Targeted Proteomics Data

We have an extensive quality control procedure that we follow when running an assay. This allows for full visibility and control over the technical performance of the assay at each step and ensures that reliable data are generated with customers' samples. 

Internal controls
Three internal controls are added to each sample to monitor the quality of assay performance, as well as the quality of individual samples: 

• An Incubation Control, which is a PEA assay for spiked-in non-human proteins.
• An Extension Control, which is an antibody molecule with both DNA-oligos attached and therefore always in proximity.
• An Amplification Control, which is a synthetic double-stranded amplicon. 
• The extension control is used to calculate the NPX, and the other two are used for quality control of the assay.

External control
We also include external control samples on each plate to normalize data and also to monitor the assay’s performance:

• A Sample Control of pooled plasma to estimate intra- and inter-assay precision. 
• A Negative Control to measure background levels, which is used to estimate LOD for all assays.
• A Plate Control of pooled plasma to calculate the NPX and normalize signal levels between different plates for each assay.

The Explore system´s raw data output are NGS counts, where each combination of an assay and sample is given an integer value based on the number of DNA copies detected. These raw data counts are converted into NPX values for use in downstream statistical analysis. 

NPX generation and Normalization
The NPX values are calculated in two main steps and then intensity normalized by performing between-plate-normalization. First, the assay counts of a sample are divided by those of the extension control for that sample and block, which then undergoes log2 transformation to normalize the data:

Steps in the NPX generation described in equation form, where “i” refers to a specific assay, j refers to a sample, and ExtNPX defines an extension normalized NPX value.
 

• Relate counts to a known standard (Extension control)
• For all assays and all samples, including negative controls, Plate Controls, and Sample controls.
• Log₂ transformation gives more normally distributed data

The result is a scale that has increasing sample values against increasing protein concentration for each assay. The median of the Plate Control is then subtracted from the normalized data:

• Perform plate standardization
• For all assays and per plate of samples

Intensity normalization of data is the default for randomized studies on multiple sample plates. In this setting, the median for a random selection of samples is more stable across plates than the three PC’s on each plate. Intensity normalization sets the median level of all assays to the same value for all plates:

• Between plate normalization (Optional but the default for multi-plate projects)
• For each assay, for all plates in the project.

Figure 1: Summary of the three steps involved in NPX generation and data normalization. Step 1 is illustrated by the red graph, step 2 by the yellow graph, and step 3 (data normalization) by the turquoise graph. GHRL refers to appetite-regulating hormone (Uniprot ID: Q9UBU3). 

Sample QC
Each sample in a block is given the status pass or warning. For a complete Explore 1536 run, this means that each sample is evaluated 16 times. There are three sample QC criteria per block. Exceeding specifications in any one of them will result in a QC warning status for that sample. The three criteria are listed below:

1. Incubation control deviation from the median may not exceed 0.3 NPX.
2. Amplification control deviation from the median may not exceed 0.3 NPX.
3. Average counts for a sample may not fall below 500 counts. 

Data from samples with a warning should be treated with caution.
 

Run QC 
In addition to specific samples that exceed specifications, the entire block can be considered failed if it meets criteria 1 and 2 below. For practical reasons, this will most likely result in the re-run of a panel. If criteria 3 below is met the panel is failed:

1. The number of samples with QC warning exceeds 16/88 or 1/6. 
2. The MAD for incubation or amplification control across all samples exceeds 0.3.
3. The Plate Control and Negative Control exceeds the criteria specified in section Control strip QC.

Control strip QC
The Plate Controls and negative controls have specific QC criteria. The median of the triplicates may not deviate more than 3 standard deviations from predefined tolerance levels for more than 10% of the assays on average for a panel. For the negative control, only positive deviations are considered. 

LOD is defined as being three standard deviations above the median NPX of negative controls. The median is set using all samples annotated as negative controls per plate. A predefined standard deviation is used (fixSD). Detectability is calculated per assay and plate and is defined by the percentage of samples above the LOD threshold. The overall detectability of the project is generated and reported in the CoA:

The CV is calculated per assay (i) using the assumption of a log-normal distribution. The average CV is then calculated across panels and included in the CoA output. 

Inclusion of Bridging Samples:

The final goal was to perform a joint analysis of datasets D01 and D02, and it was recommended to use common samples in the two batches that could help with batch correction to remove technical variation between the runs. Olink recommends the inclusion of 16-24 of these “bridging samples” across runs and the protein levels measured for these samples can be used as a common reference between batches, used to normalize and alleviate any potential bias. In this experiment, 39 samples from the D01 batch were included as bridging samples with the D02 dataset. These samples were selected using the dplyr R package to find the intersecting samples between D01 and D02 datasets. This list of overlapping samples was used as part of the olink_normalization_bridge_function from the Olink Analyze R package. 

Bridging Data and QC

The bridging procedure involved the calculation of the median of the paired NPX differences per assay between the bridging samples. This is the assay specific adjustment factor which is used to adjust NPX values between all remaining samples in the two datasets. In this process, the D01 dataset was considered the reference dataset and its NPX values remained unaltered. The D02 dataset was adjusted to the reference dataset based on the adjustment factors. 

In some cases, the bridging led to the duplication of the same sample x uniprot with two separate NPX values. In addition, there are cases in which there is a duplicate uniprot ID due to the same uniprot ID being in more than one Olink panel or there being two different Olink assays that map to that uniprot ID. In all of these cases, the mean NPX value was calculated and used in the final data product, resulting in one NPX value for every uniprot ID for every sample ID. 

The bridging and mean calculations were all conducted using R and were performed separately for cerebrospinal fluid data sets and the plasma data sets. Results of the bridging can be seen below through PCA. Figure 1a shows dataset D01 and D02 prior to bridging. Separation along PC1 shows datasets D01 and D02 as distinct, with 37.91% of the variation due to batch effects. Figure 1b shows that after performing bridging, D01 and D02 no longer show separation along PC1, indicating that bridging was successful and batch effects were removed.