods," filed November 22, 2019, U.S. Provisional Patent Application No. 62 / 924,621, entitled "Systems and Methods for Predicting Therapeutic Sensitivity," filed October 22, 2019, and U.S. Provisional Patent Application No. 62 / 944,292, entitled "Large Scale Phenotypic Organoid Analysis," filed December 5, 2019, each of which is incorporated by reference in its entirety for all purposes.

[0163] In a fourth embodiment, the above systems and methods can be utilized in conjunction with or as part of medical devices or laboratory developed tests directed to medical care and research generally. An embodiment of a laboratory developed test, particularly one enhanced by artificial intelligence, is disclosed, for example, in U.S. Patent Application No. 62 / 924,515, entitled "Artificial Intelligence Assisted Precision Medicine Enhancements to Standardized Laboratory Diagnostic Testing," filed October 22, 2019, which is incorporated herein by reference in its entirety for all purposes.

[0164] It should be understood that the above examples are illustrative and not limiting of the use of the systems and methods described herein in combination with digital and laboratory healthcare platforms.

[0165] The feasibility of real-world data (RWD) analytics is improving with technological advances and regulatory support for continuously capturing and integrating healthcare data sources. Several studies have demonstrated the ability of RWE to guide clinical development strategies, expand product labeling, and address knowledge gaps by examining clinical aspects not captured in clinical trials.

[0166] Despite recent advances and growing regulatory support, RWD from disparate structured and unstructured sources is often challenged by a variety of technological barriers: lack of standardization across electronic records, underpowered natural language processing tools, and uncontrolled external variables can impact the validity of well-sourced RWE.

[0167] Our RWD analysis followed strict qualitative criteria to generate RWE of demographics, clinical characteristics, molecular subtypes, treatment history, and survival outcomes from a large heterogeneous database. Importantly, the results were largely consistent with data from previous clinical studies, suggesting the feasibility of generating valid RWE.

[0168] We also demonstrate the value of integrating omics data with RWD through the use of whole-transcriptome analysis and predictive models for receptor status in relevant breast cancer signaling pathways.

[0169] These data provide a rationale for the use of clinical genetic databases to generate RWE and to conduct real-time hypothesis-driven analyses of large RWD cohorts in the future. Clinicians may utilize these large databases to circumvent restrictive exclusion criteria of controlled studies, clarify real-world patient needs, and aid in the development of clinical trials. Furthermore, our results suggest that molecular data may aid in addressing deficiencies in standard breast cancer diagnostic tests. [Table 4-1] [Table 4-2] [Table 5] [Table 6]

[0170] Throughout this specification, multiple instances may implement components, operations, or structures described as a single instance. Although individual operations of one or more methods have been illustrated and described as separate operations, one or more of the individual operations may be performed simultaneously, and the operations need not be performed in the order illustrated. Structures and functions presented as separate components in an example configuration may be implemented as a combined structure or component. Similarly, structures and functions presented as a single component may be implemented as a separate component or multiple components. These and other variations, modifications, additions, and improvements are included within the scope of the subject matter of this specification.

[0171] Further, certain embodiments are described herein as including logic or a number of routines, subroutines, applications, or instructions. These may constitute either software (e.g., code embodied on a machine-readable medium or in a transmission signal) or hardware. In hardware, routines, etc., are tangible units capable of performing certain operations and may be configured or arranged in a particular way. In an exemplary embodiment, one or more computer systems (e.g., standalone, client, or server computer systems) or one or more hardware modules of a computer system (e.g., a processor or processors) may be configured as hardware modules that operate by software (e.g., an application or portion of an application) to perform certain operations described herein.

[0172] In various embodiments, a hardware module can be implemented mechanically or electronically. For example, a hardware module can include dedicated circuitry or logic (e.g., a microcontroller, a special-purpose processor such as a field programmable gate array (FPGA) or an application specific integrated circuit (ASIC)) that is permanently configured to perform certain operations. A hardware module can also include programmable logic or circuitry (e.g., contained within a general-purpose processor or other programmable processor) that is temporarily configured by software to perform certain operations. It will be appreciated that the decision as to whether a hardware module is implemented mechanically, with dedicated and permanently configured circuitry, or with temporarily configured circuitry (e.g., configured by software) can be determined based on cost and time considerations.

[0173] Thus, the term "hardware module" should be understood to encompass tangible entities, that is physically constructed, permanently configured (e.g., hardwired), or temporarily configured (e.g., programmed) to operate in a certain manner or to perform certain operations described herein. Considering embodiments in which the hardware modules are temporarily configured (e.g., programmed), each of the hardware modules need not be configured or instantiated at any one instance of time. For example, if the hardware modules include a general-purpose processor configured using software, the general-purpose processor may be configured as different hardware modules at different times. Thus, the software may configure the processor, for example, to configure a particular hardware module at one time and another hardware module at another time.

[0174] Hardware modules can provide information to and receive information from other hardware modules. Thus, the described hardware modules can be considered to be communicatively coupled. When multiple such hardware modules are present simultaneously, communication can be achieved via signal transmission (e.g., via appropriate circuits and buses) connecting the hardware modules. In embodiments in which multiple hardware modules are configured or instantiated at different times, communication between such hardware modules can be achieved, for example, via storage and retrieval of information in a memory structure accessed by the multiple hardware modules. For example, a hardware module can perform an operation and store the output of that operation in a memory device to which the hardware module is communicatively coupled. An additional hardware module can then later access the memory device to retrieve and process the stored output. Hardware modules can also initiate communication with input or output devices to operate on resources (e.g., collect information).

[0175] Various operations of the example methods described herein may be performed, at least in part, by one or more processors that are temporarily (e.g., by software) or permanently configured to perform the associated operations. Such processors, whether temporarily or permanently configured, may constitute processor-implemented modules that operate to perform one or more operations or functions. Modules referred to herein may, in some example embodiments, include processor-implemented modules.

[0176] Similarly, methods or routines described herein may be at least partially processor-implemented. For example, at least some of the operations of a method may be performed by one or more processors or processor-implemented hardware modules. Certain performance of an operation may be distributed among one or more processors that are not only present in a single machine, but also deployed across several machines. In some example embodiments, one or more processors may be present in a single location (e.g., in a home environment, in a work environment, or as a server farm), while in other embodiments, the processors may be distributed across multiple locations.

[0177] Certain performance aspects of an operation may reside not only within a single machine, but may also be distributed among one or more processors deployed across several machines. In some exemplary embodiments, one or more processors or processor-implemented modules may reside in a single location (e.g., in a home environment, in a work environment, or as a server farm). In other exemplary embodiments, one or more processors or processor-implemented modules may be distributed across multiple locations.

[0178] Unless otherwise indicated, descriptions herein using words such as "processing," "computing," "calculating," "determining," "presenting," and "displaying" may refer to machine (e.g., computer) operations or processes that manipulate or transform data represented as physical (e.g., electronic, magnetic, or optical) quantities in one or more memories (e.g., volatile memory, nonvolatile memory, or a combination thereof), registers, or other mechanical components that receive, store, transmit, or display information.

[0179] As used herein, any reference to "one embodiment" or "an embodiment" means that a particular element, feature, structure or characteristic described in connection with an embodiment is included in at least one embodiment. The appearances of the phrase "in one embodiment" in various places in the specification are not necessarily all referring to the same embodiment.

[0180] Some embodiments may be described using the terms "coupled" and "connected," along with their derivatives. For example, some embodiments may be described using the term "coupled" to indicate that two or more elements are in direct physical or electrical contact with each other. However, the term "coupled" can also mean that two or more elements are not in direct contact with each other, but still co-operate or interact with each other. The embodiments are not limited in this context.

[0181] As used herein, the terms "comprises," "comprising," "includes," "including," "has," "having," or any other variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, article, or apparatus that includes a list of elements is not necessarily limited to only those elements, but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Furthermore, unless expressly stated to the contrary, "or" means an inclusive or, not an exclusive or. For example, a condition A or B is satisfied by any one of A being true (or present) and B being false (or absent), A being false (or absent) and B being true (or present), and both A and B being true (or present).

[0182] In addition, the use of "a" or "an" is used to describe elements and components of the embodiments herein. This is done merely for convenience and to give a general sense of the description. This description should be read to include one or at least one, and the singular also includes the plural unless it is clear that otherwise is meant.

[0183] This detailed description should be construed as merely exemplary and does not describe every possible embodiment, as describing every possible embodiment would be impractical, if not impossible. Many alternative embodiments can be implemented using either the present technology or technology developed after the filing date of this patent application. [1] A computer-implemented method for determining a molecular subtype of a cancer specimen, the method comprising: determining a pathway score for each of a plurality of predetermined biological pathways using gene expression data for a plurality of nucleic acids associated with the sample; generating a summary score for the plurality of biological pathways based on the pathway score for each biological pathway; comparing the summary score to one or more enrichment scores each associated with a given molecular subtype; and returning a determined molecular subtype based on a comparison of the summary score to the one or more enrichment scores. [2] receiving gene expression data corresponding to a plurality of available biological pathways; 2. The computer-implemented method of claim 1, further comprising: applying a pathway heuristic filter to identify a subset of the available biological pathways, the subset being the predetermined biological pathway. [3] The computer-implemented method of item 1, wherein the path heuristic filter is a path overlap filter. [4] The computer-implemented method of item 3, further comprising filtering the available biological pathways by applying the pathway overlap filter to identify and filter pathways that have 90% or more genes in common with the pathways retained in the subset. [5] The computer-implemented method of item 3, further comprising filtering the available biological pathways by applying the pathway overlap filter to identify and filter pathways that have 80% or more genes in common with the pathways retained in the subset. [6] The computer-implemented method of item 3, further comprising filtering the available biological pathways by applying the pathway overlap filter to identify and filter pathways that have 50% or more genes in common with the pathways retained in the subset. [7] The computer-implemented method of item 1, wherein the path-heuristic filter is a gene expression data filter. [8] The computer-implemented method of item 1, wherein the pathologic filter is a molecular subtype filter. [9] The computer-implemented method of item 1, wherein the pathway score for each biological pathway is a z-score.

[10] The computer-implemented method of claim 9, wherein the summary score is an average of the z-scores for the biological pathway.

[11] The computer-implemented method of item 10, further comprising scaling one or more of the z-scores prior to creating the average of the z-scores.

[12] The computer-implemented method of item 11, wherein scaling one or more of the z-scores comprises inverting the sign of the one or more z-scores.

[13] The computer-implemented method of item 12, further comprising: scaling one or more of the z-scores comprises inverting the sign of the one or more z-scores.

[14] The computer-implemented method of item 13, further inverting the signs of the z-scores having an average negative z-score in the group of positive gene expression samples and having an average positive z-score in the group of negative gene expression samples.

[15] The computer-implemented method of item 13, further reversing the sign of z-scores having negative scores below a negative threshold or positive scores above a positive threshold.

[16] The computer-implemented method of item 1, further comprising scaling the pathway scores for each of a plurality of biological pathways prior to determining the summary score.

[17] The computer-implemented method of item 1, wherein the gene expression data is RNA sequence gene expression data.

[18] The computer-implemented method of item 1, wherein the predetermined biological pathway is one or more of a hallmark pathway.

[19] The computer-implemented method of item 1, wherein the predetermined biological pathways are all one or more of the hallmark pathways.

[20] The computer-implemented method of item 1, wherein the predetermined biological pathway is one or more pathways related to estrogen signaling.

[21] The computer-implemented method of item 1, wherein the predetermined biological pathway is one or more of a downstream pathway of human epidermal growth factor receptor 2 (HER2), a downstream pathway of RAS, or a downstream pathway of mTOR.

[22] The computer-implemented method of item 1, wherein the predetermined biological pathway is one or more immune-related pathways.

[23] The computer-implemented method of item 1, wherein the predetermined biological pathway is one or more immune-related hallmark pathways.

[24] The computer-implemented method of item 1, wherein the one or more enrichment scores in (c) are determined by UMAP analysis.

[25] The computer-implemented method of item 1, wherein the determined molecular subtype is an HR+ subtype, an HR+ / HER2+ subtype, an HR- / HER2+ subtype, or a HER2- subtype.

[26] The computer-implemented method of item 1, wherein the determined molecular subtype is a triple negative subtype.

[27] The computer-implemented method of item 1, wherein the specimen is from a patient diagnosed with breast cancer.

[28] The computer-implemented method of item 1, wherein the specimen is a breast cancer specimen.

[29] A system having a memory and a processor, the memory storing instructions that, when executed, cause the processor to: determine a pathway score for each of a plurality of predefined biological pathways using gene expression data for a plurality of nucleic acids associated with the specimen; create a summary score for the plurality of biological pathways based on the pathway score for each biological pathway; compare the summary score to one or more enrichment scores each associated with a predefined molecular subtype; and return a determined molecular subtype based on the comparison of the summary score to the one or more enrichment scores.

[30] The memory storing further instructions that, when executed, cause the processor to: receiving gene expression data corresponding to a plurality of available biological pathways; 30. The system of claim 29, further comprising: applying a pathway heuristic filter to identify a subset of the available biological pathways, the subset being the predetermined biological pathway.

[31] A computer-implemented method for diagnosing HER2 status in a patient, the method comprising: obtaining a human epidermal growth factor receptor 2 (HER2) status for the specimen from analysis of an immunohistochemistry (IHC) image from a first sample from the patient; obtaining a HER2 status for the specimen from analysis of a fluorescent in situ hybridization (FISH) image from a second sample from the patient; identifying discordant HER2 status results between the HER2 status from IHC and the HER2 status from FISH; and in response to said identification of discordant HER2 status, diagnosing a HER2 status based on gene expression data from at least a third sample of said patient.

[32] The computer-implemented method of item 1, further comprising generating a HER2 mismatch status report indicating biological pathways in the gene expression data.

[33] The computer-implemented method of item 1, further comprising generating a HER2 mismatch status report comprising a display of the model of molecular subtype gene expression used to identify the mismatched HER2.

[34] The computer-implemented method of item 33, wherein the model of molecular subtype gene expression comprises a linear gene expression model.

[35] The computer-implemented method of claim 33, wherein the model of molecular subtype gene expression comprises a pathway gene expression model.

[36] The computer-implemented method of claim 35, further comprising generating the HER2 mismatch status report comprising a list of pathways identified by the pathway gene expression model.

[37] The computer-implemented method of item 33, wherein the model of molecular subtype gene expression comprises a multigene linear regression gene expression model.

[38] The computer-implemented method of item 37, further comprising generating the HER2 mismatch status report comprising a list of genes identified by the multigene linear regression gene expression model.

[39] The computer-implemented method of item 1, further comprising generating a HER2 mismatch status report indicating the molecular subtype determined from the gene expression data.

[40] The computer-implemented method of item 39, wherein the molecular subtype is an HR+ subtype, an HR+ / HER2+ subtype, an HR- / HER2+ subtype, or a HER2- subtype.

[41] The computer-implemented method of item 39, wherein the molecular subtype is ER+, ER-, PR+, or PR-.

[42] The computer-implemented method of item 39, wherein the molecular subtype is a triple negative subtype.

[43] The computer-implemented method of item 1, further comprising adjusting a therapeutic treatment protocol based on the pattern in the gene expression data.

[44] diagnosing HER2 status based on gene expression data from at least a third sample from the patient; determining a pathway score for each of a plurality of biological pathways in the gene expression data; generating a summary score for the plurality of biological pathways based on the pathway score for each biological pathway; comparing the summary score to one or more enrichment scores each associated with a given molecular subtype; 2. The computer-implemented method of claim 1, wherein diagnosing a HER2 status based on gene expression data from at least a third sample of the patient comprises determining a molecular subtype of the gene expression data that corresponds to the HER2 status based on a comparison of the summary score to the one or more enrichment scores.

[45] The computer-implemented method of item 1, wherein obtaining the HER2 status from analysis of the immunohistochemistry (IHC) image comprises applying the IHC image to a trained IHC classification model trained on histopathological slide image data to classify HER2 status, and obtaining the HER2 status from analysis of the FISH image comprises applying the IHC image to a trained FISH classification model trained on histopathological slide image data to classify HER2 status.

[46] The computer-implemented method of item 45, wherein the trained IHC classification model and the trained FISH classification model are convolutional neural networks.

[47] The computer-implemented method of item 45, further comprising, in response to identifying a discordant HER2 status result between the HER2 status from IHC and the HER2 status from FISH and diagnosing a HER2 status based on gene expression data from at least a third sample of the patient, providing the discordant HER2 status result and the diagnosed HER2 status to a hybrid classification model based on the at least gene expression data.

[48] ​​The computer-implemented method of item 47, wherein the hybrid classification model comprises a convolutional neural network.

[49] The computer-implemented method of item 1, wherein the first sample, the second sample, and the third sample are from a single biopsy block.

[50] The computer-implemented method of item 49, wherein the first sample, the second sample, and the third sample are different slices from a biopsy block.

[51] A system having a memory and a processor, the memory storing instructions that, when executed, cause the processor to: obtaining a human epidermal growth factor receptor 2 (HER2) status for the specimen from analysis of an immunohistochemistry (IHC) image from a first sample from the patient; obtaining a HER2 status for the specimen from analysis of a fluorescent in situ hybridization (FISH) image from a second sample from the patient; identifying discordant HER2 status results between the HER2 status from IHC and the HER2 status from FISH; and in response to said identification of discordant HER2 status, diagnosing a HER2 status based on gene expression data from at least a third sample of said patient.

[52] The system of claim 51, wherein the memory stores further instructions that, when executed, cause the processor to generate a HER2 mismatch status report indicating biological pathways in the gene expression data.

[53] The system of claim 51, wherein the memory stores further instructions that, when executed, cause the processor to generate a HER2 mismatch status report including a representation of the model of molecular subtype gene expression used to identify the mismatched HER2.

[54] The system of item 21, wherein the model of molecular subtype gene expression comprises a linear gene expression model.

[55] The system of item 21, wherein the model of molecular subtype gene expression comprises a pathway gene expression model.

[56] The system of claim 21, wherein the memory stores further instructions that, when executed, cause the processor to generate the HER2 mismatch status report, the report including a list of pathways identified by the pathway gene expression model.

[57] The system of item 21, wherein the model of molecular subtype gene expression comprises a multigene linear regression gene expression model.

[58] The system of claim 21, wherein the memory stores further instructions that, when executed, cause the processor to generate the HER2 mismatch status report comprising a list of genes identified by the multigene linear regression gene expression model.

[59] The system of claim 21, wherein the memory stores further instructions that, when executed, cause the processor to generate a HER2 mismatch status report indicating the molecular subtype determined from the gene expression data.

Claims

1. 1. A computer-implemented method for diagnosing HER2 status in a subject, the method comprising: obtaining a human epidermal growth factor receptor 2 (HER2) status for the specimen from a labeled immunohistochemistry (IHC) image from the subject sample; obtaining a HER2 status for the specimen from analysis of a labeled fluorescent in situ hybridization (FISH) image from the sample; identifying discordant HER2 status results between the HER2 status marked in an IHC image and the HER2 status marked in a FISH image; responsive to said identification of said discordant HER2 status, obtaining gene expression data corresponding to said sample, and diagnosing a HER2 status based on at least said obtained gene expression data using a molecular subtype model trained on the gene expression data. The method comprising:

2. 10. The computer-implemented method of claim 1, further comprising generating a HER2 discrepancy status report comprising a representation of the molecular subtype model used to identify the HER2 status.

3. The computer-implemented method of claim 2 , wherein the molecular subtype model comprises a linear regression model or a pathway analysis model.

4. 3. The computer-implemented method of claim 2, wherein the molecular subtype model comprises a pathway analysis model, the method further comprising generating the HER2 discordance status report comprising a list of pathways identified by the pathway analysis model.

5. The molecular subtype model comprises a multi-gene linear regression gene expression model, and the method comprises:

3. The computer-implemented method of claim 2, further comprising generating the HER2 discordance status report comprising a list of genes identified by the multi-gene linear regression gene expression model.

6. 2. The computer-implemented method of claim 1, further comprising generating a HER2 discordance status report indicating a molecular subtype determined from the gene expression data and / or indicating biological pathways in the gene expression data.

7. 7. The computer-implemented method of claim 6, wherein the molecular subtype is an HR+ subtype, an HR+ / HER2+ subtype, an HR- / HER2+ subtype, or an HER2- subtype, an ER+, ER-, PR+, PR-, or a triple negative subtype.

8. The computer-implemented method of claim 1 , further comprising adjusting a therapeutic treatment protocol based on patterns in the gene expression data.

9. diagnosing HER2 status based at least on gene expression data; determining a pathway score for each of a plurality of biological pathways in the gene expression data; generating a summary score for the plurality of biological pathways based on the pathway score for each biological pathway; comparing the summary score to one or more enrichment scores each associated with a given molecular subtype; 2. The computer-implemented method of claim 1, wherein diagnosing the HER2 status based on at least the gene expression data comprises determining a molecular subtype of the gene expression data that corresponds to the HER2 status based on a comparison of the summary score to the one or more enrichment scores.

10. 10. The computer-implemented method of claim 1, wherein the molecular subtype model trained on gene expression data is a machine learning model.

11. 2. The computer-implemented method of claim 1, wherein the sample comprises a first slice portion and a second slice portion, each obtained from a biopsy block, and wherein the labeled IHC image is an image of the first slice portion and the labeled FISH image is from the second slice portion.

12. The computer-implemented method of claim 1 , wherein obtaining gene expression data corresponding to the sample comprises obtaining gene expression data from a second sample obtained from the subject.

13. 1. A system having a memory and a processor, the memory storing instructions that, when executed, cause the processor to: obtaining a human epidermal growth factor receptor 2 (HER2) status for the specimen from a labeled immunohistochemistry (IHC) image from a sample of the subject; obtaining a HER2 status for a specimen from a labeled fluorescent in situ hybridization (FISH) image from the sample of the subject; identifying discordant HER2 status results between the HER2 status marked in the IHC image and the HER2 status marked in the FISH image; and in response to said identification of discordant HER2 status, obtaining gene expression data corresponding to said sample, and diagnosing the HER2 status based at least on the gene expression data using a molecular subtype model trained based on the gene expression data.

14. 14. The system of claim 13, wherein the memory stores further instructions that, when executed, cause the processor to generate a HER2 mismatch status report indicating biological pathways in the gene expression data.

15. 14. The system of claim 13, wherein the memory stores further instructions that, when executed, cause the processor to generate a HER2 mismatch status report including an indication of the molecular subtype model used to identify the mismatched HER2.

16. The system of claim 13 , wherein the molecular subtype model comprises a linear regression model or a pathway analysis model.

17. the molecular subtype model comprises a pathway analysis model; 14. The system of claim 13, wherein the memory stores further instructions that, when executed, cause the processor to generate a HER2 mismatch status report including a list of pathways identified by the pathway analysis model.

18. 14. The system of claim 13, wherein the molecular subtype model comprises a multigene linear regression gene expression model, and the memory storing further instructions that, when executed, cause the processor to generate a HER2 discordance status report comprising a list of genes identified by the multigene linear regression gene expression model.

19. The memory may store further instructions that, when executed, cause the processor to: determining a pathway score for each of a plurality of biological pathways in the gene expression data; generating a summary score for the plurality of biological pathways based on the pathway score for each biological pathway; and comparing the summary score to one or more enrichment scores, each associated with a predetermined molecular subtype; 14. The system of claim 13, wherein the instructions for diagnosing HER2 status based on at least the gene expression data comprise instructions that, when executed, cause the processor to determine a molecular subtype of the gene expression data as corresponding to the HER2 status based on a comparison of the summary score to the one or more enrichment scores.

20. The system of claim 13 , wherein the molecular subtype model trained on the gene expression data is a machine learning model.

Images