Fig 1.
The key technological features involved in the design of our recombinant antibody microarray technology platform, outlining the specific, individual features uniquely addressed in this study.
Table 1.
List of normalization processes evaluated in this study.
Fig 2.
Evaluation of the antibody quality in terms of spotting concentration, on-chip functionality and specificity.
(A) Two different scFvs antibodies, denoted as scFv1 (against complement protein C5) and scFv2 (against Myomesin 2), were dispensed in quintuplicate at three different concentrations, stock solution, 1:2 dilution and 1:4 dilution. The microarray was processed using a biotinylated serum sample. A scanner setting of 10 μm resolution, using 70% PMT gain and 90% laser power was used. (B) ScFv3 (anti-C1q) was spotted in 6 different concentrations, each in 12 replicate spots. The microarray was processed using pure antigen (5 nM pure Alexa 647-labeled C1q) and scanned at 10 μm resolution, using 60% PMT gain and 90% laser power. (C) Protein expression profiling of biotinylated serum sample. Several antibody clones directed towards the same antigen but against different epitopes were used, targeting Apolipoprotein-A4 (Apo-A4, n = 3), Complement factor 3 (C3, n = 6), Cystatin C (n = 4) and Monocyte Chemoattractant Protein-1 (MCP-1, n = 9). The observed signal intensities are given, in terms of fold change, in diseased vs. healthy samples.
Fig 3.
Evaluation of microarray quality in terms of array design and printing markers.
(A) Representative image of a microarray slide with 14 identical subarrays. Each subarray (n = 14) is divided into 3 segments separated by dispensed rows of reference marker replicate spots (n = 4). (B) Blank replicate spots (n = 5) detected in the marked area. No signal intensity was obtained from the dispensed blank replicate spots after scanning the microarray using a wavelength of 633 nm, while positive spots at 543 nm (Cadaverine) confirmed that spotting had occurred.
Table 2.
Sample labeling QC.
Fig 4.
Day-to-day variation before and after normalization.
(A) Unsupervised PCA analysis of un-normalized log2 transformed raw data. The samples were colored with respect to different rounds (days) of analysis. (B) Random Forest supervised classification differentiates between different days of analysis. (C) Unsupervised PCA analysis of processed data. The data was normalized using the “subtract by group mean” approach. (D) Random Forest classifier applied after normalization step. (E) Systemic technical variation, expressed in terms of CV, with respect to mean (median) inter- and intra-day as well as mean intra-slide variations.
Fig 5.
Boxplots for 5 antibody intensities over all the analyzed samples. Each data point represents one sample. A cut-off limit was established based on the mean negative control signal (PBS) across all the samples plus 2 standard deviations. Each analyte, from which the mean signal intensities were found to be below the LOD in > 70% of samples was removed from the data (e.g. FASN (3) and MAKT (2)). The red line corresponds to the cut-off limit.
Fig 6.
Evaluation of different normalization procedures.
Normal Q-Q plots, boxplots, density plots and meanSdPlots were used as qualitative measures in order to compare the log intensity distribution of samples output after normalization. (A) Un-normalized log2 data. (B) Subtract by group mean + semi-global normalization. (C) ComBat + semi-global normalization. (D) Global VSN + Combat normalization. (E) Global LOESS + Combat normalization.
Fig 7.
Evaluation of the effect of different normalization approaches on samples and variables in the data, shown in both sample mode and variable mode.
(A) Un-normalized log2 data. (B) Subtract by group mean + semi-global normalization. (C) ComBat + semi-global normalization. (D) Global VSN + Combat normalization. (E) Global LOESS + Combat normalization
Table 3.
Evaluation of different normalization methods.
To this end, microarray data was used to assess the No of statistically significant scFvs (ntotal = 195) when comparing day 1 vs. day 2, using a cut-off value of either q <0.05 or q <0.01 as well as q < 0.05 plus FC > 1.5.
Table 4.
Evaluation of different normalization processes.
To this end, microarray data for diseased vs. healthy samples was used and compared with respect to No. of down-regulated scFvs antibodies and No. of complete matches per target molecule, using a fold change (FC) filter of either FC > 1 or FC > 1.1 and with and without a cut-off value of q <0.05.
Fig 8.
Effect of the normalization on sample variability.
Un-normalized log2 data was compared with subtract by group mean normalized data as well as with subtract by group mean + Semi-global normalized data. The sample cohort included QCref samples as well as diseased and healthy samples. (A to C) QCref samples analyzed on two days. (D to F) Diseased and healthy samples analyzed on different days (mapped for day 1 vs. day 2). (G to I) Diseased vs. healthy samples analyzed on different days (mapped for diseased/healthy).
Table 5.
Evaluation of four feature selection methods for defining a condensed biomarker signature for classification of two groups, using a 3-fold cross validation scheme repeated twice.
The length of the biomarker signature was set to 25. A linear SVM classifier was used to assess the performance in all cases, except for RF (RF), where random forest was used both for the ranking and the final AUC calculation. AUC values are given together with standard deviations. The biomarker overlap was defined as the number of common biomarkers between two signatures.
Table 6.
Evaluation of four methods for defining a condensed biomarker signature for classification of two groups.
Once the biomarker signature of the optimal length was defined using the validation set, the classifier was re-trained, using both the training and the validation set and finally tested on the test set. Sample cohort 3 was used.