Serum kidney injury molecule 1 and β2-microglobulin perform as well as larger biomarker panels for prediction of rapid decline in renal function in type 2 diabetes

Aims/hypothesis As part of the Surrogate Markers for Micro- and Macrovascular Hard Endpoints for Innovative Diabetes Tools (SUMMIT) programme we previously reported that large panels of biomarkers derived from three analytical platforms maximised prediction of progression of renal decline in type 2 diabetes. Here, we hypothesised that smaller (n ≤ 5), platform-specific combinations of biomarkers selected from these larger panels might achieve similar prediction performance when tested in three additional type 2 diabetes cohorts. Methods We used 657 serum samples, held under differing storage conditions, from the Scania Diabetes Registry (SDR) and Genetics of Diabetes Audit and Research Tayside (GoDARTS), and a further 183 nested case–control sample set from the Collaborative Atorvastatin in Diabetes Study (CARDS). We analysed 42 biomarkers measured on the SDR and GoDARTS samples by a variety of methods including standard ELISA, multiplexed ELISA (Luminex) and mass spectrometry. The subset of 21 Luminex biomarkers was also measured on the CARDS samples. We used the event definition of loss of >20% of baseline eGFR during follow-up from a baseline eGFR of 30–75 ml min−1 [1.73 m]−2. A total of 403 individuals experienced an event during a median follow-up of 7 years. We used discrete-time logistic regression models with tenfold cross-validation to assess association of biomarker panels with loss of kidney function. Results Twelve biomarkers showed significant association with eGFR decline adjusted for covariates in one or more of the sample sets when evaluated singly. Kidney injury molecule 1 (KIM-1) and β2-microglobulin (B2M) showed the most consistent effects, with standardised odds ratios for progression of at least 1.4 (p < 0.0003) in all cohorts. A combination of B2M and KIM-1 added to clinical covariates, including baseline eGFR and albuminuria, modestly improved prediction, increasing the area under the curve in the SDR, Go-DARTS and CARDS by 0.079, 0.073 and 0.239, respectively. Neither the inclusion of additional Luminex biomarkers on top of B2M and KIM-1 nor a sparse mass spectrometry panel, nor the larger multiplatform panels previously identified, consistently improved prediction further across all validation sets. Conclusions/interpretation Serum KIM-1 and B2M independently improve prediction of renal decline from an eGFR of 30–75 ml min−1 [1.73 m]−2 in type 2 diabetes beyond clinical factors and prior eGFR and are robust to varying sample storage conditions. Larger panels of biomarkers did not improve prediction beyond these two biomarkers. Electronic supplementary material The online version of this article (10.1007/s00125-018-4741-9) contains peer-reviewed but unedited supplementary material, which is available to authorised users.

For high sensitivity serum metabolite estimation 30µL of aqueous standards, controls and samples are pipetted into 1.8mL polypropylene snap-top Eppendorf tubes. To each tube, 75µL of methanol/water stable isotope mixture 1 followed by 75µL of pure methanol will be added, the tubes capped, vortex mixed for 2-5seconds, and centrifuged at 21,000g at 4 o C for 6min. Supernatants, 120µL, will be transferred to a 96 deep well (2mL) polypropylene sample block, sealed, and placed in the autosampler at 8 o C ready for analysis by LC electrospray MSMS on an API5500 under Analyst 1.5.2 control. Sample supernatants (3µL) are injected automatically and chromatography performed on an AstecChirobiotic™ T HPLC column 25cm x 2.1mm, 5µm with a 2cm x 4.0mm, 5µm guard column with an isocratic running solvent (acetonitrile:water, 1:1, with 0.025% formic acid) at a flow rate of 250µl/min. Data is acquired in positive ion MRM mode for 15min. Followed by re-injection and data acquisition in negative ion MRM mode for 9min. For Serum tryptic peptide targeted proteomic analysis, 10µL of plasma controls and samples are pipetted as above. To each tube, 40µl of water, 50µL of diluted stable isotope labelled albumin T6 aqueous internal standard, 10µL of acetonitile and 10µL of 1% formic acid are added and mixed on an orbital shaker at RT for 5min. Then 6µL of 1M NH4CO3 is added to each tube and vortex mixed for 5 seconds before addition of 25µL of trypsin, vortex mixing, and incubation at 37 o C for 1h. After incubation, 200µL of running buffer (acetonitrile:water, 1:1, with 0.025% formic acid) is added to each tube, vortex mixed for 2-5seconds and centrifuged at 21,000g at 4 o C for 5min. The supernatants, 200µL, will be transferred to a 96 deep well (2mL) polypropylene sample block as above. Sample supernatants (5µL) will be injected automatically and chromatography performed on two, in series, AstecChirobiotic™ T HPLC Guard columns 2cm x 4.0mm, 5µm with an isocratic running solvent (acetonitrile:water, 1:1, with 0.025% formic acid) at a flow rate of 320µl/min. Data will be acquired in positive ion MRM mode for 10min. Data will be analysed in Analyst version 1.5.2 and MultiQuant version 2.1.

Details of the Myriad RBM platform
At Myriad RBM (MRBM) Luminex technology performs multiplexed, microsphere-based assays in a single reaction vessel by combining optical classification schemes, biochemical assays, flow cytometry and advanced digital signal processing hardware and software. Multiplexing is accomplished by assigning each analyte-specific assay a microsphere set labelled with a unique fluorescence signature. To attain distinct microsphere signatures, two fluorescent dyes, red and far red, are mixed in various combinations using various intensity levels of each dye. Each batch or set of microspheres is encoded with a fluorescent signature by impregnating the microspheres with one of these dye combinations. After the encoding process, an assay-specific capture reagent (i.e., antigens, antibodies, receptors, peptides, enzyme substrates, etc.) is conjugated covalently to each unique set of microspheres. Covalent attachment of the capture reagent to the microspheres is achieved with standard carbodiimide chemistry. After optimizing the parameters of each assay separately, Multi-Analyte Profiles are performed by mixing different sets of the microspheres in a single well of a 96-or 384-format microtiter plate. A small sample volume is added to the well and allowed to react with the microspheres. The assay-specific capture reagent on each individual microsphere binds the analyte of interest. A cocktail of assay-specific, biotinylated detecting reagents (e.g., antigens, antibodies, ligands, etc.), is reacted with the microsphere mixture, followed by a streptavidin-labelled fluorescent "reporter" molecule (typically phycoerythrin). Because the microspheres are in suspension, the assay kinetics are near solutionphase. Finally, the multiplex is washed to remove unbound detecting reagents. After washing, the mixture of microspheres is analyzed using the Luminex 100/200™ instrument. Similar to a flow cytometer, the instrument uses hydrodynamic focusing to pass the microspheres in single file through two laser beams. As each individual microsphere passes through the excitation beams, it is analyzed for size, encoded fluorescence signature and the amount of fluorescence generated in proportion to the analyte. The resulting data stream is interpreted using proprietary data analysis software developed at MRBM. Assays are run in high density multiplexed panels and the Least Detectable Dose (LDD) is determined as the mean +3 standard deviations of 20 blank readings. The LLOQ is determined by the concentration of an analyte where the measurement of analyte demonstrates a coefficient of variation (CV) of 30%. It represents the lowest concentration of analyte that can be measured with a precision better than or equal to 30%. Appropriate dilutions are made to ensure a quantitative measurement within the limits of the assay. An eight (n=8) point standard curve (S1 -S8) is used to obtain quantitative measurements for each sample. Quality Controls (QC's) are run in duplicate along different points of the curve to ensure both accuracy and precision for each analyte.

External QC measures
We undertook a limited study to capture external QC measures on our SUMMIT sample retrieval, transfer, assay and data handling processes for renal biomarkers. This included up to 13 pairs of blinded duplicate samples from the Go-DARTS and SDR sample sets run at the two principle laboratories included in this study. For this data we calculated the intra-class correlation coefficients (ICC) for all biomarkers except Troponin T. We used three levels of ICC to categorise biomarker assays -ICC-≥0.75 (good), ICC 0.4-0.74 (acceptable) and ICC <0.4 (poor). Of the 42 biomarkers included in the study 39 (93%) were either acceptable or good, 2 performed poorly (1 Luminex and 1 mass spectroscopy biomarkers) and one was not assessed (see table 1 for details).

Biomarker data cleaning and imputation
Data was imputed using a sparse iterative regression approach, where each missing measurement is repeatedly predicted from the observed values using L1-regularised linear and generalised linear models. Our imputation model doesn't make use of rapid progression status (the target variable to be predicted), in the reconstruction of the missing values. We used a flat prior for imputing the values missing at random and bounded Pareto priors for imputing continuous censored values. When available, we made use of the information concerning detection thresholds in imputing plausible values for censored entries; otherwise these were inferred from the observed values. The iterative imputation model was run 10 times, with initial values of the missing at random entries set by sampling from the marginal distribution of the observed values for each variable. Clinical covariates were imputed first. For biomarkers, the imputation model used all variables including imputed clinical covariates as well as information concerning the missingness type and the lower and upper detection limits provided by the biomarker laboratories. The dataset used in the analysis was the average of the 10 imputed sets.

Predictive performance of sparse biomarker panels in the three validation cohorts
In addition to calculating the AUROC for each model we also considered difference in test log likelihoods to evaluate the strength of evidence favouring one model over another, based on the asymptotic equivalence of model selection by cross-validation and Akaike Information Criteria (AIC) [1]. For a p-value of 0.05, in a comparison of a two nested models differing by one extra parameter, the difference in the training deviance is distributed according to chisquare distribution with one degree of freedom, and approximately equal to 4. The difference in AIC (the difference in training deviances minus twice the difference in the number of parameters), would be 2 natural log units, and the corresponding approximate difference in the test log-likelihoods will be 1 natural log unit. Therefore, classical statisticians using the likelihood-ratio test to compare training likelihoods of the two models at the significance level of 0.05, should regard a difference of 1 in test log-likelihoods as significant. The more stringent thresholds suggested for Bayes Factors use the cut-offs of 1.2 and 2.3 natural log units as "substantial" and "strong" evidence in favour of the higher-likelihood model [2]. The use of AIC rather than the training likelihoods for model selection has the advantage of penalizing complex models (which helps to prevent overfitting), while the use of the test likelihoods has the additional advantage of not needing to explicitly evaluate the model complexity.
In the SDR and new Go-DARTS datasets we evaluated the performance increment achieved with the larger multiplatform biomarker panels identified previously using the original case-control dataset [3]. The addition of the 14 or 35 biomarker panels to a model including the limited clinical covariates improved the AUROC in both cohorts, with little difference between the two panels (ESM Table 4). Overall, for SDR the best performing panel was the 35 biomarker panel. However, for Go-DARTS the best Luminex-based sparse model out-performed either of these two larger panels when assessed on top of clinical covariates. It is worth noticing that a model based only on the extended set of clinical covariates, including weighted average of past eGFR values, also improved prediction to a similar degree as the biomarker panels added to a basic set of covariates. Where units given as * the measure is semi-quantitative with the value being a ratio of the biomarker to a stable isotope of albumin T6. Where a biomarker is followed by the suffix (1) or (2), this indicates the peptide produced by tryptic digest to which the measured signal relates. a Intraclass correlation coefficient was not assessed b Intraclass correlation coefficient <0.4 (poor)