System and method for cancer treatment decisions using deep learning

Abstract

The present disclosure provides a method and system for classifying and / or monitoring a subject's cancer. The method for assessing a subject's cancer may include obtaining a dataset including image data and / or tabular data from the subject, and processing the data with one or more trained algorithms to classify the subject's cancer. Based on the result of the classification, the subject's cancer may be assessed. The assessment may include determining biomarkers that predict response to a therapeutic intervention to treat the subject's cancer.

Description

[Technical field]

[0001] cross reference

[0001] This application claims the benefit of U.S. Provisional Patent Application No. 63 / 313,505, filed February 24, 2022, which is incorporated by reference in its entirety herein. [Background technology]

[0002] Prostate cancer is the leading cause of cancer death in men. Nevertheless, international standards for prognostication of patient outcome rely on non-specific and insensitive tools that usually lead to overtreatment or undertreatment. Summary of the Invention

[0003]

[0003] Determining a patient's optimal cancer treatment is a challenging task, and oncologists must select the treatment with the highest probability of success and the lowest potential for toxicity. The difficulty of treatment selection is rooted in the enormous molecular, phenotypic, and prognostic heterogeneity that cancers exhibit. Herein, the need for accurate, globally scalable tools to aid in the personalization of cancer treatment is recognized. [Means for solving the problem]

[0004]

[0004] Radiation therapy is a common form of definitive therapy for the treatment of localized prostate cancer with curative intent. The addition of androgen deprivation therapy (ADT) to radiation therapy has been demonstrated to improve oncological outcomes. However, ADT has well-known toxicities, such as hot flashes, decreased libido and erectile function, loss of muscle mass, increased body fat, decreased bone density, and potential adverse effects on heart and brain health.

[0005]

[0005] Despite the benefits of ADT, the majority of men with localized prostate cancer treated with radiotherapy alone without ADT do not develop distant metastases. Unfortunately, there are no predictive biomarkers yet to identify which men will specifically benefit from ADT combined with radiotherapy, so current guidelines recommend the use of ADT based on estimated prognosis using National Comprehensive Cancer Network (NCCN) risk groups. Thus, there is a significant unmet need to guide the use of radiotherapy and ADT in men with localized prostate cancer.

[0006]

[0006] Thus, disclosed herein are methods and systems for identifying or monitoring cancer-related conditions by processing a biological sample obtained or derived from a subject, e.g., a cancer patient. The biological sample (e.g., a tissue sample) obtained from the subject can be analyzed to prognose clinical outcomes, which may include, e.g., distant metastasis, biochemical recurrence, death, progression-free survival, and overall survival. The biological sample obtained from the subject may be analyzed to predict response to treatment and guide treatment decisions.

[0007] In one aspect, the disclosure provides a method for assessing cancer in a subject, the method comprising: (a) obtaining a dataset including at least image data obtained or derived from the subject; (b) processing the dataset with a trained algorithm to classify the dataset into one of a plurality of categories, the classifying step comprising applying an image processing algorithm to the image data; and (c) assessing cancer based at least in part on the classification of the dataset into categories, the assessing step comprising determining biomarkers predictive of response to a therapeutic intervention for treating cancer in the subject.

[0008]

[0008] In some embodiments, the response comprises overall survival. In some embodiments, the response comprises progression-free survival. In some embodiments, the response comprises a reduction in mortality. In some embodiments, the response comprises a reduction in prostate cancer-specific mortality. In some embodiments, the response comprises metastasis-free survival. In some embodiments, the response comprises a reduction in metastases. In some embodiments, the response comprises a reduction in distant metastases. In some embodiments, the response comprises a reduction in distant metastases after 5 years.

[0009]

[0009] In some embodiments, the method further includes a step of determining whether the subject is biomarker positive or biomarker negative for the biomarker, where the step of determining whether the subject is biomarker positive or biomarker negative includes: (i) calculating a first probability that the subject will exhibit a response in the presence of the therapeutic intervention; (ii) calculating a second probability that the subject will exhibit a response in the absence of the therapeutic intervention; (iii) calculating a probability delta between the first probability and the second probability; and (iv) comparing the probability delta to a reference standard.

[0010]

[0010] In some embodiments, the method further comprises determining whether the subject is biomarker positive or biomarker negative for the biomarker, the determining whether the subject is biomarker positive or biomarker negative comprises: (i) calculating a first probability that the subject will respond in the presence of the therapeutic intervention and in the presence of radiation therapy (RT); (ii) calculating a second probability that the subject will respond in the absence of the therapeutic intervention and in the presence of radiation therapy (RT); (iii) calculating a probability delta between the first probability and the second probability; and (iv) comparing the probability delta to a reference standard. In some embodiments, if the probability delta is higher than the reference standard, the subject is biomarker positive, and if the probability delta is lower than the reference standard, the subject is biomarker negative. In some embodiments, the reference standard is determined at least in part by measuring a median probability delta from a plurality of subjects.

[0011]

[0011] In some embodiments, the method further comprises: (i) calculating a first probability that the subject will respond in the presence of the therapeutic intervention and in the presence of radiation therapy (RT); (ii) calculating a second probability that the subject will respond in the absence of the therapeutic intervention and in the presence of radiation therapy (RT); (iii) calculating a probability delta between the first and second probabilities; and (iv) comparing the probability delta to a reference standard, where if the probability delta is higher than the reference standard, the subject is biomarker positive, and if the probability delta is lower than the reference standard, the subject is biomarker negative. In some embodiments, if the probability delta is higher than the reference standard, the subject is biomarker positive, and if the probability delta is lower than the reference standard, the subject is biomarker negative. In some embodiments, the reference standard is determined at least in part by measuring the median probability delta from a plurality of subjects.

[0012] In some embodiments, the biomarker positive subject is a candidate for a therapeutic intervention. In some embodiments, the method further comprises treating the subject with a therapeutic intervention. In some embodiments, the therapeutic intervention comprises androgen deprivation therapy (ADT). In some embodiments, the ADT is short-term ADT (ST-ADT). In some embodiments, the trained algorithm is trained using self-supervised learning. In some embodiments, the trained algorithm comprises a deep learning algorithm. In some embodiments, the dataset further comprises tabulated data. In some embodiments, the trained algorithm comprises a first trained algorithm that processes image data and a second trained algorithm that processes tabulated data. In some embodiments, the trained algorithm further comprises a third trained algorithm that processes an output of the first and second trained algorithms. In some embodiments, the tabulated data comprises clinical data of the subject. In some embodiments, the clinical data comprises laboratory data, a therapeutic intervention, or a long-term outcome. In some embodiments, the cancer comprises prostate cancer, bladder cancer, breast cancer, pancreatic cancer, or thyroid cancer. In some embodiments, the cancer comprises prostate cancer. In some embodiments, the image data comprises digital histopathology data. In some embodiments, the data comprises images derived from a biopsy sample of the subject. In some embodiments, the images are obtained by microscopic examination of the biopsy sample. In some embodiments, the method further comprises processing the image data using image segmentation, image stitching, or object detection algorithms. In some embodiments, the method further comprises extracting features from the image data.

[0013]

[0013] Another aspect of the present disclosure provides a non-transitory computer-readable medium containing machine-executable code that, when executed by one or more computer processors, performs any of the methods described above or elsewhere in this specification.

[0014] Another aspect of the present disclosure provides a system including one or more computer processors and a computer memory coupled thereto, the computer memory including machine-executable code that, when executed by the one or more computer processors, performs any of the methods described above or elsewhere herein.

[0015]

[0015] Further aspects and advantages of the present disclosure will become readily apparent to those skilled in the art from the following detailed description, in which only illustrative embodiments of the present disclosure are shown and described. The present disclosure is capable of other and different embodiments, and its several details are capable of modification in various obvious respects, all without departing from the present disclosure. Accordingly, the drawings and description should be regarded as illustrative in nature, and not as restrictive. Incorporation by Reference

[0016]

[0016] All publications, patents, and patent applications mentioned in this specification are incorporated by reference herein to the same extent as if each individual publication, patent, or patent application was specifically and individually indicated to be incorporated by reference. To the extent that the publications and patents or patent applications incorporated by reference conflict with the disclosure contained herein, the specification is intended to supersede and / or take precedence over such conflicting material.

[0017]

[0017] The novel features of the invention are set forth with particularity in the appended claims. A better understanding of the features and advantages of the present invention will be obtained by reference to the following detailed description that sets forth illustrative embodiments, in which the principles of the invention are utilized, and the accompanying drawings (also referred to herein as "drawings" and "figures") in which: [Brief description of the drawings]

[0018] [Figure 1]

[0018] FIG. 1 is a diagram of a computer system programmed or otherwise configured to implement the methods provided herein. [Figure 2A]

[0019] FIG. 1 is a diagram of an example of a multimodal deep learning system and dataset, showing that the multimodal architecture consists of three parts: a tower stack that analyzes tabular clinical data, a tower stack that analyzes a variable number of digital histopathology slides, and a third tower stack that merges the resulting features and predicts binary outcomes. [Figure 2B] FIG. 1 is a diagram of an example multi-modal deep learning system and dataset for training a self-supervised model of image tower stacks. [Figure 2C] 2A-2C are diagrams of an example of a multimodal deep learning system and dataset, and are schematic representations of a multimodal AI (MMAI) system. The exemplary MMAI system shown in FIG. 2C accepts both tabular (e.g., clinical) and image (e.g., histopathology) data and outputs a delta score that characterizes the magnitude of treatment benefit to a patient. [Figure 3A]

[0020] An example of a comparison of a deep learning system with established clinical guidelines across outcomes from a clinical trial is shown. Performance results are reported using time-dependent receiver operating characteristics for area under the curve (AUC) of sensitivity and specificity for the MMAI model (blue bars) versus the NCCN model (grey bars). Comparisons are made at 5 and 10 years for binary outcomes: distant metastasis (DM), biochemical recurrence (BCR), prostate cancer-specific survival (PCaSS), and overall survival (OS). [Figure 3B] An example of a comparison of a deep learning system with established clinical guidelines is shown in the table, across outcomes from clinical trials. A summary table of the relative improvement of the AI ​​model over the NCCN model across various outcomes, split by performance on data from each trial in the test set. The relative improvement is given as (PAI-PNCCN) / PNCCN, where P is the performance of the model. [Figure 3C]An example of a comparison of a deep learning system to established clinical guidelines across outcomes from a clinical trial is shown. Results from an ablation study showing model performance when trained on a decreasing set of data inputs. NCCN refers to three variables: combined Gleason, baseline PSA, and t-stage. NCCN+3 refers to NCCN plus Gleason primary, Gleason secondary, and age. Path refers to digital histopathology images. [Figure 3D] 3A-3H show an example of a comparison of a deep learning system with established clinical guidelines across outcomes from clinical trials, with performance comparisons for individual clinical trial subsets of the test set, and Figures 3D-3H including the entire test set shown in Figure 3A. [Figure 3E] 3A-3H show an example of a comparison of a deep learning system with established clinical guidelines across outcomes from clinical trials, with performance comparisons for individual clinical trial subsets of the test set, and Figures 3D-3H including the entire test set shown in Figure 3A. [Figure 3F] 3A-3H show an example of a comparison of a deep learning system with established clinical guidelines across outcomes from clinical trials, with performance comparisons for individual clinical trial subsets of the test set, and Figures 3D-3H including the entire test set shown in Figure 3A. [Figure 3G] 3A-3H show an example of a comparison of a deep learning system with established clinical guidelines across outcomes from clinical trials, with performance comparisons for individual clinical trial subsets of the test set, and Figures 3D-3H including the entire test set shown in Figure 3A. [Figure 3H]3A-3H show an example of a comparison of a deep learning system with established clinical guidelines across outcomes from clinical trials, with performance comparisons for individual clinical trial subsets of the test set, and Figures 3D-3H including the entire test set shown in Figure 3A. [Figure 4]

[0021] Figure 1 shows an example of a pathologist's interpretation of SSL tissue clusters. A self-supervised model in a multimodal model is trained to identify whether an enhanced version of a small patch of tissue comes from the same original patch, without seeing the labels in the clinical data. Once trained, each image patch in a dataset of 10.05M image patches is passed through the model to extract a 128-dimensional feature vector, which is then clustered and visualized using the UMAP algorithm31. The pathologist is then asked to interpret the 20 image patches closest to each of the 25 cluster centers, with explanatory text provided beside the inset. For clarity, only 6 clusters (in color) are highlighted, while the remaining clusters are shown in gray. See Figure 7 for the complete pathologist annotation. [Diagram 5]

[0022] Figure 1 shows an example image quilt for four exemplary patients. The dataset includes patients with a variable number of histopathology slides. To standardize the image input to the model, we segment the tissue from each slide and stitch all tissues into a single square image of 51200x51200 pixels, divided into 200x200 patches, to represent all the histopathology data for one patient. Image quilts for four patients are shown here. [Figure 6]

[0023] Figure 1 shows an example of kernel density sampling of example image patches. The brown boxes indicate the detection of kernels used to calculate kernel density, and oversample the patches input to the self-supervised training protocol according to kernel density. Each patch is partitioned into deciles according to density, and each decile is oversampled so that the MMAI model sees the same total number of images from each decile. [Figure 7-1]

[0024] Figure 1 shows an example of pathologist-interpreted patch clusters. Using UMAP, 25 clusters have been generated from the SSL features of all pathological tissue patches in study RTOG-9202. Each row in the image corresponds to the 25 image patches closest to the cluster center point. These were then inspected by a pathologist to determine a human-interpretable description of the clusters listed in the table. [Figure 7-2] Same as above. [Figure 8]

[0025] FIG. 1 shows an example of an NCCN model algorithm. The rules-based algorithm is based on D'Amico risk groups and models the guidelines published annually by the National Cancer Center Network. [Figure 9]

[0026] Table showing statistics from processed clinical trial data, with the first 5 columns showing statistics for each trial. Column "combined" shows statistics for the final dataset using all 5 trials for training and validation. ***RTOG9413 will randomize patients in a 2x2 fashion to examine the impact of hormonal therapy timing (pre-RT vs. RT initiation) and field size (prostate-only vs. complete pelvic RT). New abbreviations will be used: disease-free survival (DFS), progression-free survival (PFS), and prostate cancer-specific mortality (PCSM). [Figure 10]

[0027] CONSORT flow diagram for processing the clinical dataset. ST-ADT = short-term androgen deprivation therapy; RT = radiation therapy. [Figure 11]

[0028] Figure 1 shows the relative feature importance of various tabular (e.g., clinical) and image (e.g., histopathology) data features as determined using the trained algorithm. Feature importance was calculated based on the average of the Shapely absolute value of each variable and normalized between features. Image importance was measured to be 37.3%, followed by Gleason Primary: 35.8%, Gleeson Secondary: 9.0%, Gleason Combined: 5.6%, T-Stage: 5.5%, Age: 3.5%, and Baseline PSA: 3.4%. [Figure 12]

[0029] Figure 1 shows the distribution of delta scores for the development cohort (left) and validation cohort (right) determined using the trained algorithm, with the vertical line indicating the 67th percentile chosen as the boundary between "biomarker positive" and "biomarker negative" subjects. [Figure 13]

[0030] FIG. 1 shows cumulative incidence curves of distant metastasis by AI biomarker subgroups ("biomarker positive" and "biomarker negative") as determined using the trained algorithm. [Figure 14]

[0031] Forest plots for distant metastasis (DM) and prostate cancer specific mortality (PCSM) across positive and negative biomarker groups. [Figure 15]

[0032] Forest plots of distant metastasis (DM) and prostate cancer-specific mortality (PCSM) in positive and negative biomarker groups for the National Comprehensive Cancer Network (NCCN) low-to-intermediate risk patient subgroup. [Figure 16]

[0033] FIG. 1 shows cumulative incidence curves of prostate cancer-specific mortality (PCSM) by AI biomarker subgroups ("biomarker positive" and "biomarker negative") as determined using the trained algorithm. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0019]

[0034] While various embodiments of the present invention have been shown and described herein, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Numerous variations, changes, and substitutions will occur to those skilled in the art without departing from the invention. It should be understood that various alternatives to the embodiments of the invention described herein may be employed.

[0020]

[0035] As used in this specification and claims, the singular forms "a," "an," and "the" include plural references unless the context clearly dictates otherwise. For example, the term "nucleic acid" includes a plurality of nucleic acids, including mixtures thereof.

[0021]

[0036] As used herein, the term "subject" generally refers to an entity or medium having genetic information that can be tested or detected. The subject can be a human, an individual, or a patient. The subject can be a vertebrate, such as a mammal. Non-limiting examples of mammals include humans, monkeys, farm animals, sport animals, rodents, pets, and the like. The subject can be a male subject. The subject can be a female subject. The subject may exhibit symptoms indicative of the subject's health or physiological state or condition, such as the subject's cancer-related health or physiological state or condition. Alternatively, the subject may be asymptomatic with respect to such health or physiological state or condition. The subject may be suspected of having a health or physiological state or condition. The subject may be at risk for developing a health or physiological state or condition. The health or physiological condition may correspond to a disease (e.g., cancer). The subject may be an individual diagnosed with a disease. The subject may be an individual at risk for developing a disease.

[0022]

[0037] As used herein, "cancer diagnosis," "diagnosing cancer," and related or derived terms include identifying cancer in a subject, determining the aggressiveness of cancer, or determining the stage of cancer.

[0023]

[0038] As used herein, "cancer prognosis," "determining a cancer prognosis," and related or derived terms include predicting a patient's clinical outcome, assessing the risk of cancer recurrence, determining treatment modality, or determining treatment efficacy.

[0024]

[0039] As used herein, the term "nucleic acid" generally refers to a polymer of nucleotides of any length, deoxyribonucleotides (dNTPs) or ribonucleotides (rNTPs), or analogs thereof. Nucleic acids may have any three-dimensional structure and may perform any function, known or unknown. Non-limiting examples of nucleic acids include deoxyribonucleic acid (DNA), ribonucleic acid (RNA), coding or non-coding regions of a gene or gene fragment, loci defined from linkage analysis, exons, introns, messenger RNA (mRNA), transfer RNA, ribosomal RNA, short interfering RNA (siRNA), short hairpin RNA (shRNA), microRNA (miRNA), ribozymes, cDNA, recombinant nucleic acids, branched nucleic acids, plasmids, vectors, isolated DNA of any sequence, isolated RNA of any sequence, nucleic acid probes, and primers. Nucleic acids may contain one or more modified nucleotides, such as methylated nucleotides and nucleotide analogs. If present, modifications to the nucleotide structure may occur before or after assembly of the nucleic acid. The sequence of nucleotides of a nucleic acid may be interrupted by non-nucleotide components. Nucleic acids can be further modified after polymerization, such as by conjugation or conjugation with a reporter agent.

[0025]

[0040] As used herein, the term "target nucleic acid" generally refers to a nucleic acid molecule in a starting population of nucleic acid molecules having a nucleotide sequence, the presence, amount, and / or sequence, or one or more changes therein, of which it is desired to determine. The target nucleic acid may be any type of nucleic acid, including DNA, RNA, and analogs thereof. As used herein, "target ribonucleic acid (RNA)" generally refers to a target nucleic acid that is RNA. As used herein, "target deoxyribonucleic acid (DNA)" generally refers to a target nucleic acid that is DNA.

[0026]

[0041] As used herein, the terms "amplifying" and "amplification" generally refer to increasing the size or amount of a nucleic acid molecule. The nucleic acid molecule may be single-stranded or double-stranded. Amplification may include generating one or more copies of the nucleic acid molecule or "amplification products." Amplification may be performed, for example, by extension (e.g., primer extension) or ligation. Amplification may include performing a primer extension reaction to generate a strand complementary to the single-stranded nucleic acid molecule, and optionally generating one or more copies of the strand and / or the single-stranded nucleic acid molecule. The term "DNA amplification" generally refers to generating one or more copies of a DNA molecule or "amplified DNA product." The term "reverse transcription amplification" generally refers to generating deoxyribonucleic acid (DNA) from a ribonucleic acid (RNA) template by the action of reverse transcriptase.

[0027] <Embodiments of the present disclosure>

[0028]

[0042] Despite its high prevalence, accurate, sensitive and specific diagnosis of prostate cancer remains challenging. Although prostate cancer is often indolent and may be cured by treatment, the adverse effects of overtreatment and undertreatment make it the leading cause of cancer-related disability worldwide and remains one of the leading causes of cancer deaths in men. Determining the optimal course of treatment for patients with prostate cancer is a challenging medical task that requires consideration of the patient's overall health, characteristics of the patient's cancer, the side-effect profile of many possible treatments, outcome data from clinical trials involving similarly diagnosed patients, and prognosticating the expected future outcome of the patient at hand. This difficulty is exacerbated by the lack of readily accessible prognostic tools to better risk stratify patients.

[0029]

[0043] Artificial intelligence (AI) is enabling insights to be gained from huge data sets that were previously difficult to interpret. Whereas standard risk stratification tools are fixed and based on a small number of variables, AI can learn from large amounts of minimally processed data across a range of modalities. AI systems are low-cost, massively scalable, and can improve over time through use.

[0030]

[0044] There is a strong need for accurate, globally scalable tools to aid in the personalization of treatment. The methods and systems disclosed herein demonstrate the personalization of prostate cancer treatment by predicting the impact of specific treatments on long-term, clinically relevant outcomes (e.g., distant metastasis, biochemical recurrence, partial response, complete response, death, relative survival, cancer-specific survival, progression-free survival, disease-free survival, 5-year survival, and overall survival) using novel multimodal deep learning models trained on digital histopathology and clinical data from prostate biopsies.

[0031]

[0045] Thus, the present disclosure provides methods, systems, and kits for identifying or monitoring cancer-related categories and / or conditions by processing a biological sample obtained from or derived from a subject (e.g., a male patient suffering from or suspected of suffering from prostate cancer). A biological sample obtained from a subject (e.g., a prostate biopsy sample) can be analyzed to identify a cancer-related category (which can include, for example, estimating the presence or absence of a cancer-related category, or a quantitative assessment (e.g., risk, predicted outcome), or estimating a predicted or observed response to a therapeutic intervention). Such subjects may include subjects with one or more cancer-related categories and subjects without a cancer-related category. Cancer-related categories or conditions may include, for example, cancer positive, cancer negative, cancer stage, observed response to cancer treatment (e.g., radiation therapy, chemotherapy, surgical intervention), observed long-term outcome (e.g., disease metastasis, biochemical recurrence, partial response, complete response, relative survival, cancer-specific survival, progression-free survival, disease-free survival, 5-year survival, or overall survival), predicted response to cancer treatment, and / or predicted long-term outcome. Assay of biological samples

[0032]

[0046] Biological samples can be obtained from or derived from human subjects (e.g., male subjects). Biological samples can be stored before processing at various temperatures (e.g., room temperature, under refrigerated or freezer conditions, 25°C, 4°C, -18°C, -20°C, or -80°C), or in various storage conditions, such as various suspensions (e.g., formalin, EDTA collection tube, cell-free RNA collection tube, or cell-free DNA collection tube). Biological samples can be obtained from subjects with or suspected of having cancer (e.g., prostate cancer), or from subjects without or not suspected of having cancer.

[0033]

[0047] The biological sample can be used to diagnose, detect or identify a disease or health or physiological condition in a subject by analyzing the biological sample. The biological sample or a portion thereof can be analyzed to determine the likelihood that the sample is positive for a disease or health condition (e.g., prostate cancer). Alternatively, or in addition, the methods described herein can include diagnosing a subject with a disease or health condition, monitoring a disease or health condition in a subject, and / or determining a subject's propensity for a health disease / condition. In some embodiments, the biological sample can be used to classify the sample and / or subject into a cancer-related category and / or to identify a subject as having a particular cancer-related condition. The cancer-related category or condition can include a diagnosis (e.g., positive or negative for cancer), a particular type of cancer (e.g., prostate cancer), a stage of cancer, a predicted outcome or prognosis, a predicted response to one or more treatments, or a combination thereof.

[0034]

[0048] The measurable substance can be a source of a sample. The substance can be a fluid, such as a biological fluid. Fluid substances can include blood (e.g., whole blood, plasma, serum), umbilical cord blood, saliva, urine, sweat, serum, semen, vaginal fluid, gastric and digestive fluids, cerebrospinal fluid, placental fluid, cavity fluid, ocular fluid, serum, breast milk, lymphatic fluid, or combinations thereof.

[0035]

[0049] The material may be a solid, for example a biological tissue. The material may include normal healthy tissue. The tissue may be associated with various types of organs. Non-limiting examples of organs may include the brain, breast, liver, lung, kidney, prostate, ovary, spleen, lymph nodes (including tonsils), thyroid, pancreas, heart, skeletal muscle, intestine, larynx, esophagus, stomach, or combinations thereof.

[0036]

[0050] The material may include a tumor. The tumor may be benign (non-cancerous), pre-malignant, or malignant (cancer), or a metastasis thereof. Non-limiting examples of tumors and associated cancers include acoustic neuroma, acute lymphoblastic leukemia, acute myeloid leukemia, adenocarcinoma, adrenal cortical carcinoma, AIDS-related cancer, AIDS-related lymphoma, anal cancer, angiosarcoma, appendix cancer, astrocytoma, basal cell carcinoma, bile duct cancer, bladder cancer, bone cancer, brain tumors such as cerebellar astrocytoma, cerebral astrocytoma / malignant glioma, ependymoma, medulloblastoma, supratentorial primitive neuroectodermal tumor, visual pathway and hypothalamic glioma, breast cancer, bronchial adenoma, Burkitt's lymphoma, cancer of unknown primary, central nervous system lymphoma, bronchogenic carcinoma, cerebellar astrocytoma, cervical cancer, childhood cancer, chondrosarcoma, Chordoma, choriocarcinoma, chronic lymphocytic leukemia, chronic myelogenous leukemia, chronic myeloproliferative disorders, colon cancer, craniopharyngioma, cutaneous T-cell lymphoma, cystadenocarcinoma, detumorable small round cell tumor, germ cell cancer, endocrine cancer, endometrial cancer, endothelial sarcoma, ependymoma, epithelial cancer, esophageal cancer, Ewing's sarcoma, fibrosarcoma, germ cell tumor, gallbladder cancer, stomach cancer, gastrointestinal carcinoid tumor, gastrointestinal stromal tumor, digestive system cancer, genitourinary system cancer, glioma, hairy cell leukemia, head and neck cancer, heart cancer, blood ductal blastoma, hepatocellular (liver) carcinoma, Hodgkin's lymphoma, hypopharyngeal carcinoma, intraocular melanoma, pancreatic islet cell carcinoma, Kaposi's sarcoma, kidney cancer, laryngeal cancer, leiomyosarcoma, lip and oral cavity cancer, liposarcoma, liver cancer, lung cancer, e.g. non-small cell lung cancer and small cell lung cancer, lung carcinoma, lymphangiosarcoma, lymphangioendothelioma, lymphoma, leukemia, macroglobulinemia, bone malignant fibrous histiocytoma / osteosarcoma, medulloblastoma, medullary carcinoma, melanoma, meningioma, mesothelioma, metastatic squamous cell neck cancer of unknown primary, oral cancer, multiple endocrine neoplasia syndrome, bone Myelodysplastic syndrome, myeloid leukemia, myxosarcoma, nasal and paranasal sinus cancer, nasopharyngeal cancer, neuroblastoma, non-Hodgkin's lymphoma, non-small cell lung cancer, oligodendroglioma, oral cavity cancer, oropharyngeal cancer, osteosarcoma / malignant fibrous histiocytoma of bone, ovarian cancer, ovarian epithelial cancer, ovarian germ cell tumor, pancreatic cancer, pancreatic islet cell cancer, papillary adenocarcinoma, papillary carcinoma, paranasal sinus and nasal cancer, parathyroid cancer, penile cancer, pharyngeal cancer, pheochromocytoma, pineal astrocytoma, pineal germinoma, pituitary adenoma, pleuropulmonary blastoma, plasma cell neoplasm, primary central nervous system lymphoma, prostate cancer,Rectal cancer, renal cell carcinoma, renal pelvis and ureter transitional cell carcinoma, retinoblastoma, rhabdomyosarcoma, salivary gland cancer, sarcoma, sebaceous gland carcinoma, seminoma, skin cancer, cutaneous Merkel cell carcinoma, small intestine cancer, soft tissue sarcoma, squamous cell carcinoma, gastric cancer, sweat gland carcinoma, synovium, T-cell lymphoma, testicular tumor, pharyngeal cancer, thymoma, thymic carcinoma, thyroid cancer, trophoblastic tumor (gestational), cancer of unknown primary site, urethral cancer, uterine sarcoma, vaginal cancer, vulvar cancer, Waldenstrom's macroglobulinemia, Wilms' tumor, or combinations thereof. The tumors may be associated with various types of organs. Non-limiting examples of organs may include brain, breast, liver, lung, kidney, prostate, ovary, spleen, lymph nodes (including tonsils), thyroid, pancreas, heart, skeletal muscle, intestine, larynx, esophagus, stomach, or combinations thereof.

[0037]

[0051] The material may include a mixture of normal healthy tissue or tumor tissue. The tissue may be associated with various types of organs. Non-limiting examples of organs may include brain, breast, liver, lung, kidney, prostate, ovary, spleen, lymph nodes (including tonsils), thyroid, pancreas, heart, skeletal muscle, intestine, larynx, esophagus, stomach, or combinations thereof. In some embodiments, the tissue is associated with the prostate of a subject. In the case of biological samples (e.g., biopsy samples) that include cells and / or tissues, the biological samples may be further analyzed or assayed. In some embodiments, the biopsy samples may be fixed, processed (e.g., dehydrated), embedded, frozen, stained, and / or examined under a microscope. In some embodiments, digital slides are generated from the processed samples.

[0038]

[0052] In some embodiments, the material may include a variety of cells, including eukaryotic cells, prokaryotic cells, fungal cells, cardiac cells, lung cells, kidney cells, liver cells, pancreatic cells, germ cells, stem cells, induced pluripotent stem cells, gastrointestinal cells, blood cells, cancer cells, bacterial cells, bacterial cells isolated from a human microbiome sample, and cells circulating in human blood. In some embodiments, the material may include the contents of a cell, e.g., the contents of a single cell, or the contents of multiple cells.

[0039]

[0053] In some embodiments, a substance may include one or more markers, the presence or absence of which is indicative of some phenomenon, such as a disease, disorder, infection, or environmental exposure. A marker may be, for example, a cell, a small molecule, a macromolecule, a protein, a glycoprotein, a carbohydrate, a sugar, a polypeptide, a nucleic acid (e.g., deoxyribonucleic acid (DNA), ribonucleic acid (RNA)), a cell-free nucleic acid (e.g., cf-DNA, cf-RNA), a lipid, a cellular component, or a combination thereof.

[0040]

[0054] Biological samples can be taken before and / or after treatment of a subject with cancer. Biological samples can be obtained from a subject during treatment or treatment regimen. Multiple biological samples may be taken from a subject to monitor the effectiveness of treatment over time. Biological samples can be taken from subjects known to have or suspected to have cancer (e.g., prostate cancer). Biological samples can be taken from subjects with unexplained symptoms, such as fatigue, nausea, weight loss, aches and pains, weakness, or bleeding. Biological samples can be taken from subjects with explainable symptoms. Biological samples can be taken from subjects at risk for developing cancer due to factors such as family history, age, hypertension or prehypertension, diabetes or prediabetes, overweight or obesity, environmental exposures, lifestyle risk factors (e.g., smoking, alcohol, or drug use), or the presence of other risk factors.

[0041]

[0055] After obtaining a biological sample from a subject, the biological sample can be processed to generate a data set indicative of a disease, condition, cancer-related category, or health status of the subject. For example, a tissue sample can be subjected to a histopathological assay (e.g., microscopy including digital image acquisition such as whole slide imaging) to generate image data based on the biological sample. Alternatively, a liquid sample or markers isolated from the sample can be subjected to testing (e.g., clinical laboratory tests) to generate tabular data. In some embodiments, the sample is assayed for the presence, absence, or amount of one or more metabolites (e.g., prostate specific antigen (PSA)).

[0042] <Data type>

[0043]

[0056] The methods and systems described herein take as input one or more datasets. The one or more datasets may include tabular data and / or image data. The tabular data and / or image data may be obtained from a biological sample of a subject. In some embodiments, the data is not derived from a biological sample.

[0044]

[0057] The data may include images of tissue samples taken from a biopsy of a subject. The image data may be obtained by microscopy of the biopsy sample. The microscopy may consist of optical microscopy, virtual or digital microscopy (such as whole slide imaging (WSI)), or any suitable microscopy technique known in the art. The microscopy images may be subjected to one or more processing steps, such as filtering, segmentation, concatenation, or object detection.

[0045]

[0058] The tabular data described herein may include any non-image data related to the subject's health or condition (e.g., disease). The tabular data may include clinical data, such as laboratory data (e.g., prostate serum antigen (PSA) level), qualitative measures of cytopathology (e.g., Gleason grade, Gleason score), structured or unstructured health data (e.g., digital rectal exam results), medical image data or results (e.g., X-ray, computed tomography (CT) scan, magnetic resonance imaging (MRI) scan, positron emission tomography (PET) scan, or ultrasound results, e.g., transrectal ultrasound results), age, medical history, previous or current cancer status (e.g., remission, metastasis) or stage, current or previous therapeutic interventions, long-term outcomes, and / or National Comprehensive Cancer Network (NCCN) classification or components thereof (e.g., combined Gleason score, t-stage, baseline PSA) at one or more time points.

[0046]

[0059] In some embodiments, the therapeutic intervention may include radiation therapy (RT). In some embodiments, the therapeutic intervention may include chemotherapy. In some embodiments, the therapeutic intervention may include surgical intervention. In some embodiments, the therapeutic intervention may include immunotherapy. In some embodiments, the therapeutic intervention may include hormonal therapy. In some embodiments, the RT may include RT with a preset use of short-term androgen deprivation therapy (ST-ADT). In some embodiments, the RT may include RT with a preset use of long-term ADT (LT-ADT). In some embodiments, the RT may include RT with a preset use of dose-escalated RT (DE-RT). In some embodiments, the surgical intervention may include radical prostatectomy (RP). In some embodiments, the therapeutic intervention may include any combination of the therapeutic interventions disclosed herein. In some embodiments, the long-term outcome may include distant metastasis (DM). In some embodiments, the long-term outcome may include biochemical recurrence (BR). In some embodiments, the long-term outcome may include partial response. In some embodiments, the long-term outcome may include complete response. In some embodiments, the long term outcome may include death. In some embodiments, the long term outcome may include relative survival. In some embodiments, the long term outcome may include cancer specific survival. In some embodiments, the cancer specific survival may include prostate cancer specific survival (PCaSS). In some embodiments, the long term outcome may include progression free survival. In some embodiments, the long term outcome may include disease free survival. In some embodiments, the long term outcome may include 5 year survival. In some embodiments, the long term outcome may include overall survival (OS). In some embodiments, the long term outcome may include any combination of the long term outcomes disclosed herein.

[0047]

[0060] Data used in the methods and systems described herein may be subjected to one or more processing steps. In some embodiments, the data (e.g., image data) is subjected to image processing, image segmentation, and / or object detection processes, coded as image processing, image segmentation, or object detection algorithms. Image processing procedures may filter, translate, scale, rotate, mirror, shear, combine, compress, split, concatenate, extract features from images, and / or smooth images before downstream processing. In some embodiments, multiple images (e.g., histopathology slides) are combined to form an image quilt. The image quilt may be converted to a representation (e.g., tensor) useful for downstream processing of the image data. Image segmentation processing may partition an image into one or more segments that contain a factor or region of interest. For example, an image segmentation algorithm may process a digital histopathology slide to determine regions of tissue that are distinct from regions of blanks or artifacts. In some embodiments, the image segmentation algorithm may include machine learning or artificial intelligence algorithms. In some embodiments, image segmentation may precede image processing. In some embodiments, image processing may precede image segmentation. The object detection process may include detecting the presence or absence of a target object (e.g., a cell or a cell part, such as a nucleus). In some embodiments, object detection may precede image processing and / or image segmentation. For example, images found by an image detection algorithm to contain one or more objects of interest may be concatenated in a subsequent image processing step. Alternatively, or in addition, image processing may precede object detection and / or image segmentation. For example, raw image data may be processed (e.g., filtered) and the processed image data may be subjected to an object detection algorithm. The image data may be subjected to multiple image processing, image segmentation, and / or object detection steps in any suitable order. In one example, the image data is optionally subjected to one or more image processing steps to improve image quality.The processed images are then run through an image segmentation algorithm to detect regions of interest (e.g., regions of tissue in a set of histopathology slides). The regions of interest are then run through an object detection algorithm (e.g., an algorithm to detect nuclei in an image of tissue) and regions found to have at least one target object are linked to produce processed image data for downstream use.

[0048]

[0061] In some embodiments, the data (e.g., tabular data) may be subjected to one or more processing steps. Processing steps may include, but are not limited to, standardization, or normalization. The one or more processing steps may, for example, discard data that contains non-true values ​​or that contains very few observations. The one or more processing steps may additionally or alternatively standardize the encoding of the data values. Different input data sets may have identical parameter values ​​that are encoded differently depending on the source of the data set. For example, "900", "900.0", "904", "904.0", "-1", "-1.0", "None", "NaN" may all encode "missing" parameter values. The one or more processing steps may recognize variations in encoding for the same value and standardize the data sets to have a uniform encoding for a given parameter value. Thus, the processing steps may reduce irregularities in the input data for downstream use. The one or more data sets may normalize the parameter values. For example, numerical data may be scaled, whitened, colored, embellished, or standardized. For example, the data may be scaled or shifted to fall into a particular interval (e.g., [0,1] or [-1,1]) and / or correlated. In some embodiments, categorical data may be encoded as one-hot vectors. In some embodiments, one or more different types of tabular (e.g., numerical, categorical) data may be concatenated. In some embodiments, the data is not subjected to processing steps.

[0049]

[0062] Data may be obtained at one or more time points. In some embodiments, data is obtained at an initial time point and a later time point. The initial time point and the later time point may be separated by any suitable length of time, such as 1 hour, 1 day, 1 week, 2 weeks, 3 weeks, 4 weeks, 6 weeks, 12 weeks, 4 months, 5 months, 6 months, 7 months, 8 months, 9 months, 10 months, 11 months, 1 year, 2 years, 3 years, 4 years, 5 years, 6 years, 7 years, 8 years, 9 years, 10 years, or more. In some embodiments, data is from three or more time points. In some embodiments, data is from 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, or more time points.

[0050] <Pre-trained algorithm>

[0051]

[0063] After using one or more assays to process one or more biological samples from a subject to generate one or more datasets indicative of the subject's cancer status (e.g., one or more cancer-related categories), a trained algorithm can be used to process one or more of the datasets (e.g., visual data and / or tabular data) to determine the subject's cancer status. For example, a trained algorithm can be used to determine the presence or absence of (e.g., prostate) cancer in a subject based on imaging data and / or laboratory data. The trained algorithm can be configured to identify the cancer status with an accuracy of at least about 50%, at least about 55%, at least about 60%, at least about 65%, at least about 70%, at least about 75%, at least about 80%, at least about 85%, at least about 90%, at least about 95%, at least about 96%, at least about 97%, at least about 98%, at least about 99%, or greater than 99% for at least about 25, at least about 50, at least about 100, at least about 150, at least about 200, at least about 250, at least about 300, at least about 350, at least about 400, at least about 450, at least about 500, or greater than about 500 independent samples.

[0052]

[0064] The trained algorithm may include an unsupervised machine learning algorithm. The trained algorithm may include a supervised machine learning algorithm. The trained algorithm may include a classification and regression tree (CART) algorithm. The supervised machine learning algorithm may include, for example, a random forest, a support vector machine (SVM), a neural network, or a deep learning algorithm. The trained algorithm may include a self-supervised machine learning algorithm.

[0053]

[0065] In some embodiments, the machine learning algorithm of the method or system described herein utilizes one or more neural networks. In some cases, a neural network is a type of computational system that can learn relationships between input data sets and target data sets. A neural network can be a software representation of a human nervous system (e.g., a cognitive system) and is intended to capture the ability of "learning" and "generalization" as used by humans. In some embodiments, the machine learning algorithm includes a neural network consisting of a CNN. Non-limiting examples of structural components of the machine learning algorithm described herein include the following: CNN, recurrent neural network, dilated CNN, fully connected neural network, deep generative model, and Boltzmann machine.

[0054]

[0066] In some embodiments, a neural network is composed of a series of layers called "neurons." In some embodiments, a neural network is composed of an input layer, where data is presented, one or more internal and / or "hidden" layers, and an output layer. The neurons are connected to neurons in other layers via connections with weights, which are parameters that control the strength of the connections. The number of neurons in each layer may be related to the complexity of the problem to be solved. The minimum number of neurons required in a layer may be determined by the complexity of the problem, and the maximum number may be limited by the neural network's ability to generalize. The input neurons may receive the data presented to them and send the data to a first hidden layer through connection weights that are modified during training. The first hidden layer may process the data and send the results to the next layer through a second set of weighted connections. Each subsequent layer may "pool" the results from the previous layer into more complex relationships. In addition, whereas traditional software programs require specific instructions to be written to perform a function, a neural network is programmed by training itself with a known set of examples and being able to modify itself during (and after) training to provide a desired output, such as an output value. After training, when a neural network is presented with new input data, it is configured to generalize what it "learned" during training and apply what it learned in training to new, previously unseen input data in order to generate an output that is related to that input.

[0055]

[0067] In some embodiments, the neural network is comprised of an artificial neural network (ANN). An ANN may be a machine learning algorithm that can be trained to map an input data set to an output data set, and includes an interconnected group of nodes organized into multiple layers of nodes. For example, an ANN architecture may be comprised of at least one input layer, one or more hidden layers, and one output layer. An ANN may include any total number of layers, and any number of hidden layers, where the hidden layers act as trainable feature extractors that allow for mapping a set of input data to an output value or a set of output values. As used herein, a deep learning algorithm (such as a deep neural network (DNN)) is an ANN that includes multiple hidden layers, e.g., two or more hidden layers. Each layer of a neural network may include a number of nodes (i.e., "neurons"). A node receives inputs directly from the input data or from the output of a node of a previous layer, and performs a particular operation, e.g., a summation operation. The connections from the inputs to the nodes have weights (or weighting coefficients) associated with them. A node may add up the pairwise products of all inputs with their associated weights. The weighted sum may be offset with a bias. The output of a node or neuron may be gated using a threshold or activation function. The activation function may be a linear or nonlinear function. The activation function may be, for example, a rectified linear unit (ReLU) activation function, a leaky ReLU activation function, or other functions such as a saturated hyperbolic tangent, identity, binary step, logistic, arctan, softsign, parametric rectified linear unit, exponential linear unit, softplus, bending identity, softexponential, sinusoid, sinc, Gaussian, or sigmoid function, or a combination thereof.

[0056]

[0068] The weighting coefficients, bias values, and thresholds, or other computational parameters of a neural network may be "taught" or "learned" in a training phase that uses one or more sets of training data. For example, the parameters may be trained using input data from a training data set and gradient descent or backpropagation techniques such that the output values ​​that the ANN computes are consistent with the examples contained in the training data set.

[0057]

[0069] The number of nodes used in the input layer of an ANN or DNN may be at least about 10, 50, 100, 200, 300, 400, 500, 600, 700, 800, 900, 1000, 2000, 3000, 4000, 5000, 6000, 7000, 8000, 9000, 10,000, 20,000, 30,000, 40,000, 50,000, 60,000, 70,000, 80,000, 90,000, 100,000, or more. In other examples, the number of nodes used in the input layer may be at most about 100,000, 90,000, 80,000, 70,000, 60,000, 50,000, 40,000, 30,000, 20,000, 10,000, 9000, 8000, 7000, 6000, 5000, 4000, 3000, 2000, 1000, 900, 800, 700, 600, 500, 400, 300, 200, 100, 50, 10, or fewer. In some examples, the total number of layers (including the input layer and the output layer) used in an ANN or DNN may be at least about 3, 4, 5, 10, 15, 20, or more. In other examples, the total number of layers may be at most about 20, 15, 10, 5, 4, 3, or less.

[0058]

[0070] In some examples, the total number of learnable or trainable parameters, e.g., weighting coefficients, biases, or thresholds used in an ANN or DNN may be at least about 10, 50, 100, 200, 300, 400, 500, 600, 700, 800, 900, 1000, 2000, 3000, 4000, 5000, 6000, 7000, 8000, 9000, 10,000, 20,000, 30,000, 40,000, 50,000, 60,000, 70,000, 80,000, 90,000, 100,000, or more. In other examples, the number of learnable parameters may be at most about 100,000, 90,000, 80,000, 70,000, 60,000, 50,000, 40,000, 30,000, 20,000, 10,000, 9000, 8000, 7000, 6000, 5000, 4000, 3000, 2000, 1000, 900, 800, 700, 600, 500, 400, 300, 200, 100, 50, 10, or fewer.

[0059]

[0071] In some embodiments of the machine learning algorithms described herein, the machine learning algorithm includes a neural network, such as a deep CNN. In some embodiments where a CNN is used, the network is constructed with any number of convolutional layers, dilation layers, or fully connected layers. In some embodiments, the number of convolutional layers is between 1 and 10, and the number of dilation layers is between 0 and 10. The total number of convolutional layers (including input and output layers) may be at least about 1, 2, 3, 4, 5, 10, 15, 20, or more, and the total number of dilation layers may be at least about 1, 2, 3, 4, 5, 10, 15, 20, or more. The total number of convolutional layers may be at most about 20, 15, 10, 5, 4, 3, or less, and the total number of dilation layers may be at most about 20, 15, 10, 5, 4, 3, or less. In some embodiments, the number of convolutional layers is between 1 and 10, and the fully connected layers are between 0 and 10. The total number of convolutional layers (including input and output layers) may be at least about 1, 2, 3, 4, 5, 10, 15, 20, or more, and the total number of fully connected layers may be at least about 1, 2, 3, 4, 5, 10, 15, 20, or more. The total number of convolutional layers may be at most about 20, 15, 10, 5, 4, 3, 2, 1, or less, and the total number of fully connected layers may be at most about 20, 15, 10, 5, 4, 3, 2, 1, or less.

[0060]

[0072] In some embodiments, the machine learning algorithm comprises a neural network including a CNN, an RNN, an augmented CNN, a fully connected neural network, a deep generative model, and / or a deep restricted Boltzmann machine.

[0061]

[0073] In some embodiments, the machine learning algorithm is composed of one or more CNNs. The CNNs may be deep ANNs and feed-forward ANNs. CNNs may also be applied to the analysis of visual images. The CNNs may include an input layer, an output layer, and multiple hidden layers. The hidden layers of the CNNs may include convolutional layers, pooling layers, fully connected layers, and normalization layers. The layers may be organized in three dimensions: width, height, and depth.

[0062]

[0074] A convolutional layer applies a convolution operation to the input and can pass the result of the convolution operation to the next layer. When processing images, the convolution operation can reduce the number of free parameters, allowing the network to be deeper with fewer parameters. In a neural network, each neuron may receive input from several locations in the previous layer. In a convolutional layer, neurons may receive input only from limited subregions of the previous layer. The parameters of a convolutional layer may consist of a set of learnable filters (or kernels). The learnable filters may have small receptive fields and span the full depth of the input volume. During the forward pass, each filter may be convolved across the width and height of the input volume to compute the dot product between the entries of the filter and the input, generating a two-dimensional activation map for that filter. As a result, the network can learn filters that activate when it detects a certain type of feature at some spatial location in the input.

[0063]

[0075] In some embodiments, the pooling layer includes a global pooling layer, which can combine the outputs of neuron clusters in one layer to a single neuron in the next layer. For example, a max pooling layer can use the maximum value from each of the neuron clusters in the previous layer. An average pooling layer can use the average value from each of the neuron clusters in the previous layer.

[0064]

[0076] In some embodiments, a fully connected layer connects every neuron in one layer to every neuron in another layer. In a neural network, each neuron may receive input from several places in the previous layer. In a fully connected layer, each neuron can receive input from every element in the previous layer.

[0065]

[0077] In some embodiments, the normalization layer is a batch normalization layer. A batch normalization layer can improve the performance and stability of a neural network. A batch normalization layer can provide zero mean / unit variance inputs to any layer of a neural network. Advantages of using a batch normalization layer can include faster networks to train, faster learning rates, easier initialization of weights, more feasible activation functions, and a simplified process of creating deep networks.

[0066]

[0078] The trained algorithm can be configured to accept a plurality of input variables and generate one or more output values ​​based on the plurality of input variables. The plurality of input variables can consist of one or more data sets indicative of cancer-related categories. For example, the input variables can include microscopy images of a biopsy sample of the subject. The plurality of input variables can also include clinical health data of the subject.

[0067]

[0079] The trained algorithm may include a classifier such that each of the one or more output values ​​includes one of a fixed number of possible values ​​(e.g., a linear classifier, a logistic regression classifier, etc.) indicating the classification of the biological sample and / or subject by the classifier. The trained algorithm may include a binary classifier such that each of the one or more output values ​​includes one of two values ​​(e.g., {0,1}, {positive, negative}, or {high risk, low risk}) indicating the classification of the biological sample and / or subject by the classifier. The trained algorithm may be another type of classifier such that each of the one or more output values ​​includes one of three or more values ​​(e.g., {0,1,2}, {positive, negative, or indeterminate}, or {high risk, intermediate risk, or low risk}) indicating the classification of the biological sample and / or subject by the classifier. The output values ​​may include descriptive labels, numerical values, or combinations thereof. Some of the output values ​​may include descriptive labels. Such descriptive labels can provide an identification or indication of the subject's disease or disorder status, and may include, for example, positive, negative, high risk, medium risk, low risk, or indeterminate. Such descriptive labels can provide an identification of a treatment for the subject's cancer-related category, and may include, for example, a therapeutic intervention, a duration of the therapeutic intervention, and / or a dosage of the therapeutic intervention suitable for treating a subject falling into a particular cancer-related category.

[0068]

[0080] Some of the output values ​​may include numerical values, such as binary, integer, or continuous values. Such binary output values ​​may include, for example, {0,1}, {positive,negative}, or {high risk,low risk}. Such integer output values ​​may include, for example, {0,1,2}. Such continuous output values ​​may include, for example, probability values ​​at least 0 and less than or equal to 1. Such continuous output values ​​may include, for example, an unnormalized probability value of at least 0. Such continuous output values ​​may indicate a cancer-related category prognosis for the subject. Some of the numerical values ​​may be mapped to descriptive labels, for example, by mapping 1 to "positive" and 0 to "negative."

[0069]

[0081] A portion of the output values ​​may be assigned based on one or more cutoff values. For example, a binary classification of a sample may assign an output value of "positive" or 1 if the sample indicates that the subject has at least a 50% chance of having a cancer-associated condition (e.g., a type or stage of cancer) or belonging to a cancer-associated category. For example, a binary classification of a sample may assign an output value of "negative" or 0 if the sample indicates that the subject has less than a 50% chance of belonging to a cancer-associated category. In this instance, a single cutoff value of 50% is used to classify the sample into one of two possible binary output values. Examples of single cutoff values ​​include about 1%, about 2%, about 5%, about 10%, about 15%, about 20%, about 25%, about 30%, about 35%, about 40%, about 45%, about 50%, about 55%, about 60%, about 65%, about 70%, about 75%, about 80%, about 85%, about 90%, about 91%, about 92%, about 93%, about 94%, about 95%, about 96%, about 97%, about 98%, and about 99%.

[0070]

[0082] As another example, a classification of a sample can be assigned an output value of "positive" or 1 if the sample indicates that the subject is at least about 50%, at least about 55%, at least about 60%, at least about 65%, at least about 70%, at least about 75%, at least about 80%, at least about 85%, at least about 90%, at least about 91%, at least about 92%, at least about 93%, at least about 94%, at least about 95%, at least about 96%, at least about 97%, at least about 98%, at least about 99% or more in a cancer-related category (e.g., cancer diagnosis or prognosis). Classification of the sample can be assigned an output value of "positive" or 1 if the sample indicates that the subject is greater than about 50%, greater than about 55%, greater than about 65%, greater than about 65%, greater than about 70%, greater than about 75%, greater than about 80%, greater than about 85%, greater than about 90%, greater than about 91%, greater than about 92%, greater than about 93%, greater than about 94%, greater than about 95%, greater than about 96%, greater than about 97%, greater than about 98%, or greater than about 99% likely to belong to a cancer-related category (e.g., long-term outcome).

[0071]

[0083] Classification of the sample can be assigned an output value of "negative" or 0 if the sample indicates that the subject has less than about 50%, less than about 45%, less than about 40%, less than about 35%, less than about 30%, less than about 25%, less than about 20%, less than about 15%, less than about 10%, less than about 9%, less than about 8%, less than about 7%, less than about 6%, less than about 5%, less than about 4%, less than about 3%, less than about 2%, or less than about 1% chance of having a cancer-related condition or belonging to a cancer-related category (e.g., positive for prostate cancer). A classification of a sample can be assigned an output value of "negative" or 0 if the sample indicates that the subject has about 50% or less, about 45% or less, about 40% or less, about 35% or less, about 30% or less, about 25% or less, about 20% or less, about 15% or less, about 10% or less, about 9% or less, about 8% or less, about 7% or less, about 6% or less, about 5% or less, about 4% or less, about 3% or less, about 2% or less, or about 1% or less of the probability that the subject has a cancer-related condition (e.g., prostate cancer).

[0072]

[0084] Classification of a sample can be done by assigning an output value of "indeterminate" or 2 if the sample is not classified as "positive", "negative", 1 or 0. In this case, a set of two cutoff values ​​is used to classify the sample into one of three possible output values. Example sets of cutoff values ​​include {1%, 99%}, {2%, 98%}, {5%, 95%}, {10%, 90%}, {15%, 85%}, {20%, 80%}, {25%, 75%}, {30%, 70%}, {35%, 65%}, {40%, 60%}, and {45%, 55%}. Similarly, n sets of cutoff values ​​can be used to classify a sample into one of n+1 possible output values, where n is any positive integer.

[0073]

[0085] The trained algorithm can be trained using multiple independent training samples. Each of the independent training samples can include a biological sample from a subject, an associated dataset obtained by assaying the biological sample (as described elsewhere herein), clinical data from the subject, and one or more known output values ​​corresponding to the biological sample and / or the subject (e.g., clinical diagnosis, prognosis, absence, or treatment efficacy of the subject's cancer-related condition). The independent training samples can include biological samples and associated datasets and outputs obtained or derived from multiple different subjects. The independent training samples can include biological samples and associated datasets and outputs obtained from the same subject at multiple different time points (e.g., periodically, such as weekly, biweekly, monthly, yearly, etc.). The independent training samples can be associated with the presence of a cancer-related condition (e.g., a training sample including biological samples and associated datasets and outputs obtained or derived from multiple subjects known to have a cancer-related condition). Independent training samples may be associated with the absence of a cancer-associated condition (e.g., training samples including biological samples and associated datasets and output obtained from or derived from multiple subjects known to have not been previously diagnosed with a cancer-associated condition or multiple subjects who have received a negative test result for a cancer-associated condition).

[0074]

[0086] The trained algorithm can be trained using at least about 5, at least about 10, at least about 15, at least about 20, at least about 25, at least about 30, at least about 35, at least about 40, at least about 45, at least about 50, at least about 100, at least about 150, at least about 200, at least about 250, at least about 300, at least about 350, at least about 400, at least about 450, or at least about 500 independent training samples. The independent training samples may include acellular biological samples and clinical data related to the presence of cancer-related categories and / or acellular biological samples and clinical data related to the absence of cancer-related categories. The trained algorithm can be trained using about 500 or less, about 450 or less, about 400 or less, about 350 or less, about 300 or less, about 250 or less, about 200 or less, about 150 or less, about 100 or less, or about 50 or less independent training samples related to the presence of cancer-related categories. In some embodiments, the biological sample and / or clinical data is independent of the sample used to train the trained algorithm.

[0075]

[0087] The trained algorithm may be trained using a first number of independent training samples associated with the presence of a cancer-associated category and a second number of independent training samples associated with the absence of a cancer-associated category. The first number of independent training samples associated with the presence of a cancer-associated category may be less than or equal to the second number of independent training samples associated with the absence of a cancer-associated category. The first number of independent training samples associated with the presence of a cancer-associated category may be equal to the second number of independent training samples associated with the absence of a cancer-associated category. The first number of independent training samples associated with the presence of a cancer-associated category may be greater than the second number of independent training samples associated with the absence of a cancer-associated category.

[0076]

[0088] Various loss functions may be used to train the algorithm. In some embodiments, the algorithm may include a regression loss function. In some embodiments, the algorithm may include a logistic loss function. In some embodiments, the algorithm may include a cross-entropy loss function. In some embodiments, the algorithm may include a (e.g., negative) log-likelihood loss function. In some embodiments, the algorithm may include a negative partial log-likelihood loss. In some embodiments, the algorithm may include a variational loss. In some embodiments, the loss function may be formulated to optimize regression loss, evidence-based lower bound, maximum likelihood, Kullback-Leibler divergence, applied with various distribution functions such as Gaussian, non-Gaussian, mixture of Gaussians, mixture of logistic functions, etc.

[0077]

[0089] In some embodiments, the algorithms described herein are trained in a multi-task manner. In such cases, the algorithm may be trained to perform multiple learning tasks simultaneously. In one example, the algorithm is trained to perform a first classification task (e.g., associating subject data with a cancer-related category). The algorithm is further trained to perform a second task, including associating subject data with a second cancer-related category. In some cases, the algorithm is further trained to calculate a delta or deviation between the likelihood of the first classification and the second classification. The first classification and the second classification may correspond to any cancer-related category, as described herein. The first classification and the second classification may be based on the same subject data. Alternatively, the classifications may be based on different or transformed subject data. In one example, the first classification is based on factual data related to one or more therapeutic interventions that the subject has undergone, while the second classification is based on counterfactual data related to one or more therapeutic interventions that the subject may undergo.

[0078]

[0090] Various optimizers can be used to train the neural network. In some embodiments, the neural network can be trained with an Adam optimizer. In some embodiments, the neural network can be trained with a stochastic gradient descent optimizer. In some embodiments, the neural network can be trained with an active learning algorithm. The neural network can be trained with various loss functions, the derivatives of which can be calculated to update one or more parameters of the neural network. The neural network can be trained with a hyperparameter search algorithm. In some embodiments, the hyperparameters of the neural network are optimized with a Gaussian process.

[0079]

[0091] The trained algorithm may identify at least about 50%, at least about 55%, at least about 60%, or at least about 70% of the cancer-related categories for at least about 5, at least about 10, at least about 15, at least about 20, at least about 25, at least about 30, at least about 35, at least about 40, at least about 45, at least about 50, at least about 100, at least about 150, at least about 200, at least about 250, at least about 300, at least about 350, at least about 400, at least about 450, or at least about 500 independent training samples. The accuracy of the trained algorithm may be calculated as the percentage of independent test samples (e.g., subjects known to be in the cancer-associated category or subjects with a negative clinical test result for the cancer-associated category) that are correctly identified or classified as belonging or not belonging to the cancer-associated category.

[0080]

[0092] The trained algorithm may be configured to identify a cancer-related category with a positive predictive value (PPV) of at least about 5%, at least about 10%, at least about 15%, at least about 20%, at least about 25%, at least about 30%, at least about 35%, at least about 40%, at least about 50%, at least about 55%, at least about 60%, at least about 65%, at least about 70%, at least about 75%, at least about 80%, at least about 81%, at least about 82%, at least about 83%, at least about 84%, at least about 85%, at least about 86%, at least about 87%, at least about 88%, at least about 89%, at least about 90%, at least about 91%, at least about 92%, at least about 93%, at least about 94%, at least about 95%, at least about 96%, at least about 97%, at least about 98%, at least about 99% or more. The PPV of identifying a cancer-associated category using the trained algorithm can be calculated as the proportion of acellular biological samples identified or classified as having a cancer-associated category that correspond to subjects that truly belong to that cancer-associated category.

[0081]

[0093] The trained algorithm may be configured to identify a cancer-related category with a negative predictive value (NPV) of at least about 5%, at least about 10%, at least about 15%, at least about 20%, at least about 25%, at least about 30%, at least about 35%, at least about 40%, at least about 50%, at least about 55%, at least about 60%, at least about 65%, at least about 70%, at least about 75%, at least about 80%, at least about 81%, at least about 82%, at least about 83%, at least about 84%, at least about 85%, at least about 86%, at least about 87%, at least about 88%, at least about 89%, at least about 90%, at least about 91%, at least about 92%, at least about 93%, at least about 94%, at least about 95%, at least about 96%, at least about 97%, at least about 98%, at least about 99% or more. The NPV of identifying a cancer-related category using a trained algorithm can be calculated as the proportion of a subject dataset identified or classified as not having a cancer-related category that corresponds to subjects who truly do not belong to that cancer-related category.

[0082]

[0094] The trained algorithm may be at least about 5%, at least about 10%, at least about 15%, at least about 20%, at least about 25%, at least about 30%, at least about 35%, at least about 40%, at least about 50%, at least about 55%, at least about 60%, at least about 65%, at least about 70%, at least about 75%, at least about 80%, at least about 81%, at least about 82%, at least about 83%, at least about 84%, at least about 85%, at least about 86%, at least about 87%, at least about 88%, at least about 89%, at least about 90%, The cancer-associated category may be configured to identify a cancer-associated category with a clinical sensitivity of at least about 91%, at least about 92%, at least about 93%, at least about 94%, at least about 95%, at least about 96%, at least about 97%, at least about 98%, at least about 99%, at least about 99.1%, at least about 99.2%, at least about 99.3%, at least about 99.4%, at least about 99.5%, at least about 99.6%, at least about 99.7%, at least about 99.8%, at least about 99.9%, at least about 99.99%, at least about 99.999%, or more. The clinical sensitivity of identifying a cancer-associated category using the trained algorithm may be calculated as the proportion of independent test samples associated with the cancer-associated category (e.g., subjects known to belong to the cancer-associated category) that are correctly identified or classified as having the cancer-associated category.

[0083]

[0095] The trained algorithm may be at least about 5%, at least about 10%, at least about 15%, at least about 20%, at least about 25%, at least about 30%, at least about 35%, at least about 40%, at least about 50%, at least about 55%, at least about 60%, at least about 65%, at least about 70%, at least about 75%, at least about 80%, at least about 81%, at least about 82%, at least about 83%, at least about 84%, at least about 85%, at least about 86%, at least about 87%, at least about 88%, at least about 89%, at least about 90%, at least about 91%, at least about 92%, at least about 93%, at least about 94%, at least about 95%, at least about 96%, at least about 97%, at least about 98%, at least about 99%, at least about 100%, at least about 101%, at least about 102%, at least about 103%, at least about 104%, at least about 105%, at least about 106%, at least about 107%, at least about 108%, at least about 109%, at least about 110%, at least about 111%, at least about 112%, at least about 113%, at least about 114%, at least about 115%, at least about 116%, at least about 117%, at least about 118%, at least about 119%, at least about 120%, at least about 121%, at least about 122%, at least about 123%, at least about 124%, at least about 125%, at least about 126%, at least about 127%, at least about 128%, at least about 129%, at least about 130%, at least about 131%, at least about 132%, at least about The cancer-associated category may be configured to identify a cancer-associated category with a clinical specificity of at least about 91%, at least about 92%, at least about 93%, at least about 94%, at least about 95%, at least about 96%, at least about 97%, at least about 98%, at least about 99%, at least about 99.1%, at least about 99.2%, at least about 99.3%, at least about 99.4%, at least about 99.5%, at least about 99.6%, at least about 99.7%, at least about 99.8%, at least about 99.9%, at least about 99.99%, at least about 99.999%, or more. Clinical specificity for identifying a cancer-associated category using a trained algorithm may be calculated as the proportion of independent test samples associated with the absence of a cancer-associated category (e.g., subjects with a negative clinical test result for the cancer-associated category) that are correctly identified or classified as not belonging to the cancer-associated category.

[0084]

[0096] The trained algorithm may be configured to distinguish a cancer-related category with an area under the curve (AUC) of at least about 0.50, at least about 0.55, at least about 0.60, at least about 0.65, at least about 0.70, at least about 0.75, at least about 0.80, at least about 0.81, at least about 0.82, at least about 0.83, at least about 0.84, at least about 0.85, at least about 0.86, at least about 0.87, at least about 0.88, at least about 0.89, at least about 0.90, at least about 0.91, at least about 0.92, at least about 0.93, at least about 0.94, at least about 0.95, at least about 0.96, at least about 0.97, at least about 0.98, at least about 0.99, or more. The AUC can be calculated as the integral of the receiver operating characteristic (ROC) curve (e.g., the area under the ROC curve) associated with a trained algorithm in classifying a dataset derived from subjects into or out of a cancer-related category.

[0085]

[0097] The trained algorithm can be adjusted or tuned to improve one or more of the following performance, accuracy, PPV, NPV, clinical sensitivity, clinical specificity, or AUC of identifying cancer-related categories. The trained algorithm can be adjusted or tuned by adjusting the parameters of the trained algorithm (e.g., a set of cutoff values ​​used to classify biological samples as described elsewhere herein, or the weights of a neural network). The trained algorithm can be adjusted or tuned continuously during the training process or after the training process is completed.

[0086]

[0098] After the trained algorithm is initially trained, a subset of inputs can be identified as the most influential or most important to include to make a high-quality classification. For example, a subset of clinical data can be identified as the most influential or most important to include to make a high-quality classification or discrimination of a cancer-related category (or a subtype of a cancer-related category). The clinical data or a subset thereof may be ranked based on a classification metric that indicates the influence or importance of each parameter in making a high-quality classification or discrimination of a cancer-related category (or a subtype of a cancer-related category). Such metrics may be used in some embodiments to meaningfully reduce the number of input variables (e.g., predictor variables) that can be used to train the trained algorithm to a desired performance level (e.g., based on a desired minimum accuracy, PPV, NPV, clinical sensitivity, clinical specificity, AUC, or a combination thereof). For example, if training a trained algorithm with a large number of input variables, including tens to hundreds in the trained algorithm, results in greater than 99% classification accuracy, instead training the trained algorithm using only a selected subset of the large number of input variables, such as about 5 or less, about 10 or less, about 15 or less, about 20 or less, about 25 or less, about 30 or less, about 35 or less, about 40 or less, about 45 or less, about 50 or less, or about 100 or less of the most influential or most significant input variables, may result in reduced classification accuracy but still provide acceptable accuracy (e.g., at least about 50 %, at least about 55%, at least about 60%, at least about 65%, at least about 70%, at least about 75%, at least about 80%, at least about 81%, at least about 82%, at least about 83%, at least about 84%, at least about 85%, at least about 86%, at least about 87%, at least about 88%, at least about 89%, at least about 90%, at least about 91%, at least about 92%, at least about 93%, at least about 94%, at least about 95%, at least about 96%, at least about 97%, at least about 98%, or at least about 99%.The subset can be selected by rank-ranking the entire number of input variables and selecting a predetermined number of input variables (e.g., about 5 or less, about 10 or less, about 15 or less, about 20 or less, about 25 or less, about 30 or less, about 35 or less, about 40 or less, about 45 or less, about 50 or less, or about 100 or less) having the best classification metric.

[0087]

[0099] The systems and methods described herein can use multiple trained algorithms to determine an output (e.g., a cancer-related category of a subject). The systems and methods can include 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, or more trained algorithms. A trained algorithm of the multiple trained algorithms can be trained on a particular type of data (e.g., image data or tabular data). Alternatively, a trained algorithm can be trained on multiple types of data. The input of a trained algorithm may include the output of one or more other trained algorithms. In addition, a trained algorithm can receive the output of one or more trained algorithms as input. Identifying or monitoring cancer-related categories or conditions

[0088]

[0100] After processing the dataset using the trained algorithm, a cancer-associated category can be identified or monitored in the subject. Identification may be based, at least in part, on quantitative or qualitative measures of the biological sample (e.g., of a histopathology slide of a biopsy sample), proteomic data including quantitative measures of proteins of the dataset in a panel of cancer-associated proteins, and / or metabolomic data including quantitative measures of a panel of cancer-associated metabolites.

[0089]

[0101] Cancer-related category can characterize the cancer-related status of the subject.As a non-limiting example, the cancer-related status can include the status that the subject has or does not have cancer (e.g., prostate cancer), the status that the subject is at risk of cancer or has a certain risk level (e.g., high risk, low risk), the expected long-term outcome of cancer (e.g., distant metastasis, biochemical recurrence, partial response, complete response, overall survival, cancer-specific survival, progression-free survival, disease-free survival, 5-year survival, death), response or acceptance to therapeutic intervention, or combinations thereof.

[0090]

[0102] A subject may be identified as belonging to a cancer-related category with an accuracy of at least about 50%, at least about 55%, at least about 60%, at least about 65%, at least about 70%, at least about 75%, at least about 80%, at least about 81%, at least about 82%, at least about 83%, at least about 84%, at least about 85%, at least about 86%, at least about 87%, at least about 88%, at least about 89%, at least about 90%, at least about 91%, at least about 92%, at least about 93%, at least about 94%, at least about 95%, at least about 96%, at least about 97%, at least about 98%, at least about 99%, or more. The accuracy of identifying an individual's cancer-related category by the trained algorithm may be calculated as the proportion of independent test samples (e.g., subjects known to belong to a cancer-related category or subjects with a negative clinical test result corresponding to a cancer-related category) that are correctly identified or classified as belonging to or not belonging to a cancer-related category.

[0091]

[0103] A subject may be determined to be in a cancer-associated category with a positive predictive value (PPV) of at least about 5%, at least about 10%, at least about 15%, at least about 20%, at least about 25%, at least about 30%, at least about 35%, at least about 40%, at least about 50%, at least about 55%, at least about 60%, at least about 65%, at least about 70%, at least about 75%, at least about 80%, at least about 81%, at least about 82%, at least about 83%, at least about 84%, at least about 85%, at least about 86%, at least about 87%, at least about 88%, at least about 89%, at least about 90%, at least about 91%, at least about 92%, at least about 93%, at least about 94%, at least about 95%, at least about 96%, at least about 97%, at least about 98%, at least about 99% or more. The PPV of identifying a cancer-associated category using a trained algorithm can be calculated as the proportion of biological samples identified or classified as belonging to a cancer-associated category that correspond to subjects that truly belong to that cancer-associated category.

[0092]

[0104] A cancer-related category may be identified in a subject with a negative predictive value (NPV) of at least about 5%, at least about 10%, at least about 15%, at least about 20%, at least about 25%, at least about 30%, at least about 35%, at least about 40%, at least about 50%, at least about 55%, at least about 60%, at least about 65%, at least about 70%, at least about 75%, at least about 80%, at least about 81%, at least about 82%, at least about 83%, at least about 84%, at least about 85%, at least about 86%, at least about 87%, at least about 88%, at least about 89%, at least about 90%, at least about 91%, at least about 92%, at least about 93%, at least about 94%, at least about 95%, at least about 96%, at least about 97%, at least about 98%, at least about 99% or more. The NPV of identifying a cancer-related category using a trained algorithm can be calculated as the proportion of biological samples identified or classified as not belonging to the cancer-related category that correspond to subjects that truly do not belong to that cancer-related category.

[0093]

[0105] Subjects are at least about 5%, at least about 10%, at least about 15%, at least about 20%, at least about 25%, at least about 30%, at least about 35%, at least about 40%, at least about 50%, at least about 55%, at least about 60%, at least about 65%, at least about 70%, at least about 75%, at least about 80%, at least about 81%, at least about 82%, at least about 83%, at least about 84%, at least about 85%, at least about 86%, at least about 87%, at least about 88%, at least about 89%, at least about 90%, at least about 100%, at least about 101%, at least about 102%, at least about 103%, at least about 104%, at least about 105%, at least about 106%, at least about 107%, at least about 108%, at least about 109%, at least about 200%, at least about 201%, at least about 202%, at least about 203%, at least about 204%, at least about 205%, at least about 206%, at least about 207%, at least about 208%, at least about 209%, at least about 300%, at least about 301%, at least about 302%, at least about 303, at least about 304, at least about 305, at least about 306, at least about 307, at least about 308, at least about 309, at least about 310, at least about 311, at least about 312, at least about 313, at least about 314, at least about 315, at least about 316, at least about 317, at least about 318. The cancer-associated category may be identified with a clinical sensitivity of at least about 91%, at least about 92%, at least about 93%, at least about 94%, at least about 95%, at least about 96%, at least about 97%, at least about 98%, at least about 99%, at least about 99.1%, at least about 99.2%, at least about 99.3%, at least about 99.4%, at least about 99.5%, at least about 99.6%, at least about 99.7%, at least about 99.8%, at least about 99.9%, at least about 99.99%, at least about 99.999%, or more. The clinical sensitivity of identifying a cancer-associated category using a trained algorithm may be calculated as the proportion of independent test samples associated with belonging to the cancer-associated category (e.g., subjects known to belong to the cancer-associated category) that are correctly identified or classified as belonging to the cancer-associated category.

[0094]

[0106] The cancer-related category is at least about 5%, at least about 10%, at least about 15%, at least about 20%, at least about 25%, at least about 30%, at least about 35%, at least about 40%, at least about 50%, at least about 55%, at least about 60%, at least about 65%, at least about 70%, at least about 75%, at least about 80%, at least about 81%, at least about 82%, at least about 83%, at least about 84%, at least about 85%, at least about 86%, at least about 87%, at least about 88%, at least about 89%, The cancer-associated category may be identified with a clinical specificity of at least about 90%, at least about 91%, at least about 92%, at least about 93%, at least about 94%, at least about 95%, at least about 96%, at least about 97%, at least about 98%, at least about 99%, at least about 99.1%, at least about 99.2%, at least about 99.3%, at least about 99.4%, at least about 99.5%, at least about 99.6%, at least about 99.7%, at least about 99.8%, at least about 99.9%, at least about 99.99%, at least about 99.999%, or more. Clinical specificity for identifying a cancer-associated category using a trained algorithm may be calculated as the proportion of independent test samples associated with not belonging to the cancer-associated category (e.g., subjects with a negative clinical test result for the cancer-associated category) that are correctly identified or classified as not belonging to the cancer-associated category.

[0095]

[0107] After a cancer-related category is identified in a subject, a subtype of the cancer-related category (e.g., selected from among a plurality of subtypes of the cancer-related category) may be further identified. The subtype of the cancer-related category may be determined based at least in part on quantitative or qualitative measures of a biological sample (e.g., of a histopathology slide of a biopsy sample), proteomic data including quantitative measures of proteins of a dataset in a panel of cancer-related proteins, and / or metabolomic data including quantitative measures of a panel of cancer-related metabolites. For example, a subject may be identified as being at risk for a subtype of prostate cancer (e.g., among several subtypes of prostate cancer). After identifying a subject as being at risk for a subtype of prostate cancer, a clinical intervention for the subject may be selected based at least in part on the subtype of prostate cancer for which the subject is identified as being at risk. In some embodiments, the clinical intervention is selected from a plurality of clinical interventions (e.g., clinically indicated for different subtypes of prostate cancer).

[0096]

[0108] Once a subject is identified as belonging to a cancer-related category, the subject may optionally be provided with a therapeutic intervention (e.g., prescribing an appropriate course of treatment to treat the subject's cancer type, subtype, or condition). The therapeutic intervention may include prescribing an effective amount of a drug or other therapy (e.g., radiation therapy, chemotherapy), a surgical intervention (e.g., radical prostatectomy), further testing or evaluation of the cancer-related category, further monitoring of the cancer-related category, or a combination thereof. If the subject is currently being treated with a course of treatment for the cancer-related category, the therapeutic intervention may include a subsequent different course of treatment (e.g., to increase the effectiveness of the treatment because the current course of treatment is not effective).

[0097]

[0109] The therapeutic intervention may include recommending a secondary laboratory test for the subject to confirm the diagnosis of the cancer-related category, which may include imaging tests, blood tests, computed tomography (CT) scans, magnetic resonance imaging (MRI) scans, ultrasound scans, x-rays, positron emission tomography (PET) scans, PET-CT scans, bone scans, lymph node biopsies, or combinations thereof.

[0098]

[0110] Analysis of biopsy samples (e.g., analysis of microscopic images of prostate tissue), proteomic data including quantitative measures of proteins of a dataset in a panel of cancer-related category-related proteins, and / or metabolomic data including quantitative measures of a panel of cancer-related category-related metabolites can be evaluated over a period of time to monitor a patient (e.g., a subject with or at risk of cancer, or a subject undergoing treatment for cancer). In such cases, the patient's dataset measures may change over the course of treatment. For example, a dataset measure of a patient whose risk of a cancer-related category has decreased due to effective treatment may shift toward a profile or distribution of a healthy subject (e.g., a subject who is cancer-free or in remission from cancer). Conversely, a dataset measure of a patient whose risk of a cancer-related category has increased due to ineffective treatment may shift toward a profile or distribution of a subject who is at higher risk of a cancer-related category or a subject whose cancer-related category is more advanced.

[0099]

[0111] The cancer-related category of the subject can be monitored by monitoring a course of treatment for treating the cancer or cancer-related condition of the subject. The monitoring may include evaluating the cancer-related category or condition of the subject at two or more time points. The evaluation may be based on at least a quantitative or qualitative measure of the biological sample (e.g., of a histopathology slide of a biopsy sample), proteomic data including a quantitative measure of the proteins of a dataset in a panel of cancer-related proteins, and / or metabolomic data including a quantitative measure of a panel of cancer-related metabolites, determined at each of the two or more time points.

[0100]

[0112] In some embodiments, differences in quantitative or qualitative measures of a biological sample (e.g., of a histopathology slide of a biopsy sample), proteomic data including quantitative measures of proteins of a dataset in a panel of cancer-associated proteins, and / or metabolomic data including quantitative measures of a panel of cancer-associated metabolites determined between two or more time points may be indicative of one or more clinical indications, such as (i) a diagnosis of a cancer-associated condition in a subject, (ii) a prognosis of a cancer-associated condition in a subject, (iii) an increased risk of a cancer-associated condition in a subject, (iv) a decreased risk of a cancer-associated condition in a subject, (v) the effectiveness of a course of treatment for treating a cancer-associated condition in a subject, and (vi) the ineffectiveness of a course of treatment for treating a cancer-associated condition in a subject.

[0101]

[0113] In some embodiments, a difference in a quantitative or qualitative measure of a biological sample (e.g., of a histopathology slide of a biopsy sample), a proteomic data including a quantitative measure of a protein of a dataset in a panel of cancer-related category-related proteins, and / or a metabolomic data including a quantitative measure of a panel of cancer-related metabolites determined between two or more time points may indicate a cancer-related condition or a diagnosis of a cancer-related category in a subject. For example, if a cancer-related condition is not detected in a subject at an earlier time point, but is detected in the subject at a later time point, the difference indicates that the cancer-related condition in the subject is a diagnosis. Based on this indication of a diagnosis of the cancer-related condition in the subject, a clinical action or decision may be made, such as, for example, prescribing a new therapeutic intervention for the subject. The clinical action or decision may include recommending a secondary laboratory test for the subject to confirm the diagnosis of the cancer-related category. The secondary laboratory tests may include imaging tests, blood tests, computed tomography (CT) scans, magnetic resonance imaging (MRI) scans, ultrasound scans, x-rays, positron emission tomography (PET) scans, PET-CT scans, bone scans, lymph node biopsies, or combinations thereof.

[0102]

[0114] In some embodiments, differences in quantitative or qualitative measures of a biological sample (e.g., of a histopathology slide of a biopsy sample), proteomic data including quantitative measures of proteins of a dataset in a panel of cancer-associated proteins, and / or metabolomic data including quantitative measures of a panel of cancer-associated metabolites determined between two or more time points may be indicative of a subject's prognosis in a cancer-associated category.

[0103]

[0115] In some embodiments, a difference in the quantitative or qualitative measure of the biological sample (e.g., of a histopathology slide of a biopsy sample), the proteomic data including a quantitative measure of proteins of a dataset in a panel of proteins associated with a cancer-related category, and / or the metabolomic data including a quantitative measure of a panel of metabolites associated with a cancer-related category determined between two or more time points may indicate that the subject has an elevated risk of a cancer-related condition. For example, if a cancer-related condition is detected in a subject at both an early time point and a later time point and the difference is a negative difference (e.g., if the quantitative or qualitative measure of the biological sample (e.g., of a histopathology slide of a biopsy sample), the proteomic data including a quantitative measure of proteins of a dataset in a panel of proteins associated with a cancer-related category, and / or the metabolomic data including a quantitative measure of a panel of metabolites associated with a cancer-related category increases from an early time point to a later time point), the difference may indicate that the subject has an elevated risk of a cancer-related condition. Based on this indication of an increased risk of a cancer-related condition, a clinical action or decision can be made, such as prescribing a new therapeutic intervention for the subject or switching therapeutic interventions (e.g., terminating the current treatment and prescribing a new treatment). The clinical action or decision can include recommending a secondary laboratory test for the subject to confirm the increased risk of the cancer-related category. The secondary laboratory test can include imaging tests, blood tests, computed tomography (CT) scans, magnetic resonance imaging (MRI) scans, ultrasound scans, x-rays, positron emission tomography (PET) scans, PET-CT scans, bone scans, lymph node biopsies, or combinations thereof.

[0104]

[0116] In some embodiments, a difference in the quantitative or qualitative measure of the biological sample (e.g., of a histopathology slide of a biopsy sample), the proteomic data including a quantitative measure of a protein of a dataset in a panel of cancer-associated proteins, and / or the metabolomic data including a quantitative measure of a panel of cancer-associated metabolites determined between two or more time points may indicate that the subject has a reduced risk of a cancer-associated condition. For example, if a cancer-associated condition is detected in a subject at both an earlier time point and a later time point and the difference is a positive difference (e.g., if the quantitative or qualitative measure of the biological sample (e.g., of a histopathology slide of a biopsy sample), the proteomic data including a quantitative measure of a protein of a dataset in a panel of cancer-associated proteins, and / or the metabolomic data including a quantitative measure of a panel of cancer-associated metabolites decreases from an earlier time point to a later time point), the difference may indicate that the subject has a reduced risk of the cancer-associated condition. Based on this indication of a reduced risk of the cancer-associated condition, a clinical action or decision (e.g., continuation or termination of a current therapeutic intervention) can be made for the subject. The clinical action or decision may include recommending a secondary laboratory test for the subject to confirm the reduction in risk of a cancer-related category, which may include an imaging test, a blood test, a computed tomography (CT) scan, a magnetic resonance imaging (MRI) scan, an ultrasound scan, an x-ray, a positron emission tomography (PET) scan, a PET-CT scan, a bone scan, a lymph node biopsy, or a combination thereof.

[0105]

[0117] In some embodiments, a difference in a quantitative or qualitative measure of a biological sample (e.g., of a histopathology slide of a biopsy sample), a proteomic data including a quantitative measure of a protein of a dataset in a panel of cancer-associated proteins, and / or a metabolomic data including a quantitative measure of a panel of cancer-associated metabolites determined between two or more time points may indicate the effectiveness of a course of treatment for treating a cancer-associated condition in a subject. For example, if a cancer-associated condition is detected in a subject at an earlier time point but not in the subject at a later time point, the difference may indicate that a course of treatment for treating a cancer-associated condition in a subject is effective. Based on this indication that a course of treatment for treating a cancer-associated condition in a subject is effective, a clinical action or decision may be made, such as to continue or terminate a current therapeutic intervention for the subject. The clinical action or decision may include recommending a secondary laboratory test for the subject to determine the effectiveness of a course of treatment for treating a cancer-associated category. The secondary laboratory tests may include imaging tests, blood tests, computed tomography (CT) scans, magnetic resonance imaging (MRI) scans, ultrasound scans, x-rays, positron emission tomography (PET) scans, PET-CT scans, bone scans, lymph node biopsies, or combinations thereof.

[0106]

[0118] In some embodiments, a difference in the quantitative or qualitative measure of the biological sample (e.g., of a histopathology slide of a biopsy sample), the proteomic data including a quantitative measure of a protein of a dataset in a panel of cancer-associated proteins, and / or the metabolomic data including a quantitative measure of a panel of cancer-associated metabolites determined between two or more time points may indicate the ineffectiveness of a course of treatment for treating the cancer-associated category of the subject. For example, if a cancer-associated condition is detected in a subject at both an early time point and a later time point and the difference is negative or is a zero difference (e.g., if the quantitative or qualitative measure of the biological sample (e.g., of a histopathology slide of a biopsy sample), the proteomic data including a quantitative measure of a protein of a dataset in a panel of cancer-associated proteins, and / or the metabolomic data including a quantitative measure of a panel of cancer-associated metabolites increases or remains at a constant level from the early time point to the later time point), and if an effective treatment was indicated at the earlier time point, the difference may indicate that a course of treatment for treating the cancer-associated condition of the subject is not effective. Based on this indication that the course of treatment for treating the subject's cancer-related condition is ineffective, a clinical action or decision can be made, such as terminating the current therapeutic intervention for the subject and / or switching to (e.g., prescribing) a different new therapeutic intervention. The clinical action or decision may include recommending a secondary laboratory test for the subject to determine the ineffectiveness of the course of treatment for treating the cancer-related condition. The secondary laboratory test may include imaging tests, blood tests, computed tomography (CT) scans, magnetic resonance imaging (MRI) scans, ultrasound scans, x-rays, positron emission tomography (PET) scans, PET-CT scans, bone scans, lymph node biopsies, or combinations thereof.

[0107] Predicting response to therapeutic intervention

[0108]

[0119] After processing the dataset using the trained algorithm, a cancer-related category can be identified or monitored in the subject. Identification may be based, at least in part, on quantitative or qualitative measures of the biological sample (e.g., of a histopathology slide of a biopsy sample), proteomic data including quantitative measures of proteins of the dataset in a panel of cancer-related proteins, and / or metabolomic data including quantitative measures of a panel of cancer-related metabolites. In some cases, the cancer-related category includes a predicted response to a therapeutic intervention.

[0109]

[0120] The predicted response may be determined by processing data (e.g., tabular data and / or image data) associated with the subject in a factual model. Under the factual model associated with the subject, the data represents a therapeutic intervention that the subject did or did not actually receive. For example, a factual model of a patient who received androgen deprivation therapy (ADT) as (e.g., at least a part of) the therapeutic intervention may include an indication that the subject was treated with ADT. In another example, a factual model of a patient who did not receive ADT as (e.g., at least a part of) the therapeutic intervention may include an indication that the subject did not receive ADT. The predicted response may be further determined by processing the data associated with the subject, or a subset thereof, again, except under a counterfactual model that indicates the opposite regarding the application of the treatment to the subject. For example, a counterfactual model for a patient who received ADT as (e.g., at least a part of) the therapeutic intervention may include an indication that the subject was not treated with ADT, and a corresponding counterfactual model for a subject who did not receive ADT as (e.g., at least a part of) the therapeutic intervention may include an indication that the subject received ADT. By working with both factual and counterfactual models, it is possible to calculate the observed or predicted benefits of a therapeutic intervention.

[0110]

[0121] The difference in the likelihood of a cancer-related state (e.g., long-term outcomes such as distant metastasis [DM] or cancer type-specific mortality) predicted from the factual and counterfactual models can be described as the delta score or treatment delta.

[0111]

[0122] In some cases, a trained algorithm as described herein may be configured to determine the predicted benefit of a therapeutic intervention by predicting a treatment delta for a subject. The trained algorithm may be further trained for delta loss as described herein. Delta loss may characterize the deviation between a predicted delta score and an expected delta score for a given training subject.

[0112]

[0123] Expected delta scores may depend on subgroups of training subjects considering their treatment type and long-term outcomes (e.g., DM).

[0113]

[0124] In some cases, the subgroup includes subjects who do not receive the therapeutic intervention and do not exhibit the cancer-associated category (e.g., are DM negative). In such cases, the treatment delta is expected to be 0 or close to 0 because the subject would not have exhibited the cancer-associated category in the absence of the therapeutic intervention (e.g., distant metastases did not occur as would be expected in the absence of treatment, even though the subject did not receive the therapeutic intervention).

[0114]

[0125] In some cases, a subgroup includes subjects who did not receive the therapeutic intervention and did not exhibit a cancer-related category (e.g., DM positive). In such cases, the treatment delta is expected to be greater than or equal to 0 because subjects may have benefited from the therapeutic intervention (e.g., subjects who did not receive the therapeutic intervention may have benefited from receiving it).

[0115]

[0126] In some cases, the subgroup includes subjects who received the therapeutic intervention and had a negative cancer-related category. In such cases, the treatment delta is expected to be greater than 0 because the cancer-related category was negative in subjects in the presence of the therapeutic intervention (e.g., subjects who received the therapeutic intervention would show a decrease in DM, indicating the effectiveness of the intervention).

[0116]

[0127] In some cases, the subgroup includes subjects who received a therapeutic intervention and have a positive cancer-related category. In such cases, the treatment delta is expected to be 0 or close to 0 because the cancer-related category is positive in subjects in the presence of the therapeutic intervention (e.g., subjects who develop DM despite receiving the therapeutic intervention indicate the ineffectiveness of the treatment). During training, the model may be penalized if the delta score does not fall within the expected range.

[0117]

[0128] After training the algorithm, the trained algorithm can be used to predict the response of previously unseen subjects to a therapeutic intervention. In some cases, the trained algorithm can predict that a subject will be positive to a therapeutic intervention ("positive" or "biomarker positive") if the subject is predicted to show a reduced risk of a cancer-related category in the presence of the therapeutic intervention. A subject can be classified as biomarker positive if the subject's predicted treatment delta exceeds a cutoff value. The cutoff can include values ​​greater than or equal to about -0.5, -0.4, -0.3, -0.2, -0.1, 0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, or 0.9. The cutoff can be determined based on a population of subjects. In one example, the cutoff can be selected based on a percentile of the distribution of treatment delta observed across a population of subjects. The cutoff corresponds to about the 10th, 20th, 30th, 40th, 50th, 60th, 70th, 80th, 90th, or higher percentile. The cutoff corresponds to at least about the 10th, 20th, 30th, 40th, 50th, 60th, 70th, 80th, 90th, or higher percentile. In some cases, the cutoff corresponds to up to about the 90th, 80th, 70th, 60th, 50th, 40th, 30th, 20th, 10th, or lower percentile. In some cases, the cutoff is between any two of these values.

[0118]

[0129] In some embodiments, the therapeutic intervention may include radiation therapy (RT). In some embodiments, the therapeutic intervention may include chemotherapy. In some embodiments, the therapeutic intervention may include surgical intervention. In some embodiments, the therapeutic intervention may include immunotherapy. In some embodiments, the therapeutic intervention may include hormonal therapy. In some embodiments, the RT may include RT with a preset use of short-term androgen deprivation therapy (ST-ADT). In some embodiments, the RT may include RT with a preset use of long-term ADT (LT-ADT). In some embodiments, the RT may include RT with a preset use of dose-escalated RT (DE-RT). In some embodiments, the surgical intervention may include radical prostatectomy (RP). In some embodiments, the therapeutic intervention may include any combination of the therapeutic interventions disclosed herein. In some embodiments, the long-term outcome may include distant metastasis (DM). In some embodiments, the long-term outcome may include biochemical recurrence (BR). In some embodiments, the long-term outcome may include partial response. In some embodiments, the long-term outcome may include complete response. In some embodiments, the long term outcome may include death. In some embodiments, the long term outcome may include relative survival. In some embodiments, the long term outcome may include cancer specific survival. In some embodiments, the cancer specific survival may include prostate cancer specific survival (PCaSS). In some embodiments, the long term outcome may include progression free survival. In some embodiments, the long term outcome may include disease free survival. In some embodiments, the long term outcome may include 5 year survival. In some embodiments, the long term outcome may include overall survival (OS). In some embodiments, the long term outcome may include any combination of the long term outcomes disclosed herein.

[0119]

[0130] In some embodiments, the therapeutic intervention comprises additional therapeutic intervention. In one example, the therapeutic intervention comprises hormone therapy in addition to radiation therapy. In some cases, the hormone therapy is androgen deprivation therapy (ADT). In some cases, the ADT is short-term ADT (ST-ADT). In some cases, the ADT is long-term ADT (LT-ADT).

[0120]

[0131] In some embodiments, a subject may be "biomarker positive" if the subject is predicted to show a reduced risk of DM if treated with a therapeutic intervention. In some embodiments, a subject may be biomarker positive if the subject is predicted to show a reduced risk of BR if treated with a therapeutic intervention. In some embodiments, a subject may be biomarker positive if the subject is predicted to show a reduced risk of cancer type-specific mortality (such as prostate cancer-specific mortality) if treated with a therapeutic intervention. In some embodiments, a subject may be biomarker positive if the subject is predicted to show a reduced risk of death if treated with a therapeutic intervention. In some embodiments, a subject may be biomarker positive if the subject is predicted to show an increased likelihood of partial response. In some embodiments, a subject may be biomarker positive if the subject is predicted to show an increased likelihood of complete response. In some embodiments, a subject may be biomarker positive if the subject is predicted to show an increased likelihood of overall survival. In some embodiments, a subject may be biomarker positive if the subject is predicted to show an increased likelihood of 5-year survival. In some embodiments, a subject may be biomarker positive if the subject is predicted to show an increased likelihood of 10-year survival. In some embodiments, a subject may be biomarker positive if the subject is predicted to exhibit an increased likelihood of 15-year survival.

[0121]

[0132] In some embodiments, a subject may be "biomarker negative" if the subject is predicted not to show a reduced risk of DM (e.g., no change or increased risk) if treated with a therapeutic intervention. In some embodiments, a subject may be biomarker negative if the subject is predicted not to show a reduced risk of BR if treated with a therapeutic intervention. In some embodiments, a subject may be biomarker negative if the subject is predicted not to show a reduced risk of cancer type-specific mortality (such as prostate cancer-specific mortality) if treated with a therapeutic intervention. In some embodiments, a subject may be biomarker negative if the subject is predicted not to show a reduced risk of death if treated with a therapeutic intervention. In some embodiments, a subject may be biomarker negative if the subject is predicted not to show an increased likelihood of partial response. In some embodiments, a subject may be biomarker negative if the subject is predicted to show an increased likelihood of complete response. In some embodiments, a subject may be biomarker negative if the subject is predicted not to show an increased likelihood of overall survival. In some embodiments, a subject may be biomarker negative if the subject is predicted not to show an increased likelihood of 5-year survival. In some embodiments, a subject may be biomarker negative if the subject is predicted not to have an increased likelihood of 10-year survival. In some embodiments, a subject may be biomarker negative if the subject is predicted not to have an increased likelihood of 15-year survival.

[0122] <Cancer-related status report output>

[0123]

[0133] After a cancer-related condition has been identified or an elevated risk of a cancer-related condition has been monitored in the subject, a report may be electronically outputted that is indicative of (e.g., identifies or provides an indication of) the subject's cancer-related condition. The subject may not exhibit a cancer-related condition (e.g., is asymptomatic of a cancer-related condition, such as the presence or risk of prostate cancer). The report is presented on a graphical user interface (GUI) of the user's electronic device. The user may be the subject, a caregiver, a physician, a nurse, or other medical professional.

[0124]

[0134] The report may include one or more clinical indicators, such as (i) a diagnosis of the subject's cancer-related condition, (ii) a prognosis of the subject's cancer-related category, (iii) an increased risk of the subject's cancer-related category, (iv) a decreased risk of the subject's cancer-related category, (v) an efficacy of a course of treatment for treating the subject's cancer-related category, (vi) an ineffectiveness of a course of treatment for treating the subject's cancer-related category, and (vii) a long-term outcome of the cancer-related category. The report may include one or more clinical actions or decisions made based on these one or more clinical indicators. Such clinical actions or decisions may be directed to a therapeutic intervention of the subject's cancer-related condition, or further clinical evaluation or testing.

[0125]

[0135] For example, a clinical indication of a diagnosis of a cancer-related condition in a subject may involve the clinical action of prescribing a new therapeutic intervention for the subject. As another example, a clinical indication of an increased risk of a cancer-related condition in a subject may involve the clinical action of prescribing a new therapeutic intervention for the subject or switching therapeutic interventions (e.g., terminating a current treatment and prescribing a new treatment). As another example, a clinical indication of a decreased risk of a cancer-related condition in a subject may involve the clinical action of continuing or terminating a current therapeutic intervention for the subject. As another example, a clinical indication of the effectiveness of a course of treatment to treat a cancer-related condition in a subject may involve the clinical action of continuing or terminating a current therapeutic intervention for the subject. As another example, a clinical indication of the ineffectiveness of a course of treatment to treat a cancer-related condition in a subject may involve the clinical action of terminating a current therapeutic intervention for the subject and / or switching (e.g., prescribing) a new, different therapeutic intervention.

[0126] <Computer System>

[0127]

[0136] The present disclosure provides computer systems programmed to implement the methods of the present disclosure. Figure 1 illustrates a computer system 101 programmed or otherwise configured to, for example, (i) train and test a trained algorithm, (ii) use the trained algorithm to process image data and / or tabular data to determine a cancer-related category or cancer-related status of a subject, (iii) assess the cancer of the subject based on the classified category, (iv) identify or monitor the cancer-related category or cancer-related status of the subject, and (v) electronically output a report indicating the cancer-related category or cancer-related status of the subject.

[0128]

[0137] The computer system 101 can coordinate various aspects of the analysis, calculation, and generation of the present disclosure, such as, for example, (i) training and testing the trained algorithm, (ii) processing the image data and / or tabular data using the trained algorithm to determine the cancer-related category or cancer-related status of the subject, (iii) assessing the cancer of the subject based on the classified category, (iv) identifying or monitoring the cancer-related category or cancer-related status of the subject, and (v) electronically outputting a report indicating the cancer-related category or cancer-related status of the subject. The computer system 101 can be a user's electronic device or a computer system located remotely with respect to the electronic device. The electronic device can be a mobile electronic device.

[0129]

[0138] The computer system 101 includes a central processing unit (CPU, also referred to herein as "processor" and "computer processor") 105, which may be a single-core or multi-core processor, or multiple processors for parallel processing. The computer system 101 also includes memory or memory locations 110 (e.g., random access memory, read-only memory, flash memory), an electronic storage unit 115 (e.g., hard disk), a communication interface 120 (e.g., network adapter) for communicating with one or more other systems, and peripherals 125, such as cache, other memory, data storage, and / or electronic display adapters. The memory 110, the storage unit 115, the interface 120, and the peripherals 125 can communicate with the CPU 105 via a communication bus (solid lines), such as a motherboard. The storage unit 115 can be a data storage unit (or data repository) for storing data. The computer system 101 can be operatively coupled to a computer network ("network") 130 with the aid of the communication interface 120. The network 130 may be the Internet, the Internet and / or an extranet, or an intranet and / or an extranet in communication with the Internet.

[0130]

[0139] In some embodiments, the network 130 is a telecommunications network and / or a data network. The network 130 may include one or more computer servers that enable distributed computing, such as cloud computing. For example, the one or more computer servers may enable cloud computing via the network 130 ("cloud") to perform various aspects of the analysis, calculation, and generation of the present disclosure, such as, for example, (i) training and testing the trained algorithms, (ii) using the trained algorithms to process data and determine the cancer-related category of the subject, (iii) determining a quantitative measure indicative of the cancer-related category of the subject, (iv) identifying or monitoring the cancer-related category of the subject, and (v) electronically outputting a report indicative of the cancer-related category of the subject. Such cloud computing may be provided, for example, by a cloud computing platform such as Amazon Web Services (AWS), Microsoft Azure, Google Cloud Platform, IBM cloud, etc. In some embodiments, the network 130 may implement a peer-to-peer network with the aid of the computer system 101, whereby devices coupled to the computer system 101 may act as clients or servers.

[0131]

[0140] CPU 105 may include one or more computer processors and / or one or more graphic processing units (GPUs). CPU 105 may execute a series of machine-readable instructions, which may be embodied in a program or software. The instructions may be stored in a memory location, such as memory 110. The instructions may be directed to CPU 105, which may then be programmed or otherwise configured to implement the methods of the present disclosure. Examples of operations performed by CPU 105 may include fetch, decode, execute, and writeback.

[0132]

[0141] The CPU 105 may be part of a circuit, such as an integrated circuit. One or more other components of the system 101 may also be included in the circuit. In some embodiments, the circuit is an application specific integrated circuit (ASIC).

[0133]

[0142] The storage unit 115 can store files such as drivers, libraries, saved programs, etc. The storage unit 115 can store user data, e.g., user preferences and user programs. In some embodiments, the computer system 101 can include one or more additional data storage units that are external to the computer system 101, such as on a remote server in communication with the computer system 101 through an intranet or the Internet.

[0134]

[0143] The computer system 101 can communicate with one or more remote computer systems via the network 130. For example, the computer system 101 can communicate with a remote computer system of a user. Examples of remote computer systems include a personal computer (e.g., a portable PC), a slate or tablet PC (e.g., an Apple® iPad®, a Samsung® Galaxy Tab), a phone, a smartphone (e.g., an Apple® iPhone®, an Android-enabled device, a Blackberry®), or a personal digital assistant. A user can access the computer system 101 via the network 130.

[0135]

[0144] The methods described herein may be implemented by machine (e.g., computer processor) executable code stored in electronic storage locations of computer system 101, such as, for example, on memory 110 or electronic storage unit 115. Machine executable or machine readable code may be provided in the form of software. In use, the code may be executed by processor 105. In some embodiments, the code may be retrieved from storage unit 115 and stored in memory 110 for ready access by processor 105. In some circumstances, electronic storage unit 115 may not be used and machine executable instructions may be stored in memory 110.

[0136]

[0145] The code may be precompiled and configured for use on a machine having a processor that executes the code, or it may be compiled at run time. The code may be provided in a programming language that can be selected to allow the code to be executed in a precompiled or as-compiled fashion.

[0137]

[0146] The embodiments of the systems and methods provided herein, such as the computer system 101, may be embodied in programming. Various aspects of the technology may be considered as a "product" or "article of manufacture" in the form of machine (or processor) executable code and / or associated data typically carried or embodied in some type of machine-readable medium. The machine executable code may be stored in an electronic storage unit such as a memory (e.g., read-only memory, random access memory, flash memory, etc.) or a hard disk. A "storage" type medium may include tangible memory of a computer, a processor, etc., or its associated modules, such as various semiconductor memories, tape drives, or disk drives, which may provide non-transitory storage at any time for software programming. All or a portion of the software may be communicated at times over the Internet or various other telecommunications networks. Such communication allows loading of the software from one computer or processor to another, such as, for example, from a management server or host computer to a computer platform of an application server. Thus, other types of media that may carry software elements include light waves, radio waves, electromagnetic waves, such as those used across physical interfaces between local devices, through wired and optical fixed line networks, and on various air links. The physical elements that carry such waves, such as wired or wireless links, optical links, etc., may also be considered media that carry software. As used herein, unless limited to non-transitory tangible "storage" media, terms such as computer or machine "readable storage medium" refer to any medium that participates in providing instructions to a processor for execution.

[0138]

[0147] Thus, a machine-readable medium such as a computer executable code can take many forms, such as a tangible storage medium, a carrier wave medium, or a physical transmission medium. Non-volatile storage media include optical or magnetic disks, such as any of the storage devices in any computer, such as may be used to implement the databases, etc., shown in the figures. Volatile storage media include dynamic memory, such as the main memory of such a computer platform. Tangible transmission media include coaxial cables, copper wire, optical fibers, and the like, including the wires that make up a bus in a computer system. Carrier wave transmission media can take the form of electric or electromagnetic signals, or acoustic or light waves such as those generated during radio frequency (RF) and infrared (IR) data communications. Thus, common forms of computer readable media include, for example, floppy disks, flexible disks, hard disks, magnetic tape, other magnetic media, CD-ROMs, DVDs or DVD-ROMs, other optical media, punch cards paper tape, other physical storage media with patterns of holes, RAM, ROM, PROMs and EPROMs, FLASH-EPROMs, other memory chips or cartridges, carrier waves carrying data or instructions, cables or links carrying such carrier waves, or other media from which a computer can read programming code and / or data. Many of these forms of computer readable media may be involved in carrying one or more sequences of one or more instructions to a processor for execution.

[0139]

[0148] The computer system 101 can include or communicate with an electronic display 135 that includes a user interface (UI) 140 to provide, for example, (i) a visual display indicating the training and testing of a trained algorithm, (ii) a visual display of data indicating the subject's cancer-related category, (iii) a quantitative measure of the subject's cancer-related category, (iv) an identification of the subject as having a cancer-related category, or (v) an electronic report indicating the subject's cancer-related category. Examples of UIs include, but are not limited to, graphical user interfaces (GUIs) and web-based user interfaces.

[0140]

[0149] The methods and systems of the present disclosure may be implemented by one or more algorithms. The algorithms may be implemented by software when executed by the central processing unit 205. The algorithms may, for example, (i) train and test a trained algorithm, (ii) use the trained algorithm to process image data and / or tabular data to determine a cancer-related category or cancer-related status of a subject, (iii) assess the cancer of the subject based on the classified category, (iv) identify or monitor the cancer-related category or cancer-related status of the subject, and (v) electronically output a report indicating the cancer-related category or cancer-related status of the subject.

[0141]

[0150] <Example 1: Personalization of prostate cancer treatment using multimodal deep learning>

[0142]

[0151] The methods and systems disclosed herein demonstrate personalization of prostate cancer treatment by predicting long-term, clinically relevant outcomes (distant metastasis, biochemical recurrence, death from prostate cancer, and overall survival) using a novel multimodal deep learning model trained on digital histopathology and clinical data of prostate biopsies. The exemplary system of the present disclosure includes a trained algorithm that was trained and validated using a dataset from five randomized Phase III multinational trials conducted across hundreds of clinical sites. Clinical and histopathology data were available for 5,654 of 7,957 patients (71.1%), resulting in 16.1 terabytes of histopathology images with 10-20 years of patient follow-up. When compared to the National Cancer Center Network (NCCN) risk groups, the most commonly used risk stratification tool, the deep learning model showed superior prognostic and discrimination performance across all outcomes examined. This artificial intelligence system may enable oncologists to computationally model the most likely outcomes for a given patient and determine the patient's optimal treatment. Equipped with a digital histopathology scanner and internet access, any clinic can offer such capabilities, allowing low-cost, universal access to vital personalization of care.

[0143]

[0152] NCCN risk groups are based on an international standard for risk stratification developed in the late 1990s and referred to as D'Amico risk groups. This system is based on tumor grade as assessed by digital rectal examination, serum prostate-specific antigen (PSA) measurements, and histopathology. Although this three-tiered system continues to form the backbone of treatment recommendations worldwide, its prognostic and discriminatory performance for risk stratifying patients has been suboptimal. This is in part because the core variables of these models are highly subjective and nonspecific in nature. For example, Gleason grading was developed in the 1960s but remains highly subjective and has unacceptable interobserver reproducibility, even among expert urologic pathologists. More recently, tissue-based genomic biomarkers have demonstrated improved prognostic performance. However, almost all of these tests lack validation in prospective randomized clinical trials in the intended use population, and international adoption has been limited due to cost and turnaround time issues. Thus, there remains an unmet clinical need for improved tools to personalize treatment for prostate cancer.

[0144]

[0153] Artificial intelligence (AI) has demonstrated impressive capabilities in many use cases in healthcare, from physician-level diagnosis to workflow optimization, and has the potential to support cancer treatment as clinical adoption of digital histopathology continues. AI is beginning to make progress in histopathology-based prognosis, for example by predicting short-term patient outcomes and improving the accuracy of Gleason-based cancer grading in post-operative surgical specimens. Whereas standard risk stratification tools are fixed and based on a small number of variables, AI can learn from large amounts of minimally processed data across various modalities. In contrast to genomic biomarkers, AI systems are low-cost, massively scalable, and incrementally improved through use. Furthermore, a key challenge for any biomarker is having optimal data to train and validate the relevant endpoints, and some commercially available prognostic biomarkers in oncology are trained on retrospective convenience sampling.

[0145]

[0154] The methods and systems disclosed herein can include a multimodal artificial intelligence (MMAI) system that meaningfully overcomes the unmet need for outcome prognostication in localized prostate cancer and can generate generalizable biomarkers with potential for global adoption. By utilizing multimodal deep learning with digital histopathology, localized prostate cancer prognostic biomarkers from five Phase III randomized clinical trials were used to train the algorithms described herein.

[0146]

[0155] A unique dataset was generated from five large multinational randomized phase III clinical trials (NRG / RTOG9202, 9408, 9413, 9910, 0126) in men with localized prostate cancer. All patients received short-term androgen deprivation therapy (ST-ADT), long-term ADT (LT-ADT), definitive radiation therapy (RT) with pre-specified use of sirolimus, and / or dose-escalated RT (DE-RT) (Figure 9). Of the 7,957 patients enrolled in these five trials, 7,752 patients had complete baseline clinical data and 5,654 patients had complete baseline and digital histopathology data. This represents 16.1TB of histopathology images from 16,204 histopathology slides of pre-treatment biopsy samples.

[0147]

[0156] The MMAI architecture can ingest both tabular (clinical) and image-based (histopathology) data, making it uniquely suited for randomized clinical trial data. The complete architecture is shown in Figure 2A. Each patient in the dataset is represented by clinical variables (including laboratory data, pathology data, therapeutic interventions, and long-term outcomes), and digital histopathology slides (median 3.5 slides). Joint learning across both data streams is complex and involved the construction of three separate deep learning pipelines: one for images, one for the tabular data, and a third to integrate them. Data was standardized across trials for consistency.

[0148]

[0157] Effectively learning relevant features from a variable number of digital histopathology slides involved several preprocessing steps to standardize the images, followed by self-supervised training. For each patient, all tissue sections in the patient's biopsy slide were segmented and composited into a single large image of fixed width and height across all patients, called the image quilt (Figure 5). A H×W grid was overlaid on top of the image quilt and sliced ​​into 256×256 pixel patches across the RGB channels. These patches were then used to train a self-supervised (SSL) model to learn histopathological features useful for downstream AI tasks. Figure 2B illustrates this part of the pipeline. Once trained, the SSL model can then ingest patches of the image quilt and output a 128-dimensional vector representation for each patch. Concatenating all these vectors in the same spatial orientation as the original patches gives us a H×W×128 tensor (the feature quilt) that compresses the initially huge image quilt into a compact representation useful for further downstream learning.

[0149]

[0158] SSL is a method that can be used to learn from unannotated datasets. A typical ML setup leverages supervised learning, where a dataset consists of data points (e.g., images) and data labels (e.g., object classes). In contrast, in SSL, synthetic data labels are extracted from the original data and used to train a general feature representation that can be used for downstream tasks. Momentum contrast is a technique that takes a set of image patches, generates augmented copies of each patch, and then trains a model to predict whether any two augmented copies are derived from the same original patch, which can be useful in learning features for medical tasks. The structural setup is shown in Figure 2B and is described in further detail elsewhere in this document.

[0150]

[0159] To guide the SSL process to patch regions that are more likely to be clinically useful, the patches in the dataset were oversampled based on nucleus density. An object detection model trained to detect nuclei was used to estimate the number of nuclei in each patch. The patches were divided into deciles based on this count, and each decile was oversampled so that the net number of images seen during one epoch of training was the same for each decile. An example image is shown in Figure 6.

[0151]

[0160] The system described herein can learn from patient-level annotations without the need for annotations in histopathology slides, and furthermore, the self-supervised learning of the image model allows it to learn from new image data without the need for annotations.

[0152]

[0161] Learning from tabular data involves two steps. First, the clinical data was standardized across trials and used to pre-train the TabNet architecture through self-supervision by masking parts of the data and training the model to learn them. Each patient's data was then passed through TabNet to extract a feature vector, which was concatenated with the output of the image pipeline. The concatenated vector was then passed through further neural layers and the model output a binary outcome for the task at hand.

[0153]

[0162] The internal data representation of the SSL model is shown in Figure 4. Image patches from the entire dataset were passed through the SSL model, and model features (128-dimensional vectors output by the model) were extracted for each patch. The Uniform Manifold Approximation and Projection (UMAP) algorithm was then applied to these features, projecting the features from 128 dimensions to 2 dimensions and plotting each patch as an individual point. Adjacent data points represent image patches that the model considered similar. UMAP grouped the feature vectors into 25 clusters, some of which are shown in different colors. The inset shows example image patches that are close to the cluster center points in the feature space. The 20 image patches closest to the cluster center points were then interpreted by a pathologist. An example interpretation is shown in Figure 4, and the full interpretation in Figure 7. Despite never being trained on clinical annotations, the SSL model learned human-interpretable image features that indicate complex aspects of cancer, such as Gleason grade or tissue type.

[0154]

[0163] Six different MMAI models were trained and tested across four endpoints (DM, BCR, PCaSS, OS) and two time frames: 5 and 10 years. The performance of these models was measured by the time-dependent area under the receiver operating characteristic curve (AUC) of sensitivity and specificity taking into account competing events. Sensitivity is defined as the proportion of correct positive predictions for indicated positive events (sensitivity = predicted_positive / num_positive), and specificity is defined as the proportion of correct negative predictions for indicated negative events (specificity = predicted_negative / num_negative). In this metric, 0.5 represents chance accuracy and 1.0 represents perfect accuracy.

[0155]

[0164] The NCCN model served as the baseline comparator, as shown in Figure 8. Patients were grouped into low-, intermediate-, and high-risk groups using three variables: Gleason score, tumor t stage, and baseline PSA.

[0156]

[0165] The results are shown in Figures 3A-H. For each outcome and time point, a separate model was trained. In Figure 3A and Figures 2D-H, blue bars represent the performance of the MMAI model trained on a particular task, and gray bars represent the performance of the corresponding NCCN model. Figure 2B shows the relative improvement of MMAI over NCCN on a subset of outcomes and test sets from five studies. The MMAI model consistently outperformed the NCCN model on all outcomes tested. The relative improvement in AUC varied from 11.45% to 19.72%. Furthermore, the subsets of studies all uniformly showed relative improvement over NCCN.

[0157]

[0166] An ablation study was performed to evaluate the effect of various data components specific to the MMAI model. Additional MMAI models were trained using the following data setups: NCCN variables only, pathology images only, pathology images + NCCN variables (combined Gleason score, t-stage, baseline PSA), pathology images + NCCN variables + three additional variables used in the model (age, Gleason 1, Gleason 2). Each additional data component improved performance, with the full setup (pathology, six clinical variables) producing the best results (Figure 2C).

[0158]

[0167] The MMAI system significantly outperformed the model-encoded NCCN risk stratification tool in predicting four important future patient outcomes: distant metastasis, biochemical recurrence, prostate cancer-specific survival, and overall survival. By creating a deep learning architecture that simultaneously incorporates multiple data types (of varying sizes) from patients, as well as clinical data, a deep learning system was built that was able to predict long-term patient outcomes with significantly greater accuracy than established clinical models.

[0159]

[0168] The methods and systems described herein can leverage robust, large-scale clinical data from five different prospective, randomized, multinational trials with 10-20 years of patient follow-up on 5,654 patients across diverse populations. Validating these prognostic classifiers with large amounts of clinical trial data (in the intended use population) uniquely positions these tools as therapeutic decision-making aids. Similarly, significant drawbacks of genomics-based assays are their high cost and long test turnaround times. AI tools do not bear such limitations, making the barriers to large-scale international adoption significantly lower. Despite nearly 60% of the world having access to the internet, only about 4% (the US population) has easy access to genomics-based assays. The ever-increasing adoption of digital histopathology, coupled with internet connectivity, may support the global proliferation of AI-based prognostic and predictive testing, enabling access to critical treatment personalization at a lower cost.

[0160]

[0169] <Method>

[0161]

[0170] Tabular Pipeline. Tabular clinical data were split into numerical and categorical variables. Numerical variables were whitened (mean subtraction + max normalization) to the range of [-1,1]. Categorical variables were treated as one-hot vectors that were embedded into 2-3 dimensional vectors according to traditional word-to-vec techniques, with the number of dimensions being determined by the formula, D=Round(1.6·num_categories 0.56 We used a TabNet model with input being a concatenation of categorical and numerical variables (parameters: learning rate 0.2, Adam optimizer with step learning rate scheduler, batch size 1024, up to 50 epochs with early stopping with patience 10 epochs).

[0162]

[0171] Image Pipeline. The ResNet50 model was used in conjunction with the MoCo-v2 training protocol (parameters: learning rate = 0.03 with cosine learning rate schedule, moco-t = 0.2, multi-layer perceptron head, batch size 256, MoCo-v2 data augmentation, 200 epochs) to train the SSL model used in the system architecture of Figure 2B. For each held-out test set in Figure 3A, the corresponding SSL model was pre-trained using only the images from the training data. Certain image patches were oversampled using kernel density sampling as described elsewhere herein. Once the SSL pre-training was complete, all W × H patches were fed into the SSL pre-trained ResNet50 model, generating a W × H × 128 feature quilt for each image quilt. The final image model used for prediction was a two-layer CNN model with batch norm and dropout, taking the feature tensors as input. The final CNN model was trained with the Adam optimizer with a step learning rate scheduler, a batch size of 32, a maximum of 150 epochs, and a learning rate of 0.01.

[0163]

[0172] Downstream Pipeline. A joint fusion approach was used to leverage information from both modalities (image features and tabular features). Images were featurized into feature tensors and fed into the final image model to generate a feature vector, and the tabular features were fed separately into the TabNet model to generate another feature vector. Two fully connected layers processed the concatenated feature vectors from each pipeline and output a prediction probability. For patients with missing histopathology data, the image-based feature vectors were zeroed out before concatenation.

[0164]

[0173] Dataset preparation. All patient-level baseline clinical data, digital histopathology slides of prostate biopsies, and longitudinal outcomes from five landmark, large-scale, prospective, randomized, international clinical trials involving 5,654 patients, 16,204 histopathology slides, and a median patient follow-up of 10 to 20 years were used. These trials were RTOG 9202, 9408, 9910, 0126, and 9413 (Figure 2C). These trials randomized patients with various combinations of radiation therapy (RT) and androgen deprivation therapy (ADT): RT+short-term ADT (RTS), RT+intermediate ADT (RTM), RT+long-term ADT (RTL), and RT dose and volume levels (RT+). Slides were digitized by NRG Oncology using a Leica Biosystems Aperio AT2 digital pathology scanner at 20x resolution over a 2-year period. Histopathology images were manually reviewed for image quality and clarity. Digital slides were converted into a single image quilt of 200 × 200 patches for each unique patient prior to model training. The clinical variables collected in each trial were subtly different. Six clinical variables that were available in all trials (combined Gleason, Gleason 1, Gleason 2, t-stage, baseline PSA, age) and digital histopathology were used to train and validate the models.

[0165]

[0174] Tissue Segmentation.After slicing slides into 256x256 pixel patches at 10x zoom, an artifact classifier was developed by training a ResNet-18 to classify whether a patch showed usable tissue or was blank or showed artifacts. The artifact classifier was trained for 25 epochs and optimized using SGD with a learning rate of 0.001. The learning rate was decreased by 10% every 7 epochs. 3661 patches (tissue vs. non-tissue) were manually annotated, of which the classifier was trained on 3366, achieving a validation accuracy of 97.6% on the remaining 295. This artifact classifier was then used to segment tissue sections during image quilting.

[0166]

[0175] Nuclear Density Sampling. Due to the large variability in staining intensity and stain degradation, readily available pre-trained models for nuclei detection and segmentation were unable to accurately detect nuclei in the majority of slides. To overcome this, we trained a nuclear detector using the object detection method from YOLOv5 (github.com / ultralytics / yolov5).

[0167]

[0176] To train the YOLOv5 model, a selection of representative 34 sample slides were manually labeled using the QuPath image analysis platform. First, tissues were segmented using the "Simple tissue detection" module. Next, cells were segmented using the "Watershed cell detection" module, with parameters manually tuned for each slide. A YOLOv5-Large model was then trained on annotations from 29 slides and evaluated on the remaining 5. The model was trained using 256 × 256 patches at 10x zoom.

[0168]

[0177] Model performance metrics (AUC). For each model and each outcome, we estimated the time-dependent receiver operating characteristic that takes into account competing events using the R-package timeROC. This is a time-dependent sensitivity and specificity curve, calculated by sweeping a threshold t in the interval [0,1] and defining the model prediction as

[0169]

number

[0170]

[0178] NCCN Model. The NCCN model was coded according to the algorithm in Figure 8 and used three clinical variables, namely Gleason, t-stage, and baseline PSA, to stratify patients into low, intermediate, and high risk groups.

[0171]

[0179] Example 2 - Predictive AI biomarkers for prostate cancer treatment

[0172]

[0180] A study was conducted to determine biomarkers predicting response to androgen deprivation therapy (ADT) in addition to radiation therapy. The study acquired, digitized, trained, and validated five international phase III randomized trials with long-term follow-up to predict which men with localized prostate cancer would derive greater benefit from the addition of ADT to radiation therapy.

[0173]

[0181] <Method>

[0174]

[0182] Patients and studies. A predefined analysis plan was approved through NRG Oncology, a National Clinical Trials Network (NCTN) group funded by the National Cancer Institute (NCI). Studies were included if they were randomized phase III trials, had a median follow-up of more than 8 years, were conducted in men with localized prostate cancer, were treated with radiotherapy with or without ADT, and had pathology slides stored in the NRG biobank. Studies that examined the use of chemotherapy were excluded. This identified five randomized phase III trials (NRG / RTOG9202, 9408, 9413, 9910, and 0126) [refs 7-11].

[0175]

[0183] Details of each trial can be found in Table 1. Briefly, NRG / RTOG9202 enrolled men with intermediate- and high-risk prostate cancer and randomized patients to receive 4 months of radiation therapy with ADT versus 24 months of radiation therapy with ADT. NRG / RTOG9408 enrolled men with low-, intermediate-, and high-risk prostate cancer and randomized patients to receive 4 months of radiation therapy with ADT versus no ADT. NRG / RTOG9413 was a 2 × 2 factorial trial that enrolled men with intermediate- and high-risk prostate cancer and randomized patients to receive a 4-month ADT sequence and pelvic nodal radiation therapy. NRG / RTOG9910 enrolled men with intermediate-risk prostate cancer and randomized patients to receive 4 months of radiation therapy with ADT versus 9 months of radiation therapy with ADT. NRG / RTOG0126 randomized intermediate-risk patients to receive low-dose versus high-dose radiation therapy alone. Trials including ADT use consisted of androgen deprivation combined with LHRH agonists and antiandrogens. Patients with missing data (clinical data or available histopathology slides) were excluded from all analyses.

[0176] [Table 1-1] [Table 1-2] [Table 1-3]

[0177]

[0184] In some embodiments, the examples describe the development and validation of predictive biomarkers that can identify the differential benefit of adding ADT to radiotherapy in localized prostate cancer.One endpoint that defines benefit is the time to distant metastasis, which is measured from randomization to the onset of distant metastasis or last follow-up.Additional endpoints include prostate cancer-specific mortality (PCSM), metastasis-free survival (MFS) and overall survival (OS).

[0178]

[0185] Histopathology Pipeline. Pre-treatment biopsy slides from the NRG Oncology Biospecimen Bank were independently digitized without access to clinical outcome data. Slides were digitized at 20x resolution using a Leica Biosystems Aperio AT2 digital pathology scanner. Histopathology images were manually reviewed for image quality and clarity.

[0179]

[0186] Digital histopathology images were sent to a team of AI scientists for feature extraction, blinded to the clinical data at this point. For each unique patient, tissues from all of their digital slides were cropped and stitched into a single image quilt of size 51,200 × 51,200 pixels. The image quilt was divided into 256 × 256 pixel patches, resulting in 200 × 200 patches per image quilt. A YOLOv5 object detection model was trained to identify nuclei in the histopathology images and calculate the number of nuclei per patch for downstream training.

[0180]

[0187] A Resnet-50 object classification model was trained on these image patches using self-supervised learning (SSL) [Ref. 16]. The training protocol of MoCo-v2 was employed without access to clinical or outcome data [Ref. 17]. Over 6 million tissue patches from the entire training dataset were passed through the model over 100 times. Image patches with high nuclei counts were oversampled to learn salient histopathological features for downstream prediction.

[0181]

[0188] Model development and validation. The training cohort for model development included patients from NRG / RTOG9910 and 0126. Because NRM / RTOG9910 and each patient contributed to one treatment type of interest (RT+ST-ADT vs. RT only, respectively), inverse probability weighting (IPTW) of treatment was used to ensure that selected baseline clinical characteristics, such as age, baseline PSA, T stage, Gleason score, and primary / secondary pattern, were comparable in patients from the two treatment types. The model development cohort was then further stratified by treatment type and randomly split into a training (60%) and tuning (40%) set for model training and hyperparameter tuning, respectively [18, 19]. Clinical data, imaging data, and treatment type (rx) were used as inputs to the multimodal predictive model architecture (schematically illustrated in Figure 2C).

[0182]

[0189] Clinical data preprocessing. Categorical clinical variables (T stage, Gleason score, primary / secondary Gleason pattern) and binary treatment type (radiotherapy [RT] alone = 0, RT with ST-ADT = 1) were passed through a neural network embedding layer to generate continuous vector embeddings. Groupings were as follows: total Gleason (≤6, 7, 8, ≥9), both primary and secondary Gleason pattern (≤3, 4, 5), T stage (Tx, T0, T1a, T1b, T1c or T1, T2a, T2b, T2c or T2, T3a, T3b, T3c or T3, T4a, T4b, T4). Continuous clinical variables (age, baseline PSA) were standardized based on the mean and standard deviation of the training data.

[0183]

[0190] Image feature extraction model development. For each patch in a patient's histopathology image, a 128-dimensional feature vector was extracted using a self-supervised pre-trained Resnet-50 image feature extraction model and standardized based on the mean and standard deviation of the training data. All patch-level feature vectors from the same patient were stacked to form an image feature tensor, which was fed into the downstream prediction model.

[0184]

[0191] Inverse probability of treatment weighting. Because the development set included two phase III randomized trials (NRG / RTOG9910 and 0126), inverse probability of treatment weighting (IPTW) was used to ensure that patients had comparable clinical baseline characteristics in the two treatment types. Propensity scores were calculated using a logistic regression model with elastic net penalties, and treatment type was regressed against patient age, baseline PSA, Gleason score, Gleason primary / secondary pattern, and T stage variables. To mitigate the high variability introduced by large weights, IPTW weights were trimmed based on the 1st and 99th percentiles.

[0185]

[0192] Downstream prediction model development

[0186]

[0193] The downstream prediction model took as input the image feature tensor, preprocessed clinical data, and treatment type (rx) for each patient (Figure 2C). An attention multi-instance learning network was employed to learn weights for each patch from a patient. From each patient's image feature tensor, a single 128-D image vector was generated by taking the weighted sum of the image vectors of all patches from the same patient, where the weights were learned by an attention mechanism. This single 128-D image vector, preprocessed clinical data, and treatment type were concatenated and further processed in a joint fusion pipeline to effectively learn a predictive feature encoding of the differential treatment benefit of adding ST-ADT to radiotherapy.

[0187]

[0194] The multimodal prediction model was trained in a multitask manner. The first task was to predict the relative risk of DM using the facts rx ("Task 1" in Figure 2C). The image, clinical, and facts rx vectors were concatenated and passed through a fully connected neural network with several layers to generate a continuous score for each patient that estimates the relative risk of DM (sometimes referred to herein as the "fact model prediction score"). We used the negative log partial likelihood as the training objective for the first task, and the fact model prediction score was the estimated relative log hazard.

[0188]

[0195] The negative log partial likelihood loss is parameterized by the model weights θ and formulated as follows:

[0189]

number

[0190]

[0196] Here, the value T i , E i , x i are the event time or last follow-up time, an indicator variable for whether an event was observed, and the model input for the i-th observation, respectively. The function f θrepresents the fact branch of the multimodal model, and f θ (x) is the estimated relative risk given the input x. Value N E=1 represents the number of patients with an observable event. The set of patients with an observable event is E i =1. Risk Set

number

[0191]

[0197] Based on the relative risk estimated in the first task, the second task was to predict the delta score, defined as the difference between the factual model prediction score and the counterfactual model prediction score ("Task 2" in Figure 2C). For this purpose, a counterfactual rx variable was created by switching the patient's factual rx (RT for patients who received RT together with short-term ADT and vice versa). The counterfactual rx variable was passed through the same rx embedding layer and concatenated with the image and clinical vectors. The concatenated vector was then passed through the same fully connected neural network layer to obtain another continuous score (sometimes referred to herein as "counterfactual model prediction score"). For patients who received RT alone, delta was the factual model prediction score minus the counterfactual prediction score, and for patients who received RT+ST-ADT treatment, delta was the counterfactual prediction score minus the factual model prediction score. Delta indicates the magnitude of treatment benefit for each patient, with a larger delta suggesting a greater benefit from the addition of ST-ADT and vice versa.

[0192]

[0198] For this prediction task, delta loss was designed and used as a training objective. Delta loss characterized the deviation between predicted delta score and expected delta score. The expected delta score was according to which of the four subgroups the patient was included in based on the treatment type and DM outcome: (a) subgroup A, patients without metastases who received RT alone; (b) subgroup B, patients with DM who received RT alone; (c) subgroup C, patients without metastases who received RT with ST-ADT; and (d) subgroup D, patients with DM who received RT with ST-ADT. In subgroup A, delta was expected to be close to 0 because patients were free of DM when receiving RT alone and additional ST-ADT would not affect their risk of DM. In subgroup B, delta was expected to be greater than 0 because patients may benefit from additional ST-ADT treatment; in subgroup C, delta was expected to be greater than 0 because patients did not show DM even with additional ST-ADT. Finally, delta for subgroup D was expected to be close to 0 since patients had DM even after receiving additional ST-ADT treatment. During training, models were penalized if their delta scores did not fall within the expected ranges mentioned above. The training objective for the prediction task was defined using the softplus function.

[0193]

[0199] During training, a weighted sum of both losses from the prognostic task and the predictive task was fitted, and each data point was weighted by its IPTW weight. After the model was trained, a cutoff was selected at the 67th percentile of the delta score in the development set, and all patients in the validation set with a delta score greater than the cutoff were considered biomarker positive and predicted to benefit from the addition of ST-ADT, and patients with a delta score less than the cutoff were considered biomarker negative. The final model was selected based on the minimum ratio of IPTW-weighted hazard ratios between the biomarker-positive and -negative subgroups in the tuning set.

[0194]

[0200] The multimodal prediction model optimized the difference in magnitude of ADT benefit outputting the continuous score delta described herein above. The 67th percentile of delta score in the development set was selected as the cutoff threshold because it maximized the difference in treatment effect of biomarker subgroups in the tuning set and resulted in reasonably sized biomarker subgroups for clinical utility. Patients with delta scores greater than the cutoff were identified as biomarker positive and vice versa (Figure 9). After the biomarkers were locked, they were provided to an independent biostatistician to perform clinical validation of the biomarkers in NRG / RTOG9408.

[0195]

[0201] <Statistical analysis>

[0196]

[0202] Validation cohort characteristics according to biomarker status were reported and compared using chi-squared tests or Fisher's exact tests for low cell counts for categorical variables and Wilcoxon rank-sum tests for continuous variables. Time to DM and PCSM was analyzed using cumulative incidence functions with death without corresponding events as the competing risk. Outcomes in patients treated with and not treated with ST-ADT were compared using Gray's test [Ref. 20]. Fine and Gray regression was also performed to estimate subdistribution hazard ratios (sHRs) and 95% confidence intervals (CIs) for ST-ADT treatment effects [Ref. 21]. Tests of biomarker-treatment interactions were performed to evaluate this predictive biomarker. Treatment effects in biomarker-positive and -negative subgroups were evaluated as in the entire validation cohort to measure the relative treatment effects between groups. 15-year restricted mean survival times were reported to provide alternative estimates in cases where nonproportional hazards were observed [Ref. 10].

[0197]

[0203] Exploratory subgroup analyses were performed based on NCCN risk stratification, and the primary analysis was reanalyzed in NCCN low- and intermediate-risk patients. Statistical analyses were performed using R version 3.5.1 (R Foundation for Statistical Computing, Vienna, Austria). All statistical tests were two-sided, and a p value of <0.05 was considered statistically significant.

[0198]

[0204] result

[0199]

[0205] Patient characteristics. Of 7,752 eligible patients enrolled in the five phase III randomized trials, 5,825 (75.1%) patients had available pre-procedure biopsy tissue in the NRG Biospecimen Bank. Of these patients, 5,727 (98.3%) were eligible with high-quality digital pathology data. Of these patients, 39 patients with TURP specimens were further excluded from the validation cohort (NRG / RTOG9048).

[0200]

[0206] The downstream prediction model development cohort included 2,024 patients with a median follow-up of 10.6 years, with 1,050 (52%) patients receiving radiotherapy alone and 974 (48%) patients receiving RT with ST-ADT (Table 2). Median PSA was 9 ng / mL (interquartile range [IQR], 6–13), 87% had intermediate-risk disease, and median age was 71 years (IQR, 65–74). The validation cohort (NRG / RTOG9408) included 1,594 patients with a median follow-up of 14.9 years, with balanced treatment type (RT alone = 806 patients, RT + ST-ADT = 788 patients; Figure 10 and Table 3). Median PSA was 8 ng / mL (IQR, 6–12), 56% had intermediate-risk disease, and median age was 71 years (IQR, 66–74). There were no significant differences in baseline characteristics between treatment types, and the evaluable cohort of NRG / RTOG9408 was representative of cohorts across published trials (Table 3).

[0201] [Table 2]

[0202] [Table 3-1] [Table 3-2]

[0203]

[0207] ST-ADT predictive biomarkers

[0204]

[0208] The final rock model included 128 image features from digital pathology slides and seven clinical variables, including age, combined Gleason score, primary and secondary Gleason, PSA, T stage, and treatment type (RT with or without ST-ADT). Treatment type was used only for model development. Histopathology image features contributed nearly 40% of the model predictions (Figure 11).

[0205]

[0209] Applying the locked prediction model to the validation set, 543 patients (34%) were biomarker positive (predicted to benefit most from ST-ADT) and 1,051 patients (66%) were biomarker negative (predicted to benefit least from ST-ADT). Patients with biomarker positive disease were less likely to have a Gleason Score of 7 (24% vs. 30%; p=0.02) compared with patients with biomarker negative disease (Table 4).

[0206] [Table 4]

[0207]

[0210] In the entire validation cohort, ST-ADT significantly extended the time to DM (sHR 0.64, 95% CI [0.45-0.90], p=0.01; Figure 13). When applying the locked AI-derived biomarkers to the validation set, the biomarker-treatment interaction was significant for time to DM (p-interaction=0.01; Figure 14). In patients with biomarker-positive disease, the addition of ST-ADT significantly reduced the risk of DM compared with radiotherapy alone (sHR 0.34, 95% CI [0.19-0.63], p<0.001). In contrast, there was no significant difference between groups in the biomarker-negative subgroup (sHR 0.92, 95% CI [0.59-1.43], p=0.71). The absolute benefit of ST-ADT, measured as the difference in DM risk between treatment arms at 15 years after randomization, was 10.5% in biomarker-positive patients (95% CI 5.4% to 15.5%; Figure 13). In contrast, in patients with biomarker-negative disease, the addition of ADT reduced the 15-year DM risk by 0.5% (95% CI -2.8% to 3.7%).

[0208]

[0211] In an exploratory subset analysis, results remained similar when the analysis was restricted to only patients with low- and intermediate-risk disease. Furthermore, a similar biomarker x treatment interaction for DM was measured in this subgroup of patients (p-interaction = 0.009; Figure 15).

[0209]

[0212] The secondary endpoint, PCSM, was also evaluated (Figures 14 and 15). In the entire validation cohort, ST-ADT significantly prolonged time to events (sHR 0.52, 95% CI [0.35-0.78], p=0.001; Figure 15). Patients with biomarker-positive disease who received ST-ADT had significantly improved PCSM compared with radiotherapy alone (sHR 0.28, 95% CI 0.14-0.57, p<0.001). In contrast, despite the biomarker-negative group being significantly larger with more events than the biomarker-positive group, PCSM was not significantly improved with the addition of ADT compared with RT alone, sHR 0.74, 95% CI 0.45-1.22, p=0.24.

[0210]

[0213] The use of AI in oncology may still remain in its infancy with limited clinical deployment. Using a novel deep learning methodology and leveraging imaging data from over 5,000 patients across five randomized Phase III trials, we successfully trained and validated predictive biomarkers to guide the use of ADT in men treated with radiation therapy for localized prostate cancer. These findings are a long-awaited milestone in prostate cancer and can be used to rapidly advance precision medicine for this common disease.

[0211]

[0214] Across cancer types, histopathological information is primarily used for diagnostic and tumor grading purposes. In prostate cancer, Gleason grading can be modest in prognosis and has not been shown to function as a predictive biomarker for ADT use [Ref. 22]. Thus, a number of tissue-based gene expression, serum, and imaging biomarkers have been generated to fill this unmet need [Ref. 12]. Although some have shown improvements in risk stratification and prognosis [Ref. 23], none have been shown to function as predictive biomarkers in randomized trial validation.

[0212]

[0215] As patients' prognosis worsens (e.g., from low to high NCCN risk), the recommendation to add ADT to radiation therapy strengthens, despite evidence that NCCN risk group is not predictive of ADT benefit [5]. Thus far, this example demonstrates that biomarker-positive and -negative patients have similar baseline PSA, T stage, and NCCN risk group distributions, with only minor differences in Gleason grade, reinforcing the notion that historical measures of tumor aggressiveness appear to have little correlation with which patients will derive differential relative benefit from ADT.

[0213]

[0216] Despite the limitations of Gleason grading, histopathology images can be of significant value. There is a wealth of information in histopathology images that far exceeds the current five grade groups. However, to leverage this information, this example demonstrates that AI can be used to unlock this vast amount of quantifiable data from digital pathology without limiting models to only established human-interpretable features.

[0214]

[0217] In general, models may encounter difficulties with possible overfitting and validation failures. Therefore, independent validation may be required to prove the performance of the biomarker. In the specific case of a predictive biomarker that aims to provide information about which patients will derive relatively higher or lower benefit, this can be performed within the context of a randomized trial of the treatment of interest to avoid confounding and bias between groups. Here, NRG / RTOG9408 was chosen because it is the largest published trial of radiotherapy with or without ST-ADT with a very long follow-up. Although this trial found a clear benefit of ADT in unselected patients, the majority of the enrolled patients did not have a demonstrable benefit. The results indicate that more than 60% of the intermediate-risk patients enrolled in NRG / RTOG9408 could avoid unnecessary treatment with ADT and potentially be spared the morbidity and costs associated with ADT.

[0215]

[0218] The primary endpoint, time to DM, was carefully selected to train ST-ADT predictive biomarkers. Other endpoints, such as biochemical recurrence, metastasis-free survival (MFS), and OS, are all clinically relevant but have notable limitations for biomarker development in the context of localized prostate cancer. Because ADT suppresses PSA production, it is expected to delay biochemical recurrence regardless of subgroup. Furthermore, the majority of biochemical recurrence events do not result in metastasis or death [Ref. 22]. Therefore, it may be a less optimal endpoint for biomarker training to determine intrinsic tumor-specific benefit from ADT.

[0216]

[0219] MFS and OS are important endpoints for determining the net effect of a given treatment and are the gold standard for clinical trial design. However, they may be suboptimal endpoints for the development of prostate cancer-specific biomarkers, as 78% of deaths in the validation cohort were not due to prostate cancer, and only 12% of events in the MFS endpoint were due to metastatic events. Thus, the strongest predictive models for MFS and OS may actually be driven by variables (e.g., comorbid conditions) associated with death from causes other than prostate cancer. Importantly, even though the biomarkers were trained for DM, the cancer-driven endpoint PCSM showed clearly differential effects of ADT depending on biomarker status.

[0217]

[0220] As with any model, generalizability can be critical. Difficulties can arise with AI models derived from a limited number of sites and less diverse cohorts. In contrast, NRG / RTOG enrolled patients from over 500 sites, primarily school, community, and Veterans Affairs sites across the United States and Canada, with a good representation of racial minorities (20% of patients in the validation set were African American). This significant real-world diversity strengthens the generalizability of our results.

[0218]

[0221] Similar to prognostic and predictive biomarkers used in current clinical practice, the ST-ADT predictive biomarkers described herein (e.g., in this example) could be developed and validated as part of novel, prospective, biomarker-only trials. This approach is supported by Simon et al., and the use of a randomized trial of radiation therapy with or without ADT strengthens the credibility and level of evidence of our study [Ref. 24]. The use of advanced molecular imaging was virtually nonexistent during the time of this study's implementation and follow-up. Stage migration due to changes in the Gleason grading system may have also influenced the stratification of patients into NCCN risk groups. However, any potential bias introduced by this is likely to be random and affect both groups, and many of the models are driven by raw histopathology images and may not be affected by changes in grading definitions over time. Information on other prognostic clinicopathological variables, such as percentage Gleason pattern 4 or percent positive biopsy cores, was not available. Therefore, alternative risk classification schemes for exploratory analyses were not performed [Ref. 25, 26].

[0219]

[0222] This example describes the successful use of a novel AI-derived digital pathology-based platform to develop predictive biomarkers to guide the combination of radiotherapy and ADT in localized prostate cancer, which were independently validated in a phase III randomized trial, which could have safely avoided the morbidity and economic disadvantage associated with unnecessary ADT treatment in the majority of patients.

[0220]

[0223] <References>

[0221]

[0224] 1. Jones, CU et al. Adding Short-Term Androgen Deprivation Therapy to Radiation Therapy in Men With Localized Prostate Cancer: Long-Term Update of the NRG / RTOG 9408 Randomized Clinical Trial. Int. J. Radiat. Oncol. Biol. Phys. (2021) doi:10.1016 / j.ijrobp.2021.08.031, which is incorporated by reference in its entirety.

[0222]

[0225] 2. Pilepich, MV et al. Androgen suppression adjuvant to definitive radiotherapy in prostate carcinoma--long-term results of phase III RTOG 85-31. Int. J. Radiat. Oncol. Biol. Phys. 61, 1285-1290 (2005) (which is incorporated by reference in its entirety).

[0223]

[0226] 3. D'Amico, AV, Chen, M.-H., Renshaw, A., Loffredo, M. & Kantoff, PW Long-term Follow-up of a Randomized Trial of Radiation With or Without Androgen Deprivation Therapy for Localized Prostate Cancer. JAMA vol. 314 1291 (2015) (which is incorporated by reference in its entirety).

[0224]

[0227] 4. Bolla, M. et al. Short Androgen Suppression and Radiation Dose Escalation in Prostate Cancer: 12-Year Results of EORTC Trial 22991 in Patients With Localized Intermediate-Risk Disease. J. Clin. Oncol. 39, 3022-3033 (2021) (which is incorporated by reference in its entirety).

[0225]

[0228] 5. Kishan, AU et al. Androgen deprivation therapy use and duration with definitive radiotherapy for localised prostate cancer: an individual patient data meta-analysis. Lancet Oncol. 23, (2022) (which is incorporated by reference in its entirety).

[0226]

[0229] 6. Nguyen, PL et al. Adverse effects of androgen deprivation therapy and strategies to mitigate them. Eur. Urol. 67, 825-836 (2015) (which is incorporated by reference in its entirety).

[0227]

[0230] 7. Horwitz, EM et al. Ten-year follow-up of radiation therapy oncology group protocol 92-02: a phase III trial of the duration of elective androgen deprivation in locally advanced prostate cancer. J. Clin. Oncol. 26, 2497-2504 (2008) (which is incorporated by reference in its entirety).

[0228]

[0231] 8. Roach, M., 3rd et al. Phase III trial comparing whole-pelvic versus prostate-only radiotherapy and neoadjuvant versus adjuvant combined androgen suppression: Radiation Therapy Oncology Group 9413. J. Clin. Oncol. 21, 1904-1911 (2003) (which is incorporated by reference in its entirety).

[0229]

[0232] 9. Pisansky, TM et al. Duration of androgen suppression before radiotherapy for localized prostate cancer: radiation therapy oncology group randomized clinical trial 9910. J. Clin. Oncol. 33, 332-339 (2015) (which is incorporated by reference in its entirety).

[0230]

[0233] 10. Jones, CU et al. Adding Short-Term Androgen Deprivation Therapy to Radiation Therapy in Men With Localized Prostate Cancer: Long-Term Update of the NRG / RTOG 9408 Randomized Clinical Trial. Int. J. Radiat. Oncol. Biol. Phys. (2021) doi:10.1016 / j.ijrobp.2021.08.031, which is incorporated by reference in its entirety.

[0231]

[0234] 11. Michalski, JM et al. Effect of Standard vs Dose-Escalated Radiation Therapy for Patients With Intermediate-Risk Prostate Cancer. JAMA Oncology vol. 4 e180039 (2018) (which is incorporated by reference in its entirety).

[0232]

[0235] 12. Schaeffer, E. et al. NCCN Guidelines Insights: Prostate Cancer, Version 1.2021: Featured Updates to the NCCN Guidelines. J. Natl. Compr. Canc. Netw. 19, 134-143 (2021) (which is incorporated by reference in its entirety).

[0233]

[0236] 13. Tolkach, Y., Dohmgorgen, T., Toma, M. & Kristiansen, G. High-accuracy prostate cancer pathology using deep learning. Nature Machine Intelligence 2, 411-418 (2020) (which is incorporated by reference in its entirety).

[0234]

[0237] 14. Nagpal, K. et al. Development and Validation of a Deep Learning Algorithm for Gleason Grading of Prostate Cancer From Biopsy Specimens. JAMA Oncol 6, 1372-1380 (2020) (which is incorporated by reference in its entirety).

[0235]

[0238] 15. Pantanowitz, L. et al. An artificial intelligence algorithm for prostate cancer diagnosis in whole slide images of core needle biopsies: a blinded clinical validation and deployment study. Lancet Digit Health 2, e407-e416 (2020) (which is incorporated by reference in its entirety).

[0236]

[0239] 16. He, Kaiming et al. Deep Residual Learning for Image Recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 27-30, 2016, ieeexplore.ieee.org / document / 7780459, which is incorporated by reference in its entirety.

[0237]

[0240] 17. Chen, X., Fan, H., Girshick, R. & He, K. Improved Baselines with Momentum Contrastive Learning (2020) (which is incorporated by reference in its entirety).

[0238]

[0241] 18. Hutter, F., Kotthoff, L. & Vanschoren, J. Automated Machine Learning: Methods, Systems, Challenges (Springer, 2019) (which is hereby incorporated by reference in its entirety).

[0239]

[0242] 19. Claesen, M. & De Moor, B. Hyperparameter Search in Machine Learning (2015) doi:10.48550 / arXiv.1502.02127, which is incorporated by reference in its entirety.

[0240]

[0243] 20. Gray, R. J. A Class of K-Sample Tests for Comparing the Cumulative Incidence of a Competing Risk. Ann. Stat. 16, 1141-1154 (1988), which is incorporated by reference in its entirety.

[0241]

[0244] 21. Fine, JP & Gray, RJ A Proportional Hazards Model for the Subdistribution of a Competing Risk. J. Am. Stat. Assoc. 94, 496-509 (1999) (which is incorporated by reference in its entirety).

[0242]

[0245] 22. Jones, CU et al. Radiotherapy and short-term androgen deprivation for localized prostate cancer. N. Engl. J. Med. 365, (2011) (which is incorporated by reference in its entirety).

[0243]

[0246] 23. Spratt, DE et al. Development and Validation of a Novel Integrated Clinical-Genomic Risk Group Classification for Localized Prostate Cancer. J. Clin. Oncol. (2017) doi:10.1200 / JCO.2017.74.2940, which is incorporated by reference in its entirety.

[0244]

[0247] 24. Simon, RM, Paik, S. & Hayes, DF Use of archived specimens in evaluation of prognostic and predictive biomarkers. J. Natl. Cancer Inst. 101, 1446-1452 (2009) (which is incorporated by reference in its entirety).

[0245]

[0248] 25. Cooperberg, MR et al. The UCSF Cancer of the Prostate Risk Assessment (CAPRA) Score: a straightforward and reliable preoperative predictor of disease recurrence after radical prostatectomy. J. Urol. 173, 1938 (2005) (which is incorporated by reference in its entirety).

[0246]

[0249] 26. A New Risk Classification System for Therapeutic Decision Making with Intermediate-risk Prostate Cancer Patients Undergoing Dose-escalated External-beam Radiation Therapy. Eur. Urol. 64, 895-902 (2013) (which is incorporated by reference in its entirety).

[0247]

[0250] 27. Ilse, M., Tomczak, J. & Welling, M. Attention-based Deep Multiple Instance Learning. in Proceedings of the 35th International Conference on Machine Learning (eds. Dy, J. & Krause, A.) vol. 80 2127-2136 (PMLR, 10--15 Jul 2018) (which is incorporated by reference in its entirety).

[0248]

[0251] 28. Austin, PC & Stuart, EA Moving towards best practice when using inverse probability of treatment weighting (IPTW) using the propensity score to estimate causal treatment effects in observational studies. Stat. Med. 34, 3661-3679 (2015) (which is incorporated by reference in its entirety).

[0249]

[0252] 29. Cole, S. R. & Hernan, M. A. Constructing Inverse Probability Weights for Marginal Structural Models. Am. J. Epidemiol. 168, 656-664 (2008) (which is incorporated by reference in its entirety).

[0250]

[0253] 30. Lee, B. K., Lessler, J. & Stuart, E. A. Weight Trimming and Propensity Score Weighting. PLoS One 6, e18174 (2011) (which is incorporated by reference in its entirety).

[0251]

[0254] 31. Katzman, JL et al. DeepSurv: personalized treatment recommender system using a Cox proportional hazards deep neural network. BMC Med. Res. Methodol. 18, 1-12 (2018) (which is incorporated by reference in its entirety).

[0252]

[0255] 32. Breslow, N. Covariance analysis of censored survival data. Biometrics 30, 89-99 (1974) (which is incorporated by reference in its entirety).

[0253]

[0256] 33. Leen, TK, Dietterich, TG & Tresp, V. Advances in Neural Information Processing Systems 13: Proceedings of the 2000 Conference (MIT Press, 2001) (which is incorporated by reference in its entirety).

[0254]

[0257] 34. Shapley, LS 17. A Value for n-Person Games. in Contributions to the Theory of Games (AM-28), Volume II (eds. Kuhn, HW & Tucker, AW) 307-318 (Princeton University Press, 1953) (which is incorporated herein by reference in its entirety).

[0255]

[0258] While preferred embodiments of the present invention have been shown and described herein, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. The present invention is not intended to be limited by the specific examples provided herein. Although the present invention has been described with reference to the foregoing specification, the description and illustration of the embodiments herein should not be construed in a limiting sense. Numerous variations, changes, and substitutions will occur to those skilled in the art without departing from the present invention. Furthermore, it should be understood that all aspects of the present invention are not limited to the specific depictions, configurations, or relative proportions set forth herein which depend upon a variety of conditions and variables. It should be understood that various alternatives to the embodiments of the present invention described herein may be employed in the practical application of the present invention. Accordingly, the present invention is intended to cover such alternatives, modifications, variations, or equivalents. That is, it is intended that the following claims define the scope of the present invention, and that methods and structures within the scope of these claims and their equivalents are hereby covered.

Claims

1. A method for assessing cancer in a subject, comprising: (a) obtaining a dataset comprising at least image data obtained from or derived from the subject; (b) processing the dataset with a trained algorithm to determine an output indicative of a classification of the dataset into one of a plurality of categories, the processing comprising applying an image processing algorithm to the image data; (c) assessing the cancer based at least in part on the classification of the dataset into the categories, wherein the assessing step comprises determining biomarkers that predict response to a therapeutic intervention to treat the cancer in the subject; A method comprising:

2. The method described in claim 1, wherein the response includes overall survival or progression-free survival.

3. The method described in claim 1, wherein the response includes a decrease in mortality rate.

4. The method of claim 1, wherein the response includes metastasis-free survival, reduced metastasis, or reduced distant metastasis.

5. The method further comprises a step of determining whether the subject is biomarker-positive or biomarker-negative for the biomarker, wherein the step of determining whether the subject is biomarker-positive or biomarker-negative comprises: (i) calculating a first probability that the subject will exhibit the response in the presence of the therapeutic intervention; (ii) calculating a second probability that the subject will exhibit the response in the absence of the therapeutic intervention; and (iii) calculating a probability delta between the first probability and the second probability; (iv) comparing said probability delta with a reference standard; The method of claim 1 , comprising:

6. The method described in claim 5, wherein the subject is biomarker positive if the probability delta is higher than the reference standard, and the subject is biomarker negative if the probability delta is lower than the reference standard.

7. The method described in claim 5, wherein the reference standard is determined at least in part by measuring the median probability delta from multiple subjects.

8. The method described in claim 5, further comprising a step of treating the subject with the therapeutic intervention.

9. The method described in claim 8, wherein the therapeutic intervention includes androgen deprivation therapy (ADT).

10. The method according to claim 9, wherein the ADT is short-term ADT (ST-ADT).

11. The method of claim 1, wherein the trained algorithm comprises a self-supervised learning or deep learning algorithm.

12. The method of claim 1, wherein the dataset further includes tabular data, and the trained algorithms include a first trained algorithm that processes the image data and a second trained algorithm that processes the tabular data.

13. The method of claim 12, wherein the tabular data includes clinical data of the subject.

14. The method of claim 13, wherein the clinical data includes laboratory data, therapeutic intervention, or long-term outcome.

15. The method of claim 1, wherein the cancer comprises prostate cancer, bladder cancer, breast cancer, pancreatic cancer, or thyroid cancer.

16. The method described in claim 15, wherein the cancer includes prostate cancer.

17. The method of claim 1, wherein the image data includes digital histopathology data.

18. The method of claim 1, further comprising processing the image data using an image segmentation, image stitching, or object detection algorithm.

19. A non-transitory computer-readable medium comprising machine-executable code that, when executed by one or more computer processors, implements a method for assessing cancer in a subject, the method comprising: (a) obtaining a dataset comprising at least image data obtained from or derived from the subject; (b) processing the dataset with a trained algorithm to determine an output indicative of a classification of the dataset into one of a plurality of categories, the processing comprising applying an image processing algorithm to the image data; (c) assessing the cancer based at least in part on the classification of the dataset into the categories, wherein the assessing step comprises determining biomarkers that predict response to a therapeutic intervention to treat the cancer in the subject; 1. A non-transitory computer-readable medium, comprising:

20. A system including one or more computer processors and computer memory coupled to said computer processors, comprising: the computer memory comprises machine-executable code that, when executed by the one or more computer processors, implements a method for assessing cancer in a subject; The method comprises: (a) obtaining a dataset comprising at least image data obtained from or derived from the subject; (b) processing the dataset with a trained algorithm to determine an output indicative of a classification of the dataset into one of a plurality of categories, the processing comprising applying an image processing algorithm to the image data; (c) assessing the cancer based at least in part on the classification of the dataset into the categories, wherein the assessing step comprises determining biomarkers that predict response to a therapeutic intervention to treat the cancer in the subject; Including, the system.

Images