Prognostic methods and systems for chronic kidney disease

Abstract

Methods, systems and products for determining a prognosis for a subject with chronic kidney disease are disclosed, comprising obtaining the values of a plurality of proteomic features and / or clinical features for the subject and classifying the subject between a plurality of classes including at least a first class and a second class, the first class having better prognosis than the second class Methods of identifying groups of subjects with chronic kidney disease, and biomarkers of such groups are also described.

Description

[0001] PROGNOSTIC METHODS AND SYSTEMS FOR CHRONIC KIDNEY DISEASE

[0002] Field of Disclosure

[0003] The present disclosure relates to methods of determining a prognosis for a subject with chronic kidney disease, using clinical and / or proteomic signatures. Related methods, products and systems are also described.

[0004] Background

[0005] Chronic kidney disease (CKD) is a long-term condition, resulting in gradual loss of kidney function. The primary driver of CKD initiation in patients with diabetes, known as diabetic kidney disease (DKD) is chronic hyperglycemia (high blood sugar), damaging kidney blood vessels overtime, causing the blood vessels to narrow and resulting in restricted blood flow to the kidney. In the UK alone, an estimated 7.2 million adults have CKD, a figure which is estimated to increase to 7.6 million by 2033 (Kidney Research UK 2023). The global burden of DKD is increasing alongside that of diabetes, with the number of deaths attributed to DKD increasing by 94% between 1990 and 2012 (Alicic, Rooney, and Tuttle 2017).

[0006] The disease progression of DKD is a dynamic and fluctuating condition as opposed to a linear, predictable process (Alicic, Rooney, and Tuttle 2017). As disease progresses, the estimated glomerular filtration rate (eGFR) typically reduces, with eGFR levels lower than 15 ml / min per 1.73m2classifying end-stage renal disease (ESRD). Once patients develop ESRD, they are likely to require renal replacement therapy (RRT), including either dialysis or kidney transplant. The global burden of RRT is considerable, requiring specialised equipment and healthcare teams, as well as extensive facilities, with estimated costs of dialysis treatment at 2%-3% of annual healthcare budget in high-income countries (Li et al. 2021 ; Vanholder et al. 2017). In the UK, the annual cost of dialysis is estimated at £1.05 billion, equating to 0.53% of the NHS budget (Kidney Research UK 2023).

[0007] Identifying patients most at risk of developing ESRD and requiring RRT would enable potential preventative treatment interventions to be administered, reducing the morbidity, mortality and economic burden of DKD.

[0008] of the Disclosure

[0009] Identifying groups of patients at risk of severe CKD / DKD and adverse outcomes, such as dialysis, ensures that patients receive any appropriate, available treatment interventions. One computational biology strategy for supporting precision medicine drug discovery involves identifying subgroups (also known as endotypes) of patients with disease, with the goal of administering targeted therapies according to the patients’ underlying molecular profiles and risk of disease progression. Once patient endotypes are characterised, robust methodologies are required for assigning new patients to these endotypes to ensure appropriate treatment. The present inventors have developed classifier models for stratifying patients into multiple DKD patient endotypes that capture differential disease severity trajectories (herein referred to as prognostic models). Using baseline clinical information and high-throughput plasma proteomic profiling, they have identified novel combinations of biomarkers that can predict disease risk progression with high accuracy. In particular, they demonstrated that small combinations of proteins, measurable in blood, can stratify patients, enabling identification of DKD patients with elevated risk of rapid deterioration and disease progression. Furthermore, they showed that patients can also be stratified into categories associated with disease progression using routinely collected clinical measurements.

[0010] Thus, according to a first aspect, there is provided a method of determining a prognosis for a subject that has been identified as having or being likely to have chronic kidney disease, the method comprising performing one or both of: (i) obtaining the values of a first plurality of clinical features associated with the subject, wherein the first plurality of clinical features comprise at least two clinical features selected from a first set of clinical features comprising: a feature indicative of creatinine concentration in a blood or urine sample from the subject, a feature indicative of estimated glomerular filtration rate associated with a blood sample from the subject, a feature indicative of age of the subject, a feature indicative of a diastolic blood pressure measurement for the subject, a feature indicative of a systolic blood pressure measurement for the subject, a feature indicative of the presence or concentration of albumin in a blood sample from the subject, a feature indicative of phosphate concentration in a blood sample from the subject, a feature indicative of the ratio of protein to creatinine in a urine sample from the subject, and a feature indicative of whether the subject is undergoing treatment with calcium channel blockers; and classifying the subject between a plurality of classes including at least a first class and a second class, the first class having better prognosis than the second class using the values of the first plurality of clinical features and a classifier model that has been trained to classify subjects with chronic kidney disease between the first class and the second class using training data comprising values of the first plurality of clinical features for a plurality of training subject and associated labels indicating whether each training subject belongs to the first or second class; and (ii) obtaining the values of a second plurality of clinical features associated with the subject, wherein the second plurality of clinical features comprise at least two clinical features selected from a second set of clinical features comprising a feature indicative of creatinine concentration in a blood or urine sample from the subject, a feature indicative of age of the subject, a feature indicative of the presence or concentration of albumin in a blood sample from the subject, a feature indicative of whether the subject is undergoing treatment with calcium channel blockers, a feature indicative of CRP concentration in a blood sample from the subject, a feature indicative of calcium concentration in a blood sample from the subject, and a feature indicative of total cholesterol concentration in a blood sample from the subject; classifying the subject between a plurality of classes including at least a third class and a fourth class, the third class having better prognosis than the fourth class using the values of the second plurality of clinical features and a classifier model that has been trained to classify subjects with chronic kidney disease between the third class and the fourth class using training data comprising values of the second plurality of clinical features for a plurality of training subject and associated labels indicating whether each training subject belongs to the third or fourth class. The method may have any one or more of the following optional features.

[0011] The first plurality of clinical features may include at least the feature indicative of creatinine concentration in a blood or urine sample from the subject, and / or the feature indicative of estimated glomerular filtration rate associated with a blood sample from the subject. The first plurality of clinical features may include at least the feature indicative of creatinine concentration in a blood or urine sample from the subject, and the feature indicative of the ratio of protein to creatinine in a urine sample from the subject. The second plurality of clinical features may include at least the feature indicative of creatinine concentration in a blood or urine sample from the subject.

[0012] The first plurality of clinical features may include: (a) the feature indicative of creatinine concentration in a blood or urine sample from the subject, and the feature indicative of age of the subject; or (b) the feature indicative of creatinine concentration in a blood or urine sample from the subject, the feature indicative of estimated glomerular filtration rate associated with a blood sample from the subject, the feature indicative of age of the subject, the feature indicative of a diastolic blood pressure measurement for the subject, the feature indicative of the presence or concentration of albumin in a blood sample from the subject, the feature indicative of phosphate concentration in a blood sample from the subject, and the feature indicative of whether the subject is undergoing treatment with calcium channel blockers; (c) the feature indicative of creatinine concentration in a blood or urine sample from the subject, the feature indicative of the ratio of protein to creatinine in a urine sample from the subject, and the feature indicative of age of the subject, or (d) all of the features in the first set of features.

[0013] The second plurality of clinical features may include: (a) the feature indicative of whether the subject is undergoing treatment with calcium channel blockers, and the feature indicative of age of the subject; or (b) all of the features in the second set of features.

[0014] In embodiments, in the first set of clinical features: the feature indicative of creatinine concentration in a blood or urine sample from the subject, the feature indicative of a diastolic blood pressure measurement for the subject, the feature indicative of a systolic blood pressure measurement for the subject, the feature indicative of phosphate concentration in a blood sample from the subject, and the feature indicative of whether the subject is undergoing treatment with calcium channel blockers are expected to be higher in subjects in the second class than in subjects in the first class; and the feature indicative of estimated glomerular filtration rate associated with a blood sample from the subject, the feature indicative of age of the subject, and the feature indicative of the presence or concentration of albumin in a blood sample from the subject are expected to be lower in subjects in the second class than in subjects in the first class.

[0015] In embodiments, in the second set of clinical features, the feature indicative of creatinine concentration in a blood or urine sample from the subject, the feature indicative of whether the subject is undergoing treatment with calcium channel blockers, and the feature indicative of total cholesterol concentration in a blood sample from the subject are expected to be higher in subjects in the fourth class than in subjects in the third class; and the feature indicative of age of the subject, the feature indicative of the presence or concentration of albumin in a blood sample from the subject, the feature indicative of CRP concentration in a blood sample from the subject, and the feature indicative of calcium concentration in a blood sample from the subject are expected to be lower in subjects in the fourth class than in subjects in the third class.

[0016] The method of any aspect may be computer implemented. According to any aspect, obtaining protein levels may comprise receiving protein level data that has previously been acquired.

[0017] According to a second aspect, there is provided a method of determining a prognosis for a subject that has been identified as having or being likely to have chronic kidney disease, the method comprising performing one or both of: (i) obtaining the values of a first plurality of proteomic features associated with the subject, wherein the first plurality of proteomic features comprise at least two proteomic features selected from a first set of proteomic features comprising features indicative of the protein level in a sample from the subject of at least two different proteins selected from: PTGDS or PGF orTNFRSF14, F11R or CD40, CST3 or IGFBP4, SDC1 or EPHA2, PILRA or PILRB, FAS or IGFBP6 or ESAM (optionally FAS or ESAM, or FAS), CD160 or JAM2 or NCRI , TFPI or PROSI orAPCS (optionally TFPI orAPCS, optionally TFPI), CPA1 or CPBI , CDH3 orTNFRSF14, GFOD2, SUSD1 or PLXNA4, ADH4 or UPB1 , KLK11 or NECTIN4, ATRAID or EFCAB14, DBI or TMSB10, TNFRSF14 or TNFRSF1A, DSC2 orTNFRSFIB, CD274 or TNFRSFIB, TNFRSF11 A or TNFRSF1 A, GPR37 or lGFBP4, THBD or ESAM, ENTPD5, NPDC1 , HSPG2, IGFBP6, PGF, LY6D, BTN2A1 , AGRN, TXNDC5, JAM2, PIK3IP1 , RELT, IGFBP4, TNFRSF19, WFDC2, ATRAID, EFCAB14, TFPI, HAVCR1 , and MXRA8; and classifying the subject between a plurality of classes including at least a first class and a second class, the first class having better prognosis than the second class using the values of the first plurality of proteomic features and a classifier model that has been trained to classify subjects with chronic kidney disease between the first class and the second class using training data comprising values of the first plurality of proteomic features for a plurality of training subject and associated labels indicating whether each training subject belongs to the first or second class; and (ii) obtaining the values of a second plurality of proteomic features associated with the subject, wherein the plurality of proteomic features comprise at least two proteomic features selected from a second set of proteomic features comprising features indicative of the protein level of at least two proteins selected from: LAIR2 orTREH, SELPLG or HAO1 , FGFR2 or NBL1 , TNFRSF11 A or TNFRSF1 A, PM20D1 or TOP1 in a sample from the subject; and classifying the subject between a plurality of classes including at least a third class and a fourth class, the third class having better prognosis than the fourth class using the values of the second plurality of proteomic features and a classifier model that has been trained to classify subjects with chronic kidney disease between the third class and the fourth class using training data comprising values of the second plurality of proteomic features for a plurality of training subject and associated labels indicating whether each training subject belongs to the third or fourth class.

[0018] The method may have any one or more of the following optional features.

[0019] The first plurality of proteomic features may comprise features indicative of the protein level of:

[0020] PTGDS and at least one protein different from PTGDS selected from F11R, CD40, CST3, IGFBP4, SDC1 , EPHA2, PILRA, PILRB, FAS, IGFBP6, ESAM, CD160, JAM2, NCR1 , TFPI, PROS1 , APCS, CPA1 , CPB1 , CDH3, TNFRSF14, GFOD2, SUSD1 , PLXNA4, ADH4, UPB1 , KLK11 , NECTIN4, ATRAID, EFCAB14, DBI, TMSB10, TNFRSF14, TNFRSF1A, DSC2, TNFRSF1B, CD274, TNFRSF1B, TNFRSF11A, TNFRSF1A, GPR37, IGFBP4, THBD, ESAM, ENTPD5, NPDC1 , HSPG2, IGFBP6, PGF, LY6D, BTN2A1 , AGRN, TXNDC5, JAM2, PIK3IP1 , RELT, IGFBP4, TNFRSF19, WFDC2, ATRAID, EFCAB14, TFPI, HAVCR1 , and MXRA8;

[0021] two or more or all of: PTGDS or PGF, F11 R or CD40, SDC1 or EPHA2, PILRA or PILRB, two or more or all of: CD160 or JAM2, and TFPI or PROSI orAPCS (optionally TFPI orAPCS, optionally TFPI); or two or more or all of: SDC1 , TFPI, PTGDS, CD160, PILRA, and F11 R; or

[0022] two or more or all of: CD160 or JAM2, and F11R or CD40; or

[0023] two or more or all of: two or more or all of: CD160 and F11 R; or

[0024] two or more or all of: SDC1 or EPHA2, TNFRSF11A orTNFRSFIA, and CD160 or JAM2; or two or more or all of: PTGDS or PGF, CPA1 or CPB1 , CDH3 of TNFRSF14, SUSD1 or PLXNA4, GFOD2, ADH4 or UBP1 , and KLK11 or NECTIN4, optionally wherein the two or more include at least PTGDS or PGF, and / or at least GFOD2 and / or at least CPA1 or CPAB1 , and / or at least KLK11 or NECTIN4; or

[0025] two or more or all of: PTGDS or PGF or TNFRSF14, TNFRSF14, DBI or TMSB10, THBD or ESAM, DSC2 or TNFRSF1B, TNFRSF11A or TNFRSF1A, GPR37 or IGFBP4, and ENTDP5, optionally wherein the two or more include at least PTGDS or PGF or TNFRSF14, and / or at least DBI or TMSB10, and / or at least THBD or ESAM, and / or at least DSC2 or TNFRSF1B, and / or at least TNFRSF11A or TNFRSF1A; or

[0026] two or more or all of: PTGDS or TNFRSF14, CD160 or NCR1 , ATRAID or EFCAB14, DBI or TMSB10, TNFRSF14 orTNFRSFIA, DSC2 orTNFRSFIB, and CD274 orTNFRSF14, optionally wherein the two or more include at least PTGDS or TNFRSF14, and / or at least DSC2 orTNFRSFIB; or ATRAID, and PTGDS; or

[0027] PTGDS, and CD160 or JAM2; or

[0028] PTGDS, and DBI; or

[0029] PTGDS, and TNFRSF14; or

[0030] ATRAID, and CD160 or JAM2; or

[0031] ATRAID, and DBI; or

[0032] ATRAID, PTGDS, and DBI; or

[0033] PTGDS, and DSC2; or

[0034] ATRAID, PTGDS, and CD160 or JAM2; or

[0035] PTGDS, and CD274; or

[0036] PTGDS, HSPG2, and NDPC1 ; or

[0037] PTGDS, ATRAID, BTN2A1 and PIK3IP1 , or

[0038] at least 3 proteins selected from NPDC1 , HSPG2, PTGDS, IGFBP6, PGF, LY6D, BTN2A1 , TNFRSF14, AGRN, TXNDC5, JAM2, PIK3IP1 , RELT, TNFRSF1A, IGFBP4, TNFRSF19, NECTIN4, WFDC2, ATRAID, and EFCAB14.

[0039] The second plurality of proteomic features may comprise features indicative of the protein level of: (i) LAIR2 or TREH, SELPLG or HAO1 , FGFR2 or NBL1 , PM20D1 or TOP1 ; or (ii) PM20D1 , SELPLG, LAIR2, and FGFR2; or (iii) LAIR2 or TREH, and PM20D1 or TOPI ; or (iv) PM20D1 and LAIR2; or (v) SELPLG or HAO1 , and LAIR2 or TREH.

[0040] The first plurality of proteomic features may comprise features indicative of the protein level of one or more or all of: (i) PTGDS or PGF or TNFRSF14, (ii) TNFRSF11 A or TNFRSF1 A, (iii) DBI or TMSB10, (iv) ENTPD5, DSC2 or TNFRSFI B.

[0041] In embodiments, the first plurality of proteomic features further comprises features indicative of the protein level of any one or more of: ATRAID, EFCAB14, DBI, TMSB10, TNFRSF14, TNFRSF1B, DSC2, CD274, ENTPD5, PTPRS, ANG, CXADR, TNFRSF11A, GKN1 , DSG4, PVALB, DSG3, KDR, WFDC2, MIA, DDA1 , COX6B1 , TOMM20, CFP, EFNB2, LCN15, CHCHD10, SCRIB, AMN, CXCL5, RARRES2 and INHBC. In embodiments, the second plurality of proteomic features further comprises features indicative of the protein level of any one or more of: TSHB, ANGPTL1 , CA3, ANG, SCGB1A1 , EPHB6, ADAM22, CCL2, FUT3 / FUT5, INPP1 , CRNN, HS6ST1 , GFRA1 , SIAE, RNF149, MAN1A2, GRP, ITGBL1 , LAMP1 , SCGB3A1 , CD72, TGOLN2, ITIH1 , CFP, ELAVL4, SCT, MCEE, PCDH12, PITHD1 , SNX2, and PHLDB1.

[0042] In embodiments, in the first set of proteomic features, the features indicative of the protein level of PTGDS, PGF, F11R, CD40, CST3, IGFBP4, SDC1 , EPHA2, PILRA, PILRB, FAS, IGFBP6, ESAM, CD160, JAM2, TFPI, PROS1 , CDH3, GFOD2, KLK11 , NECTIN4, TNFRSF14, CD160, NCR1 , ATRAID, EFCAB14, DBI, TMSB10, DSC2, TNFRSF1B, TNFRSF1A, CD274, GPR37, IGFBP4, THBD, TNFRSF11A, APCS, NPDC1 , HSPG2, IGFBP6, PGF, LY6D, BTN2A1 , AGRN, TXNDC5, JAM2, PIK3IP1 , RELT, IGFBP4, TNFRSF19, WFDC2, ATRAID, EFCAB14, TFPI, HAVCR1 , and MXRA8 are expected to be higher in subjects in the second class than in subjects in the first class. In embodiments, the features indicative of the protein level of CPA1 , CPB1 , SUSD1 , PLXNA4, ADH4, UPB1 , and ENTPD5 are expected to be lower in subjects in the second class than in subjects in the first class. In the context of the present disclosure, features indicated as expected to be higher in a class compared to another class refers to the features having a higher average value in the class compared to the average value of the feature in the other class across subjects in training data. In embodiments, in the second set of proteomic features, the features indicative of the protein level of LAIR2, TREH, FGFR2, NBL1 , TNFRSF11A and TNFRSF1A are expected to be higher in subjects in the fourth class than in subjects in the third class; and the features indicative of the protein level of SELPLG, HAO1 , PM20D1 and TOP1 are expected to be lower in subjects in the fourth class than in subjects in the third class. In embodiments, the first plurality of proteomic features comprises features indicative of the protein level of each of the proteins in a feature subset listed in Table 7 (protein A1 vs A2) or proteins listed as correlating with one of these proteins in Table 6.

[0043] In embodiments, the first plurality of proteomic features comprises features indicative of the protein level of each of the proteins in a feature set listed in Table 9 or proteins listed as correlating with one of these proteins in Table 10. In embodiments, the first plurality of proteomic features comprises features indicative of the protein level of each of the proteins in a feature set listed in Table 11 or proteins listed as correlating with one of these proteins in Table 14.

[0044] In embodiments, the first plurality of proteomic features comprises features indicative of the protein level of each of the proteins in a feature set listed in Table 13 or proteins listed as correlating with one of these proteins in Table 14.

[0045] In embodiments, the first plurality of proteomic features comprises features indicative of the protein level of each of the proteins in a feature set listed in Table 15 or proteins listed as correlating with one of these proteins in Table 16.

[0046] In embodiments, the first plurality of proteomic features comprises features indicative of the protein level of at least 3 proteins listed as “component 1” in Table 19. In embodiments, the first plurality of proteomic features comprise features indicative of the protein level of one or more or all of PTGDS, ATRAID, BTN2A1 and PIK3IP1. In embodiments, the first plurality of proteomic features comprises features indicative of the protein level of one or more or all of NPDC1 , HSPG2 and PTGDS. In embodiments, the first plurality of proteomic features comprises features indicative of the protein level of one or more or all of: PTGDS, ATRAID, BTN2A1 , and EFCAB14, and optionally one or more or all of: AMN, CXCL5, RARRES2 and INHBC.

[0047] Also described according to a third aspect is a method of determining a prognosis for a subject that has been identified as having or being likely to have chronic kidney disease, the method comprising performing the method of any embodiment of the first aspect and the method of any embodiment of the second aspect.

[0048] In embodiments, the method comprises: (i) obtaining the values of the first plurality of clinical features associated with the subject; and classifying the subject between a plurality of classes including at least the first class and the second class, the first class having better prognosis than the second class using the values of the first plurality of clinical features and the classifier model that has been trained to classify subjects with chronic kidney disease between the first class and the second class using training data comprising values of the first plurality of clinical features for a plurality of training subject and associated labels indicating whether each training subject belongs to the first or second class; and (ii) when the subject is classified in the second class and the method further comprises obtaining the values of a second plurality of proteomic features associated with the subject; and classifying the subject between a plurality of classes including at least the third class and the fourth class, the third class having better prognosis than the fourth class using the values of the second plurality of proteomic features and the classifier model that has been trained to classify subjects with chronic kidney disease between the third class and the fourth class using training data comprising values of the second plurality of proteomic features for a plurality of training subject and associated labels indicating whether each training subject belongs to the third or fourth class. The methods of any aspect may comprise classifying the subject between a plurality of classes including at least the first class and the second class, and when the subject is classified in the second class, classifying the subject between a plurality of classes including at least the third class and the fourth class. Embodiments of any of the preceding aspects may have any of the following features. In embodiments, all classifiers are binary classifiers and / or each plurality of classes consists of two classes. In embodiments, the classifier model is a binary classifier. In embodiments, the classifier model is a machine learning model that has been trained to classify subjects with chronic kidney disease between the first class and the second class. In embodiments, the classifier model is a decision tree model, or a gradient boosted decision tree model. In embodiments, the first class (good prognosis) comprises subjects that have better prognosis than the second class (intermediate / poor prognosis), and the third class (intermediate prognosis) comprises subjects that have better prognosis than the fourth class (poor prognosis) and poorer prognosis than the first class. In embodiments, a poorer prognosis is associated with faster disease progression than a better prognosis, and / or a poorer prognosis is associated with a shorter expected time to a disease worsening event than a better prognosis. In embodiments, each class is associated with a different prognosis in terms of an expected time to disease progression, wherein disease progression is associated with a diagnosis of end stage renal disease, dialysis or renal replacement therapy. The chronic kidney disease may be diabetic kidney disease. In embodiments, the subject is a human, and / or the subject is an adult, and / or the subject is a human of 16 years or over. In embodiments, the method further comprises selecting a subject classified in the second class for further diagnostic testing. In embodiments, the further diagnostic testing comprises obtaining the value of the second plurality of clinical features and / or second plurality of proteomic features and classifying the subject between a plurality of classes including at least the third class and the fourth class. The method may further comprise selecting a subject classified in the second, third or fourth class for more regular monitoring than a subject classified in the first class; or selecting a subject classified in the second class for more regular monitoring than a subject classified in the first class; or selecting a subject classified in the fourth class for more regular monitoring than a subject classified in the third class. In embodiments of methods of the third aspect, the method comprises performing the method of any embodiment of the first aspect and the further diagnostic testing comprises performing the method of any embodiment of the second aspect. In embodiments of any aspect, the method comprises selecting a subject classified in the second class for inclusion in a clinical trial, or selecting a subject classified in the second class for treatment with a first predetermined treatment and selecting a subject classified in the first class for treatment with a second treatment different from the first predetermined treatment, optionally wherein the first predetermined treatment comprises treatment with a predetermined therapeutic compound and the second treatment comprises treatment with a different therapeutic compound.

[0049] In embodiments, the classifier model is a machine learning model that has been trained to classify subjects with chronic kidney disease between the first class and the second class using training data comprising values of the first plurality of proteomic features for a plurality of training subject and associated ground truth labels indicating whether each training subject belongs to the first or second class, wherein the ground truth labels were obtained by: obtaining clinical data for a plurality of subjects with chronic kidney disease, said clinical data comprising for each of one or more biomarkers, a plurality of measurements of the biomarker at a plurality of timepoints forming a temporal trajectory for the biomarker for the subject, the one or more biomarkers comprising one or more biomarkers of kidney function, optionally creatinine, and / or one or more biomarkers of anaemia, optionally haematocrit or haemoglobin; and identifying a first group and a second group of subjects in the plurality of subjects, wherein subjects in the first and second groups have different temporal trajectories of the one or more biomarkers, wherein the first group of subjects are associated with a ground truth label indicating that the subjects in the first group belong to the first class, and the second group of subjects are associated with a ground truth label indicating that the subjects in the second group belong to the second class. In embodiments, identifying a first group and a second group of subjects comprises: for each biomarker separately, obtaining pairwise distances between all of the temporal trajectories of the subjects thereby obtaining a tensor of pairwise distances, optionally using a Dynamic Time Warping (DTW) algorithm; and identifying clusters of subjects that have different temporal trajectories using the tensor of pairwise distances, optionally by applying a matrix decomposition method to the tensor of pairwise distances to obtain a matrix of predetermined dimensions, applying a dimensionality reduction algorithm to the matrix of predetermined dimensions to obtain coordinates for each subject in a two dimensional space, and applying a clustering algorithm to the coordinates. In embodiments, the classifier model uses as input proteomic features indicative of the protein level in a sample from the subject of at least two different proteins that are significantly differentially expressed between the subjects in the first group of subjects and the subjects in the second group of subjects. Embodiments of an aspect may further comprise training the classifier model.

[0050] Also described according to a fourth aspect is a method of treating a subject that has been identified as having or being likely to have chronic kidney disease, the method comprising determining a prognosis for the subject using the method of any embodiment of the first, second or third aspects, and (a) (i) treating the subject with a first therapy, when the subject is classified in the fourth class; and (ii) treating the subject with a second therapy different from the first therapy, when the subject is not classified in the fourth class; or (b) (i) treating the subject with a first therapy, when the subject is classified in the second class; and (ii) treating the subject with a second therapy different from the first therapy, when the subject is classified in the first class.

[0051] Also described according to a fifth aspect is a computer-implemented method of identifying a plurality of endotypes of subjects with chronic kidney disease, the method comprising: obtaining clinical data for a plurality of subjects with chronic kidney disease, said clinical data comprising for each of one or more biomarkers, a plurality of measurements of the biomarker at a plurality of timepoints forming a temporal trajectory for the biomarker for the subject, the one or more biomarkers comprising one or more biomarkers of kidney function, optionally creatinine, and / or one or more biomarkers of anaemia, optionally haematocrit or haemoglobin; and identifying a first group and a second group of subjects in the plurality of subjects, wherein subjects in the first and second groups have different temporal trajectories of the one or more biomarkers. In embodiments, identifying a first group and a second group of subjects comprises: for each biomarker separately, obtaining pairwise distances between all of the temporal trajectories of the subjects thereby obtaining a tensor of pairwise distances (e.g. a tensor of dimensions n x n x b where n is the number of subjects and b is the number of biomarkers), optionally using a Dynamic Time Warping (DTW) algorithm; and identifying clusters of subjects that have different temporal trajectories using the tensor of pairwise distances, optionally by applying a matrix decomposition method (e.g. Tucker decomposition) to the tensor of pairwise distances to obtain a matrix of predetermined dimensions (e.g. a matrix of dimensions n x c where c is the dimension of a decomposition core tensor), applying a dimensionality reduction algorithm (e.g. UMAP, PCA) to the matrix of predetermined dimensions to obtain coordinates for each subject in a lower dimensional space (e.g. a 2 dimensional space, a 3 dimensional space), and applying a clustering algorithm (e.g. K-means, DBSCAN) to the coordinates. In embodiment, the method further comprises validating the plurality of groups of subjects using a hazard model, wherein a plurality of groups of subjects are considered to be validated when the hazard model indicates a statistically significant difference in time to a disease relevant event between the groups. The disease relevant event may be selected from e.g. diagnosis of ESRD (end-stage renal disease) or occurrence of RRT (renal replacement therapy).

[0052] Also described according to a sixth aspect is a computer-implemented method of training a machine learning model to provide a prognosis for a subject with chronic kidney disease, the method comprising: identifying a plurality of endotypes of subjects with chronic kidney disease using a method according to any embodiment of the fifth aspect using clinical data for a plurality of subjects with chronic kidney disease thereby identifying a first group and a second group of subjects in the plurality of subjects; obtaining values of a first plurality of clinical features associated with the subjects and / or values of a first plurality of proteomic features associated with the subjects; and training a machine learning classifier to classify subjects with chronic kidney disease between the first class and the second class using training data comprising values of the first plurality of proteomic features for the plurality of subjects and / or values of the first plurality of clinical features for the plurality of subjects, and associated ground truth labels indicating whether each training subject belongs to the first or second class, wherein the first group of subjects are associated with a ground truth label indicating that the subjects in the first group belong to the first class, and the second group of subjects are associated with a ground truth label indicating that the subjects in the second group belong to the second class. In embodiments, the first plurality of clinical features associated with the subjects and / or the first plurality of proteomic features associated with the subjects are selected from respective sets of candidate clinical features or candidate proteomic features using a feature selection method. In embodiments, the feature selection method comprises identifying clinical features I proteomic features that have statistically significantly different values between the first and second groups. In embodiments, the feature selection method comprises using regularisation when training the machine learning model. In other words, the feature selection may comprise using a regularised machine learning model.

[0053] Also described according to a seventh aspect is a method comprising performing the method of embodiment of any preceding aspect, and prior to said performing, obtaining the values of the clinical features from the subject or from samples previously obtained from the subject, and / or obtaining the values of the proteomic features from a sample previously obtained from the subject.

[0054] Also described according to an eighth aspect is a system comprising a processor and one or more computer readable media storing instructions that, when executed by the processor, cause the processor to implement any computer-implemented method described herein, such as the method of any embodiment of the first, second, third, fifth and / or sixth aspects.

[0055] Also described according to a ninth aspect is a computer readable medium or media storing instructions that, when executed by a processor, cause the processor to implement any computer-implemented method described herein, such as the method of any embodiment of the first, second, third, fifth and / or sixth aspects.

[0056] Also described according to a tenth aspect is computer program product comprising instructions that, when executed by a processor, cause the processor to implement any computer-implemented method described herein, such as the method of any embodiment of the first, second, third, fifth and / or sixth aspects.

[0057] Also described according to an eleventh aspect is a kit comprising reagents specific for measuring a protein level of at least two proteins selected from: PTGDS orPGF orTNFRSF14, F11R or CD40, CST3 or IGFBP4, SDC1 or EPHA2, PILRA or PILRB, FAS or IGFBP6 or ESAM (optionally FAS or ESAM, further optionally FAS), CD160 or JAM2 or NCRI , TFPI or PROSI orAPCS (optionally TFPI orAPCS, further optionally TFPI), CPA1 or CPB1 , CDH3 or TNFRSF14, GFOD2, SUSD1 or PLXNA4, ADH4 or UPB1 , KLK11 or NECTIN4, ATRAID or EFCAB14, DBI or TMSBIO, TNFRSF14 orTNFRSFIA, DSC2 or TNFRSF1B, CD274 or TNFRSF1B, TNFRSF11A or TNFRSF1A, GPR37 or IGFBP4, THBD or ESAM, ENTPD5, NPDC1 , HSPG2, IGFBP6, PGF, LY6D, BTN2A1 , AGRN, TXNDC5, JAM2, PIK3IP1 , RELT, IGFBP4, TNFRSF19, WFDC2, ATRAID, EFCAB14, TFPI, HAVCR1 , and MXRA8, in a sample from a subject; and / or reagents specific for measuring a protein level of at least two proteins selected from: LAIR2 or TREH, SELPLG or HAO1 , FGFR2 or NBL1 , TNFRSF11 A or TNFRSF1 A, PM20D1 or TOP1 in a sample from a subject. The kit may comprise reagents specific for measuring a protein level for any of the genes in Table 0.

[0058] The kit may comprise reagents specific for measuring protein expression levels for a plurality of proteins consisting or consisting essentially of: (i) proteins selected from Table 0, or (ii) proteins selected from: PTGDS and / or PGF and / or or TNFRSF14, F11R and / or CD40, CST3 and / or IGFBP4, SDC1 and / or EPHA2, PILRA and / or PILRB, FAS and / or IGFBP6 and / or ESAM (optionally FAS and / or ESAM, further optionally FAS), CD160 and / or JAM2, TFPI and / or PROS1 and / or APCS (optionally TFPI and / or APCS, further optionally TFPI),CPA1 and / or CPB1 , CDH3 and / or TNFRSF14, GFOD2, SUSD1 and / or PLXNA4, ADH4 and / or UPB1 , KLK11 and / or NECTIN4, ATRAID and / or EFCAB14, DBI and / or TMSB10, TNFRSF14 and / or TNFRSF1A, DSC2 and / or TNFRSF1B, CD274 and / or TNFRSF1B, TNFRSF11A and / or TNFRSF1A, GPR37 and / or IGFBP4, THBD and / or ESAM, ENTPD5, NPDC1 , HSPG2, IGFBP6, PGF, LY6D, BTN2A1 , AGRN, TXNDC5, JAM2, PIK3IP1 , RELT, IGFBP4, TNFRSF19, WFDC2, ATRAID, EFCAB14, TFPI, HAVCR1 , and MXRA8; and / or (ii) LAIR2 or TREH, SELPLG or HAO1 , FGFR2 or NBL1 , TNFRSF11 A or TNFRSF1 A, PM20D1 or TOP1.

[0059] In embodiments, the kit comprises reagents specific for measuring a protein level of each of the proteins in a feature subset listed in Table 7 (protein A1 vs A2) or proteins listed as correlating with one of these proteins in Table 6.

[0060] In embodiments, the kit comprises reagents specific for measuring a protein level of each of the proteins in a feature set listed in Table 9 or proteins listed as correlating with one of these proteins in Table 10. In embodiments, the kit comprises reagents specific for measuring a protein level of each of the proteins in a feature set listed in Table 11 or proteins listed as correlating with one of these proteins in Table 14. In embodiments, the kit comprises reagents specific for measuring a protein level of each of the proteins in a feature set listed in Table 13 or proteins listed as correlating with one of these proteins in Table 14. In embodiments, the kit comprises reagents specific for measuring a protein level of each of the proteins in a feature set listed in Table 15 or proteins listed as correlating with one of these proteins in Table 16. In embodiments, the kit comprises reagents specific for measuring a protein level of each of at least 3 proteins listed as “component 1” in Table 19. In embodiments, the kit comprises reagents specific for measuring a protein level of each of one or more or all of PTGDS, ATRAID, BTN2A1 and PIK3IP1. In embodiments, the kit comprises reagents specific for measuring a protein level of each of one or more or all of NPDC1 , HSPG2 and PTGDS. In embodiments, the kit comprises reagents specific for measuring a protein level of each of one or more or all of: PTGDS, ATRAID, BTN2A1 , and EFCAB14, and optionally one or more or all of: AMN, CXCL5, RARRES2 and INHBC.

[0061] Also described according to a twelfth aspect is a system comprising the kit of any embodiment of the eleventh aspect, and a computer program product according to the tenth aspect, a computer readable medium according to the ninth aspect, or a system according to the eighth aspect.

[0062] Also described herein according to a further aspect is a method of analysing a sample from a subject, wherein the subject is a subject who has or is likely to have chronic kidney disease, the method comprising: contacting the sample with reagents specific for measuring a protein level of a first plurality of signature proteins comprising or consisting essentially of at least 2 proteins selected from: PTGDS or PGF, F11 R or CD40, CST3 or IGFBP4, SDC1 or EPHA2, PILRA or PILRB, FAS or IGFBP6 or ESAM (optionally FAS or ESAM), CD160 or JAM2 or NCR1 , TFPI or PROS1 or APCS (optionally TFPI or APCS) CPA1 or CPBI , CDH3 orTNFRSF14, GFOD2, SUSD1 or PLXNA4, ADH4 or UPB1 , KLK11 or NECTIN4, ATRAID or EFCAB14, DBI or TMSBW, TNFRSF14 or TNFRSFIA, DSC2 or TNFRSFIB, CD274 or TNFRSF1B, TNFRSF11A or TNFRSF1A, GPR37 or IGFBP4, THBD or ESAM, ENTPD5, NPDC1 , HSPG2, IGFBP6, PGF, LY6D, BTN2A1 , AGRN, TXNDC5, JAM2, PIK3IP1 , RELT, IGFBP4, TNFRSF19, WFDC2, ATRAID, EFCAB14, TFPI, HAVCR1 , and MXRA8; and / or at least two proteins selected from: LAIR2 or TREH, SELPLG or HAO1 , FGFR2 or NBL1 , TNFRSF11A or TNFRSF1A, PM20D1 or TOP1 The reagents may comprise reagents specific for measuring: a protein level for any of the proteins in Table 0.

[0063] The reagents may further comprise reagents specific for measuring a protein level of one or more control protein.

[0064] According to any embodiment of any aspect, the protein levels may have been measured or may be measured using an immune assay, optionally selected from a proximity extension assay, an ELISA assay, or a bead-based immunoassay

[0065] The disclosure also encompasses embodiments including any combinations of features described in relation to any of the above aspects, unless such features are clearly incompatible.

[0066] Brief Description of Figures

[0067] Fig. 1A is a flow diagram showing, in schematic form, a method of providing a prognosis or treatment recommendation according to the disclosure.

[0068] Fig. 1B is a flow diagram showing, in schematic form, a method of identifying patient groups according to the disclosure.

[0069] Fig. 2 shows an embodiment of a system for implementing methods of the disclosure.

[0070] Fig. 3 shows patient clusters identified from longitudinal creatinine and haematocrit data, and Cox survival curves for these clusters and comparative clinical metrics. Figs. 3A-B show UMAP representation of the patient clusters. Each point represents a patient with points coloured according to which cluster they are assigned to. Fig. 3A shows data stratification using K-means clustering and Fig.

[0071] 3B shows data stratification using DBSCAN. Hazard curves are shown for Clustering A (Fig. 3C) and Clustering B (Fig. 3D), with events defined as either ESRD or RRT. Figs. 3E, F, G, H Cox survival curves show the performance of clinically used risk stratification methods for CKD (Figs.3E, F, G), and comparative hazard curves associated with clusters according to the disclosure (i.e. newly discovered endotypes - Fig.3H), in the SKS cohort. Outcome is a composite of renal replacement therapy (dialysis or kidney transplant), eGFR dropping below 15 (ESRD; sustained for >1 month), eGFR dropping to >40% of starting eGFR (sustained), or death from metabolic syndrome related causes. Fig. 3E shows Cox survival curves for CKD stage determined by baseline (study start) eGFR levels. Fig. 3F shows Cox survival curves for KDIGO risk categories determined by baseline UACR and eGFR. Fig.3G shows Cox survival curves for the CKD-PC determined by age, sex UACR and eGFR. Fig. 3H shows Cox survival curves for the patient clusters described herein and on Figs. 3A and 3C.

[0072] Fig. 4 shows boxplots, receiver operating characteristic (ROC) curves, and Cox regression analysis curves for Clustering A (Figs. 4A-C) and Clustering B (Figs. 4D-F) for clinical prognostic models predicting cluster labels. Fig. 4A shows boxplots showing the predicted probability of patients (represented by points) of being A2 with true cluster labels shown on the x-axis and the thresholds of classification represented with red dashed horizontal lines. Fig. 4B shows a ROC curve for the A2 classification. Fig. 4C shows Cox regression analysis curves showing the estimated proportion of patients in each of the A1 and A2 clusters reaching outcomes over time, with outcomes defined as either ESRD or RRT. Fig. 4D shows boxplots showing the predicted probability of patients (represented by points) of being B3 with true cluster labels shown on the x-axis and the thresholds of classification represented with red dashed horizontal lines. Fig. 4E shows a ROC curve for the B3 classification. Fig.

[0073] 4F shows Cox regression analysis curves showing the estimated proportion of patients in each of the B1 , B2 and B3 clusters reaching outcomes overtime, with outcomes defined as either ESRD or RRT.

[0074] Fig.5 shows the distributions of clinical features included in a clinical prognostic model for distinguishing between A1 vs. A2. Boxplots are shown for all features except calcium channel blockers which is displayed as a bar chart coloured by true cluster. Boxplots show true clusters on the x-axis and numeric values of each clinical feature on the y-axis

[0075] Fig. 6 shows the distributions of clinical features included in a clinical prognostic model for distinguishing between B2 vs. B3. Boxplots are shown for all features except calcium channel blockers which is displayed as a bar chart coloured by true cluster. Boxplots show true clusters on the x-axis and numeric values of each clinical feature on the y-axis.

[0076] Fig. 7 shows boxplots, ROC curves, and Cox regression analysis curves for clustering A (Figs. 7A-C) and clustering B (Figs. 7D) for proteomic prognostic models. Fig. 7A shows boxplots showing the predicted probability of patients (represented by points) of being A2 with true cluster labels shown on the x-axis and the thresholds of classification represented with red dashed horizontal lines. Fig. 7B shows a ROC curve for the A2 classification. Fig. 7C shows Cox regression analysis curves showing the estimated proportion of patients in each of the A1 and A2 clusters reaching outcomes over time, with outcomes defined as either ESRD or RRT. Fig. 7D shows boxplots showing the predicted probability of patients (represented by points) of being B3 with true cluster labels shown on the x-axis and the thresholds of classification represented with red dashed horizontal lines. Fig. 7E shows a ROC curve for the B3 classification. Fig. 7F shows Cox regression analysis curves showing the estimated proportion of patients in each of the B1 , B2 and B3 clusters reaching outcomes overtime, with outcomes defined as either ESRD or RRT.

[0077] Fig. 8 shows distributions of protein features included in a proteomic prognostic model for distinguishing between A1 vs. A2 (batch 1). Boxplots show true clusters on the x-axis and NPX values of each protein feature on the y-axis.

[0078] Fig. 9 shows distributions of protein features included in a proteomic prognostic model for distinguishing between B2 vs. B3. Boxplots show true clusters on the x-axis and NPX values of each protein feature on the y-axis.

[0079] Fig. 10 shows results of testing of a prognostic clinical model in the CHUIMI external cohort for clustering A (Figs. 10A-B) and clustering B (Figs. 10D-F). Fig. 10A is a bar plot (showing the proportion of patients in each predicted cluster in clustering A that required RRT with the number of patients in each group printed on the bar plots in white text. Fig. 10B shows boxplots showing creatinine slopes split by the predicted cluster in clustering A. Fig. 10C shows boxplots showing eGFR slopes split by the predicted cluster in clustering A. Fig. 10D is a bar plot (showing the proportion of patients in each predicted cluster in clustering B that required RRT with the number of patients in each group printed on the bar plots in white text. Fig. 10E shows boxplots showing creatinine slopes split by the predicted cluster in clustering B. Fig. 10F shows boxplots showing eGFR slopes split by the predicted cluster in clustering B.

[0080] Fig. 11 shows associations between predicted clusters (Clustering A - Figs. 11A-B; Clustering B -Figs. 11C-D) and outcomes in the UKBB cohort. Fig. 11 A is a bar plot showing the proportion of patients in each predicted cluster in clustering A to reach either ESRD or dialysis, with raw numbers printed on the plot in white text. Fig. 11B shows Cox regression curves show differences in time to outcomes (ESRD or dialysis) between clusters A1 and A2. Fig. 11C is a bar plot showing the proportion of patients in each predicted cluster in clustering B to reach either ESRD or dialysis, with raw numbers printed on the plot in white text. Fig. 11D shows Cox regression curves show differences in time to outcomes (ESRD or dialysis) between clusters B1 , B2 and B3.

[0081] Fig. 12 shows results from applying an 8-protein A1 vs A2 patient stratifying model (PSM) according to the present disclosure using the features on Fig. 8 to plasma protein profiles of patients in batch 2. Fig.

[0082] 12A shows the A2 prediction scores from the model (probability of belonging to A2 cluster output by an XGBoost using the 8 proteins on Fig. 8), split according to the true cluster labels (from clustering process on Fig. 3). The dashed red line represents the threshold of classification derived from the discovery data (batch 1) - see Fig. 7A. Fig. 12B shows a ROC curve for the A2 prediction scores vs. the true cluster labels.

[0083] Fig. 13 shows batch 2 data for the 8 proteins in the 8-protein A1 vs A2 model from batch 1 (Fig. 8), with each point representing a sample (separated by true cluster labels from the clustering process on Fig.

[0084] 3A. Normalised NPX values returned by Olink are shown on the y-axis. Point clouds and overlaid boxplots are shown indicating the median, first and third quartiles, and whiskers extending to 1.5 x interquartile range, for each distribution.

[0085] Fig. 14 shows Cox Hazard curves showing predicted time to event for the A1 (labelled “stable / slow”) and A2 (labelled “faster progressors”) endotypes in a discovery and a validation cohort. Events were defined as either eGFR < 15, eGFR dropping > 40% from baseline, renal replacement therapy (RRT) or death from metabolic syndrome related causes. Fig. 14A shows data for the discovery cohort which was from SKS (Chinnadurai et al. 2023). Fig. 14B shows data for the validation cohort which was from SHARE (McKinstry et al., 2017). Endotypes were derived from longitudinal creatinine and haematocrit data. Fig. 15 shows distributions of protein features included in a proteomic prognostic model for distinguishing between A1 vs. A2 (batch 2). Boxplots show true clusters on the x-axis and NPX values (normalised values from Olink) of each protein feature on the y-axis. Each point represents a sample.

[0086] Fig. 16 shows distributions of protein features included in a proteomic prognostic model for distinguishing between A1 vs. A2 (batches 1+2 cross validation experiments). Boxplots show true clusters on the x-axis and NPX values (normalised values from O-link) of each protein feature on the y-axis. Each point represents a sample.

[0087] Fig. 17 shows distributions of protein features included in a proteomic prognostic model for distinguishing between A1 vs. A2 (batches 1+2). Boxplots show true clusters on the x-axis and NPX values (normalised values from O-link) of each protein feature on the y-axis. Each point represents a sample.

[0088] Fig. 18 shows distributions of UPCR values. Boxplots show true clusters on the x-axis and UPCR values on the y-axis. Each point represents a sample. Fig. 18A shows distributions in the SKS cohort. Fig.

[0089] 18B shows distributions in the SKS cohort.

[0090] Figure 19A shows UMAP visualisations of clusters from the Discovery cohort (SKS) Each point represents an individual, with the input data to UMAP being the output from Tucker decomposition. Figure 19B, C show UMAP visualisations of clusters from two validation cohorts (SHARE - B; CHUIMI - C). Each point represents an individual, with the input data to UMAP being the output from Tucker decomposition.

[0091] Figure 20A, B show trajectories of creatinine and haematocrit overtime for individuals in SKS-slow and SKS-fast in the discovery cohort. The dashed line indicates when eGFR first drops below 30 (±2) mL / min / 1 ,73m2.

[0092] Figure 21A shows results of stratification of patients using creatinine at baseline shown in violin plot format. Patients were stratified using a cut-off of 189 pmol / L.

[0093] Figure 21 B shows the association between creatinine-stratified patients and time to composite outcome as a Cox regression plot (p-value: 0.607).

[0094] Figures 22A, B show Cox survival curves for the two validation cohorts, SHARE (A) and CHUIMI (B). For SHARE, outcome is renal replacement therapy (dialysis or kidney transplant), eGFR dropping below 15 (sustained for >1 month), eGFR dropping to >40% of starting eGFR (sustained), or death from metabolic syndrome related cause, and for CHIUMI it is dialysis.

[0095] Figures 23A, B shows correlations between Haemoglobin (g / L) and Haematocrit (L / L) for SKS (A) and SHARE (B) showing all measurements recorded across time on a per person, matched date basis. Figure 24 shows a UMAP visualisation of endotypes identified in CHUIMI with points (representing patients) coloured according to whether patients had DKD or CKD, showing no stratification by CKD / DKD.

[0096] Figure 25 shows results from multi-omic integration using DIABLO showing per-patient scores for components 1 and 2 split across omics and coloured by endotypes.

[0097] Figure 26 shows an overview of the biological processes disrupted in fast progressors (SKS-fast) vs. slow progressors (SKS-slow) identified from multi-omic analysis using DIABLO.

[0098] Detailed Description

[0099] The present disclosure relates broadly to protein and clinical signatures that are indicative of prognosis in subjects with chronic kidney disease.

[0100] In describing the present invention, the following termswill be employed, and are intended to be defined as indicated below.

[0101] A “sample” as used herein may be a sample of bodily fluid previously obtained from a subject. The sample may be a plasma sample, a blood sample, a serum sample, or a urine sample. The sample may be one which has been freshly obtained from a subject or may be one which has been processed and / or stored prior to making a determination (e.g. frozen, fixed, mixed with one or more preservatives, such as e.g. EDTA and / or heparin or subjected to one or more purification, enrichment or extractions steps). As such, a sample as described herein may refer to any type of sample from which protein levels can be obtained. The sample is preferably from a mammalian (such as e.g. a mammalian cell sample or a sample from a mammalian subject, including in particular a model animal such as mouse, rat, etc. or a pet such as a cat, dog or horse). The subject is preferably a human. The subject may be an adult subject. The sample may be transported and / or stored, and collection may take place at a location remote from the expression data acquisition location, and / or the computer-implemented method steps may take place at a location remote from the sample collection location and / or remote from the transcriptomic data acquisition (e.g. sequencing) location (e.g. the computer-implemented method steps may be performed by means of a networked computer, such as by means of a “cloud” provider). Chronic kidney disease (CKD), also referred to as chronic kidney failure, is a condition in which kidney function is gradually lost. Chronic kidney disease can progress to end stage kidney failure (also referred to as end stage renal disease, ESRD), which requires dialysis or a kidney transplant (renal replacement therapy). Multiple diseases and conditions can cause chronic kidney disease, including e.g. diabetes (type 1 or type 2 diabetes), high blood pressure, glomerulonephritis, interstitial nephritis, polycystic kidney disease, prolonged obstruction of the urinary tract (e.g. due to enlarged prostate, kidney stones or some cancers), vesicoureteral reflux, and pyelonephritis. CKD that arises as a consequence of diabetes is referred to as diabetic kidney disease. The present disclosure relates to prognosis prediction and treatment selection for subjects with CKD. In embodiments, the subject has CKD as a consequence of diabetes, i.e. DKD. Thus, in embodiments, the chronic kidney disease is diabetic kidney disease. The terms “patient” and “subject” are used herein interchangeably. The subject may be an adult subject, e.g. a human over the age of 16, or over the age of 18. The subject may be male orfemale. The subject may be male. The subject may be a post-menopausal woman. The subject may be a subject with any stage of CKD / DKD. The subject may be a subject who does not have ESRD. The subject may be a subject who has been diagnosed as having stage 3 or 4 CKD / DKD.

[0102] The methods of the present disclosure are based at least in part on the discovery of multiple subgroups of patients with CKD (more specifically, identified in a cohort of patients with DKD, but validated in multiple cohorts including CKD and DKD patients) that have different disease progressions. These subgroups may be referred to as “endotypes”. In particular, the present inventors identified 2 distinct subgroups of patients with distinct disease progression, referred to herein as a group with good prognosis (referred to as A1 or B1), and a group with intermediate / poor prognosis (referred to as A2 or B2+B3). The former may be referred to as a first group, while the latter may be referred to as a second group. Two subgroups of patients were further identified within the group with intermediate / poor prognosis (second group), referred to herein as a group with intermediate prognosis (referred to as B2) and a group with poor prognosis (referred to as B3). The former may be referred to as a third group, while the latter may be referred to as a fourth group, the third and fourth group together forming the second group. The group with good prognosis comprises patients with CKD or DKD who have a better prognosis than the patients in the groups with intermediate prognosis. The group with poor prognosis comprises patients with CKD or DKD who have a worse prognosis than the patients in the groups with good or intermediate prognosis. The groups were identified by having different disease trajectories in terms of creatinine and haematocrit or haemoglobin values. Creatinine is a metabolic by-product that is normally filtered out by the kidneys and evacuated as a waste product in urine. Measurements of creatinine in the blood (serum) or urine provide an indication of kidney function. Indeed, increased levels of creatinine may be a sign of decreased kidney function. Measurements of creatinine used in the present examples were obtained in blood samples (specifically serum), but urine samples would be equally usable. A haematocrit test measures the proportion of red blood cells in the blood, as a proportion of total cell count (complete blood count). A haemoglobin test measures the amount of haemoglobin in the blood, typically as a concentration (e.g. g / L of blood). The haemoglobin concentration is proportional to the red blood cells count in a blood sample. Anaemia (lower red blood cell count) is common in patients with CKD, and is a risk factor in developing ESRD. The groups identified based on these biomarkers were shown to have different prognosis in terms of risk of progressing to more severe disease (ESRD, dialysis or RRT). Thus, poor, intermediate and / or good prognosis can be quantified in terms of time to a disease worsening event (also referred to as progression to kidney failure). Progression to kidney failure may be identified as a diagnosis of end stage renal disease, a requirement for dialysis or a requirement for renal replacement therapy. Thus, individuals in the first group (good prognosis) may have a longer expected time to ESRT, dialysis or RRT than individuals in the second group (intermediate / poor prognosis). Similarly, individuals in the third group may have a longer expected time to ESRT, dialysis or RRT than individuals in the fourth group. Note that although the patient subgroups were identified using creatinine and haematocrit or haemoglobin data, this was based on detailed time course data that is not routinely available in the clinic. By contrast, the methods of the present disclosure provide means to determine prognosis using clinical variables and / or levels of signature proteins at a single point in time. This means that the methods of the present disclosure can be performed as simple prognostic tests at any point during the course of a patient’s disease, using very limited data that is either available as part of routine clinical care, or that has been collected using a simple target proteomic test. Advantageously, the methods of the present disclosure can be used at one or more time points including at an early stage of the disease. This is particularly useful as at that point little information about disease progression is available, and the methods of the present disclosure can therefore be used to better plan monitoring and / or treatment of the patient based on their predicted disease course. The term “early stage of the disease” may refer to a clinically defined stage, e.g. stages G1 or G2 according to an eGFR test (where stage G1 may be defined as eGFR above 90ml / min, stage G2 may be defined as eGFR 60 to 89 ml / min with other signs of kidney damage, stage G3a may be defined as eGFR 45 to 59 ml / min, stage G3b may be defined as eGFR 30 to 44 ml / min, stage G4 may be defined as eGFR 15 to 29 ml / min, and stage G5 may be defined as eGFR below 15 ml / min), stage A1 according to an ACR test (urine albumin to creatinine ratio; where stage A1 may be defined as ACR less than 3mg / mmol. Stage A2 may be defined as ACR of 3 to 30mg / mmol, and stage A3 may be defined as ACR of more than 30mg / mmol), or to a time defined stage, such as e.g. within 1 , 2, 3, 4, 5, 6, 12 or 18 months of the patient having been diagnosed with CKD.

[0103] Signature proteins are proteins whose level in plasma has been identified as being indicative of prognosis in chronic kidney disease. Signature proteins of the disclosure are listed in Table 0. The protein sequence for each gene as disclosed at the Uniprot ID number indicated in Table 0, on 11 November 2024 is expressly incorporated herein by reference.

[0104] Signature clinical features are clinical features that have been identified as being indicative of prognosis in chronic kidney disease. Signature clinical features are listed in Table 0. The variable “creatinine” refers to the creatinine level in a sample of blood or urine from the subject. The creatinine level can be expressed in units of pmol / L, mg / L or any multiple thereof, such as e.g. mg / dL. Any concentration unit may be used provided that the classification model used has been trained using these units, or that the values for this feature are converted prior to use to the units that have been used for model training prior to use, using predetermined conversion factors. Similarly, either of blood samples or urine samples may be used provided that the classification model used has been trained using the same type of sample, or that the values for this feature are converted prior to use to equivalent values that would have been obtained using the type of samples that have been used for model training, using predetermined conversion factors (e.g. empirically determined conversion factors). eGFR refers to the estimated glomerular filtration rate of a subject. eGFR values are typically provided in mL of cleansed blood per minute per body surface. The eGFR value can be estimated using the creatinine blood level (pmol / L or mg / dL), age and biological sex of the subject. Multiple equations can be used, the current standard being the 2021 CKD-EPI Creatinine equation. The term mGFR refers to the measured glomerular filtration rate. The mGFR can be measured either as a urinary clearance during constant infusion of a GFR marker (e.g. inulin, iothalamate or iohexol) or as a plasma clearance performed during either constant infusion (assuming constant urinary excretion and GFR) or as a single injection with calculation of GFR based on injected dose and an estimated AUC (area under the curve, assuming a constant GFR over many hours). The term “eGFR” is used herein generally to refer to an estimated or measured GFR, in blood, plasma or urine, all of which could be used in the context of the present disclosure. The age at study start refers to the age of the subject, i.e. the age of the subject at the time the test is performed, typically expressed in years (although other encodings such as age range in decades can be accommodated provided that the model has been trained using the age feature encoded in this manner). Diastolic and systolic blood pressure are measured as known in the art, and typically expressed in mmHg. Albumin refers to the presence or concentration of albumin in a urine or blood sample in the subject. In embodiments, albumin refers to the presence or concentration of albumin in a blood sample in the subject. Unless indicated otherwise, the variable albumin as used herein is measured in a blood sample in the subject. The albumin variable can be expressed as a binary variable (e.g. absence presence of albumin, presence of albumin at a concentration above a predetermined threshold), or as a concentration (in any concentration units provided that the model was trained using the same concentration units or the concentration is converted prior to use to units used in training the model using a predetermined conversion factor). Phosphate refers to the concentration of phosphates in a blood sample from the subject. This can be expressed in mmol / L or any other units of concentration, provided that the model was trained using the same concentration units or the concentration is converted prior to use to units used in training the model using a predetermined conversion factor. Calcium channel blockers is a variable indicative of whether the subject is under treatment with calcium channel blockers (e.g. statins, ace I arb inhibitors, EPO, metformin etc.). Calcium channel blockers are a class of medication that prevent calcium from entering heart and blood vessel cells. They are commonly used for the treatment of high blood pressure (hypertension), chest pain caused by heart disease, abnormal heart rhythms, and migraines. Calcium channel blocker may be encoded as a binary variable, i.e. a variable that takes a first value (e.g. 1) if the subject is under treatment with a calcium channel blocker, and a second value (e.g. 0) if the subject is not under treatment with a calcium channel blocker. Total cholesterol refers to the total amount of cholesterol in a blood sample from the subject. This can be expressed in mmol / L or any other units of concentration, provided that the model was trained using the same concentration units or the concentration is converted prior to use to units used in training the model using a predetermined conversion factor. CRP (C-reactive protein) refers to the concentration of CRP in a blood sample from the subject. This can be expressed in mg / L or any other units of concentration, provided that the model was trained using the same concentration units or the concentration is converted prior to use to units used in training the model using a predetermined conversion factor. Calcium refers to the concentration of calcium in a blood sample from the subject. This can be expressed in mmol / L or any other units of concentration, provided that the model was trained using the same concentration units or the concentration is converted prior to use to units used in training the model using a predetermined conversion factor. The variable “UPCR” refers to the ratio of the amount of protein to the amount of creatinine in a sample of urine from the subject. The UPCR can be expressed in mg / g or any multiple thereof, such as e.g. mg / mg. Any concentration unit may be used provided that the classification model used has been trained using these units, or that the values for this feature are converted prior to use to the units that have been used for model training prior to use, using predetermined conversion factors. All non-factorial variables (i.e. concentrations or blood pressures) can be expressed as continuous or discrete variables. For example, for any variable expressed in units of concentration below (e.g. creatinine, eGFR, albumin, total cholesterol, CRT, calcium, UPCR, and / or phosphates), the variables may be used as measured concentrations or as discrete variables associated with respective ranges of concentration values. For example, values of concentrations within a first predetermined range for a concentration variable can be associated with a first value, and values of concentrations within a second, higher predetermined range for a concentration variable can be associated with a second value. Multiple such nonoverlapping ranges and associated values may be used., i.e. each of a plurality of predetermined ranges of concentrations for a concentration variable may be associated with a respective discrete value. Similarly, blood pressure values may be expressed as continuous variables (i.e. measured blood pressure) or discrete variables where each of a plurality of predetermined ranges of blood pressure are associated with a respective discrete value. Any clinical variables may be pre-processed to be expressed in the same units as those with which the classifier was trained. Clinical variable data may be scaled to the corresponding healthy range for the test that was used to generate the clinical variable data.

[0105] >

[0106] >

[0107] >

[0108]

[0109] >

[0110] >

[0111] >

[0112] >

[0113] >

[0114]

[0115] >

[0116] >

[0117] >

[0118] >

[0119] >

[0120] >

[0121] >

[0122] >

[0123]

[0124] >

[0125]

[0126] >

[0127] >

[0128] >

[0129]

[0130] >

[0131]

[0132] >

[0133]

[0134]

[0135]

[0136]

[0137]

[0138]

[0139]

[0140]

[0141]

[0142] Table 0. Signature proteins and clinical markers. Core model=XGBoost, expanded model=Lasso. Core model unless indicated otherwise. Batch 1 model=XGBoost model trained on a first batch of samples (see Examples). Batch 2 model=XGBoost model trained on a second batch of samples (see Examples). Batch 1 + batch 2 model= XGBoost model trained on a combination of the first and second batch of samples (see Examples). Note all models were shown to satisfactorily classify samples in both batches. The Lasso model was trained on batch 1 data (and was also shown to be applicable to batch 2).

[0143] Protein level data refers to data indicative of the level of one or more signature proteins in a sample. The one or more signature proteins may be selected from those listed in Table 0, and homologs thereof. The level of a protein in a sample may be measured using any proteomic assay known in the art. The level of a protein in a sample may be measured using an affinity based assay, such as a multiplex affinity based assay, or any assay using antigen binding reagents (e.g. antibodies or fragments thereof) to specifically detect and / or quantify the presence of one or more proteins. The level of a protein in a sample may be measured using an ELISA assay, Luminex assay or an Olink Proximity Extension Assay (PEA). The level of “FUT3 / FUT5” refers to the level of one or both of proteins FUT3 and FUT5. This can be measured using an affinity-based assay that uses one or more affinity reagents that recognise both proteins. Alternatively, a measurement that is specific to either FUT3 or FUT5 may be used. Protein level data comprising protein levels (also referred to as “protein expression level” and “plasma protein level”, when measured in a plasma sample) for a plurality of signature proteins may be normalised, for example by reference to one or more reference signals, such as e.g. signals associated with one or more reference protein levels. A reference signal may be a positive control or a negative control. A positive control may refer to a signal associated with a reference protein that is expected to be present in a sample. A negative control may refer to a signal associated with a protein that is not expected to be present in a sample. A reference protein level may be the protein level of a control gene (or a summarised protein level across a plurality of control protein, such as e.g. a mean or median protein level) in the same sample. A reference protein may be the level of a protein that is expected to be present in all samples from subjects with different prognosis (i.e. a positive control signal), whether it is present in the biological sample itself or added to the sample as a spike-in control prior to measurement. A reference protein may be a protein that is expected to have a constant level across samples from subjects with different prognosis, when measured using the same assay. For example, levels of one or more housekeeping proteins may be used. Thus, a reference protein level may be a level of a housekeeping or other positive control protein, and a reference signal may be a signal associated with a housekeeping protein or other positive control protein. As another example, expression levels of one or more spike in targets may be used. Thus, a reference protein level may be a level of a spike in target (and a reference signal may be a signal associated with a spike in target). Spike in targets are proteins that are included in a sample to be analysed, at a known predetermined concentration. A reference protein level may be a summary protein level over a plurality of proteins, which may comprise one or more signature proteins and may also comprise additional proteins not part of the signature. For example, a reference protein level may be a mean or median protein level across a plurality of proteins measured in the sample, comprising the signature proteins measured and a plurality of additional genes. Protein expression levels may be expressed as unitless concentrations by normalisation. Log transformation or any other type of transformation may be applied prior to or after normalisation, provided that the same transformation is applied to the data used in training the model used. Protein expression levels may be normalised by obtaining a z-score for each signature protein measured. A z-score centres and standardises the measured protein level for a protein using the average and standard deviation of protein levels across a plurality of protein levels including the protein level that is being standardised. The plurality of protein levels may be protein levels from a plurality of proteins for the same sample, or protein levels for the same protein across a cohort of samples. Multiple normalisation approaches may be combined. For example, a protein level may be normalised by reference to a control (e.g. housekeeping protein or spike in), then by reference to a plurality of normalised protein levels measured in the same sample (e.g. by obtaining a z-score). Protein level data may have been measured together with a negative control (e.g. a sample that is known to have none of the proteins present), a positive control (a recombinant or natural sample that is known to be detectable in the assay), and / or one or more spike controls (e.g. to spike in recombinant or natural protein into samples and verify the amount spiked in is what is read out). Kits described herein may comprise any one or more of any of the above.

[0144] Embodiments of the present disclosure make use of machine learning models. The machine learning models described herein are mathematical models trained (i.e. parameterised) to classify observations between a plurality of classes, i.e. classifiers (also referred to as “classification model”). A classification model may provide as output a classification label and / or one or more probabilities of an observation belonging to respective one or more classes. A binary classification model may provide as output a single probability or score indicating the probability that the observation belongs to a positive class (e.g. a class with a poorer prognosis, rather than a class with a better prognosis). A predetermined threshold may be applied to the one or more probabilities or the score, to assign a class label. The threshold may be determined based on desired characteristics of the classification, such as e.g. a desired level of specificity, sensitivity or any accuracy performance that combines aspects of specificity (precision) and sensitivity (recall), such as accuracy and F1 score or balanced versions thereof that take into account the proportions of observations in the training data in each of the classes. Alternatively, an observation may simply be assigned to the class with highest probability. When using a binary classifier this may be equivalent to assigning the class for which the probability is over 50%. Embodiments of the present disclosure make use of one or more binary classifiers. The one or more binary classifiers may each output a score or probability. Multiple machine learning classifiers known in the art may be used in the context of the present disclosure. Indeed, the present inventors have demonstrated that useful predictions can be obtained when using any classifier from a simple linear model producing a score compared to a threshold, to a non-linear tree-based model such as a gradient boosted tree model (e.g. XG Boost).

[0145] The term “machine learning algorithm” or “machine learning method” refers to an algorithm or method that trains and / or deploys a machine learning model. The machine learning models of the present disclosure are trained by supervised learning, i.e. using training data comprising observations for each of set of predictive variables for each of a plurality of training subjects, and a class (also referred to as subgroup or endotype) label (also referred to as ground truth label). The ground truth label indicates the known prognosis group that a training subject belongs to, based on disease progression data for the training subject. Disease progression data can include one or more of time to a disease worsening event (including e.g. time to a treatment intervention such as renal replacement therapy), time course of disease progression, and time course of creatinine and / or eGFR measurements. Supervised learning refers to the parameterisation of a model so as to optimise an optimality criterion based on the comparison of predicted and ground truth labels for observations in the training data. An optimality criterion may be the minimisation of a loss function that quantifies the model prediction error based on the observed (ground truth) and predicted values of the predicted variables. Suitable loss functions for use in training machine learning models are known in the art and include the d error, the mean absolute error, and regularised versions thereof. Any of these can be used according to the present disclosure. Regularised loss functions are functions that include a loss function as described above, and one or more terms penalizing model complexity in order to reduce the risk of overfitting. Overfitting is a phenomenon that occurs where a machine learning model is trained to very closely reproduce the features of a training data set, resulting in poorer performance on other datasets that do not have the same characteristics (i.e. poor generalizability). L1 regularisation (also known as “Lasso” in the context of regression) adds a regularization term to the loss function that penalizes models based on the sum of absolute value of the coefficients of the model. L2 regularisation (also known as “Ridge” in the context of regression) adds a regularization term to the loss function that penalizes models based on the sum of squared value of the coefficients of the model. L1 regularisation can be used as a feature selection method as it minimizes the coefficients associated with less informative predictive features. In embodiments, the machine learning model is a regularized model. In embodiments, the machine learning model is a regularized tree-based model. In embodiments, the machine learning model is a regularized gradient boosted decision tree model. Examples of such models are available in the XGBoost software library (xgboost.ai / ). Such models may be referred to as “XGBoost” models, although any other implementation of regularized gradient boosted models may equally be used.

[0146] A machine learning model as described herein may be selected from: decision trees and variants thereof including regularised and / or gradient boosted decision trees and random forest models, regularised discriminant analysis, logistic regression models, generalised linear models, artificial neural networks (ANNs) including multilayer perceptrons (with linear or non-linear activation functions) and deep learning models (e.g. long short-term memory networks (LSTMs), Recurrent neural networks (RNNs)), naive Bayes classifiers, and support vector machines (SVM, using linear or non-linear kernels such as radial basis function). Non-linear models, such as decision trees and variants thereof (including in particular random forests and gradient boosted trees), SVM with a non-linear kernel, and ANNs (e.g. multilayer perceptrons) with non-linear activation functions may be particularly performant. Relatively simple ANNs such as MLPs are advantageous for the purpose of feature-based classification (As is the case here) because more complex deep learning models such as LSTMs and RNNs typically have more trainable parameters.

[0147] In embodiments, a machine learning model comprises an ensemble of models whose predictions are combined. Alternatively, a machine learning model may comprise a single model. Random forest models and gradient boosted tree models (such as XGBoost) are ensemble models. Ensemble versions of any models can be constructed. Ensemble models are expected to result in better prediction performance than single models and may therefore be advantageous in the context of the methods described herein. For example, the machine learning model may be a random forest classifier or a gradient boosted decision tree model. A random forest classifier is a model that comprises an ensemble of decision trees and outputs a class that is the average prediction of the individual trees. Decision trees perform recursive partitioning of a feature space until each leaf (final partition sets) is associated with a single value of the target. Gradient boosting is a machine learning method that forms an ensemble of weak prediction models (e.g. decision trees) from which a combined strong prediction is obtained. The algorithm iteratively adds new weak predictors to improve the prediction obtained by combining the outputs of the weak predictors. By contrast, random forest iteratively trains a set number of trees using random subsets of the training data. Conversely, very simple models such as linear regression models (also referred to herein as “generalised linear models”) or a “simple score” model can be useful to enable implementation by clinicians. A simple score model is a linear model in which values for all features are added up with a coefficient equal to +1 or -1. The coefficient value is chosen as positive or negative depending on whether the value is expected, based on training data, to increase in subjects in a poorer prognosis group compared to subjects in a better prognosis group (positive coefficient) or decrease in subjects in a poorer prognosis group compared to subjects in a better prognosis group (negative coefficient). This results in a simple score that increases with likelihood of poorer prognosis. The coefficient value may be chosen as positive or negative depending on whether the value is expected, based on training data, to increase in subjects in a poorer prognosis group compared to subjects in a better prognosis group (positive coefficient) or decrease in subjects in a poorer prognosis group compared to subjects in a better prognosis group (negative coefficient) by determining whether the average of values for a feature in the poorer prognosis group is higher (positive coefficient) or lower (negative coefficient) compared to the average of values for the feature in the better prognosis group. Alternatively, the coefficient value may be chosen as positive or negative by fitting a multivariate generalised linear model (e.g. a logistic regression model) classifying subjects between the poorer prognosis group and the better prognosis group to training data for the features included in the simple score model and determining whether the coefficient for the feature in the generalised linear model is positive or negative. In such embodiments, a positive coefficient can be used for a feature in the simple score model configured to output a score indicative of the likelihood of a subject belonging to the poorer prognosis when the feature has a positive coefficient in the generalised linear model predicting the probability of a subject having poorer prognosis. Similarly, a negative coefficient can be used for a feature in the simple score model configured to output a score indicative of the likelihood of a subject belonging to the poorer prognosis when the feature has a negative coefficient in the generalised linear model predicting the probability of a subject having poorer prognosis. As the skilled person understands, the opposite coefficients can be used when the simple score model and the generalised linear models are configured in opposite ways, i.e. the simple score increases with the risk of poorer prognosis whereas the generalised linear model predicts the likelihood of good prognosis, and vice versa. The use of the signs of coefficients from a generalised linear model comprising the same features as the simple scores model is advantageous as it considers the effect of each feature in a multivariate setting rather than in isolation. This method is used by default herein unless indicated otherwise. As the skilled person understands, a feature that is expected to be higher in the poorer prognosis group than in the better prognosis group (in the sense that the average of values for the feature in the poorer prognosis group is higher compared to the average of values for the feature in the better prognosis group) typically but not necessarily also has a positive coefficient in a generalised linear model predicting the probability of a subject having poorer prognosis using a plurality of features comprising said features. Indeed, the present inventors have found this to be the case in the vast majority of features tested, with the exception of CST3 and TNFRSF14, both of which had higher average values in a poorer prognosis group (labelled A2 in the examples below) than in a better prognosis group (labelled A1 in the examples below), but negative coefficients in multivariate generalised linear models trained to predict a probability of a subject belonging to the poorer prognosis group. The variables values may be used after normalisation to be in the same range (e.g. normalisation between 0 and 1), or may be used in their original range without normalisation between 0 and 1. This may also be referred to as a model that has been trained, although coefficient values are not strictly speaking fitted to training data using a loss function.

[0148] The method described herein may be used to identify a treatment for a subject with chronic kidney disease, and in the context of methods of treating patients with chronic kidney disease.

[0149] As used herein "treatment" refers to reducing, alleviating or eliminating one or more symptoms of the disease which is being treated, relative to the symptoms prior to treatment. Treatment may use one or more therapeutic compositions. A composition as described herein may be a pharmaceutical composition which additionally comprises a pharmaceutically acceptable carrier, diluent or excipient. The pharmaceutical composition may optionally comprise one or more further pharmaceutically active polypeptides and / or compounds. Such a formulation may, for example, be in a form suitable for intravenous infusion.

[0150] A subject who has been identified as having poor prognosis may be selected for treatment with a first therapy or combination of therapies. A subject who has been identified as not having poor prognosis may be selected for treatment with a second therapy or combination of therapies. The first and / or second therapies may be therapies that are currently used in the clinic but indicated for different types of subjects or diseases, or experimental therapies.

[0151] The systems and methods described herein may be implemented in a computer system, in addition to the structural components and user interactions described. As used herein, the term “computer system” includes the hardware, software and data storage devices for embodying a system or carrying out a method according to any of the described embodiments. For example, a computer system may comprise a central processing unit (CPU), input means, output means and data storage, which may be embodied as one or more connected computing devices. Preferably the computer system has a display or comprises a computing device that has a display to provide a visual output display. The data storage may comprise RAM, disk drives or other computer readable media. The computer system may include a plurality of computing devices connected by a network and able to communicate with each other over that network. It is explicitly envisaged that a computer system may consist of or comprise a cloud computer.

[0152] The methods described herein may be provided as computer programs or as computer program products or computer readable media carrying a computer program which is arranged, when run on a computer, to perform the method(s) described herein. As used herein, the term “computer readable media” includes, without limitation, any non-transitory medium or media which can be read and accessed directly by a computer or computer system. The media can include, but are not limited to, magnetic storage media such as floppy discs, hard disc storage media and magnetic tape; optical storage media such as optical discs or CD-ROMs; electrical storage media such as memory, including RAM, ROM and flash memory; and hybrids and combinations of the above such as magnetic / optical storage media.

[0153] Figure 1 A is a flow diagram showing, in schematic form, a method of providing a prognosis or treatment recommendation for a subject according to the disclosure. At optional step 10, a sample is obtained from a subject. The step of obtaining a sample from a subject may comprise physically obtaining the sample from the subject. Alternatively, the sample may have been previously obtained and no interaction with the subject may be required. In other words, obtaining a sample may comprise receiving a previously acquired sample. At optional step 12, protein level data is obtained for the sample comprising protein levels for a plurality of signature proteins as described herein. Instead or in addition to protein level data, clinical data may be obtained at step 12. The step of obtaining protein level data from a sample may comprise determining a protein level for each of the plurality of proteins using any assay known in the art, such as e.g. an immune assay (ELISA, Proximity Extension Assay, Luminex). The step of obtaining clinical data may comprise obtaining the clinical data from the subject or from samples previously obtained from the subject. Alternatively, protein level data and / or clinical data may have been previously obtained. Thus, obtaining protein level data and / or clinical data may comprise receiving the data from one or more databases, or from a user through a user interface. At step 14, it is determined whether the subject has a good or intermediate / poor prognosis, using methods described herein. This may use a trained model that has been obtained using a method as described by reference to Fig. 1B below, using training data comprising protein level data and / or clinical data for a cohort of training subjects comprising subjects who have been identified as belonging to a first group with good prognosis and subjects who have been identified as belonging to a second group with poorer prognosis (e.g. intermediate / poor prognosis) using a method as described by reference to Fig. 1B. A subject that has been identified at step 14 as having intermediate / poor prognosis may be subject to a further classification step to determine whether the subject has an intermediate or poor prognosis, at step 16. This is optional, and the further determination at step 16 may not be performed. When performed, step 16 may use a trained model that has been obtained using a method as described by reference to Fig.

[0154] 1B below, using training data comprising protein level data and / or clinical data for a cohort of training subjects comprising subjects who have been identified as belonging to a third group with intermediate prognosis and subjects who have been identified as belonging to a fourth group with poor prognosis using a method as described by reference to Fig. 1B.

[0155] At optional step 18, a particular course of treatment (which may comprise one or more different individual therapies) may be identified based on the results of step 14, and / or the results of step 16 in embodiments in which this step is performed. For example, a subject who has been identified as having a good prognosis at step 14 may be treated with a standard of care treatment for CKD / DKD (e.g. standard of care for the stage of CKD / DKD that the subject has been diagnosed as having). For example, a subject who has CKD / DKD stage 1 and who has been identified as having a good prognosis at step 14 may be treated with a standard of care treatment for CKD / DKD stage 1. As another example, a subject who has been identified as having an intermediate / poor prognosis at step 14 may be recommended for further testing at step 16. Instead or in addition to this, a subject who has been identified as having an intermed iate / poor prognosis at step 14 may be recommended for a treatment that is more intensive than standard of care for the stage of CKD / DKD that the subject has been identified as having, and / or may be recommended for inclusion in a clinical trial. For example, a subject who has CKD / DKD stage 1 and who has been identified as having a poor prognosis at step 14 may be treated with a more intensive treatment than the standard of care treatment for CKD / DKD stage 1. The more intensive treatment may include e.g. increased monitoring, and / or treatment with an experimental drug e.g. as part of a clinical trial. A subject who has been identified at step 16 as having intermediate prognosis may be treated using a treatment that is more intensive than standard of care for the stage of CKD / DKD that the subject has been identified as having, and / or may be recommended for inclusion in a clinical trial. A subject who has been identified at step 16 as having poor prognosis may be treated a treatment that is more intensive than standard of care for the stage of CKD / DKD that the subject has been identified as having, or more intense than the treatment recommended for a subject in the intermediate prognosis group and / or may be recommended for inclusion in a clinical trial.

[0156] At step 20, one or more results of this analysis may optionally be provided to a user through a user interface. At optional step 22, the subject may be treated with the therapy identified at step 18.

[0157] The prognosis categories used in the methods described by reference to Figure 1A may correspond to prognosis groups that have been previously identified using a method as illustrated on Figure 1B.

[0158] Indeed, Figure 1B illustrates schematically a method of identifying prognosis groups of patients with CKD. Each group, also referred to herein as “endotype” corresponds to a cohort of patients that have different disease progression compared to patients in other groups. Thus, also described herein are methods comprising performing any of the methods described by reference to Figure 1B to identify two or more groups of CKD patients with different prognosis, and to use these groups to train models to classify patients, and using the trained models to classify patients as described by reference to Figure 1A. At step 110, clinical data fora plurality of subjects with CKD / DKD is obtained comprising longitudinal measurements of one or more biomarkers of kidney function and / or one or more biomarkers of anaemia. In embodiments, the clinical data for the plurality of subjects with CKD / DKD comprise longitudinal measurements of at least one biomarker of kidney function and at least one biomarker of anaemia. The plurality of subjects may be referred to as “training subjects” or “discovery cohort” in that data for these subjects may be used to identify prognosis groups. Proteomic and / or clinical data associated with these subjects may be referred to as “training data” in that data for these subjects may be used to train a model as described herein, where the identified prognosis groups are used as ground truth data fortraining a model as described herein. The one or more biomarkers of kidney function can comprise creatinine values. As described above, measurements of creatinine may be measurements obtained from blood (e.g. serum) or urine samples. In embodiments, all creatinine measurements used are measurements obtained from blood samples. The one or more biomarkers of anaemia can comprise haematocrit values and / or haemoglobin values. Haematocrit measurements are obtained from blood samples. Haemoglobin measurements are obtained from blood samples. The cohort can comprise at least 100, at least 150, at least 200, at least 250 or at least 300 subjects. The cohort can comprise at least 10%, at least 15%, at least 20% or at least 25% of subjects who are known to have progressed to ESRD or RRT. The cohort may comprise at most 50%, at most 40% or at most 35% of subjects who are known to have progressed to ESRD or RRT. Longitudinal data may comprise a plurality of measurements obtained over a period of at least 1 year, at least 2 years, at least 3 years, at least 4 years, at least 5 years for each of a plurality of the subjects. Longitudinal data may comprise measurements obtained at least once a month, once every two months, once a quarter, once every 6 months or once a year for each biomarker in each of the plurality of subjects. At step 115, the clinical data is optionally preprocessed. Preprocessing may comprise, for each biomarker, performing one or more of: normalising, summarising, outlier removal, and censoring. Normalising may comprise one or both of: scaling data to a healthy reference range, scaling to an observed range, and scaling the data on a per subject basis. Scaling may be performed using any method known in the art, such as e.g. z-scoring (standardisation), min-max normalization, or mean normalisation. Scaling may comprise z-scoring relative to a reference range for a sex matched healthy population. Summarising may comprise obtaining a summary value for a plurality of values for a subject over a predetermined period of time. For example, monthly, bimonthly (i.e. once every 2 months) or quarterly summarised values may be obtained as the average value over a plurality of values obtained for a subject in each such predetermined period of time. Outlier removal may be performed using any method known in the art. At step 130, the longitudinal clinical data is used to identify a plurality of groups of subjects that have different temporal trajectories in terms of the one or more biomarkers. As a result of step 130, each subject is associated with a group label. The optionally preprocessed biomarker measurements for each subject may be referred to as the subject’s temporal trajectory. Thus, the data may comprise, for each of the plurality of subject, one or more temporal trajectories each comprising a set of measurements for one of the biomarkers, and associated time stamps. The time stamps may be expressed in terms of time from the first measurement available. In other words, the time stamps may be expressed in relative time units, where a time t=0 corresponds to the first measurement available (optionally after outlier removal) for a subject. The first measurement for a subject may be referred to as “baseline” measurement, and the corresponding time may be referred to as “baseline”. Step 130 may comprise step 130A of obtaining pairwise distances between all of the temporal trajectories in the data. For example, in the case of a dataset comprising measurements for b biomarkers (where b can be e.g. 2, e.g. one biomarker of kidney function such as creatinine and one biomarker of anaemia such as haematocrit or haemoglobin) for each of n subjects, step 130A may comprise calculating n x n x b pairwise distances between respective temporal trajectories. Distances between temporal trajectories may be calculated at step 130A using any method known in the art for measuring similarity between data series. Advantageously, distances between temporal trajectories may be calculated at step 130A using a method for measuring similarity between time series, such as e.g. Dynamic Time Warping (DTW). Dynamic Time Warping (DTW) advantageously enables determination of similarity between two time series even if they vary in speed. Step 130 may further comprise step 130B of identifying clusters of subjects that have different temporal trajectories using the distances obtained in step 130A. This may comprise using a matrix decomposition method (e.g. Tucker decomposition) applied to a matrix of pairwise distances between patients (e.g. matrix of size n x n x b) to identify a matrix of size n x c (n being the number of patients and cthe dimension of the decomposition core tensor used on the Tucker decomposition) which identifies the combined effect of differences in the plurality of biomarkers (b). The matrix may then be subject to clustering, optionally after dimensionality reduction. Any clustering algorithm known in the art may be used, such as e.g. K-means or DBSCAN. Any dimensionality reduction known in the art may be used, such as e.g. UMAP or PCA, where coordinates for each subject along the first 1 , 2, 3 or more dimensions (optionally the first 2 dimensions) can be used as inputs for clustering. At optional step 140, hazard modelling is optionally used to validate the clusters identified at step 130. Specifically, step 140 comprises fitting one or more hazard models to determine whether there is a statistically significant difference in time to a disease relevant event between the groups identified at step 130. The disease relevant event may be selected as a diagnosis of ESRD and / orthe occurrence of RRT. A hazard model may be a proportional hazard model, specifically a Cox proportional hazard model. The hazard model may use one or more covariates including one or more of: age, sex and eGFR at baseline (i.e. at the time of the first measurement in the temporal trajectory for a subject). At optional step 150, the (optionally validated) group labels obtained at step 140 are used as ground truth to train a machine learning model configured to classify subjects between two or more of the groups identified at step 130, using proteomic data and / or clinical data for the subjects, as described herein. The trained machine learning model may then be used in a method as illustrated by reference to Figure 1 A.

[0159] Thus, also described herein is a method (e.g. a computer-implemented method, including e.g. a method of identifying a plurality of groups of CKD / DKD subjects with different prognosis, and / or a method of training a machine learning model to provide a prognosis for a subject with CKD / DKD) comprising: obtaining clinical data for a plurality of subjects with CKD / DKD, said clinical data comprising longitudinal measurements of one or more biomarkers comprising one or more biomarkers of kidney function and / or one or more biomarkers of anaemia; and identifying a plurality of groups of subjects that have different temporal trajectories in terms of the one or more biomarkers. The plurality of groups of subjects may have statistically significant differences in time to a disease relevant event. For example, the plurality of groups of subjects may comprise a first group (e.g. good prognosis group) and second group (e.g. poorer prognosis group), where the first group has a statistically significant different disease progression trajectory in terms of time to a disease relevant event (e.g. ESRD or RRT). The method may further comprise validating the plurality of groups of subjects using a hazard model, wherein a plurality of groups of subjects are considered to be validated when the hazard model indicates a statistically significant difference in time to a disease relevant event between the groups. The method may further comprise training a machine learning model configured to classify subjects between two or more of the groups using as input proteomic data and / or clinical data for the subjects, using ground truth labels for the subjects corresponding to the groups identified. Training the machine learning model may comprise identifying input features of the machine learning model using any feature selection method known in the art. Training the machine learning model may comprise identifying input features of the machine learning model as proteomic and / or clinical variables that are significantly different between at least two groups that the machine learning model is trained to distinguish (e.g. two groups of the identified plurality of groups, the machine learning model being configured to classify subjects between the two groups), where the two groups are identified in a cohort of training subjects using a method as described herein. Training the machine learning model may comprise training a generalised linear model to classify subjects between two groups and using the coefficients of the trained generalised linear model to set coefficients of a simple score model using the same input features as the trained generalised linear model, as described herein. Training the machine learning model may comprise training a regularised model. A regularised model inherently includes a feature selection process (e.g. an XGBoost model training can be used to filter input features by selecting only features that have non-zero gain in the trained model). However, input features for regularised model may have additionally been filtered prior to training, e.g. based on being statistically significant between two groups to be distinguished. Statistical significance may be assessed using any statistical significance test known in the art suitable for comparing two distributions, such as e.g. a t-test, and a chosen level of confidence.

[0160] Figure 2 shows an embodiment of a system for providing a prognosis or treatment recommendation, according to the present disclosure. The system comprises a computing device 1 , which comprises a processor 101 and computer readable memory 102. In the embodiment shown, the computing device 1 also comprises a user interface 103, which is illustrated as a screen but may include any other means of conveying information to a user such as e.g. through audible or visual signals. The computing device 1 is communicably connected, such as e.g. through a network, to protein level data acquisition means 3, such as an immune assay platform, protein array, etc., and / or to one or more databases 2 storing protein level data and / or clinical data. For example, the computing device 1 may be communicably connected to an electronic health records database. The one or more databases 2 may further store one or more of: signatures information, parameters (such as e.g. thresholds or parameters of a machine learning algorithm trained to predict prognosis), patient and / or sample related information, etc. The computing device may be a smartphone, tablet, personal computer or other computing device. The computing device is configured to implement a method as described herein. In alternative embodiments, the computing device 1 is configured to communicate with a remote computing device (not shown), which is itself configured to implement a method as described herein. In such cases, the remote computing device may also be configured to send the result of the method to the computing device. Communication between the computing device 1 and the remote computing device may be through a wired or wireless connection, and may occur over a local or public network 6 such as e.g. over the public internet. The gene expression data acquisition means may be in wired connection with the computing device 1 , or may be able to communicate through a wireless connection, such as e.g. through WiFi and / or over the public internet, as illustrated. The connection between the computing device 1 and the protein level data acquisition means 3 and / or databases 2 may be direct or indirect (such as e.g. through a remote computer). The protein level data acquisition means 3 are configured to acquire protein level data from samples, for example plasma samples. In some embodiments, the samples may have been subject to one or more preprocessing steps such as plasma separation, protein extraction, etc. Preferably, the sample has not been subject to steps other than plasma separation. Any sample preparation process that is suitable for use in the determination of a protein levels in a sample of bodily fluid may be used within the context of the present invention. The protein level data acquisition means may be an ELISA assay, Luminex assay or an Olink Proximity Extension Assay (PEA). The following is presented by way of example and is not to be construed as a limitation to the scope of the claims.

[0161] Examples

[0162] Introduction

[0163] Diabetic kidney disease (DKD) is a common complication of diabetes, resulting in progressive loss of kidney function overtime. Identifying groups of patients at risk of severe DKD and adverse outcomes, such as dialysis, ensures that patients receive any appropriate, available treatment interventions. In the work described in these examples, the inventors demonstrated that small combinations of proteins, measurable in blood, can stratify patients, enabling identification of CKD / DKD patients with elevated risk of rapid deterioration and disease progression. Furthermore, the inventors showed that patients can also be stratified into categories associated with disease progression using routinely collected clinical measurements. Classifier models were developed for stratifying patients into multiple CKD / DKD patient endotypes that capture differential disease severity trajectories (herein referred to as prognostic models). Using baseline clinical information and high-throughput plasma proteomic profiling, the inventors identified novel combinations of biomarkers that can predict disease risk progression with high accuracy.

[0164] The inventors stratified a cohort of CKD / DKD patients into distinct groups using two clustering approaches: Clustering A (which identified two groups, referred to herein as A1 , A2) and Clustering B (which identified three groups, referred to herein as B1 , B2, B3), with each cluster exhibiting different disease trajectories, including variations in the emergence and timing of outcomes. Two types of prognostic models were developed, based on either clinical or proteomic features measured at baseline. In the training sets, these models successfully stratified patients into either A1 or A2, and B2 or B3, achieving area under the receiver operating curve (AUCs) of 91% and 83% for the clinical models (respectively for clustering A and clustering B), and 99% for both proteomic models for the batch 1 cohort. These models were further tested on external, independent cohorts, where associations between the predicted clusters and disease progression mirrored trends observed in the original patient clusters. The proteomics-based A1 vs A2 model was further validated in a separate cohort of DKD patients, batch 2, and additional proteomics-based A1 vs A2 models were developed using batch 2 data. The inventors further demonstrated that the A1 and A2 patient endotypes could be recreated using two validation cohorts, SHARE (McKinstry et al., 2017) and CHUIMI (Boronat et al., 2015).

[0165] Example 1: Identification of patient clusters and discovery of prognostic models

[0166] In the work described in the present example, the inventors stratified patients into distinct groups using two clustering approaches: Clustering A (A1 , A2) and Clustering B (B1 , B2, B3) and developed prognostic models using baseline clinical data and plasma proteomic measurements separately. Other strategies for developing models for stratification were also tested, which exhibited poorer performance. Methods

[0167] Cohort descriptions. Discovery of prognostic models was performed using a prospective cohort composed of patients with all-cause kidney disease, from the SKS cohort (described in Chinnadurai et al. 2023). The SKS recruits patients aged >18 years old with non-dialysis CKD who have been referred to the renal services at Salford Royal NHS Foundation Trust in the United Kingdom. The following inclusion criteria were used for patients from the Salford Kidney Study (SKS): (i) Diagnosis of Type 2 Diabetes (T2D) (or haemoglobin A1C (HbA1C) > 6.5% (48 mmol / mol)) at study entry; (ii) Diagnosis of chronic kidney disease (CKD) stage 3a, 3b or 4 at study entry; (iii) No diagnosis of Type 1 Diabetes (T1D) or cancer at study entry; (iv) DNA sample available; (v) At least two timepoints of clinical information available: one at study entry and another at least three years later. If the participant dies or is diagnosed with ESRD / kidney failure within three years, the time of death or diagnosis will serve as the second timepoint; (vi) Minimum level of clinical information available: month and year of birth; biological sex; smoking history; history of cardiovascular, cerebrovascular or thromboembolic diseases; systolic and diastolic blood pressure; heart rate; body mass index (BMI); acute kidney injury (AKI) events; haemoglobin; haematocrit; creatinine; estimated glomerular filtration rate (eGFR); sodium; potassium; albumin; phosphate; corrected calcium; urine protein : creatinine ratio (uPCR); serum lipid panel; HbA1C; concomitant medication use by drug class or individual drug.

[0168] Endotype validation was performed in a cohort of patients recruited at Centro Hospitalario Universitario Insular Materno Infantil (referred to herein as the CHUIMI cohort) in Las Palmas de Gran Canaria, Spain and is described in the present example. An additional validation was performed in the SHARE cohort, described in Example 7. SHARE is an ongoing voluntary register of individuals who are interested in taking part in health research (McKinstry et al., 2017). To identify a suitable participant cohort from SHARE for endotype validation, various inclusion criteria were applied to the SHARE registry to mimic the SKS cohort (described in Example 7 - Methods). Individuals in the CHUIMI cohort were recruited as part of the Characterization of Chronic Kidney Disease Associated with Diabetes (CERCA)-Diabetes Study which aimed to identify the sociodemographic, clinical, and genetic characteristics that determine the development and progression of advanced chronic kidney disease related to T2D mellitus in the population of the southern area of Gran Canaria (Boronat et al., 2015). Patient biochemical measurements were recorded at baseline and during follow-up visits. Follow-up was carried out from the time they were included in the study until their entry into dialysis, death or for a maximum of 4 years. Two groups of patients in the CHUIMI cohort were included in the analyses presented here; cases, and controls. Cases were defined as patients diagnosed with T2DM who attended the advanced chronic kidney disease (ACKD) consultation of the Nephrology Service of HUIGC with the clinical diagnosis of CKD secondary to diabetic nephropathy. According to the usual practice of the Nephrology Service, patients are referred to this consultation when they present with chronic kidney failure in stages 4 or 5 according to the National Kidney Foundation classification (GFR <30 mL / min / 1.73 m2). Controls were defined as patients who attended the ACKD consultation of the Nephrology Service of HUIGC who do not have diabetes mellitus. Cases and controls will be referred to as DKD and CKD, respectively. The prognostic models were evaluated in two cohorts of patients: the CHUIMI cohort; and the UK BioBank (www.ukbiobank.ac.uk / , referred to herein as the UK biobank cohort), a large-scale population study. The UK BioBank cohort generally has less advanced stages of disease than the discovery cohort, since it is a population study. This serves as a good test of the generalisability of the method.

[0169] Further cohort characteristics are detailed in Table 1a and in Figs. 5-6 (which show characteristics of the discovery (SKS) cohort). Model discovery in the SKS cohort is described in the present example. Model validation in the CHUIMI and UK biobank cohorts is described in Example 2.

[0170] Table 1a. Overview of the three cohorts used to validate and test the performance of the prognostic models for stratifying DKD patients based on disease trajectories.

[0171]

[0172] Patient endotype identification. Patient endotypes were identified using longitudinal creatinine and haematocrit values from patients in the SKS cohort. Longitudinal data was normalised on a per-patient per-test basis (i.e. haematocrit and creatinine separately) including scaling the data, collapsing to single month values (median values per month), removing outlier points, and removing data points in the last 3 months of life. In particular, the following steps were taken: (i) for each test, duplicate readings taken on the same day were aggregated using the median, followed by a weekly median calculation for each patient; (ii) outliers were removed (if they were determined to be outside of a normal range for that individual), then (iii) data was collapsed to single month values (following outlier removal, monthly median calculations were taken for each patient and each test.), and (iv) data in the last 3 months of life was removed. Data was scaled to the healthy reference range for the particular lab test (to avoid skewing the distance metrics for patients who were undergoing serious biological changes - in dynamic time warping, the start and end of sequences tend to contribute disproportionally to the estimated similarity, so removing data points from the last 3 months of life attempts to reduce the phenomenon of “total failure” state). The biochemical measurements were then normalised to reference ranges (Table 1b). The mean and the standard deviation were calculated for each reference range for each biochemical measurement, and these values were used to calculate z-scores from the biochemical data as follows:

[0173] z = value — meari / sd where value is the monthly median calculation for each patient for each test. This process was done separately for males and females if they had different reference ranges.

[0174] Table 1b. Reference ranges used to normalise biochemical measurements. Reference ranges reported by the NHS were used.

[0175]

[0176] Patients were filtered based on their number of observations for each biochemical test. The minimum observation threshold was 6 for each of creatinine and haematocrit. Patients with fewer observations than this threshold for both tests were excluded from the analysis.

[0177] A bespoke method was developed to cluster patients using real world longitudinal clinical data. Following data normalisation, dynamic time warping (Sakoe and Chiba 1990) was used to calculate the distance between each pair of patients, with creatinine and haematocrit values analysed separately. An n xn xt tensor (where n is the number of patients, and t is the number of tests, here t=2 because there are two clinical measurements, creatinine and haematocrit (or haemoglobin for CHUIMI) was created to store these pairwise distances. Tucker decomposition (Tucker 1966) was applied to this tensor. This process identifies groups of patients that have different temporal trajectories in terms of creatinine and haematocrit (i.e. the combination of both of these). UMAP was subsequently applied to the n x2 factor matrix output from Tucker decomposition (where the 2 comes from the “core size” parameter of the Tucker decomposition, which is the dimension of the core tensor- representing the underlying structure or essence of the original tensor, its size dictates the amount of information retained after the decomposition), and patient clusters were identified from UMAP 1 and UMAP 2. K-means (Jin and Han 2011) with k=2 identified two clusters (A1 and A2) and DBSCAN (Ester et al. 1996) identified three clusters (B1, B2 and B3).

[0178] The association between cluster labels and disease trajectory was explored through fitting Cox proportional hazard regression models, with age, sex and eGFR at baseline accounted for as covariates in the models (Andersen and Gill 1982; Therneau and Grambsch 2010). A diagnosis of ESRD (endstage renal disease) or occurrence of RRT (renal replacement therapy) was used as the event of interest in these Cox regression models. Associations between endotypes and clinical traits at baseline were tested using one-way ANOVA tests for continuous traits and Chi-squared tests for categorical traits. Associations between endotypes and classes of medications were also tested using the Chi-squared test. All p-values were adjusted for multiple testing using the Benjamini-Hochberg adjustment (Benjamini & Hochberg, 1995). Evaluating performance of clinically-used CKD stratifying tools in the Discovery Cohort. Using the discovery cohort (SKS), the inventors evaluated the performance of the following clinically-used CKD stratifying tools: CKD staging, KDIGO (Kidney Disease: Improving Global Outcomes) CKD guidelines (KDIGO CKD Work Group, 2024; Stevens et al., 2024), and the CKD-PC (Chronic Kidney Disease Prognosis Consortium) Risk Tool (Nelson et al., 2019). For CKD staging, patients were grouped using eGFR at study start as follows: Stage 3a: eGFR 45-59; Stage 3b: eGFR 30-44; Stage 4: eGFR 15-29; Stage 5: eGFR < 15). For the KDIGO risk categories, the urine albumin to creatinine ratio (UACR) was estimated using the urine protein to creatinine ratio (UPCR) taken at study start. Predicted UACR (pACR) was estimated using the “Crude Model” described in Sumida et al. (2020) as follows:

[0179] pACR = exp (5.3920 + 0.3072 x log (min (UPCR / 50, 1)) + 1.5793 x log (max(min(UPCR / 500, 1),0.1)) + 1.1266 x log (max (UPCR / 500, 1))).

[0180] The 2-year and 5-year CKD-PC risk scores were calculated using the equations described in Tangri et al. (2016) and as follows:

[0181] 2-year risk score: 1 - 0.9750Aexp (-0.2201 x (Age / 10 - 7.036) + 0.2467 x (Sex - 0.5642) - 0.5567 x (eGFR / 5 - 7.222) + 0.4510 x (log(UACR+1xia6) - 5.137))

[0182] 5-year risk score: 1 - 0.9240Aexp (-0.2201 x (Age / 10 - 7.036) + 0.2467 x (Sex - 0.5642) - 0.5567 x (eGFR / 5 - 7.222) + 0.4510 x (log(UACR+1xia6) - 5.137)).

[0183] The association between patients in the different risk categories and a disease-relevant composite outcome was tested using Cox proportional hazard models using the survival R package (Therneau, 2024), using time to the composite outcome (if > 1 outcome was reached, the first outcome was taken) since the study start date. The composite outcome was as follows: renal replacement therapy (dialysis or kidney transplant), eGFR dropping below 15 (sustained for >1 month; ESRD), eGFR dropping to >40% of starting eGFR (sustained), or death from metabolic syndrome related causes. Age (at study start), sex and eGFR (at study start) were included in the models as covariates to adjust for.

[0184] Endotype validation. Endotypes were validated in two independent validation cohorts - SHARE (further described in Example 7) and CHUIMI (described in the present Example). For SHARE, longitudinal biochemical measurements of creatinine and haematocrit were pre-processed using the same method as described for SKS. For CHUIMI, haematocrit data was unavailable so haemoglobin data was used instead. Data was scaled to the healthy ranges used in SKS (Table 1b). In SHARE, a different filtering approach was applied due to a wider range in the number of observations compared to SKS, reflecting the nature of electronic health record data. Individuals were excluded if they had fewer than 6 creatinine or 4 haematocrit measurements, or if their number of observations exceeded the 90th percentile for each test. The lower threshold for haematocrit was chosen because it was generally measured less frequently than creatinine. In CHUIMI, a maximum of 5 data points (including baseline) were available for creatinine and haemoglobin per person, as such the more stringent thresholds for the minimum number of observations that were applied to SKS and SHARE were not applied. A threshold of at least 3 observations (including baseline) was used instead. For both datasets, dynamic time warping was applied to all individuals for longitudinal creatinine and haematocrit / haemoglobin separately. Tucker decomposition was applied to the two similarity matrices, followed by UMAP and k-means clustering (k=2). Differences in disease trajectories were explored between the clusters in both datasets separately using Cox proportional hazard models. For CHUIMI, dialysis was the only outcome available, however for SHARE the outcomes used were a composite of renal replacement therapy (dialysis or kidney transplant), eGFR dropping below 15 (ESRD; sustained for >1 month), eGFR dropping >40% from starting eGFR (sustained), or death from metabolic syndrome related cause (as used in SKS). As in SKS, in both validation cohorts, associations between endotypes and clinical traits at baseline were tested using one-way ANOVA tests for continuous traits and Chi-squared tests for categorical traits. All p-values were adjusted for multiple testing using the Benjamini-Hochberg adjustment (Benjamini & Hochberg, 1995).

[0185] Plasma proteomics data generation and normalisation. Characterisation of the plasma proteome was performed using the Olink® Explore 3072 / 384 panel (Olink, Uppsala, Sweden) which measures over 2900 known proteins. Proteomic profiling was performed using plasma samples from 125 patients in the SKS cohort. Data normalisation was performed as per the Olink NPX transformation (Assarsson et al. 2014), a normalisation value that takes into account technical variation whilst Iog2 transforming the data, resulting in relative rather than absolute protein quantification values. In short, the counts are scaled relative to an extension control and a plate control. The Extension Control (an antibody linked to two matched oligonucleotides) monitors the extension and readout steps independent of antigen binding, and is used for data normalization across samples. An external inter-plate control (IPC), is included on each plate and is used in a second normalization step. This control is made up of a pool of probes similar to the Extension control (Ext Ctrl), but generated with 92 matching oligonucleotide pairs. This improves inter-assay precision and allows for optimal comparison of data derived from multiple runs. The frequency with which protein measurements were below the limit of detection (LOD) was quantified for patients stratified according to whether they had ESRD or not. Proteins with values less than LOD in over 50% of patients with ESRD, and in over 50% of patients without ESRD were removed. Calculation of the LOD rate in both ESRD categories was performed to ensure that proteins displaying real biological differences were not inadvertently removed. Samples were randomised across two plates. There were no proteins significantly differentially expressed (SDE) between plates, with differential expression analysis performed using Limma (Ritchie et al., 2015). As a quality control check, proteins SDE between males and females were identified using Limma (Ritchie et al., 2015). SDE proteins were input into principal component analysis (PCA) and principal components (PC) 1 and 2 visualised with samples coloured according to their clinically recorded sex. Two samples were excluded as they were clustering with the opposite sex to their clinically recorded sex.

[0186] Plotting trends per endotype. To generate the results on Fig. 20, first, the inventors defined the first date of eGFR reaching 30 as the first time patients’ eGFR was between 28-32 based on the longitudinal eGFR data. Then, all lab dates were anchored to that first date by calculating the time of that test date from the time of the first eGFR reaching 30, with time measured in half years. The inventors then took the median value per individual within that time range. Then, grouping by endotype, they computed the mean and standard error of the mean per time point. The inventors selected only those time points that had at least 10 patients. Splitting baseline creatinine into two groups. To generate the results on Fig. 21, the inventors took baseline creatinine values for all patients and took the Euclidian distance between them and then ran hierarchical clustering using ward.D and a cut tree with k = 2 to generate two clusters. This resulted in a cluster with creatinine < 189 pmol / L and a cluster with creatinine > 189 pmol / L.

[0187] Discovery of prognostic models. Prognostic models were developed using baseline clinical data and plasma proteomic measurements separately. Data (either baseline clinical or plasma proteomic) was input into XGBoost (Chen and Guestrin, 2016) with in-sample performance evaluated through calculating the area under the receiver operating characteristic (ROC) curve (AUC) and sensitivity and specificity, using cluster labels derived using the previously described patient endotype identification model. Both the clinical and proteomic models for which results are shown had the following parameters (although as the skilled person understands this is likely to be fine tuned for any particular training data): eta = 0.3; max_depth = 6; lambda = 1 ; gamma = 5; min_child_weight = 1 ; subsample = 1 ; colsample_bytree = 1. Features were filtered prior to XGBoost as follows. For the baseline clinical traits, clinical features were removed if they were missing in over 50% of patients and if they were not measured in the CHUIMI cohort as well as the SKS cohort. For the Olink baseline proteomics, proteins were only retained for XGBoost models if they met the following criteria:

[0188] 1. They were significantly differentially abundant between any of the clusters (e.g., A1 vs. A2, B1 vs. B2, B1 vs. B3, B2 vs. B3, etc) at an unadjusted p-value of 0.05 (linear model t-test- Limma, see Ritchie et al. 2015). This ensured that protein variation driven by confounding factors (age, sex, plate, baseline eGFR) were removed from the potential biomarker candidates. An unadjusted p-value threshold was used to limit overfitting.

[0189] 2. They had evidence of expression in the blood. This involved either being annotated as “secreted into blood” in Human Protein Atlas (HPA; (Thul and Lindskog 2018); Accessed July 2024), or having a recorded abundance value in mass spectrometry or immunoassay data from blood in HPA.

[0190] 3. They were also measured in the Olink panel included in the UK BioBank. This is to ensure that the model could be tested in UK BioBank.

[0191] Results - Endotype identification

[0192] Patient clusters. The inventors developed a comprehensive framework to leverage real world clinical biochemical data to identify sub-populations using unsupervised techniques (see Methods). After preprocessing data, including applying a threshold for the minimum number of observations, longitudinal clustering was applied to data from 335 patients with both DKD (n=303) and CKD (n=32). The inventors used all measurements of creatinine and haematocrit that were routinely collected at clinical visits for the discovery cohort (Table 17). As the data was collected during real visits and not for a study designed specifically for this research, the lab tests are taken at different intervals and frequencies for each individual. As such, the inventors utilised an approach that handles multi-dimensional data in a scale free manner and looks for patients who are most similar to each other based on their clinical trends (see Methods). Clustering patients from the SKS cohort (see Methods) based on longitudinal creatinine and haematocrit values resulted in two different clustering combinations depending on the clustering algorithm used, with Clustering A composed of A1 (n=159) and A2 (n=175) patients (Fig. 3A), and Clustering B composed of B1 (n=159), B2 (n=92) and B3 (n=83) patients (Fig. 3B). Cluster A1 was identical to cluster B1 , with cluster A2 being split into B2 and B3. Cox regression analysis revealed significant differences in the time to event (defined as either RRT or ESRD) between clusters A1 and A2 (p-value: 1.1x1 O’11; Fig. 3C) and between clusters B1 , B2 and B3 (p-value: <2x10-16; Fig. 3D). The number of patients with ESRD or requiring RRT was higher in cluster A2 (n=93; 53% of patients in A2) than A1 (n=5; 3% of patients in A1). Similarly, the number of patients in B3 with ESRD or requiring RRT was highest in cluster B3 (n=73; 89% of patients in B3) than both B2 (n=19; 21% of patients in B2) and B1 (n=5; 3% of patients in B1), increasing in a stepwise manner from B3 to B2, to B1 (Fig. 3D). A reanalysis of the same data with a minor outlier detection error corrected showed the same two groups, A1 (n=162) & A2 (n=173), with clear significant differences in their survival curves for events (Cox Hazard ratio p-value: 1.01x10-13, Fig. 3H) and visible stratification (Fig. 19A). As such, these groups are also referred to below as SKS-slow (A1) and SKS-fast (A2), respectively. When contrasting the trajectories of creatinine and haematocrit between the patients in the two endotypes, the inventors observed that individuals in SKS-fast trended towards more rapidly increasing creatinine levels, in comparison to individuals in SKS-slow who showed steady, flatter trajectories of creatinine over time (Fig. 20A). For haematocrit, individuals in SKS-slow displayed an increasing trajectory overtime whilst individuals in SKS-fast displayed a decreasing trajectory overtime (Fig. 20B).

[0193] The inventors evaluated whether there were differences in baseline clinical traits taken at enrolment into the study between SKS-slow and SKS-fast, with significant differences observed in age, diastolic blood pressure, and age of T2D onset (Table 1 C). As for baseline biochemical tests, creatinine / eGFR, albumin, haemoglobin, phosphate and UPCR were all significantly different (FDR < 0.05, Table 1 C).

[0194] When medication classes were looked at, there were significant differences between endotypes in rates of EPO, the number of antihypertensives, Vitamin D analogs and calcium channel blockers at baseline (FDR < 0.05, Table 1 C), with n=44 (25%) of SKS-fast on EPO, vs. n=13 (8%) of SKS-slow on EPO, n=32 (20%) of SKS-fast receiving Vitamin D analogs, vs. n=9 (6%) of SKS-slow receiving Vitamin D analogs, and with n=92 (56%) of SKS-fast receiving CCBs, vs. n=59 (40%) of SKS-slow receiving CCBs. For all of these baseline variables, the inventors did not see clustering of these traits within the clusters identified, indicating that the clustering is not driven by these factors. Furthermore, despite the associations between age and eGFR and endotypes, when including age and eGFR at study start in the Cox model, there is still a significant association between endotypes and time to outcome (Cox Hazard model p-value: 4.95x10-12).

[0195] SKS, the cohort that underwent endotype discovery, was mainly composed of patients with DKD (n=303), however there were some patients with CKD (n=32) that were also endotyped. Patients did not cluster into DKD / CKD groups using the longitudinal endotyping method, indicating that these findings could be expanded into CKD.

[0196] The inventors additionally investigated whether creatinine and haematocrit taken at baseline were able to capture the same information as the proposed approach using the full trajectories of these two biochemical measurements. Whilst levels of creatinine were significantly different at baseline between our endotypes (FDR 3.70x1 O'24), levels of haematocrit at baseline were not significantly different. Given the significant differences in creatinine at baseline between the identified endotypes, the inventors further explored baseline creatinine on its own, simply stratifying the patients based on high vs. low creatinine into two groups, using the cutoff of 189 pmol / L (see Methods, Fig. 21). The survival differences here were not significant (Cox Hazard model p-value: 0.607), the overlap between this simplistic approach and the identified clusters was not one to one (n=125 (77%) SKS-slow in the low creatinine group and n=125 (72%) SKS-fast in the high creatinine group) and the longitudinal endotypes described herein showed better split in survival (compared Fig. 3H vs. Fig. 21 B).

[0197] Table 1C. Results showing comparisons performed between fast progressor (SKS-fast, SHARE-fast, CHIUMI-fast) and slow progressor endotypes (SKS-slow, SHARE-slow, CHIUMI-slow) endotypes in baseline clinical traits. FDR = false discovery rate calculated using Benjamini-Hochberg (Benjamini & Hochberg, 1995). Test statistics and FDRs are shown for SKS, CHUIMI / CERCA-Diabetes and SHARE endotypes. Proportion tests were applied to categorical traits and Kruskal-Wallis tests were applied to continuous traits. Only variables with significant p-values (FDR < 0.05) are shown. RRT: renal replacement therapy; ESRD: end-stage renal disease; eGFR: estimated glomerular filtration rate; UPCR: urine protein to creatinine ratio; EPO: erythropoietin; T2DM: type 2 diabetes mellitus; FDR: false-discovery rate; NS: not significant.

[0198] <

[0199]

[0200]

[0201]

[0202]

[0203] Evaluation of clinically-used risk stratification methods. The inventors started by evaluating the performance of the clinical tools used currently for risk stratifying CKD patients in the discovery cohort (Table 17). When using the entire cohort with baseline eGFR data available (n=437; Table 17), there were not significant associations between CKD stage (using eGFR at study start; Fig. 3E) and KDIGO risk categories (using eGFR and uACR at study start; Fig. 3F) and time to outcome (defined as a composite of RRT, sustained drop in eGFR > 40% from baseline eGFR levels, death from metabolic syndrome related cause, or sustained reduction in eGFR < 15). When restricting to only those with more advanced CKD at baseline but prior to ESRD (CKD stages 3a, 3b, 4) there were significant associations between time to composite outcome and CKD stage 4 vs. CKD stage 3a and 3b (Cox Hazard ratio p-value: 0.018). When we applied the CKD-PC risk score, there was a significant association between time to composite outcome and the 2-year (Cox Hazard ratio p-value: 1.45x10-5; Fig. 3G) and 5-year (Cox Hazard ratio p-value: 1.75x10-6) CKD-PC risk score.

[0204] Table 17. Clinical and demographic characteristics of individuals included in each of the cohorts, with values variables shown at study start (baseline), unless otherwise noted.

[0205]

[0206] Replication and validation of endotypes. The inventors aimed to replicate and validate these findings in two other independent validation cohorts, SHARE and CHUIMI. The SHARE cohort is an ongoing voluntary register of individuals registered in Scotland pooling electronic healthcare records, enabling the framework to be recreated. In the SHARE cohort, patients were selected to match the discovery cohort (see Example 7 - Methods), resulting in data from n=824 participants being included. The inventors applied the same framework to this cohort and identified 2 clusters, S1 (n=375) and S2 (n=449) (Fig. 19B). The inventors find clear differences in frequency and time to outcome events (using the composite described in Methods) in these two groups confirming that there are also stable / slow progressors and fast progressors in this independent cohort (Cox Hazard model p-value: 3.83x10-27, Fig. 22A), these are referred to as SHARE-slow and SHARE-fast. Of the n=449 participants in SHARE-fast, n=241 (54%) reached an outcome in comparison to just 37 (10%) of the n=375 participants in SHARE-slow.

[0207] Further validation was performed in a second independent validation cohort, CHUIMI, which was composed of patients from the Canary Islands, including both CKD (n=114) and DKD (n=73) patients. The CHUIMI dataset is not composed of longitudinal real world clinical data, but rather baseline lab values and yearly follow-up values for patients in the study for up to 4 years. As haematocrit was not collected in CHUIMI, and the inventors saw in both the discovery cohort and in SHARE that it was highly correlated with haemoglobin, which was collected in CHUIMI, the inventors used haemoglobin and creatinine instead of haematocrit and creatinine. The correlation between haematocrit and haemoglobin measurements taken on the same date across all follow-up data (on a per-person basis) was 0.964 (Pearson’s correlation; p-value < 2.2x10-16; Fig. 23A) in SKS and 0.957 in SHARE (p-value < 2x10-16;

[0208] Fig. 23B). The inventors applied the same framework to CHUIMI using data from patients that had at least s observations (including baseline; n=187). Despite the lower number of observations in CHUIMI than both SKS and SHARE, two distinct clusters are observed (C1 , n=70; C2, n=117; Fig. 19C) and there is a significant difference in time to outcome (dialysis) observed between these two clusters (Cox Hazard model p-value: 1.72x10-4), providing further evidence and support for the endotypes described herein (Fig. 22B). These are therefore referred to as CHUIMI-slow and CHUIMI-fast. Of the n=117 participants in CHUIMI-fast, n=67 (56%) required dialysis in comparison to just 13 (19%) of the n=70 participants in CHUIMI-slow, with a Chi-Squared p-value 5.1x10-7. Interestingly, as seen in SKS, no stratification of CKD and DKD was observed between these clusters (Fig. 24), suggesting that these findings could be expanded into CKD as well.

[0209] As performed in SKS, the associations between various baseline traits and cluster labels were explored in both SHARE and CHUIMI. Age, creatinine, eGFR, UPCR, albumin, haemoglobin, age at T2D diagnosis and sex were significantly different between endotypes in SHARE (Table 1 C), with all except sex also significantly different in SKS. Age, eGFR and albumin were also significantly different between endotypes in CHUIMI, alongside diastolic blood pressure and phosphate levels taken at study start (all were significant in SKS; Table 1 C).

[0210] Discussion - Endotype identification

[0211] CKD is a heterogeneous disease in which some individuals progress much more rapidly than others. Current clinical risk stratification models attempt to identify those at higher risk of more rapid disease progression, providing moderate performance but with room to improve. By harnessing real world longitudinal clinical data, the inventors were able to stratify the patient cohort into two groups - SKS-fast and SKS-slow - with significant differences in disease progression. Compared to current clinical risk stratification models, the new endotyping approach results in considerably improved separation by composite outcome endpoint (ESRD, RRT, sustained eGFR drop > 40% of starting eGFR, or metabolic syndrome related cause of death), exemplified in Fig. 3 and shown by substantially lower Cox Hazard ratio p-values.

[0212] The two endotypes were discovered through combining the trajectories of haematocrit and creatinine measurements overtime, and using the same approach, they have been replicated in two independent validation cohorts. Despite the differences between the discovery cohort and the two validation cohorts, the new endotypes have been successfully replicated with significant associations between endotypes and outcomes (Cox hazard ratio p-value: 3.83x10-27and 1.72x10-4for SHARE and CHUIMI, respectively). In the CHUIMI cohort, the inventors not only see that these endotypes are replicated but they also find that even when limited to a smaller number of time points, the endotypes are preserved and are also applicable to not only DKD but also CKD. Furthermore, although the inventors attempted to match the SHARE cohort to the SKS cohort, the SHARE participants are likely at an earlier stage of disease. This is expected, as SHARE is a population-based study rather than a patient-based one, and is reflected in the higher median eGFR (52.8 vs. 29.6 mL / min / 1 ,73m2). Nonetheless, the new endotypes were successfully replicated in the SHARE cohort, suggesting they may be generalisable to earlier stages of DKD.

[0213] The fast progressor endotype described here is characterised by faster increases in creatinine combined with decreasing haematocrit trajectories over time. Creatinine is an established marker of kidney function, indicating that these individuals have more rapid decline in kidney function than those in the slow progressor endotypes. Haematocrit measures the proportion of red blood cells in the blood, with low haematocrit indicative of anaemia. As kidney disease progresses and kidney function declines, anaemia is a common complication of CKD, caused in part by the kidneys’ reduced ability to produce erythropoietin (EPO), the hormone that triggers production of red blood cells. Whilst anaemia has been widely associated with CKD, limited studies have explored haematocrit trajectories in CKD, with one study reporting an association between lower haematocrit at baseline and 17-year risk of ESRD (Iseki et al., 2003). Since both elevated creatinine and anaemia are indicative of kidney function, disentangling the cause of rapid disease progression from potential symptoms is complicated in this case, however it provides further insights into the underlying biological pathways disrupted in individuals at risk of worse disease outcomes.

[0214] In conclusion, the inventors have harnessed the power of real-world clinical data from over 1 ,300 individuals with CKD (and DKD), to identify two novel endotypes of CKD, with meaningful differences in time to disease-relevant outcomes and variations in underlying molecular make-up. Individuals in the fast progressor endotypes (SKS-fast, SHARE-fast, CHUIMI-fast), who are characterised by increasing creatinine, decreasing haematocrit, and an omic profile dominated by increased TGF- and IGF-mediated inflammation and lipid, glucose and kidney structure dysregulation, would greatly benefit from precision medicine targeted therapies designed to retain kidney function and reduce time to outcome. Precision medicine therapies combined with novel patient stratification approaches to identify fast progressors at an earlier stage of disease has the potential to revolutionise clinical care of patients with CKD, reducing the rates of patients going onto require costly, invasive and life-changing renal replacement therapy.

[0215] Results - Prognostic models

[0216] Discovery of prognostic models. Prognostic models were developed using baseline clinical data (data collected when the patient entered the study which is also when the plasma samples were collected that are used for proteomics) and plasma proteomic data separately, with two prognostic models developed for each type of data. The data was corrected for baseline eGFR (used in disease staging) prior to downstream analysis, by including eGFR at study start (i.e. baseline) as a covariate for example in the Cox regression analyses and when identifying proteins that are significantly different between clusters in the A1 vs. A2 model. Firstly, a model was identified for stratifying between patients in A1 vs. A2. Since cluster A1 is identical to cluster B1 , an additional classification model for identifying patients in cluster B1 was not required. As such, a model for stratifying between patients in clusters B2 vs. B3 was developed.

[0217] Clinical models. The clinical A1 vs. A2 model (Fig. 4) was composed of 8 baseline clinical features (creatinine, eGFR, age at study start, diastolic blood pressure, systolic blood pressure, albumin, phosphate, calcium channel blockers; Table 2, Fig. 5). These features were selected from a set of over 50 clinical features as best performing in the training data after iterations of training with an XGBoost model. All of these 8 baseline clinical features had non-zero gain indicating that they contribute to the prediction made by the XGBoost model. When the model was applied to the training set with cluster labels predicted, the model achieved an AUC of 91% (95% confidence interval: 88%-94%) with a sensitivity of 84% and specificity of 87% (Fig. 4B). The clinical B2 vs. B3 model (Fig. 4) was composed of 7 baseline clinical features (Table 3, Fig. 6). When the model was applied to the training set with cluster labels predicted, the model achieved an AUC of 83% (95% confidence interval: 77%-89%) with a sensitivity of 67% and specificity of 88% (Fig. 4E). Cox regression analysis applied to the predicted labels (using eGFR at baseline (i.e. study start), age and sex as covariates) revealed significant differences in time to outcome between clusters with a p-value of 3.2x10-5for the A1 vs. A2 model (Fig.

[0218] 4C), and a p-value of 8.9x10-9for the B2 vs. B3 model (Fig. 4F).

[0219] Table 2. Features included in the prognostic clinical model for distinguishing between A1 vs. A2. Gain represents the relative importance of each feature in the model. Note all measurements were obtained in serum but albumin in urine in mmol / 24 hours may also be used. The same is true for creatinine.

[0220]

[0221]

[0222] Table 3. Features included in the prognostic clinical model for distinguishing between B2 vs. B3. Gain represents the relative importance of each feature in the model.

[0223]

[0224] Proteomic models. The proteomic A1 vs. A2 model (Fig.7) was composed of 8 proteins (PTGDS, F11 R, CST3, SDC1 , PILRA, FAS, CD160, TFPI; Table 4; Fig. 8). When the model was applied to the training set with cluster labels predicted, the model achieved an AUC of 99% (95% confidence interval: 98%-100%) with a sensitivity of 93% and specificity of 98% (Fig. 7B). The proteomic B2 vs. B3 model was composed of 5 proteins (LAIR2, SELPLG, FGFR2, TNFRSF11A, PM20D1 ; Table 5; Fig. 9). When the model was applied to the training set with cluster labels predicted, the model achieved an AUC of 99% (95% confidence interval: 98%-100%), with a sensitivity of 91% and a specificity of 100% (Fig. 7E). Cox regression analysis applied to the predicted labels revealed significant differences in time to outcome between clusters with a p-value of 2.3x10-5for the A1 vs. A2 model (Fig. 7C), and a p-value=1.6x10-7for B2 vs. B3 (Fig. 7F).

[0225] Table 4. Features included in the prognostic proteomic model for distinguishing between A1 vs. A2.

[0226] Gain represents the relative importance of each feature in the model.

[0227]

[0228]

[0229] Table 5. Features included in the prognostic proteomic model for distinguishing between B2 vs. B3.

[0230] Gain represents the relative importance of each feature in the model .

[0231]

[0232] Further exploration of models for stratification. The inventors explored the possibility of a three-way multiclass XGBoost classifier for stratifying patients into B1 , B2 and B3. This method returns three probabilities, one for each cluster, with the highest probability indicating the cluster to which the patient is most likely assigned. A multiclass model using baseline clinical data included 16 variables and assigned 51% (n=149), 1% (n=3), and 48% (n=139) of patients in the training set to B1 , B2 and B3 respectively, corresponding to AUCs of 79%, 50% and 73%, respectively. This indicated that a multiclass model using baseline clinical data did not as effectively stratify patients into the three groups, possibly due to higher similarity between B2 and B3 patients in comparison to B1 patients. This trend was also observed in the proteomic data in an equivalent model, in which the AUCs were far lower than for a two-step approach including two binary classifiers (described in Example 4).

[0233] Increasing the number of features in a diagnostic test increases the complexity and cost of such a test, particularly if the features are molecular features such as proteins. The inventors attempted to introduce regularisation into the XGBoost models to identify smaller yet still discriminatory models. When high regularisation was introduced in the XGBoost models (e.g., resulting in models containing few protein features - by increasing the alpha and lambda parameters setting the L1 and L2 regularisation terms on weights, together controlling model complexity), model performance was compromised. This was particularly visible when the performance of the B2 vs. B3 model was tested in an external cohort (UK BioBank for proteomics). When a B2 vs B3 model composing two proteins identified from the training dataset composed of all SKS patients, was tested in the UK BioBank cohort (independent validation cohort as shown in Example 2 below), it did not stratify patients into groups that differed based on their time to ESRD or RRT using Cox regression models. Relaxing the regularisation parameters meant that more features entered the model (4 being enough for significant results, 5 further improving separation and proving to be the optimal feature number in this particular training data), and significant associations were observed between predicted clusters and time to disease outcome.

[0234] The inventors also explored using different classification algorithms to identify prognostic models, including Lasso and Elastic Net. Both approaches led to considerably larger models being identified, as shown by Lasso models for expanding the protein biomarker list (composed of 22 proteins for A1 vs. A2, and 33 proteins for B2 vs. B3) described in Example 3. XGBoost provided a better alternative, as smaller models were identified achieving equivalent model performance.

[0235] Various steps were performed to filter the proteins input into the proteomic XGBoost models. The first of these filters (see Methods) involved filtering proteins based on their significance between clusters when accounting for age, sex, plate and baseline eGFR. When this step was omitted, the proteins that were included in the models were often confounded with one of these variables, meaning differences were also driven by a confounding variable in addition to cluster. This issue was resolved once the filtering step was included.

[0236] Example 2: Testing prognostic models in external cohorts

[0237] In the work described in the present example, the prognostic models of Example 1 were tested in two external cohorts.

[0238] Methods

[0239] The prognostic models were evaluated in two cohorts of patients: the CHUIMI cohort and UK BioBank cohort (see Example 1). In particular, the clinical prognostic models were tested in the CHUIMI cohort. Predicted clusters were assigned using the thresholds derived from the SKS discovery data, with the threshold the point that maximises sensitivity and specificity simultaneously. Fisher’s exact tests were used to test the association between predicted clusters and the proportion of patients in the cluster requiring RRT (p-values < 0.05 were considered significant). Patients in the CHUIMI cohort had up to 4 measurements of creatinine and eGFR overtime, allowing the slope of both of these variables to be calculated. The association between creatinine and eGFR slope and predicted cluster labels was tested using the ANOVA test (Girden 1992) followed by the Tukey’s honest difference test (Tukey 1949), with p-values < 0.05 considered significant. The proteomic models were tested in the Olink data derived from DKD patients present in the UK BioBank cohort (data first accessed July 2023). Patients with eGFR < 30 were excluded from the analyses to focus on patients earlier in the course of disease. Cox proportional hazard regression models were fitted to test the association between predicted clusters and time to disease outcome, with the disease outcomes as either ESRD or dialysis (falling within the scope of RRT - the information on transplant not being available in the UK biobank data).

[0240] Results

[0241] The prognostic models were tested in two external cohorts, with the clinical models tested in the CHUIMI dataset and the proteomic models tested in the UK BioBank. A total of 151 DKD patients were evaluated in the CHUIMI dataset (Table 1), of which 82 patients reached RRT. The spectrum of outcome information available in CHUIMI did not fully reflect the outcome information available in SKS, with ESRD data and time to outcome unavailable. As such, the association between predicted clusters and three surrogate outcomes was evaluated: whether patients required RRT (dialysis or kidney transplant); creatinine slope over up to 4 years (1 measurement per year); eGFR slope over up to 4 years (1 measurement per year).

[0242] For the A1 vs. A2 clinical prognostic model, 31 CHUIMI patients were predicted as A1 and 120 were predicted as A2. There was an over-representation of patients requiring RRT in cluster A2 compared to cluster A1 (Fig. 10A; n=73 vs. n=9), with a Fisher’s exact p-value of 2.158x1 O'3. A significant association was observed between creatinine slope over time (calculated using a linear model) and clustering A (Fig. 10B), with a p-value of 0.035 (ANOVA). There were no significant differences between eGFR slope and any of the predicted clustering A labels in CHUIMI (Fig. 10C - eGFR measurements over ~4 years were plotted per person and the slope evaluated, then the association between slope and cluster is tested and that is not significant). This is likely at least in part because there are a limited number of eGFR measurements over time for this cohort - indeed, significant outcome differences can still be seen.

[0243] To stratify CHUIMI patients according to clustering B, patients labelled as A1 by the A1 vs. A2 model were labelled as B1 , with patients labelled as A2 stratified into B2 or B3, resulting in 31 patients predicted as B1 , 51 predicted as B2, and 69 predicted as B3. The proportion of patients requiring RRT increased from cluster B1 (n=9), to B2 (n=23), to B3 (n=50), with a Fisher’s exact p-value of 7.06x1 O'5(Fig. 10D). A significant association was observed between clustering B labels and creatinine slope (ANOVA p-value: 0.0106), which when broken down into pairwise associations, showed that the significance was driven by the B1 vs. B3 comparison (Tukey’s HSD p-value: 9.017x1 O'3). There were no significant differences between eGFR and any of the predicted clustering B labels in CHUIMI (Fig.

[0244] 10F).

[0245] The proteomic models were tested in the UK BioBank with all DKD patients predicted as either A1 or A2, and B1 , B2 or B3, using the stepwise approach described above (i.e. first classifying patients between A1 and A2, then classifying A2 patients between B2 and B3). A total of 314 patients were predicted as A1 and 748 were predicted as A2. The proportion of patients reaching either ESRD or RRT was higher in predicted cluster A2 than A1 (Fig. 11 A; n=90 vs. n=11), with a Fisher’s exact p-value of 5.44x10-6. Cox regression analysis revealed a significant difference in disease trajectories between the predicted clusters (p-value: 0.03) with A2 showing faster disease progression than A1 (Fig. 11B). The 314 patients that were predicted as A1 were also predicted as B1 , with the remaining 748 predicated as A2 split into B2 (n=314) and B3 (n=434). As with Clustering A, a significant association was observed between Clustering B and the proportion of patients reaching ESRD or dialysis (Fisher’s exact p-value: 5.38x10-8), with a stepwise increase from B1 (n=11) to B2 (n=23) to B3 (n=67; Fig. 11C).

[0246] Cox regression analysis demonstrated a significant difference between predicted clusters and disease trajectories (p-value: 0.002) with B3 showing fastest disease progression, followed by B2, and then B1 (Fig. 11 D). Example 3: Evaluating other methods combining features, and robustness evaluation of prognostic models

[0247] In this example, other methods of combining features were evaluated. The robustness of prognostic models was also explored by testing model performance using subsets of features and patients. Methods

[0248] Evaluating methods of combining features. Since XGBoost was used to not only identify the key discriminatory features, but to also combine the features into a single score for each patient, the inventors explored the influence of changing the machine learning method for combining features. Two methods were used. Firstly, a generalised linear model (GLM) was used to calculate feature weights for the features selected by XGBoost, which were then multiplied by the feature raw values themselves resulting in a single per-patient value (GLM method). Secondly, the per-patient sum of feature values with positive coefficients (as calculated by GLM) were subtracted from the per-patient sum of feature values with negative coefficients to return a single per-patient risk score (simple risk score method). The output from both of these methods was evaluated through calculating the AUC, sensitivity and specificity in relation to the known cluster labels.

[0249] Evaluating robustness of prognostic models. Robustness of prognostic models was explored through testing model performance using subsets of features and patients. To evaluate model performance with a subset of features, all sub-combinations of features were combined using the aforementioned GLM method, with the performance evaluated using the AUC, sensitivity and specificity, contrasting prediction scores to the true cluster labels.

[0250] For exploring the robustness in the context of patients, Monte Carlo cross-validation was performed in which 25 random data partitions were created, with 30% of samples left out in each iteration. Using the original XGBoost models, cluster prediction scores were assigned to patients and subsequently the AUC, sensitivity and specificity were calculated, contrasting prediction scores to the true cluster labels.

[0251] Results

[0252] Exploring other methods of combining features. The robustness of the clinical and omic prognostic models was explored through evaluating the performance when using other methods of combining the features. The GLM method involved calculating per-feature coefficients and multiplying these by the feature raw values themselves. For the A1 vs. A2 clinical model, the GLM method achieved an AUC of 88% (95% confidence interval: 84%-91%), a sensitivity of 73%, and specificity of 88%. The coefficients calculated by GLM are as follows - Creatinine: 0.00339980098775507; eGFR: -0.0128251705484681 ; age at study start: -0.00592423037404431 : diastolic blood pressure: 0.00420612641329168; systolic blood pressure: 0.00146277806277998; Albumin: -0.0110250871942654; Phosphate: 0.221058567103687; calcium channel blockers: 0.158843575263741. The simple risk score method, which involved adding the total values of features expected to increase in the A2 group compared to the A1 group (creatinine, diastolic blood pressure, systolic blood pressure, phosphate, calcium channel blockers), and subtracting the total values of features expected to decrease (eGFR, age at start of study, albumin), achieved an AUC of 84% (95% confidence interval: 80%-88%), a sensitivity of 77%, and specificity of 80%. Despite the GLM and simple risk score methods leading to 3% and 7% decreases in AUC in comparison to the original XGBoost model, the AUC for both of these models was over 80%, indicating that the combination of features retains its discriminatory ability without using XGBoost.

[0253] For the B2 vs. B3 clinical model, the GLM method had an AUC of 65% (95% confidence interval: 57%-74%), a sensitivity of 75%, and specificity of 54%. The coefficients calculated by GLM are as follows -total cholesterol: 0.0745601905295237; Creatinine: 0.00179723648543065; age at study start: -0.00517550698187641 ; Albumin: -0.0183904193925215; CRP: -0.00990733138734592; Calcium: -0.348894684475709; calcium channel blockers: 0.211655648651676. The simple risk score method (features expected to increase: total cholesterol, creatinine, calcium channel blockers; features expected to decrease: age at start of study, albumin, CRP, calcium) had an AUC of 60% (95% confidence interval: 51%-68%), a sensitivity of 36%, and specificity of 86%. The AUC for the original XGBoost model was 83%, and the GLM and simple risk score methods led to 18% and 23% decreases in AUC, respectively. The original XGBoost clinical prognostic model for B2 vs. B3 showed the lowest (albeit still high) performance out of all XGBoost models. This indicates that although all clinical models do provide significant classifications, proteomics-based classification is particularly valuable for the B2 vs B3 classification.

[0254] For the proteomic A1 vs. A2 model, the GLM method returned an AUC of 95% (95% confidence interval: 92%-99%), a sensitivity of 91%, and specificity of 89%. The coefficients calculated by GLM are as follows - SDC1 : 0.203254049143174; FAS: 0.145137068715139; TFPI: 0.178290023367982; CST3: -0.0247089452302634; PTGDS: 0.0309645819536599; CD160: 0.156257792712778; PILRA: 0.102949836057438; F11R: 0.198327997840357. The simple risk score method (features expected to increase: PTGDS, F11 R, PILRA, FAS, CD160, SDC1 , TFPI; features expected to decrease: CST3 -although the CST3 level is higher in the A2 group than in the A1 group, and so the feature is individually expected to increase, in the context of a multivariate model including the above features, the feature corresponding to CST3 level was associated with a negative coefficient and can therefore be seen as a feature expected to decrease for the purpose of the simple score method) achieved an AUC of 95% (95% confidence interval: 91%-99%), a sensitivity of 91%, and specificity of 82%. The GLM and simple risk score methods led to drops in AUC of just 4% each, demonstrating that the discriminatory qualities of these features were robust to changes in methodology.

[0255] For the proteomic B2 vs. B3 model, the GLM method returned an AUC of 86% (95% confidence interval: 75%-96%), a sensitivity of 74%, and a specificity of 85%. The coefficients calculated by GLM are as follows: PM20D1 : -0.0552436728608182; SELPLG: -0.849822150165859; TNFRSF11A: 0.647161621968939; LAIR20: 0.0829563737100124; FGFR2: 0.164732921548166. The simple risk score (features expected to increase: LAIR2, TNFRSF11A, FGFR2; features expected to decrease: SELPLG, PM20D1) had an AUC of 82% (95% confidence interval: 69%-96%), a sensitivity of 91%, and a specificity of 70%. Using the GLM method and the simple risk score method led to a 13% and 17% drop in AUC, respectively, in comparison to the XGBoost method. Despite this drop, both AUCs were above 82%, indicating that the models still retained discriminatory performance.

[0256] Testing sub-combinations of features. To evaluate how robust the combinations of features are, the inventors tested the discriminatory performance of all sub-combinations of features in the four models. Features were combined using the GLM method, with performance evaluated using AUC, sensitivity and specificity.

[0257] For the clinical A1 vs. A2 model, the AUCs of sub-combinations of clinical features (i.e. any combination of between 2 and 7 of the 8 features, covering all possible combinations) ranged from 58% to 87% (i.e. all better than random guess). Full results are shown in Table 7 below. All sub-combinations that included either eGFR or creatinine had AUCs over 80%, whilst sub-combinations without either of these variables never achieved AUCs over 69%, indicating that eGFR and / or creatinine are very important to the model performance. The sub-combination with the highest AUC was a 7-feature model composed of the following clinical variables: creatinine, eGFR, age at study start, diastolic blood pressure, albumin, phosphate, and calcium channel blockers. This model had an AUC of 87%. The trade-off between model size and performance was explored through scaling the AUC based on the number of features in the model. The combination with the highest scaled AUC was creatinine and age at study start, which had an AUC of 84%, a sensitivity of 75%, and a specificity of 82%.

[0258] The range of AUCs for the sub-combinations of features included in the clinical B2 vs. B3 model was 54% to 69%. The disparity between the sub-combinations including eGFR or creatinine was smaller compared to what was observed between the A1 and A2 models. The minimum AUC for models that included creatinine, and those without them, was around 54%. The model with the highest AUC scaled for model size was composed of age at study start and calcium channel blockers with an AUC of 62%, sensitivity of 63% and specificity of 61%. Despite some sub-combinations showing reduced performance, all sub-combinations for the A1 vs. A2 and B2 vs. B3 clinical models had AUCs over 50%, indicating that the discriminatory power of these features is better than chance.

[0259] For the proteomic A1 vs. A2 model, all sub-combinations of proteins (i.e. up to 7 proteins in all subcombinations) led to AUC values over 86%, with AUCs ranging from 86% to 96%. The subcombination with the highest AUC was SDC1 , TFPI, PTGDS, CD160, PILRA, F11R. The trade-off between model size and performance was explored through scaling the AUC based on the number of features in the model. The combination with the highest scaled AUC was CD160 and F11 R, which had an AUC of 93%, a sensitivity of 86% and specificity of 84%. This analysis demonstrates that the A1 vs. A2 model is stable to changes in the composition of the features in the model.

[0260] The range of AUCs for the sub-combinations of proteins in the B2 vs. B3 was 69% to 89%, with the sub-combination with the highest AUC composed of PM20D1 , SELPLG, LAIR2, and FGFR2. Once AUCs were scaled based on the number of features in the model, the combination of PM20D1 and LAIR2 performed best, with an AUC of 83%, sensitivity of 83% and specificity of 80%. The subcombinations of the proteomic B2 vs. B3 model performed slightly worse than the sub-combinations of the proteomic A1 vs. A2 model, suggesting that all 5 features in the B2 vs. B3 model are important for model performance.

[0261]

[0262]

[0263]

[0264]

[0265]

[0266]

[0267]

[0268]

[0269]

[0270] _

[0271] Table 7. Testing sub-combinations of features. All values provided are AUC from GLM models with the indicated combinations of features. BP=blood pressure. CCB=calcium channel blockers.

[0272] Testing subsets of patients. The robustness of the models was further evaluated through testing model performance in various subsets of the patients. Across 25 iterations in which a random subset of 30% of the patients were excluded, the range of AUCs for the clinical A1 vs. A2 prognostic model was 88% to 93% (vs. 91% for the full set of patients), the range of sensitivities was 81% to 92% (vs. 84% for the full set of patients), and the range of specificities was 77% to 90% (vs. 87% for the full set of patients). The range of AUCs for the clinical B2 vs. B3 prognostic model was 81% to 87% (vs. 83% for the full set of patients), the range of sensitivities was 64% to 76% (vs. 67% for the full set of patients), and the range of specificities was 78% to 91% (vs. 88% for the full set of patients).

[0273] Across 25 iterations in which a random subset of 30% of the patients were excluded, the range of AUCs for the 8-protein A1 vs. A2 prognostic model was 98% to 100% (vs. 99% for the full set of patients), the range of sensitivities was 90% to 100% (vs. 93% for the full set of patients), and the range of specificities was 88% to 100% (vs. 98% for the full set of patients). For the 5-protein B2 vs. B3 prognostic model, the range of AUCs was 99% to 100% (vs. 99% for the full set of patients), the range of sensitivities was 87% to 100% (vs. 91% for the full set of patients), and the range of specificities was 92% to 100% (vs.

[0274] 100% for the full set of patients).

[0275] The AUC, sensitivity and specificity values from the patient subsets were similar to the equivalent values for the full sets of patients, suggesting that the performance of the models is not influenced by a small portion of the patients evaluated. Example 4: Combining clinical and proteomic models to classify patients

[0276] In the work described in this example, the inventors identified an expanded list of protein biomarker discriminatory features, providing redundancy in the selection of biomarkers. The clinical and proteomic models were applied sequentially to stratify patients categorised as B3.

[0277] Methods

[0278] Expanding the list of potential protein biomarker candidates. It is not uncommon for protein biomarker candidates to fail to replicate across platforms. This is not only caused by variations in patient populations but also due to technical considerations such as levels of abundance and sensitivity of immunoassays. In order to introduce some redundancy into the selection of protein biomarker candidates, an expanded list of protein targets was assembled. The most correlated protein to every protein in the A1 vs. A2 and the B2 vs. B3 models were selected. Lasso models for both A1 vs. A2 and B2 vs. B3 were developed (alpha = 1) with the proteins with non-zero coefficients included in the expanded list of protein biomarker candidates.

[0279] Classifying B3 patients using clinical and proteomic models sequentially. The patients identified as B3 are most likely to be the patient population of interest due to their faster disease progression than patients in clusters B1 or B2. Employing two protein based diagnostic tests to identify a small proportion of the patient population would lead to significant costs. A more practically viable approach could be stratifying patients into A1 and A2 initially using the clinical model, and then using the protein B2 vs. B3 model to identify from the patients predicted as A2, the patients that are categorised as B3. The inventors evaluated this approach by first using the clinical A1 vs. A2 model to identify patients predicted as A2. These were then further split into B2 and B3 using the proteomic B2 vs. B3 model, but only for patients with Olink data available. The performance of this approach was evaluated through calculating the AUC for identifying patients in each of the B cluster groups (e.g., B1 vs. B2+B3, B2 vs. B1+B3, B3 vs. B1+B3).

[0280] Results

[0281] Expanding lists of discriminatory features. Multiple steps were taken to identify an expanded list of potential protein biomarker discriminatory features, in order to add redundancy in case any protein biomarker candidates did not validate in other platforms. This is especially relevant for proteins, as proteins measured by Olink may not all validate on alternative platforms. In contrast, the clinical models involve routinely measured clinical variables, so it is unlikely that any of these will be excluded from future models due to challenges in measuring them.

[0282] The most highly correlated proteins with each protein in the models were identified in the SKS Olink data using the proteins from the respective models as inputs to then identify their most correlated proteins (Table 6), identifying 8 proteins for the proteomic A1 vs. A2 model (PILRB, PGF, IGFBP4, JAM2, CD40, EPHA2, IGFBP6, PROS1) and 5 proteins for the proteomic B2 vs. B3 model (TOP1 , TNFRSF1A, NBL1 , TREH, HAO1). Furthermore, an alternative method (Lasso; Tibshirani 1996) was used to derive the model itself. For A1 vs. A2, a Lasso model composed of 22 proteins was identified, and for B2 vs. B3, a Lasso model composed of 33 proteins was identified (Table 6, Lasso models). A total of 5 proteins identified from Lasso models were already included in the XGBoost models (SDC1 , CD160, LAIR2, SELPLG, TNFRSF11A; i.e. multiple classification models selected these proteins). Furthermore, two proteins were included in both the A1 vs. A2 Lasso model and the B2 vs. B3 Lasso model (ANG, CFP). In total, 61 further proteins were identified as an expanded selection of discriminatory features for A1 vs. A2 and B2 vs. B3.

[0283] Table 6. The expanded list of proteins. Proteins were selected if they were highly correlated to features in the original models or if they were included in the Lasso models (A1 vs. A1, B2 vs. B3). ** denotes features included in the original XGBoost models.

[0284]

[0285]

[0286]

[0287] Combining clinical and proteomic models. Patients identified as B3 are most likely to be of interest due to faster disease progression than patients in clusters B1 or B2. A total of 83 patients out of 334 were originally classified as B3, representing only 25% of the total patient population. Employing two protein-based diagnostic tests to identify only 25% of the patient population would lead to significant costs. A practically viable approach would be stratifying patients into A1 and A2 initially using the clinical model, then using the 5-protein B2 vs. B3 model to identify patients categorised as B3 from the patients predicted as A2. Using the clinical model first is advantageous, as these variables are collected routinely regardless of any diagnostic test, reducing the overheads and logistical challenges.

[0288] The A1 vs. A2 clinical prognostic model was applied to the training data, and then the B2 vs. B3 proteomic prognostic model was applied to patients predicted as A2. A total of 138 of the 159 patients in B1 were correctly identified (AUC of 77%, 95% confidence interval: 70%-83%). A total of 16 of the 32 patients in B2 with Olink data were correctly identified (AUC of 73%, 95% confidence interval: 64%-82%). A total of 20 of the 34 patients in B3 with Olink data were correctly identified (AUC of 79%, 95% confidence interval: 71%-88%).

[0289] Further validations. The proteins identified in the prognostic models will now be quantified in immunoassays, using either ELISA or Luminex. This will indicate whether the changes observed between patients are replicated using a different platform as although the same associations were observed with the UK BioBank patients, this was also carried out using the same Olink panel.

[0290] Example 5: Evaluation of protein biomarker candidates in an additional patient cohort

[0291] The inventors validated the proteomics-based A1 vs A2 model developed in Example 1 using a separate group of patients from the SKS cohort (referred to in the examples below as “batch 2”). They then identified proteins highly correlated with the proteins included in the 8-protein A1 vs A2 model, in batch 2 data. Finally, the inventors further validated the Lasso model developed in Example 4 in the batch 2 cohort.

[0292] Methods

[0293] Patient cohort and endotype identification. The inventors evaluated the protein biomarker candidates identified in Example 1 using an additional 71 patients from the SKS cohort (described in Chinnadurai et al. 2023 - see Example 1). These include patients which were part of the 334 patients used to identify endotypes in Example 1 , but for whom plasma proteome data was not available at the time of developing the proteomic models described in preceding examples, and additional patients. Plasma proteomes were profiled for this additional set of 71 patients using the Olink® Explore 3072 / 384 panel (Olink, Uppsala, Sweden) as explained in Example 1 , which measures over 2900 known proteins. The same normalisation pipeline was followed as described in Example 1.

[0294] Of the n=71 patients included in batch 2, n=50 patients had been clustered using the patient endotyping method described in Example 1 (applying dynamic time warping (DTW) to longitudinal creatinine and haematocrit measurements), with n=29 clustered in the A1 cluster, and n=21 in the A2 cluster. Samples from patients in batch 2 were not included in batch 1 used for discovery of the protein patient stratifying models in the preceding examples. As such, the protein-based patient stratifying models (PSMs) described above can be validated using these n=50 patients and the performance of the models can be assessed.

[0295] Results

[0296] Validation of the 8-protein A1 vsA2 model. The inventors sought to validate the XGboost 8-protein A1 vs. A2 model identified in Example 1 in batch 2. The model was applied to patient samples in batch 2 to validate the model (Fig. 12A). As explained above, this model is composed of the following proteins: PTGDS, F11R, CST3, SDC1 , PILRA, FAS, CD160, TFPI. When applied to the new samples, this model returned an AUC of 80% (95% Cl: 67%-92%; Fig. 12B), sensitivity of 71% and specificity of 79%. When the threshold of classification derived from batch 1 (see Examples 1 and 2) was applied to batch 2 to classify patients into either A1 or A2, 22 / 29 of A1 patients were correctly classified and 15 / 21 of A2 patients were correctly classified, corresponding to a sensitivity of 71%, a specificity of 79%, and an AUC of 74% (95% Cl: 61%-86%). This confirms that the 8-protein model described in the preceding examples generalises to unseen patients, and indicates that alternatives based on correlated proteins and subsets of proteins in this model are likely to validate as well. Evaluation of proteins highly correlated with potential biomarker candidates. In the work described in Example 4, the proteins most highly correlated with each of the proteins included in the 8-protein model for stratifying A1 vs A2 patients were identified (see Example 4 Results, Expanding lists of discriminatory features; Table 6). Using the batch 2 data, the inventors evaluated whether these strong correlations persisted between each protein pair (Table 8). 5 of the 8 protein pairs (specifically PILRA-PIRLB, PTGDS-PGF, CST3-IGFBP4, CD160-JAM2, and F11R-CD40) retained correlations over 70%. Three pairs (SDC1-EPHA2, FAS-IGFBP6 and TFPI-PROS1) showed correlations below 70%. For SDC1 , FAS, and TFPI, the inventors investigated whether alternative highly correlated proteins could be identified from the Olink data from both batches pooled together. Data was filtered to only include patients classified as A1 or A2, and the top protein with the highest correlation was selected for each of these 3 proteins. ESAM (UniProt: Q96AP7) had the highest correlation (67%) with FAS; and APCS (UniProt: P02743) had the highest correlation with TFPI (63%). Whilst these correlations were not as high as the correlations for the other protein pairs, they were higher than the proteins identified in Example 4. For SDC1 , EPHA2 remained the protein with the highest correlation, showing a slightly higher correlation after both datasets were merged compared to batch 2 alone (67% vs 65%), confirming that EPHA2 remains the best alternative to SDC1.

[0297] Table 8. Correlations in the batch 2 data between the proteins identified in Example 4 as top-correlated in batch 1 with a respective protein from the 8-protein A1 vs A2 patient stratification model.

[0298]

[0299] Evaluation of proteins included in the expanded list of potential biomarkers. The Lasso model trained in Example 4 on batch 1 data (see Example 4 Methods and Results), which was composed of 22 proteins, was tested in batch 2. An AUC of 74% (95% Cl: 60%-89%) was returned alongside a specificity of 66% and sensitivity of 81%, suggesting that this combination of proteins retained its discriminatory capacity in a new dataset. Example 6: Further discovery of protein biomarker candidates.

[0300] In the work described in the present example, the inventors sought to expand the protein biomarker models developed in the previous examples by using batch 2 Olink data (alone or in combination with batch 1 data) to discover further protein biomarkers.

[0301] Results

[0302] Rediscovery of protein biomarker models - batch 2. To expand on the protein biomarker models (and biomarker candidates) discovered in the work described in the previous examples, batch 2 Olink data was used to discover further protein biomarker models. Following the methodology utilised in batch 1 (See Example 1 Methods, Discovery of prognostic models), Limma was first used to identify significantly differentially abundant proteins, including age, sex and eGFR at study start in the models. Proteins with unadjusted p-values <0.05 and those with evidence of secretion into blood were used as input data into the XGBoost model. This led to the identification of a 7-protein patient stratifying model, composed of the following proteins: PTGDS, CPA1 , CDH3, SUSD1 , GFOD2, ADH4, KLK11 (Fig. 15). Levels of PTGDS, CDH3, GFOD2 and KLK11 individually are expected to increase in A2 vs A1 , with levels of CPA1 , SUSD1 and ADH4 individually expected to decrease in A2 vs A1. When all proteins are combined in a multivariate generalised logistic regression model, PTGDS, CDH3, GFOD2 and KLK11 have positive coefficients, and SUSD1 , CPA1 , and ADH4 have negative coefficients (i.e. coefficients signs in the multivariate model were in line with individual behaviour of variables). Of the 7 proteins, one protein (PTGDS) was also included in the A1 vs A2 model originally identified from batch 1 (see Example 1). When applied to the batch 2 data (the data that the model was trained on), this new 7-protein model had an AUC of 99% (95% Cl: 97%-100%), specificity of 93% and sensitivity of 95%. Across 25 iterations in which a random subset of 30% of the patients in batch 2 were excluded, the range of AUCs for the 7-protein A1 vs. A2 prognostic model was 96% - 100%, the range of sensitivities was 83% - 100%, and the range of specificities was 86% - 100%, indicating that the high performance of the 7-protein model is robust and not driven by outlier samples. When applied to batch 1 data (testing data), the model had an AUC of 86% (95% Cl: 78%-94%), sensitivity of 86% and specificity of 80%. The threshold of classification in batch 2 that optimised sensitivity and specificity was applied to the batch 1 data, resulting in 35 / 45 of A1 patients correctly classified, and 37 / 43 of A2 patients correctly classified, corresponding to a sensitivity of 86%, a specificity of 77%, and an AUC of 82% (95% Cl: 74%-90%).

[0303] The performance of all sub-combinations of proteins was evaluated. AUCs ranged from 69% to 98% indicating that all subcombinations of at least 2 of these proteins are able to classify patients between the A1 and A2 groups. Results for this are shown in Table 9 below. The sub-combination with the highest AUC was a 6-protein model composed of PTGDS, SUSD1 , CPA1 , GFOD2, KLK11 and ADH4 which had an AUC of 98%, sensitivity of 86% and specificity of 100%. A “scaled AUC” was calculated as AUC / n features (n being the number of features in the model). The protein combination that had the highest scaled AUC was PTGDS and GFOD2 which had an (unsealed) AUC of 90%, sensitivity of 90% and specificity of 76%, showing that these are the two most effective proteins when combined. All combinations including PTGDS had AUCs of at least 87%. All combinations including GFOD2 had AUCs of at least 77%. All combinations involving CPA1 had AUCs of at least 77%. All combinations involving KLK11 had AUCs of at least 78%. There was only one combination with AUC under 75% (SUSD1 + ADH4, 69%). This indicates that any combination of 2 or more of PTGDS, CPA1 , CDH3, SUSD1 , GFOD2, ADH4, KLK11 can be used, with any combination involving at least 3 proteins and / or involving at least PTGDS, GFOD2, CPA1 , or KLK11 being particularly useful. Amongst these, any combinations involving PTGDS are particularly performant.

[0304] The most highly correlated protein to each of the proteins in the 7-protein model was identified from A1 and A2 patients in both Olink batches. Results of this are shown in Table 10 below. TNFRSF14 was the most highly correlated protein for both PTGDS and CDH3. The highest correlated protein for GFOD2 was WARS1 with a correlation of only 0.3469004 indicating that the Olink panel did not contain a protein that shows changes similar to those seen in this protein. This indicates that any of these proteins can be replaced by a correlated protein in Table 10 apart from GFOD2.

[0305] Table 9. Testing sub-combinations of features in model from batch 2 data. All values provided are AUC from GLM models with the indicated combinations of features.

[0306] >

[0307]

[0308]

[0309] Table 10. Correlations between the 7 proteins included in the 7-protein model derived from batch 2 data, and their most highly correlated protein. Correlations are shown using Pearson’s R.

[0310]

[0311] Expansion of biomarker candidates - batches 1 and 2 combined. The inventors combined batch 1 and batch 2 to form one dataset in order to identify further potential biomarker candidates. First, differential expression analysis was applied to the merged dataset, including age, sex, batch and baseline eGFR as covariates in the model (as explained in Example 1). Including cofounding variables (age, sex, plate) in the models when identifying a list of potentially discriminatory proteins reduces the possibility that these discriminatory proteins are driven by the confounding covariates. A relaxed threshold of unadjusted p-value <0.1 was used to allow more potential proteins into the pool that feature selection was run on, since the differential expression analysis model used accounted for covariates. This resulted in 261 proteins.

[0312] Repeated 5-fold cross validation (10 repeats) was performed, corresponding to 28 random samples (patients) being left out for each fold, and leading to 50 potential models derived from different subsets of batch 1 and batch 2 combined (Table 11). The inventors explored the number of times each unique combination of proteins was selected across the 50 iterations in an XGBoost model (10 repeats, 5 folds each). No single combination was selected more than once across the 50 iterations. This is not surprising as exact match between the complete set of proteins selected by XGBoost between cross-validation folds is a high bar to meet. However, there was significant recurrence of many proteins and sets of proteins. To quantify this, sub-combinations were evaluated, considering any sub-combination of 2 or more proteins (Table 12). PTGDS, CD160, ATRAID, DBI, TNFRSF14, DSC2 and CD274 were the most commonly selected proteins in combinations of 2 or more proteins. ATRAID and PTGDS was the most commonly selected sub-combination, being selected in 32 / 50 iterations. Amongst the commonly selected proteins were PTGDS and CD160, both of which were included in the A1 vs A2 biomarker model derived from batch 1 data (see Example 1). ATRAID, DBI, TNFRSF14, DSC2 and CD274 were newly identified as potential biomarker candidates. Any of these could therefore be added to a panel of biomarker proteins for classification of subjects between the A1 and A2 groups. PTGDS was included in 47 / 50 iterations, providing further indication that this is a reliable biomarker candidate. The following further biomarkers were selected in over ten iterations: ATRAID (35 / 50); CD160 (15 / 50); TNFRSF14 (13 / 50); DBI (13 / 50); DSC2 (10 / 50).

[0313] The most commonly selected proteins (PTGDS, CD160, ATRAID, DBI, TNFRSF14, DSC2 and CD274) were combined into a 7-protein model, (specifically, an XGBoost model) which when applied to all samples across batch 1 and 2, achieved an AUC of 96%, a sensitivity of 92% and a specificity of 91%. In batch 1 samples alone, this model achieved an AUC of 97%, a sensitivity of 98% and a specificity of 89%. In batch 2 samples alone, this model achieved an AUC of 94%, a sensitivity of 86% and a specificity of 93%. When taken alone, all proteins in this model increased in A2 vs A1 (Figure 16), but when combined in a multivariate logistic regression model, TNFRSF14 had a negative coefficient. Across 25 iterations in which a random subset of 30% of the patients were excluded, the range of AUCs for the 7 protein A1 vs. A2 prognostic model was 95% - 98%, the range of sensitivities was 91% - 100%, and the range of specificities was 87% - 94%, indicating that the high performance of the 7-protein model is robust and not driven by outlier samples.

[0314] Table 11. Features in XGBoost models obtained in cross-validation folds, batches 1 and 2 combined.

[0315]

[0316]

[0317] Table 12. Top 10 most commonly selected sub-combinations of proteins selected across 50 cross-validation iterations (Table 11).

[0318]

[0319] The performance of all sub-combinations of proteins was evaluated (see Table 13), resulting in AUCs ranging from 85% to 91%, sensitivities ranging from 72% to 94%, and specificities ranging from 72% to 91%. This indicates that any combination of 2 or more of these proteins can be used to classify patients. The sub-combination with the highest AUC was ATRAID, PTGDS, CD160, DBI, TNFRSF14 and DSC2 which had an AUC of 91%, a sensitivity of 82% and a specificity of 86%. When a scaled AUC was calculated (AUC I number proteins in model), PTGDS and DSC2 was the sub-combination with the highest scaled AUC, with an (unsealed) AUC of 90%, a sensitivity of 86% and a specificity of 84%.

[0320] Table 13. Testing sub-combinations of features in model from cross-validation on combined batch 1 + batch 2 data. All values provided are AUC from GLM models with the indicated combinations of features.

[0321]

[0322]

[0323] _

[0324] The most highly correlated protein to each of the proteins in the 7-protein model was identified from A1 and A2 patients in both Olink batches (Table 14). TNFRSF1B was the most highly correlated protein for both DSC2 and CD274. All of the proteins in Table 14 have high correlations to their original protein, indicating that they could be used to replace the original protein.

[0325] Table 14. Correlations between the 7 proteins included in the 7-protein model derived from repeated cross-validation, and their most highly correlated protein. Correlations are shown using Pearson’s R.

[0326]

[0327] Finally, the inventors applied XGBoost to the merged dataset (batch 1 + batch 2), without any cross-validation or data splitting, to determine the optimal combination when evaluated using all samples. First, proteins were filtered to only include those with an unadjusted p-value <0.1 as explained in Example 1 (resulting in 261 proteins). Using the same XGBoost parameters as used in Example 1 , a model composed of 8 proteins was identified. The 8 proteins were: PTGDS, TNFRSF14, DBI, THBD, DSC2, TNFRSF11A, GPR37 and ENTPD5. This model had an AUC of 96%, sensitivity of 92% and specificity of 85%. TNFRSF14, DBI, ENTPD5 and DSC2 were also identified in the above cross-validation analyses, and PTGDS and TNFRSF11 A were in the initial A1 vs A2 protein models identified in Example 1. THBD, and GPR37 were newly discovered protein biomarker candidates identified by this analysis. Taken individually, levels of DBI, DSC2, GPR37, PTGDS, THBD, TNFRSF11A and TNFRSF14 increased in A2 vs A1 , with levels of ENTPD5 decreasing in A2 vs A1 (Figure 17). When combined in a multivariate logistic regression model, PTGDS, DBI, THBD, DSC2, TNFRSF11A and GPR37 all had positive coefficients, whilst ENTPD5 and TNFRSF14 had negative coefficients. Across 25 iterations in which a random subset of 30% of the patients were excluded, the range of AUCs for the 8 protein A1 vs. A2 prognostic model was 94% - 97%, the range of sensitivities was 76% - 98%, and the range of specificities was 83% - 96%, indicating that the high performance of the 8-protein model is robust and not driven by outlier samples.

[0328] The performance of all sub-combinations of the 8 proteins was tested, with AUCs ranging from 77% to 92%, sensitivities ranging from 67% to 97%, and specificities ranging from 61% to 97% (Table 15). This indicates that any two or more of these proteins could be used to classify patients between A1 and A2. All combinations of 2 or more proteins had AUCs above 83% apart from GPR37 + ENTPD5 (which still achieved an AUC of 77%). Therefore, any combinations of two or more of these proteins that includes at least PTGDS, TNFRSF14, DBI, THBD, DSC2, or TNFRSF11A achieved at least 83% AUC. Combinations of two or more proteins including at last PTGDS all achieved AUCs of 89% or more. Combinations of two or more proteins including at least TNFRSF14 all achieved AUCs of 87% or more. Combinations of two or more proteins including at least THBD all achieved AUCs of 87% or more. Combinations of two or more proteins including at least TNFRSF11A all achieved AUCs of 85% or more. The sub-combination with the highest AUC was PTGDS, TNFRSF14, THBD, DSC2, TNFRSF11 A, GPR37 and ENTPD5 with an AUC of 92%, sensitivity of 86%, and specificity of 86%, and the sub-combination with the lowest AUC was GPR37 and ENTPD5 with an AUC of 77%, sensitivity of 80%, and specificity of 61%. AUCs were scaled to account for size of model (AUC I n proteins in model), with PTGDS and ENTPD5 having the highest scaled AUC, with an (unsealed) AUC of 90%, a sensitivity of 75%, and specificity of 93%.

[0329] Table 15. Testing sub-combinations of features in model from combined batch 1 + batch 2 data. All values provided are AUC from GLM models with the indicated combinations of features.

[0330]

[0331]

[0332]

[0333] The most highly correlated protein to each of the proteins in the 8-protein model was identified from A1 and A2 patients in both Olink batches (Table 16). TNFRSF1A was the most highly correlated protein for both DSC2 and TNFRSF11 A. The data indicate that all proteins apart from ENTPD5, and preferably also apart from GPR37 could be replaced by the corresponding correlated protein in Table 16. Table 16. Correlations between the 8 proteins included in the 8-protein model derived from applying XGBoost to both batches of Clink data without any cross-validation. Correlations are shown using Pearson’s R.

[0334]

[0335] Example 7: Evaluation of previously-identified clinical biomarker candidates in SHARE The inventors applied the same endotyping protocol applied to SKS (see Example 1 Methods, Patient endotype identification) to individuals with diabetic kidney disease (DKD) from the Scottish Health Research register (SHARE). SHARE is an ongoing voluntary register of individuals who are interested in taking part in health research (McKinstry et al., 2017). Since SHARE is a population study rather than a patient study, the inventors first identified a suitable sub-cohort of participants with DKD that could be used to validate their endotyping procedure.

[0336] The inventors identified a cohort of 824 participants from SHARE that met the following criteria:

[0337] i. Diagnosis of type 2 diabetes (T2D) and chronic kidney disease (CKD), with the first recorded diagnosis of T2D occurring before the first recorded CKD diagnosis;

[0338] ii. No diagnosis of type 1 diabetes (T 1 D) and no diagnosis of other kidney diseases (such as acute kidney injury (AKI), nephropathy, polycystic kidneys);

[0339] Hi. Annotated cause of death was neither AKI nor cancer;

[0340] iv. Highest recorded Haemoglobin A1C (HbA1C) was > 6.5% (48 mmol / mol) and highest recorded estimated glomerular filtration rate (eGFR) was > 25 mL / min / 1.73 m2during clinical records; v. Sustained drop in eGFR, described as eGFR < 60 mL / min / 1.73 m2for at least 3 consecutive months;

[0341] vi. Complete height and weight measurements.

[0342] For participants with cancer recorded, individuals were excluded if their first cancer admission date occurred < 1 year after their eGFR cut-off date. For patients whose first recorded cancer admission date was > 1 year after their eGFR cut-off date, an additional threshold was applied, excluding clinical data from the year before their first recorded cancer admission. If their last recorded cancer admission occurred > 1 year before their cut-off, their eGFR cut-off date remained unchanged and the patients were included. For participants whose last recorded cancer admission date occurred after their eGFR cut-off date, the cut-off date was adjusted to the first instance > 1 year after the last recorded cancer admission date when eGFR fell below 60 mL / min / 1.73 m2. The same approach was applied to participants whose last recorded cancer admission date was < 1 year before their eGFR cut-off date. Unlike SKS, the SHARE study harnesses electronic healthcare records, so there was no defined “study start date”. To curate a validation cohort (derived from SHARE) that was suitably matched to the discovery cohort (derived from SKS), participant clinical data was censored according to eGFR levels. The data start date (referred to as the eGFR cut-off date) was taken as the first date where eGFR was less than 60 mL / min / 1.73 m2, with a follow-up eGFR data point taken within 1 year.

[0343] The inventors applied their longitudinal clustering approach (as described in Example 1 - with k-means, k=2) to creatinine and haematocrit values available for the cohort of 824 participants, identifying two endotypes: A1 (n=375) and A2 (n=449). Using Cox regression analysis, the inventors tested whether there were significant associations between endotype label and time to disease outcome. Disease outcomes were defined as either: death due to metabolic syndrome; renal replacement therapy (dialysis and / or renal transplant); estimated glomerular filtration rate (eGFR) < 15 (mL / min / 1 ,73m2); or >40% reduction in eGFR from starting point. A p-value of <2x10-16was returned from the Cox models, with starting eGFR, age and sex included in the model. The data show that the patient endotypes discovered in SKS were successfully validated in SHARE (Fig. 14). As a result of this analysis, “true endotype labels” were identified for the 824 DKD participants in SHARE. Therefore, the inventors aimed to test the performance of their clinical patient stratifying model (PSM), originally identified in SKS using baseline clinical data in this cohort (see Example 1).

[0344] The 8-parameter clinical PSM identified in SKS for stratifying between A1 vs A2 used the following clinical variables: creatinine, eGFR, age at study start, diastolic blood pressure, systolic blood pressure, albumin, phosphate, and calcium channel blockers. In SKS, this PSM had an AUC of 91% (95% confidence interval: 88%-94%) when used to stratify between A1 vs A2 (see Example 1 Results, Clinical models). The relevant information for the 824 DKD participants from SHARE was assembled to predict endotype labels, in order to evaluate the performance of the 8-parameter clinical model in SHARE. As explained above (see Methods), the start date for endotyping (longitudinal creatinine and haematocrit data) was taken as the first recorded date when eGFR fell below 60 mL / min / 1 ,73m2. Consequently, for the clinical PSM, the nearest available data within one year of this cut-off date (based on eGFR) was used. As a result, not all patients had data for each of the 8 parameters.199 patients (24%; n=86 A1 and n=113 A2) of the cohort (n=824) had complete input data for the clinical PSM. Applying the 8-parameter clinical PSM to this subset of DKD participants from SHARE with complete data led to an AUC of 56% (95% Cl: 48%-64%), specificity of 95% and sensitivity of 21%. Using the threshold of classification obtained from SKS, 85 of the 86 participants in A1 in SHARE with complete data (99%) were correctly classified by the 8-parameter clinical model, whereas only 8 of the 113 (7%) of the A2 participants in SHARE with complete data were correctly classified. The results may be different when using a different threshold that more strongly priorities sensitivity over specificity. Indeed, the threshold identified in the SKS cohort was designed to optimise sensitivity and specificity combined. In this validation cohort this results in very good specificity, but low sensitivity. If sensitivity was more important than specificity, the threshold determination method can be chosen accordingly.

[0345] The inventors returned to the baseline clinical data from SKS to establish if there was an alternative PSM composed of baseline clinical traits. When the inventors identified the 8-parameter clinical model in Example 1 , they only used a subset of variables that were all present in CHUIMI, which is the cohort they were planning to validate the model in. In this work, the inventors went back to the SKS data and included all of the possible variables e.g. not filtering based on what was in CHUIMI (or SHARE), then re-ran XGBoost on this. A PSM composed of creatinine, age and UPCR (Urine protein-to-creatinine ratio - UPCR was higher in the A2 than in the A1 group, see Fig. 18A,B) was identified as a viable alternative PSM, with an AUC of 90% (95% Cl: 87%-94%), a specificity of 83%, and sensitivity of 83%. This model was applied to the DKD participants identified from SHARE. 128 of the SHARE DKD participants had complete UPCR, age and creatinine data (n=49 for the A1 cluster, and n=79 for the A2 cluster). This returned an AUC of 69% (95% Cl: 63%-76%), specificity of 96%, and sensitivity of 40%. All A1 patients had low A2 prediction scores, but only a small number of A2 patients had high A2 prediction scores. Using the threshold of classification obtained from SKS, 49 (100%) of the 49 participants in the A1 endotype in SHARE (with complete data) were correctly classified by this model, but only 8 (10%) of the 79 participants in the A2 endotype in SHARE (with complete data) were correctly classified.

[0346] While the sensitivity of these models is not as high as in the SKS cohort, these models are still very useful in that they have very high specificity, indicating that patients identified as A2 with these models can be confidently considered to be at high risk of faster progression.

[0347] Example 8: Multi-Omic analysis

[0348] To investigate the molecular mechanisms driving fast progressors identified in the SKS discovery cohort (Example 1), the inventors performed integrative multi-omic analysis of plasma proteomics (Olink), metabolomics (Biocrates) and DNA methylation profiling (EPIC) of samples from patients in the discovery cohort (n=87).

[0349] Methods

[0350] Omic datasets were generated using samples from patients in the SKS cohort (described in Example 1), including plasma proteomics (Olink, Olink Explore 3072), LC-MS / MS metabolomics (Biocrates, Biocrates MxP 500), and DNA methylation profiling (Illumina, Infinium MethylationEPIC v2.0 Kit). Preprocessing and normalisation was performed on each omic dataset individually and multi-omics integration was performed.

[0351] Plasma proteomics (Olink). See Example 1 . LC-MS / MS metabolomics (Biocrates). _Plasma samples underwent metabolic profiling using the Biocrates MxP 500 panel at Joanneum Research (Graz, Austria) with samples split across two plates. The Biocrates MxP 500 panel is a LC-MS / MS based metabolomics platform, quantifying up to 1 ,019 metabolites. As a QC check, creatinine levels measured using Biocrates were correlated (using Pearson correlation) to creatinine levels measured clinically at matched time points. The rates of values <LOD and > upper limit of quantification (ULOQ) were calculated for each metabolite, split by plate and split by ESRD (true / false). Missingness rates were also calculated for each metabolite, stratified by plate and ESRD status. Metabolites were excluded if > 20% of samples (within each plate and ESRD group) had values below the limit of detection (LOD), if > 80% had values below the upper limit of quantification (ULOQ), or if > 50% of values were missing (NA). <LOD, >ULOQ and NA values for metabolites that were not excluded were replaced with the median value for each metabolite. Data was merged across plates and plate effects excluded using ComBat from the sva package (Leeket al. 2012). Finally, data was Iog2 transformed.

[0352] Epigenomic sequencing (EPIC)._DN / samples underwent Genome-wide DNA methylation profiling at AIT (Vienna, Austria), performed using the Illumina Infinium MethylationEPIC BeadChip array. Raw IDAT files from two batches were processed using the SeSAMe R / Bioconductor package (Zhou et al., 2018). The EPIC hg38 manifest file was used to map probe addresses to genomic coordinates. Signal detection and masking of unreliable probes were performed using SeSAMe's default quality masking function. Background subtraction and dye-bias correction were applied using the noobO normalization method. Post-normalization quality metrics were calculated to assess probe- and sample-level performance, including bisulfite conversion efficiency and the number of detected probes. Probes were then filtered to retain only those passing SeSAMe’s masking criteria, and further restricted to autosomal loci. Beta values were calculated and logit-transformed to M-values fordownstream analysis.

[0353] Integrated multiomic analysis. DIABLO (Data Integration Analysis for Biomarker discovery using Latent components) was used for supervised data integration to uncover endotype-associated multi-omic patterns. All metabolite features were used as input due to the smaller number of features. For proteomics and DNA methylation, the top 2000 most variable features were used as input. DNA methylation probes were further aggregated to the promoter level by taking the mean signal across all probes mapped to the same promoter region, resulting in 1 ,266 promoter-level features. To identify the optimal number of features for each omic dataset, a total of 4,608 models were fitted across a grid search. The weight design matrix was set as 0.7 based on the average correlation between the latent components of each omics block. The number of components was selected by performing repeated 10-fold cross-validation. The optimal number of components was then chosen as the one minimizing the overall balanced error rate using the "maximum distance" classification method. The final DIABLO model extracts a set of latent components capturing the shared variation across omics that best distinguishes SKS endotypes, enabling the identification of correlated proteomic, metabolomic, and epigenetic features contributing to endotype stratification. Gene Ontology & Kidney Disease Gene Set Enrichment Analysis. Pathway analysis was run on the combined selected proteomic and gene level epigenetic features for each component using GOrilla (Eden et al., 2007, 2009). Enrichment with known kidney disease gene lists were also run on the combined selected proteomic and gene level epigenetic features in each component using MSigDB lists and a One Way Fisher’s Exact Test. Gene ontology enrichment was done using the GOrilla web tool (https: / / cbl-gorilla.cs.technion.ac.il / ). The background set was all proteins and genes included in the DIABLO run. The sets of interest were run for component 1 , 2 and 3 separately, combining the proteins and epigenetic genes that came up. For Kidney disease gene sets, first the inventors curated kidney related gene sets from MSigDB. They then ran for each component and kidney disease list a One-Way Fisher’s Exact Test. P-values were corrected for each component using the Benjamini & Hochberg (Benjamini & Hochberg, 1995) to give the FDR.

[0354] Results

[0355] Multiomic analyses. DIABLO (Data Integration Analysis for Biomarker discovery using Latent components; Singh et al., 2019) was used for supervised data integration to uncover endotype-associated multi-omic patterns. Following model optimization (see Methods), DIABLO was used to perform a supervised analysis discriminating SKS-fast from SKS-slow patients, with each omic achieving visible separation between SKS-fast and SKS-slow when stratified by each omic (Fig. 25) and when all features were combined (data not shown). DIABLO identified a total of 110 proteins, 55 promoter DNA methylation sites and 9 metabolites that are predictive of endotype. These features are split across 3 components (Table 18) which represent distinct multi-omic patterns, with component 1 (made up of 20 proteins, 5 promoter DNA methylation sites and 3 metabolites), capturing the most variance, followed by component 2 (made up of 50 proteins, 25 promoter DNA methylation sites and 3 metabolites), and finally component 3 (made up of 40 proteins, 25 promoter DNA methylation sites and 3 metabolites.

[0356] Table 18. List of full proteins, epigenetic genes and metabolites in multiomics analysis components.

[0357]

[0358]

[0359] Component 1 captures signatures reflective of kidney function and structural damage (Table 18, Fig.

[0360] 26), such as creatinine and symmetric dimethylarginine (SDMA) which are both established markers of renal function with SDMA also suggesting renal fibrosis. The multi-omic signature is enriched for tumor necrosis factor (TNF)-receptor signaling, cell death, and insulin-like growth factor (IGF)-II binding pathways. Key proteins selected in component 1 are central to inflammatory signaling and fibrosis, including various IGF binding proteins which modulate IGF signaling. WFDC2 and TXNDC5, both linked to renal fibrosis (Fig. 26), further support this component’s role in capturing immune-fibrotic processes relevant to disease progression. TXNDC5 may mediate its profibrotic effect via TGF-p signalling. When enriching component 1 with known kidney disease gene lists (see Methods), the inventors found it to be significantly enriched with a set of genes associated with DKD progression in Olink plasma (One-Way Fisher Exact test FDR: 2.88x10-9) (Lal Gurung et al., 2024). Component 2 reflects lipid metabolic and inflammatory dysregulation, indicated by the presence of apolipoproteins such as APOE, APOC1 , APOB, and APOM, all key proteins for lipid transport and known to be altered in CKD-associated dyslipidemia. LDLR, the low-density lipoprotein receptor, further reinforces the lipid regulatory axis of this component. Gene ontology enrichment analysis supports the involvement of lipid metabolism and lipoprotein-related pathways (data not shown). VMO1 , which has been previously associated with declining kidney function (Makhammajanov et al., 2024) is also present in component 2, alongside HAVCR1 , a synonym for kidney injury molecule 1 (KIM-1), a well known biomarker of kidney injury (Han et al., 2002). RARRES2 (chemerin), also present in component 2, links metabolic stress to inflammatory signaling. When enriching component 2 with known kidney disease gene lists (see Methods), the inventors found it to be significantly enriched with a set of genes associated with podocyte differentiation (One-Way Fisher Exact test FDR: 8.20x10-3and 2.17x10-3) (Yoshida et al., 2023).

[0361] Component 3 highlights molecular processes associated with dysglycemia and cytoskeletal disruption in DKD. It includes phosphatidylserines, which have been linked to insulin secretion and signalling alongside oxidative stress. Notably, proteins such as STX4, involved in insulin-stimulated glucose uptake, and ENO2, a glycolytic enzyme, suggest altered glucose metabolism. Pro-apoptotic and stressresponse mediators like CASP9, SRC, and MAP2K6 are also present, supporting the role of cellular injury and inflammation in this component. Several cytoskeletal regulators (IQGAP2, RHOC, VAV3, TACC3) point to structural remodeling processes in glomerular and tubular cell, with further evidence from the nominal enrichment for pathways related to actin cytoskeleton organization and cyclic nucleotide binding (data not shown) both of which are implicated in insulin signaling and cytoskeletal dynamics under hyperglycemic conditions.

[0362] Investigation of proteins in component 1. Three components were identified from DIABLO - components 1 , 2 and 3 - with 20, 40 and 50 proteins included in each component, respectively. Since component 1 explains the most variance with respect to the endotype labels, the inventors first focused on the 20 proteins identified in component 1 , calculating the performance of each individual protein in predicting endotype labels (Table 19). AUCs for proteins in component 1 ranged from 81% - 87%, with a median AUC of 85% (where AUC here is for each of the 20 proteins in component 1 individually, using the normalised protein value of the particular protein to predict the endotype label). PTGDS was the only protein identified in component 1 that was also in the original 8-protein stratification model for classifying endotypes.

[0363] XGBoost was applied to the 20 proteins included in component 1 to identify a sparse combination of proteins capable of stratifying between A1 vs A2. Out of the 20 proteins, a 4-protein signature was identified composed of PTGDS (UniProt: P41222), ATRAID (UniProt: Q6UW56), BTN2A1 (UniProt: Q7KYR7) and PIK3IP1 (UniProt: Q96FE7). This 4-protein signature had an AUC of 93% (95% confidence interval: 88%-97%), a specificity of 81% and a sensitivity of 95%. All 20 proteins included in component 1 were significant in differential expression analysis contrasting A1 to A2 (Table 19). All 20 proteins were higher in the A2 endotype compared to the A1 endotype (positive log fold change).

[0364] The performance of subsets of proteins at stratifying between endotypes was performed, with subsets of between 4-19 proteins explored. A total of 1 ,048,364 combinations were explored, all with AUCs between 84%-90%. As the number of proteins included increased, the median AUC also increased, as expected (Table 20). When evaluated on a per-protein basis, the median AUC for all models including each protein ranged from 89.89% to 89.34.% (Table 21), indicating remarkably consistent high performance. CRP achieved the highest median AUC (89.34%) across all models in which it was included. A scaled AUC was calculated, dividing the AUC by the number of features included in each model. The model with the highest scaled AUC was composed of NPDC1 , HSPG2 and PTGDS, with an AUC of 89% and a scaled AUC of 30%.

[0365] Table 19. Proteins selected by DIABLO in components 1 , 2 and 3. Benjamini-Hochberg (Benjamini & Hochberg, 1996) adjusted p-values and log fold change values calculated through differential expression analysis are shown, in addition to single protein AUC values. Positive log fold-change values indicate that the proteins are higher in A2 vs A1 .

[0366]

[0367]

[0368]

[0369]

[0370] >

[0371] Table 20. The median AUC of models constructed from subsets of the 20 proteins selected by DIABLO in Component 1. The number of proteins indicates the number of proteins included in each model with all subsets of the 20 proteins tested.

[0372]

[0373]

[0374] Table 21. The median AUC of models constructed from subsets of the 20 proteins selected by DIABLO in Component 1. Results are shown for each protein with the median AUC across all models that the

[0375]

[0376] Investigation of proteins in components 2 & 3. Across components 2 and 3, a further 90 proteins were identified by DIABLO. The AUC of each of these individual proteins was calculated (Table 19) with proteins in component 2 having AUCs ranging from 49% - 62% with a median AUC of 54%, and proteins in component 3 having AUCs ranging from 48% - 69% with a median AUC of 57% (Table 19). Only three proteins (TFPI, HAVCR1 and MXRA8) from component 3 were significant in differential expression analysis contrasting A1 to A2 (each with a positive log-fold change indicating higher values in A2 compared to A1 , Table 19) and no proteins from component 2 were significant.

[0377] XGBoost was applied to all 110 proteins included in components 1-3 to identify an additional sparse signature. Out of the 110 proteins identified, an 8-protein signature was identified composed of PTGDS (UniProt: P41222), ATRAID (UniProt: Q6UW56), BTN2A1 (UniProt: Q7KYR7), EFCAB14 (UniProt: 075071), AMN (UniProt: Q9BXJ7), CXCL5 (UniProt: P42830), RARRES2 (UniProt: Q99969) and INHBC (UniProt: P55103). This 8-protein model had an AUC of 96% (95% confidence interval: 93%-99%), a specificity of 89% and a sensitivity of 91 %.

[0378] Discussion

[0379] The three components found in the multi-omic analysis reveal the pathways and mechanisms that could underpin the fast and slow progression of disease identified herein. Each component captures slightly different information, with component 1 capturing fibrosis, inflammation and known DKD progression signal, component 2 capturing more of a lipid and podocyte signature, and component 3 capturing the disrupted cytoskeleton and dysglycemic control. When put together, the signal from the three components starts to disentangle the complexity behind DKD (Fig. 26), with TGF- and IGF-mediated inflammation and dysglycaemia contributing to renal fibrosis and kidney damage that ultimately leads to a decline in renal function, captured by the increasing levels of creatinine over time in fast progressors.

[0380] Associations between lipid dysregulation and changes in erythrocyte morphology have been reported, potentially providing a link between the multi-omic signals of altered erythrocyte life cycle and dyslipidemia. Component 3 contains three phosphatidylserine (PS) metabolites, potentially providing insight into the anemia observed in these patients (via decreasing haematocrit), as increased exposure of PS on red blood cells (RBC) promotes macrophage recognition and clearance of abnormal erythrocytes, thus decreasing RBC survival time. This process has been reported as occurring more frequently in CKD. Furthermore, FIS1 is included in component 2, and plays a key role in progression of erythropoiesis, with a previous study showing that overexpression of FIS1 resulted in both the inhibition of haemoglobin biosynthesis and the arrest of erythroid differentiation. The presence of these features provides a potential mechanism for the reduction in haematocrit observed overtime in the fast progressor endotype.

[0381] Further, the multi-omic analyses identified additional proteins that can be used to build robust prognostic proteomics models to identify patient’s endotypes.

[0382] Conclusion

[0383] The work in these examples demonstrates the stratification of a cohort of DKD patients into distinct groups: Clustering A (A1 , A2) and Clustering B (B1 , B2, B3), with each cluster exhibiting different disease trajectories, including variations in the emergence and timing of outcomes. Two types of prognostic models were developed, based on either clinical or proteomic features measured at baseline. In the batch 1 training sets, these models successfully stratified patients into either A1 or A2, and B2 or B3, achieving AUCs of 91% and 83% for the clinical models, and 99% for both proteomic models. These models were further tested on external, independent cohorts, where associations between the predicted clusters and disease progression mirrored trends observed in the original patient clusters. The proteomics-based A1 vs A2 model was further validated in a separate cohort of DKD patients, batch 2, and additional proteomics-based A1 vs A2 models were developed using batch 2 data. The inventors also demonstrated that the A1 and A2 endotypes could be recreated in two separate validation cohorts (SHARE and CHIUMI) and provided evidence that the endotypes would also be applicable to CKD.

[0384] References

[0385] Alicic, Radica Z., Michele T. Rooney, and Katherine R. Tuttle. 2017. “Diabetic Kidney Disease:

[0386] Challenges, Progress, and Possibilities.” Clinical Journal of the American Society of Nephrology : CJASN 12 (12): 2032-45.

[0387] Andersen, P. K., and R. D. Gill. 1982. “Cox’s Regression Model for Counting Processes: A Large Sample Study.” Annals of Statistics 10 (4): 1100-1120.

[0388] Assarsson, Erika, Martin Lundberg, Goran Holmquist, Johan Bjbrkesten, Stine Bucht Thorsen, Daniel Ekman, Anna Eriksson, et al. 2014. “Homogenous 96-Plex PEA Immunoassay Exhibiting High Sensitivity, Specificity, and Excellent Scalability.” PloS One 9 (4): e95192.

[0389] Benjamini & Hochberg, 1995. Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing. Journal of the Royal Statistical Society: Series B (Methodological). Volume57, Issuel , 1995, Pages 289-300.

[0390] Boronat, M. et al. (2015). Prevalence and determinants of the metabolic syndrome among subjects with advanced nondiabetes-related chronic kidney disease in Gran Canaria, Spain. Renal Failure, 38(2), 198-203.

[0391] Chen, Tianqi, and Carlos Guestrin. 2016. “XGBoost: A Scalable Tree Boosting System.” In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. New York, NY, USA: ACM. https: / / doi.org / 10.1145 / 2939672.2939785.

[0392] Chinnadurai, Rajkumar, Henry H. L. Wu, Jones Abuomar, Sharmilee Rengarajan, David I. New, Darren Green, and Philip A. Kalra. 2023. “Antihypertensive Prescribing Patterns in Non-Dialysis Dependent Chronic Kidney Disease: Findings from the Salford Kidney Study.” World Journal of Nephrology 12 (5): 168-81.

[0393] Ester, M., H. Kriegel, J. Sander, and Xiaowei Xu. 1996. “A Density-Based Algorithm for Discovering Clusters in Large Spatial Databases with Noise.” Knowledge Discovery and Data Mining, August, 226-31.

[0394] Girden, Ellen R. 1992. Anova. London, England: SAGE Publications.

[0395] Han WK, et al. Kidney Injury Molecule-1 (KIM-1): a novel biomarker for human renal proximal tubule injury. Kidney Int. 2002 Jul;62(1):237-44.

[0396] Iseki, K. et al. Proteinuria and the risk of developing end-stage renal disease, Kidney International, Volume 63, Issue 4, 2003, Pages 1468-1474.

[0397] Jin, Xin, and Jiawei Han. 2011. “K-Means Clustering.” In Encyclopedia of Machine Learning, 563-64.

[0398] Boston, MA: Springer US. Kidney Research UK. 2023. “Kidney Disease: A UK Public Health Emergency The Health Economics of Kidney Disease to 2033.” https: / / www.kidneyresearchuk.org / wp-content / uploads / 2023 / 06 / Economics-of-Kidney-Disease- full-report_accessible.pdf.

[0399] Resham Lal Gurung, et al. Proteomics profiling and association with cardiorenal complications in type 2 diabetes subtypes in Asian population, Diabetes Research and Clinical Practice, Volume 214, 2024

[0400] Jeffrey T. Leek, et al, The sva package for removing batch effects and other unwanted variation in high- throughput experiments, Bioinformatics, Volume 28, Issue 6, March 2012, Pages 882-883 Li, Philip Kam-Tao, Gordon Chun-Kau Chan, Jianghua Chen, Hung-Chun Chen, Yuk-Lun Cheng, Stanley L-S Fan, John Cijiang He, et al. 2021. “Tackling Dialysis Burden around the World: A Global Challenge.” Kidney Diseases (Basel, Switzerland) 7 (3): 167-75.

[0401] Makhammajanov, Z., Kabayeva, A., Auganova, D. et al. Candidate protein biomarkers in chronic kidney disease: a proteomics study. Sci Rep 14, 14014 (2024).

[0402] McKinstry B, Sullivan FM, Vasishta S, Armstrong R, Hanley J, Haughney J, Philip S, Smith BH, Wood A, Palmer CN. Cohort profile: the Scottish Research register SHARE. A register of people interested in research participation linked to NHS data sets. BMJ Open. 2017 Feb 1 ;7(2):e013351. Nelson, R.G., et al. (2019). Development of Risk Prediction Equations for Incident Chronic Kidney Disease. JAMA, 322(21), p.2104.

[0403] NHS. “Haematology Reference Ranges.” Gloucestershire Hospitals NHS Foundation Trust, 2024, www.gloshospitals.nhs.uk / our-services / services-we-offer / pathology / haematology / haematology- reference-ranges / .

[0404] Ritchie ME, Phipson B, Wu D, Hu Y, Law CW, Shi W, Smyth GK (2015). “limma powers differential expression analyses for RNA-sequencing and microarray studies.” Nucleic Acids Research, 43(7), e47. doi:10.1093 / nar / gkv007.

[0405] Sakoe, Hiroaki, and Seibi Chiba. 1990. “Dynamic Programming Algorithm Optimization for Spoken Word Recognition.” In Readings in Speech Recognition, 159-65. Elsevier.

[0406] Singh, A., Shannon, C.P., Gautier, B., Rohart, F., Vacher, M., Tebbutt, S.J. and Le Cao, K.-A. (2019).

[0407] DIABLO: an integrative approach for identifying key molecular drivers from multi-omics assays. Bioinformatics, [online] 35(17), pp.3055-3062.

[0408] Stevens, P.E., et alf. (2024). KDIGO 2024 Clinical Practice Guideline for the Evaluation and Management of Chronic Kidney Disease. Kidney International, 105(4), pp.S117-S314. doi : https : / / do i .o rg / 10.1016 / j.kint.2023.10.018.

[0409] Sumida K., et al; for the Chronic Kidney Disease Prognosis Consortium. Conversion of Urine Protein- Creatinine Ratio or Urine Dipstick Protein to Urine Albumin-Creatinine Ratio for Use in Chronic Kidney Disease Screening and Prognosis: An Individual Participant-Based Meta-analysis. Ann Intern Med.2020;173:426-435.

[0410] Tangri N, Grams ME, Levey AS, et al. Multinational Assessment of Accuracy of Equations for Predicting Risk of Kidney Failure: A Meta-analysis. JAMA. 2016;315(2):164-174.

[0411] Therneau, Terry, and Patricia Grambsch. 2010. Modeling Survival Data: Extending the Cox Model. 1st ed. Statistics for Biology and Health. New York, NY: Springer.

[0412] Therneau, 2024. A package for survival analysis in R Package vignette available at https: / / ftp2.uib.no / cran / web / packages / survival / vignettes / survival.pdf

[0413] Thul, Peter J., and Cecilia Lindskog. 2018. “The Human Protein Atlas: A Spatial Map of the Human Proteome.” Protein Science: A Publication of the Protein Society 27 (1): 233-44.

[0414] Tibshirani, R. 1996. “Regression Shrinkage and Selection via the Lasso.” Journal of the Royal Statistical Society Series B-Methodological 58 (1): 267-88.

[0415] Tucker, L. R. 1966. “Some Mathematical Notes on Three-Mode Factor Analysis.” Psychometrika 31 (3):

[0416] 279-311.

[0417] Tukey, J. W. 1949. “Comparing Individual Means in the Analysis of Variance.” Biometrics 5 (2): 99-114. Yoshida, Teruhiko et al. Transcriptomic Analysis of Human Podocytes In Vitro: Effects of Differentiation and APOL1 Genotype. Kidney International Reports, Volume 8, Issue 1 , 164 - 178 Vanholder, Raymond, Lieven Annemans, Edwina Brown, Ron Gansevoort, Judith J. Gout-Zwart, Norbert Lameire, Rachael L. Morton, et al. 2017. “Reducing the Costs of Chronic Kidney Disease While Delivering Quality Health Care: A Call to Action.” Nature Reviews. Nephrology 13 (7): 393- 409.

[0418] Zhou W, Triche TJ Jr, Laird PW, Shen H. SeSAMe: reducing artifactual detection of DNA methylation by Infinium BeadChips in genomic deletions. Nucleic Acids Res. 2018 Nov 16;46(20):e123. All references cited herein are incorporated herein by reference in their entirety and for all purposes to the same extent as if each individual publication or patent or patent application was specifically and individually indicated to be incorporated by reference in its entirety.

[0419] The specific embodiments described herein are offered by way of example, not by way of limitation. Various modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the technology as described. Any sub-titles herein are included for convenience only and are not to be construed as limiting the disclosure in any way. Unless context dictates otherwise, the descriptions and definitions of the features set out above are not limited to any particular aspect or embodiment of the invention and apply equally to all aspects and embodiments which are described. The methods of any embodiments described herein may be provided as computer programs or as computer program products or computer readable media carrying a computer program which is arranged, when run on a computer, to perform the method(s) described above.

[0420] Throughout the specification and claims, the following terms take the meanings explicitly associated herein, unless the context clearly dictates otherwise. The phrase “in one embodiment” as used herein does not necessarily refer to the same embodiment, though it may. Furthermore, the phrase “in another embodiment” as used herein does not necessarily refer to a different embodiment, although it may. Thus, various embodiments of the invention may be readily combined, without departing from the scope or spirit of the invention. It must be noted that, as used In the specification and the appended claims, the singular forms “a,” “an,” and “the” include plural referents unless the context clearly dictates otherwise.

[0421] Ranges may be expressed herein as from “about” one particular value, and / or to “about” another particular value. When such a range is expressed, another embodiment includes from the one particular value and / or to the other particular value. Similarly, when values are expressed as approximations, by the use of the antecedent “about,” it will be understood that the particular value forms another embodiment. The term “about” in relation to a numerical value is optional and means for example + / -10%.

[0422] Throughout this specification, including the claims which follow, unless the context requires otherwise, the word “comprise” and “include”, and variations such as “comprises”, “comprising”, and “including” will be understood to imply the inclusion of a stated integer or step or group of integers or steps but not the exclusion of any other integer or step or group of integers or steps. Other aspects and embodiments of the invention provide the aspects and embodiments described above with the term “comprising” replaced by the term “consisting of’ or “consisting essentially of’, unless the context dictates otherwise, “and / or” where used herein is to be taken as specific disclosure of each of the two specified features or components with or without the other. For example “A and / or B” is to be taken as specific disclosure of each of (i) A, (ii) B and (iii) A and B, just as if each is set out individually herein, “and / or” where used herein is to be taken as specific disclosure of each of the two specified features or components with or without the other. For example “A and / or B” is to be taken as specific disclosure of each of (i) A, (ii) B and (iii) A and B, just as if each is set out individually herein.

[0423] The features disclosed in the foregoing description, or in the following claims, or in the accompanying drawings, expressed in their specific forms or in terms of a means for performing the disclosed function, or a method or process for obtaining the disclosed results, as appropriate, may, separately, or in any combination of such features, be utilised for realising the invention in diverse forms thereof.

Claims

1. CLAIMS1. A computer-implemented method of determining a prognosis for a subject that has been identified as having or being likely to have chronic kidney disease, the method comprising:3.obtaining the values of a first plurality of proteomic features associated with the subject, wherein the first plurality of proteomic features comprise at least two proteomic features selected from a first set of proteomic features comprising features indicative of the protein level in a sample from the subject of at least two different proteins selected from:4.PTGDS or PGF or TNFRSF14,5.F11R or CD40,6.CST3 or IGFBP4,7.SDC1 or EPHA2,8.PILRA or PILRB,9.FAS or ESAM,10.CD160 orJAM2 or NCRI ,11.TFPI orAPCS,12.CPA1 or CPBI ,13.CDH3 or TNFRSF14,14.GFOD2,15.SUSD1 or PLXNA4,16.ADH4 or UPBI,17.KLK11 or NECTIN4,18.ATRAID or EFCAB14,19.DBI or TMSBW,20.TNFRSF14 orTNFRSFIA,21.DSC2 or TNFRSFI B,22.CD274 orTNFRSFI B,23.TNFRSF11 A or TNFRSF1 A,24.GPR37 or IGFBP4,25.THBD or ESAM, and26.ENTPD5,27.NPDC1 ,28.HSPG2,29.IGFBP6,30.PGF,31.LY6D,32.BTN2A1 ,33.AGRN,34.TXNDC5,35.JAM2, PIK3IP1 ,36.RELT,37.IGFBP4,38.TNFRSF19,39.WFDC2,40.ATRAID,41.EFCAB14,42.TFPI,43.HAVCR1 , and44.MXRA8; and45.classifying the subject between a plurality of classes including at least a first class and a second class, the first class having better prognosis than the second class using the values of the first plurality of proteomic features and a classifier model that has been trained to classify subjects with chronic kidney disease between the first class and the second class using training data comprising values of the first plurality of proteomic features for a plurality of training subject and associated labels indicating whether each training subject belongs to the first or second class.

2. The method of claim 1 , wherein the first plurality of proteomic features comprises features indicative of the protein level of:47.two or more or all of: PTGDS or PGF, F11 R or CD40, SDC1 or EPHA2, PILRA or PILRB, two or more or all of: CD160 or JAM2, and TFPI orAPCS; or48.two or more or all of: SDC1 , TFPI, PTGDS, CD160, PILRA, and F11 R; or49.two or more or all of: CD160 or JAM2, and F11R or CD40; or50.two or more or all of: CD160 and F11 R; or51.two or more or all of: SDC1 or EPHA2, TNFRSF11 A or TNFRSF1 A, and CD160 or JAM2; or PTGDS or PGF, and F11 R or CD40; or52.two or more or all of: PTGDS or PGF, CPA1 or CPB1 , CDH3 of TNFRSF14, SUSD1 or PLXNA4, GFOD2, ADH4 or UBP1 , and KLK11 or NECTIN4, optionally wherein the two or more include at least PTGDS or PGF, and / or at least GFOD2 and / or at least CPA1 or CPAB1 , and / or at least KLK11 or NECTIN4; or53.two or more or all of: PTGDS or PGF or TNFRSF14, TNFRSF14, DBI or TMSB10, THBD or ESAM, DSC2 or TNFRSF1 B, TNFRSF11 A or TNFRSF1 A, GPR37 or IGFBP4, and ENTDP5, optionally wherein the two or more include at least PTGDS or PGF or TNFRSF14, and / or at least DBI or TMSB10, and / or at least THBD or ESAM, and / or at least DSC2 or TNFRSF1 B, and / or at least TNFRSF11 A or TNFRSF1 A; or54.two or more or all of: PTGDS or TNFRSF14, CD160 or NCR1 , ATRAID or EFCAB14, DBI or TMSB10, TNFRSF14 or TNFRSF1A, DSC2 or TNFRSF1B, and CD274 or TNFRSF14, optionally wherein the two or more include at least PTGDS or TNFRSF14, and / or at least DSC2 or TNFRSFIB; or PTGDS and at least one protein different from PTGDS selected from: F11R, CD40, CST3, IGFBP4, SDC1 , EPHA2, PILRA, PILRB, FAS, IGFBP6, ESAM, CD160, JAM2, NCR1 , TFPI, PROS1 , APCS, CPA1 , CPB1 , CDH3, TNFRSF14, GFOD2, SUSD1 , PLXNA4, ADH4, UPB1 , KLK11 , NECTIN4, ATRAID, EFCAB14, DBI, TMSB10, TNFRSF14, TNFRSF1A, DSC2, TNFRSF1B, CD274, TNFRSF1B, TNFRSF11A, TNFRSF1A, GPR37, IGFBP4, THBD, ESAM, ENTPD5, NPDC1 , HSPG2, IGFBP6, PGF, LY6D, BTN2A1 , AGRN, TXNDC5, JAM2, PIK3IP1 , RELT, IGFBP4, TNFRSF19, WFDC2, ATRAID, EFCAB14, TFPI, HAVCR1 , and MXRA8; or55.ATRAID, and PTGDS; or56.PTGDS, and CD160 or JAM2; or57.PTGDS, and DBI; or58.PTGDS, and TNFRSF14; or59.ATRAID, and CD160 or JAM2; or60.ATRAID, and DBI; or61.ATRAID, PTGDS, and DBI; or62.PTGDS, and DSC2; or63.ATRAID, PTGDS, and CD160 or JAM2; or64.PTGDS, and CD274; or65.PTGDS, HSPG2, and NDPC1 ; or66.PTGDS, ATRAID, BTN2A1 and PIK3IP1 , or67.at least 3 proteins selected from NPDC1 , HSPG2, PTGDS, IGFBP6, PGF, LY6D, BTN2A1 , TNFRSF14, AGRN, TXNDC5, JAM2, PIK3IP1 , RELT, TNFRSF1A, IGFBP4, TNFRSF19, NECTIN4, WFDC2, ATRAID, and EFCAB14.

3. The method of claim 1 or claim 2, wherein the first plurality of proteomic features comprises features indicative of the protein level of one or more or all of:69.PTGDS or PGF or TNFRSF14,70.TNFRSF11 A or TNFRSF1 A,71.DBI or TMSBW,72.ENTPD5,73.DSC2 orTNFRSFIB.

4. The method of any of claims 1 to 3, wherein the first plurality of proteomic features further comprises features indicative of the protein level of any one or more of: ATRAID, EFCAB14, DBI, TMSB10, TNFRSF14, TNFRSF1B, DSC2, CD274, ENTPD5, PTPRS, ANG, CXADR, TNFRSF11A, GKN1 , DSG4, PVALB, DSG3, KDR, WFDC2, MIA, DDA1 , COX6B1 , TOMM20, CFP, EFNB2, LCN15, CHCHD10, SCRIB, AMN, CXCL5, RARRES2 and INHBC.

5. The method of any of claims 1 to 4, wherein:76.in the first set of proteomic features, the features indicative of the protein level of PTGDS, PGF, F11R, CD40, CST3, IGFBP4, SDC1 , EPHA2, PILRA, PILRB, FAS, ESAM, CD160, JAM2, TFPI, CDH3, GFOD2, KLK11 , NECTIN4, TNFRSF14, CD160, NCR1 , ATRAID, EFCAB14, DBI, TMSB10, DSC2, TNFRSF1B, TNFRSF1A, CD274, GPR37, IGFBP4, THBD, TNFRSF11A, APCS, NPDC1 , HSPG2, IGFBP6, PGF, LY6D, BTN2A1 , AGRN, TXNDC5, JAM2, PIK3IP1 , RELT, IGFBP4, TNFRSF19, WFDC2, ATRAID, EFCAB14, TFPI, HAVCR1 , and MXRA8 are expected to be higher in subjects in the second class than in subjects in the first class and the feature indicative of the protein level of CPA1 , CPB1 , SUSD1 , PLXNA4, ADH4, UPB1 , and ENTPD5 are expected to be lower in subjects in the second class than in subjects in the first class.

6. A computer-implemented method of determining a prognosis for a subject that has been identified as having or being likely to have chronic kidney disease, the method comprising:78.obtaining the values of a first plurality of clinical features associated with the subject, wherein the first plurality of clinical features comprise at least two clinical features selected from a first set of clinical features comprising: a feature indicative of creatinine concentration in a blood or urine sample from the subject, a feature indicative of estimated glomerular filtration rate associated with a blood sample from the subject, a feature indicative of age of the subject, a feature indicative of a diastolic blood pressure measurement for the subject, a feature indicative of a systolic blood pressure measurement for the subject, a feature indicative of the presence or concentration of albumin in a blood sample from the subject, a feature indicative of phosphate concentration in a blood sample from the subject, a feature indicative of the ratio of protein to creatinine in a urine sample from the subject, and a feature indicative of whether the subject is undergoing treatment with calcium channel blockers; and79.classifying the subject between a plurality of classes including at least a first class and a second class, the first class having better prognosis than the second class using the values of the first plurality of clinical features and a classifier model that has been trained to classify subjects with chronic kidney disease between the first class and the second class using training data comprising values of the first plurality of clinical features for a plurality of training subject and associated labels indicating whether each training subject belongs to the first or second class.

7. The method of claim 6, wherein the first plurality of clinical features include at least the feature indicative of creatinine concentration in a blood or urine sample from the subject, and / or the feature indicative of estimated glomerular filtration rate associated with a blood sample from the subject, and / or the feature indicative of the ratio of protein to creatinine in a urine sample from the subject.

8. The method of claim 6 or claim 7, wherein the first plurality of clinical features include at least the feature indicative of creatinine concentration in a blood or urine sample from the subject, or wherein the first plurality of clinical features include: the feature indicative of creatinine concentration in a blood or urine sample from the subject, the feature indicative of the ratio of protein to creatinine in a urine sample from the subject, and the feature indicative of age ofthe subject.

9. The method of any of claims 6 to 8, wherein the first plurality of clinical features include:83.(a) the feature indicative of creatinine concentration in a blood or urine sample from the subject, and the feature indicative of age of the subject; or84.(b) the feature indicative of creatinine concentration in a blood or urine sample from the subject, the feature indicative of estimated glomerular filtration rate associated with a blood sample from the subject, the feature indicative of age of the subject, the feature indicative of a diastolic blood pressure measurement for the subject, the feature indicative of the presence or concentration of albumin in a blood sample from the subject, the feature indicative of phosphate concentration in a blood sample from the subject, and the feature indicative of whether the subject is undergoing treatment with calcium channel blockers; or85.(c) the feature indicative of creatinine concentration in a blood or urine sample from the subject, the feature indicative of the ratio of protein to creatinine in a urine sample from the subject, and the feature indicative of age of the subject; or86.(d) all of the features in the first set of features.

10. The method of any of claims 6 to 9, wherein:88.in the first set of clinical features: the feature indicative of creatinine concentration in a blood or urine sample from the subject, the feature indicative of a diastolic blood pressure measurement for the subject, the feature indicative of a systolic blood pressure measurement for the subject, the feature indicative of phosphate concentration in a blood sample from the subject, and the feature indicative of whether the subject is undergoing treatment with calcium channel blockers are expected to be higher in subjects in the second class than in subjects in the first class; and89.the feature indicative of estimated glomerular filtration rate associated with a blood sample from the subject, the feature indicative of age of the subject, and the feature indicative of the presence or concentration of albumin in a blood sample from the subject are expected to be lower in subjects in the second class than in subjects in the first class.

11. The method of any preceding claim, wherein the classifier model is a binary classifier and / or wherein the plurality of classes consists of two classes and / or wherein the classifier model is a machine learning model that has been trained to classify subjects with chronic kidney disease between the first class and the second class, optionally wherein the classifier model is a decision tree model, or a gradient boosted decision tree model.

12. The method of any preceding claim, wherein the first class (good prognosis) comprises subjects that have better prognosis than the second class (intermediate / poor prognosis).

13. The method of any preceding claim, wherein a poorer prognosis is associated with faster disease progression than a better prognosis, and / or wherein a poorer prognosis is associated with a shorter expected time to a disease worsening event than a better prognosis.

14. The method of any preceding claim, wherein each class is associated with a different prognosis in terms of an expected time to disease progression, wherein disease progression is associated with a diagnosis of end stage renal disease, dialysis or renal replacement therapy.

15. The method of any preceding claim, wherein the chronic kidney disease is diabetic kidney disease.

16. The method of any preceding claim, wherein the subject is a human, and / or wherein the subject is an adult, and / or wherein the subject is a human of 16 years or over.

17. The method of any preceding claim, further comprising selecting a subject classified in the second class for further diagnostic testing, optionally wherein the method comprises performing the method of any of claims 6 to 10 and the further diagnostic testing comprises performing the method of any of claims 1 to 5, selecting a subject classified in the second class for inclusion in a clinical trial, or selecting a subject classified in the second class for treatment with a first predetermined treatment and selecting a subject classified in the first class for treatment with a second treatment different from the first predetermined treatment, optionally wherein the first predetermined treatment comprises treatment with a predetermined therapeutic compound and the second treatment comprises treatment with a different therapeutic compound.

18. The method of any preceding claim, further comprising selecting a subject classified in the second class for more regular monitoring than a subject classified in the first class.

19. The method of any preceding claim, wherein the classifier model is a machine learning model that has been trained to classify subjects with chronic kidney disease between the first class and the second class using training data comprising values of the first plurality of proteomic features for a plurality of training subject and associated ground truth labels indicating whether each training subject belongs to the first or second class, wherein the ground truth labels were obtained by:97.obtaining clinical data fora plurality of subjects with chronic kidney disease, said clinical data comprising for each of one or more biomarkers, a plurality of measurements of the biomarker at a plurality of timepoints forming a temporal trajectory for the biomarker for the subject, the one or more biomarkers comprising one or more biomarkers of kidney function, optionally creatinine, and / or one or more biomarkers of anaemia, optionally haematocrit or haemoglobin; and98.identifying a first group and a second group of subjects in the plurality of subjects, wherein subjects in the first and second groups have different temporal trajectories of the one or more biomarkers, wherein the first group of subjects are associated with a ground truth label indicating that the subjects in the first group belong to the first class, and the second group of subjects are associated with a ground truth label indicating that the subjects in the second group belong to the second class.

20. The method of claim 19, wherein identifying a first group and a second group of subjects comprises:100.for each biomarker separately, obtaining pairwise distances between all of the temporal trajectories of the subjects thereby obtaining a tensor of pairwise distances, optionally using a Dynamic Time Warping (DTW) algorithm; and101.identifying clusters of subjects that have different temporal trajectories using the tensor of pairwise distances, optionally by applying a matrix decomposition method to the tensor of pairwise distances to obtain a matrix of predetermined dimensions, applying a dimensionality reduction algorithm to the matrix of predetermined dimensions to obtain coordinates for each subject in a two dimensional space, and applying a clustering algorithm to the coordinates.

21. The method of claim 19 or claim 20, the classifier model uses as input proteomic features indicative of the protein level in a sample from the subject of at least two different proteins that are significantly differentially expressed between the subjects in the first group of subjects and the subjects in the second group of subjects.

22. The method of any preceding claim, further comprising training the classifier model.

23. A computer-implemented method of identifying a plurality of endotypes of subjects with chronic kidney disease, the method comprising:105.obtaining clinical data fora plurality of subjects with chronic kidney disease, said clinical data comprising for each of one or more biomarkers, a plurality of measurements of the biomarker at a plurality of timepoints forming a temporal trajectory for the biomarker for the subject, the one or more biomarkers comprising one or more biomarkers of kidney function, optionally creatinine, and / or one or more biomarkers of anaemia, optionally haematocrit or haemoglobin; and106.identifying a first group and a second group of subjects in the plurality of subjects, wherein subjects in the first and second groups have different temporal trajectories of the one or more biomarkers;107.optionally wherein identifying a first group and a second group of subjects comprises:108.for each biomarker separately, obtaining pairwise distances between all of the temporal trajectories of the subjects thereby obtaining a tensor of pairwise distances, optionally using a Dynamic Time Warping (DTW) algorithm; and109.identifying clusters of subjects that have different temporal trajectories using the tensor of pairwise distances, optionally by applying a matrix decomposition method to the tensor of pairwise distances to obtain a matrix of predetermined dimensions, applying a dimensionality reduction algorithm to the matrix of predetermined dimensions to obtain coordinates for each subject in a two dimensional space, and applying a clustering algorithm to the coordinates.

24. A computer-implemented method of training a machine learning model to provide a prognosis for a subject with chronic kidney disease, the method comprising:110.identifying a plurality of endotypes of subjects with chronic kidney disease using a method according to claim 23 using clinical data for a plurality of subjects with chronic kidney disease thereby identifying a first group and a second group of subjects in the plurality of subjects;111.obtaining values of a first plurality of clinical features associated with the subjects and / or values of a first plurality of proteomic features associated with the subjects; and training a machine learning classifier to classify subjects with chronic kidney disease between the first class and the second class using training data comprising values of the first plurality of proteomic features for the plurality of subjects and / or values of the first plurality of clinical features for the plurality of subjects, and associated ground truth labels indicating whether each training subject belongs to the first or second class, wherein the first group of subjects are associated with a ground truth label indicating that the subjects in the first group belong to the first class, and the second group of subjects are associated with a ground truth label indicating that the subjects in the second group belong to the second class25. A method of treating a subject that has been identified as having or being likely to have chronic kidney disease, the method comprising determining a prognosis for the subject using the method of any of claims 1 to 22, and performing one of: (i) treating the subject with a first therapy, when the subject is classified in the second class; and (ii) treating the subject with a second therapy different from the first therapy, when the subject is not classified in the second class.

26. A method comprising performing the computer implemented method of any preceding claim, and prior to said performing, obtaining the values of the clinical features from the subject or from samples previously obtained from the subject, and / or obtaining the values of the proteomic features from a sample previously obtained from the subject.

27. A system comprising a processor and one or more computer readable media storing instructions that, when executed by the processor, cause the processor to implement the method of any of claims 1 to 23.

28. A computer readable medium or media storing instructions that, when executed by a processor, cause the processor to implement the method of any of claims 1 to 23.

29. A kit comprising reagents specific for measuring a protein level of at least two proteins selected from:117.PTGDS or PGF or TNFRSF14,118.F11R or CD40,119.CST3 or IGFBP4,120.SDC1 or EPHA2, PILRA or PILRB,121.FAS or ESAM,122.CD160 or JAM2 or NCRI ,123.TFPI or CPA 1 or CPBI ,124.CDH3 or TNFRSF14,125.GFOD2,126.SUSD1 or PLXNA4,127.ADH4 or UPBI ,128.KLK11 or NECTIN4,129.ATRAID or EFCAB14,130.DBI or TMSBW,131.TNFRSF14 or TNFRSFIA,132.DSC2 or TNFRSFI B,133.CD274 or TNFRSFI B,134.TNFRSF11 A or TNFRSF1 A,135.GPR37 or IGFBP4,136.THBD or ESAM,137.ENTPD5,138.NPDC1 ,139.HSPG2,140.IGFBP6,141.LY6D,142.BTN2A1 ,143.AGRN,144.TXNDC5,145.PIK3IP1 ,146.RELT,147.IGFBP4,148.TNFRSF19,149.WFDC2,150.ATRAID,151.EFCAB14,152.TFPI,153.HAVCR1 , and154.MXRA8155.in a sample from a subject.

30. The method of any of claims 1 to 5 or 11 to 23 or the kit of claim 29, wherein the protein levels have been measured or are measured using an immune assay, optionally selected from a proximity extension assay, an ELISA assay, or a bead-based immunoassay.

Images