Participant eligibility criteria

All but one cohort had existing associated stored participant SF samples. Inclusion criteria were: i) evidence of a confirmed diagnosis of knee OA, or history of recent knee injury, ii) associated basic clinical information including (as a minimum) age at sampling, sex and indication of OA disease status, iii) a minimum volume of SF (90 μl, ideally 200 μl) and iv) SF had been centrifuged between 1800-3000g, prior to supernatant storage at -80°C. Exclusion criteria were: i) additional forms of arthritis e.g. gout, rheumatoid arthritis, psoriatic arthritis, as determined by host investigator; ii) confounding medical conditions e.g. concurrent infection, cancer; iii) confounding treatments e.g. index knee surgery in the preceding 6 months, index knee steroid injection in preceding 3 months; iv) chemotherapy and; v) significant deviation in storage procedure (e.g. freezer drop-out defined by host investigator).

Sample processing and SomaLogic assay

Consortium samples: 1746 SF samples were eligible and processed for STEpUP OA (S1 File). Archived samples were shipped to Oxford over a 20-month period (received: 27^th November 2019 to 26^th August 2021). No authors had access to participant identifiable information. A STEpUP participant ID number (PIN) and related unique sample identification number (SIN) were generated for each participant and their associated sample(s). Sample processing was performed in Oxford in four tranches over a 24-month period. For analysis by SomaLogic, SF enzymatic digestion, using hyaluronidase, was carried out. Briefly, sufficient bovine testicular hyaluronidase (4mg/ml; Sigma-Aldrich) for the entire project was reconstituted as a single batch from a single lot number and frozen in aliquots until use. For each tranche, batch processing was performed over consecutive working days, i.e. over as short a time as possible. Briefly, a batch of SFs was thawed, centrifuged at 3000g at 20°C for 25 minutes and 175 μl of SF supernatant diluted 1:2 with the same volume of hyaluronidase solution and agitated at room temperature for 1 hour, followed by further centrifugation for 5 minutes [19]. Supernatants were aliquoted and stored at -80°C and transferred on dry ice by temperature-controlled shipping to SomaLogic (single shipment per tranche).

Consortium controls/QC samples: Equal volumes of SF samples from 6 participants per group were used to generate single batches of hyaluronidase-treated ‘pooled samples’ for each of OA and knee injury at the start of project. Subaliquots of these then acted as internal QC controls, being run on each SomaLogic plate, enabling calculation of intra-assay and inter-assay coefficients of variation (CVs), as well as assessing effects of freeze thaw (a multiple freeze-thawed aliquot), frozen storage of hyaluronidase (an untreated aliquot freshly treated with frozen hyaluronidase during processing of each tranche) and centrifugation (an additional unspun pooled sample, from 6 unspun OA SF samples). A further 18 samples (split at the time of collection, with the paired aliquot remaining ‘unspun’) were included to further examine the effects of centrifugation. 42 other ‘comparator’ samples were included (disease-free controls from non-painful knees or from normal joints at amputation/post-mortem; samples from individuals with definite inflammatory arthritis). Three samples from three separate participants were re-processed under 3 different temperature conditions and re-analysed to examine the effects of laboratory re-processing. A subgroup of the freshly collected samples were processed specifically to test generalizability to OA SF which had not been centrifuged (‘unspun’) (n = 18). 235 unspun OA samples were subsequently included in the replication analysis.

Samples were assayed in three ‘dilution bins’ (to allow accurate quantification of high and low abundant proteins simultaneously), on the SomaLogic SomaScan Discovery Plex V4.1 by SomaLogic, in Boulder, US. All samples from all 4 tranches were processed as a single batch on twenty-two sequential 96-well plates in January 2022. All samples were randomised within and between plates whilst ensuring appropriate controls on each plate. Each plate included 83 participant SF single samples; one pooled OA sample; one pooled knee injury sample; five plasma calibrator samples; three plasma QC samples and three blanks per plate. The SomaScan platform quantified 7,596 synthetic DNA SOMAmers (for 7289 human targets) (Slow Off-rate Modified Aptamers [53, 54]) that bound to 6596 unique human proteins. The generated SomaScan protein quantification was securely transferred from SomaLogic to Oxford as.adat files.

Additional sample metadata, both from collecting sites (where available) and those generated in Oxford and at SomaLogic, included: sample blood staining (by visual staining defined by host investigator at time of collection); initial centrifugation of the sample; number of previous freeze thaws; the date of laboratory processing; batch and order of processing; the plate and position of the sample. These sample metadata were defined as technical confounders in our QC pipeline (S2 Table).

Clinical data

Pseudonymised associated participant clinical data were transferred from participating sites to Oxford, linked to their consortium PIN, mapped to variables where necessary and uploaded to a REDCap database (Research Electronic Data Capture, Vanderbilt University, US) [55], hosted by the University of Oxford. Data integrity and completeness were ensured using a data dictionary, data entry constraints and a combination of automated, systematic and random checks by two of the study team.

A consortium working group oversaw all aspects of data management including definition of variables and associated data dictionary, data harmonisation and design of the database (S1 Fig). Informed by their relative clinical importance, by data availability and by iterative review, a core clinical dataset (a subset of the data dictionary) was defined: the first phenotype release “Pheno 1” (demographic data and harmonised measures of radiographic disease severity) and the second release “Pheno 2” (dichotomous and continuous harmonised patient-reported outcome measures for knee pain [56]). Characteristics of cohorts and participant samples, including details of the radiographic scoring employed (Kellgren-Lawrence [57] grading where available), are described in S1 Table with core variable definitions also provided (S2 Table and S1 File).

Data QC approach

Methods to develop QC and data analysis pipelines prior to the primary discovery analysis were pre-defined in the Quality Assurance plan (https://www.kennedy.ox.ac.uk/oacentre/stepup-oa/stepup-oa). This QC pipeline aimed to validate methods for standardisation of the data, through a series of normalisation steps (given that SF was a non-standard matrix on SomaScan), correction for technical confounders and filtering based on pre-defined quality thresholds for SOMAmers, proteins and samples. The approach included pre-defined data exploration, though where issues were found, these were iteratively investigated and findings used to refine the QC pipeline. This approach was informed by our prior published work [19], SomaLogic expertise, initial consortium pilot work on 435 samples previously assessed on an earlier version (4.0) of the SomaScan platform, and subsequent QC work within this dataset.

The usual SomaScan analysis pipeline for plasma involves a series of standardisation procedures to reduce nuissance variance, using plasma calibrator and plasma QC samples included on plates to reduce the effect of technical factors across samples and plates [58]. This routine standardisation of the SomaScan relative fluorescence units (RFU), adjusted by the protein’s dilution factor used in the SomaScan assay (the “dilution bin”), was applied in a stepwise way, using i) hybridisation control normalisation (to remove well-to-well variation due to different rates of hybridisation between SOMAmers and fluorescence probes using spiked-in control SOMAmers); ii) plate scaling using plasma calibrators (to remove variation in overall intensity between plates); iii) median signal normalisation (to decrease variation due to total fluorescence intensity between samples); and iv) calibration (using plasma calibrator samples of known concentration to rescale each protein and reduce assay differences between runs). Each normalisation step was tested in a sequential manner.

After optimised standardisation, the R package limma was used to adjust proteins for a number of pre-specified continuous covariates, as described below. The batch correction method ComBat (in sva R package) was applied [59] to adjust the mean and variance of each protein for batch effects in datasets where the batch covariate, such as plate, is known [43] (again, as described below).

Based on our initial QC assessments, to quantify and adjust for differences in the contribution of intracellular proteins to the proteome, we defined an Intracellular Protein Score (IPS) for each sample i, as a weighted sum of log protein concentrations using the equation where d_p is the Cohen’s d for the difference in log concentration for protein p between paired spun and unspun samples, and C_ip is the log concentration of protein p in sample i.

Checks of assay performance and biological validity were carried out by measuring repeatability using the pooled samples on each plate, use of metadata for prespecified technical confounders, and comparison with previously generated quantitative immunoassay data (R&D or Meso Scale Discovery, available for 60 OA and injury SF samples (without hyaluronidase treatment) for 9 overlapping proteins (S3 Table).

Statistics and analysis

Principal Components Analysis (PCA) was used to visualise proteome-wide patterns of variation in the data, with further visualisation of Principal Components (PCs) with Uniform Manifold Approximation and Projection (UMAP) 2-dimensional plots (UMAP applied to the set of top PCs that explained >80% of total variation). Various other bioinformatic, descriptive and statistical techniques were employed to test the quality of the data. We checked:

• Inter-assay repeatability–the %CV of each protein, i.e. the ratio of the standard deviation of the concentration to the mean of the concentration within repeated samples, and the proportion of non-technical variation (R²) for each protein, estimated as one minus the square of the ratio of the variance in repeated samples to the variance in non-repeated SF samples.
• The effect of freeze-thawing on normalised RFU signal (measured by %CV between repeatedly freeze-thawed and non-freeze-thawed samples).
• The effect of centrifugation on normalised RFU signal (by estimates of correlation between, and differential abundance of, proteins in unspun and spun samples, using correlation tests and paired t-tests respectively).
• Assay accuracy–comparing SomaScan normalised RFU signal with existing quantitative immunoassay data (by Pearson correlation coefficients).
• Effects of each technical confounder on standardised RFU signal (from combined cohorts). Linear regression analyses were applied to identify the most significant principal components (PCs) and proteins associated with technical confounders (Bonferonni adjusted p<0.05).

In addition, SOMAmers and samples of insufficient quality were removed as follows:

SOMAmer filtering:

• SOMAmers which were highly associated with pre-specified confounders (Bonferroni adjusted p<0.05) were removed (S4 Table).
• SOMAmers from non-human organisms or control SOMAmers (including Spuriomer, hybridisation control elution, deprecated, non-biotin, and non-cleavable) were excluded.
• The estimated proportion of non-technical (i.e. biological) variation R² was calculated using pooled SF samples (OA, knee injury), defined as (1 –V_Pooled/V_total)², where V_Pooled and V_total are the variances (V) in pooled and non-pooled (including all individual) samples respectively. If R² for a given SOMAmer accounted for less than 50% of total variation in either OA or knee injury it was removed.

Sample filtering:

• If a sample had more than 25% of protein values above or below the upper or lower limits of detection respectively, the sample was removed.
• (This was applying lower and upper limits of detection (LOD) (defined by SomaLogic), where lower LOD was: [median concentration of blanks] + 4.9 x [median absolute deviation of blanks], based on three blanks (i.e. buffer only) per plate. Upper LOD was defined as 80,000 RFU).
• Identification of outliers by PCA: samples that were beyond 5 standard deviations (SDs) from the center of the principal component space (made up of PCs that explained at least 80% of variation) were removed.
• Identification of total signal intensity outliers: samples that were beyond 5 SDs from the mean in total RFU distribution were removed. Total RFU of per sample was defined as the sum of all the RFU values of that sample.
• Samples that were flagged by SomaLogic’s in-house QC process were removed.

Pearson correlation coefficients (with 95% confidence intervals) are given for correlations and all analyses were carried out in R (version 4.3.1), unless otherwise stated. For additional details, see the data analysis plan (listed in Appendix 1): https://www.kennedy.ox.ac.uk/oacentre/stepup-oa/stepup-oa.

Selection of each normalisation procedure

To define the optimal QC pipeline, selection of appropriate standardisation procedures was needed. These had been previously optimised by SomaLogic for plasma samples. We set out to test how well these procedures performed in all 1746 SF samples with 7596 SOMAmer features.

We measured the impact of technical variation in each normalisation step (see methods) on assay performance in a sequential manner, firstly by measuring effects on the mean %CV and the mean R² across all proteins for the pooled sample replicates across all plates (S2A and S2B Fig). Comparing each to the raw RFU data, %CVs ranged from 11.65% to 16.49% for pooled OA samples (n = 22) and 10.56% to 16.63% for pooled injury samples (n = 22). Mean R² ranged from 77.70% to 88.98% and 82.59% to 89.10% for pooled OA and injury samples respectively, suggesting a high level of repeatability. A decreasing trend in mean %CVs suggested that the routine normalisation steps improved measure repeatability with the exception of median normalisation (S2A and S2B Fig). Removing median normalisation from the standardisation procedure resulted in a mean %CV of 13.6% and a mean R² of 88.98% for pooled OA samples and %CV of 14.34% and a mean R² of 89.1% for pooled injury samples (S2A and S2B Fig).

Correlation of the SomaScan assay data with nine selected analytes measured in the same SF sample by immunoassay was generally high for most analytes (S2C and S2D Fig). However, upon normalisation steps, we saw a similar effect, with median signal normalisation reducing the correlation with these validation measurements, particularly in the injury samples (S2D Fig). Based on these combined data, we chose to use SomaLogic’s existing standardisation procedures, though omitting median normalisation from our SF standardisation pipeline (i.e. employing hybridisation normalisation, plate scaling, and plate calibration using SomaLogic’s plasma calibrators).

Identification and correction of confounding factors

After conducting standardisation, principal component (PC) analysis identified one dominant component (PC1) that explained 48% of variation in the data (Fig 1A). This principal component was positively correlated with almost all proteins measured, with a median correlation coefficient of 0.70, with the highest correlations seen with low abundance proteins (Fig 1B). We were able to rule out a total protein effect, due to a low correlation (-0.039) with standard high-abundance markers such as albumin concentration. Examining the proteins that drove this signal, we found that lower protein abundance was the strongest independent predictor of correlation (S5 Table) with PC1 (p< 2.23e-308), with the next most significant predictors being whether proteins were predicted not to be secreted (p = 3.98e-10) and proteins that were identified as nuclear and not secreted (p = 1.64e-9). This led to the hypothesis that PC1 was capturing an effect of intracellular proteins, perhaps reflecting cell turnover or due to the presence of microvesicles. A strong intracellular signal was confirmed by showing that PC1 was consistently reduced in spun samples when comparing paired SF samples (from the same parent SF, n = 18) that had been split into two and either spun or left unspun immediately after joint aspiration (Fig 1C).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1.

(A) Variation explained (%) by the top 10 PCs derived from the standardised log abundance proteomic data. (B) Correlation between PC1 and protein abundance, with two high-abundance proteins (albumin, a soluble serum protein, and LDH, an intracellular protein) marked. Protein abundance is calculated as the standardised RFU for each protein adjusted by the protein’s dilution factor used in the SomaScan assay (the "dilution bin"). (C) Comparison of variation explained (%) by PC1 between 18 pairs of SF samples that were centrifuged (spun) or not (unspun) after aspiration and prior to freezing, with paired samples from the same participant joined by separate lines. Red lines show samples that had an increased PC1 prior to spinning, and the green line where it was decreased. Correlation between PC1 and intracellular protein score (D) before and (E) after IPS adjustment. (F) Variation explained by the top 10 PCs derived from the batch corrected and IPS adjusted log abundance proteomic data. In all cases, correlation is measured using the Pearson correlation coefficient. IPS, Intracellular Protein Score; PC, principal component; LDH, Lactate dehydrogenase.

https://doi.org/10.1371/journal.pone.0309677.g001

To quantify the contribution of intracellular proteins, we derived an Intracellular Protein Score (IPS) as the weighted sum of relative protein concentrations. For weights, we calculated a Cohen’s d from the 18 paired spun and unspun samples. This score correlated very highly with PC1 (Fig 1D). We used this score in a linear regression model adjusting for the contribution of intracellular proteins, which removed the correlation between IPS and PC1 (Fig 1E), reduced the variance explained by PC1 to 16% (Fig 1F), and removed the correlation of “non-secreted nuclear protein” with PC1 (S5 Table).

This intracellular contribution to the SF proteome did not correlate strongly with any of our pre-defined technical confounders. However, it explained a large proportion of variation in our data, which ran the risk of swamping more subtle protein signatures or molecular endotypes, if present. We thus decided to include, in addition to the standardised data set without IPS adjustment, a co-primary dataset, applying IPS adjustment to each protein as part of our STEpUP OA Data Analysis Plan (https://www.kennedy.ox.ac.uk/oacentre/stepup-oa/stepup-oa).

We also found a strong ‘bimodal’ signal on PC2 of the data (Fig 2A and 2B) whereby a large number of SOMAmers (N = 4030 at Benjamini–Hochberg (BH) adjusted p<0.05) were present at either very low or very high relative signal in a given sample. Further investigation showed that PC2 was highly correlated with the technical variable ‘laboratory processing batch’ (p<2.2E-308). Investigation of exemplar proteins displaying this behaviour showed that the bimodal signal followed sample processing order, usually (but not always) between laboratory processing batches (Fig 2C). The effect became stronger over time (Fig 2C). Re-analysing (at SomaLogic) previously laboratory processed (hyaluronidase treated) samples gave the same result. However, when new aliquots of three sequential samples which had differing bimodal status were reprocessed by the Oxford laboratory and re-analysed, all three reprocessed results had a shared bimodal status. This indicated that this was due, in some way, to our laboratory sample processing (hyaluronidase treatment) (Fig 2D). We hypothesised this might be due to sample temperature differences prior to hyaluronidase treatment, but further experiments did not corroborate this. Neither was this thought to be due to the stability of frozen hyaluronidase enzyme as the pooled OA sample, freshly processed during each tranche with frozen stored hyaluronidase, showed little variability over time (S3 Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2.

(A) Distribution of the second principal component (PC2) derived from the standardised log abundance data, showing a bimodal distribution. (B) UMAP visualisation of two reduced dimensions (D1 and D2) of the top PCs of the standardised log abundance data. (C) Example of a strongly bimodal protein measurement, TSG101, RFU (y-axis) against Oxford laboratory processing order (x-axis) and coloured by laboratory processing batch (with only points within the same processing batch connected by lines). Note that the ‘flipping’ between high and low signal status occurred primarily when processing batch changed, and only rarely within processing batch. This effect was particularly strong among sample batches that were processed later in processing order. (D) The same example protein measurement for three independent SF samples before (original) and after they were re-processed and re-assayed, showing that bimodal status changed after laboratory re-processing. (E) Distribution of PC2 derived from standardised log abundance data, showing the two probability density functions of the Gaussian Mixture Model used to classify samples into the two bimodal signal status groups. (F) UMAP visualisation of two reduced dimensions (D1 and D2) of the top PCs of the standardised log abundance data, colored by the inferred bimodal signal status. (G) Histogram of PC2 of the batch corrected log abundance data, with the now near-identical distributions of the two bimodal signal status groups shown as colored lines, (H) UMAP visualisation on two reduced dimensions (D1 and D2) of the top PCs of the batch corrected log abundance data, colored by the inferred bimodal signal status. RFUs, relative fluorescence units; PC, Principal Component; TSG101, Tumor susceptibility gene 101 protein; UMAP, Uniform Manifold Approximation and Projection.

https://doi.org/10.1371/journal.pone.0309677.g002

We applied a Gaussian Mixture Model to PC2 to classify samples into high or low protein status, reflecting their bimodal signal (Fig 2E), which produced visually plausible assignments on PCA and UMAP (Fig 2E and 2F). To attempt to reduce this undesired variance, we carried out batch correction by samples’ PC2 bimodal signal status using the ComBat method [59].This correction reduced the impact of the bimodal signal considerably (Fig 2G and 2H) and was adopted into our QC pipeline.

We also discovered a significant influence of plate on a number of proteins (n = 1927) at BH adjusted p<0.05). Our samples were randomised to plate, so this was unlikely to cause significant confounding in downstream analyses, but to reduce technical variation we also applied batch correction for plate by ComBat at the same time as correcting for the bimodal signal.

We assessed the impact of these adjustments described above using the immunoassay comparison data. While the IPS-adjustment reduced the dominance of PC1, it also had a negative impact on the correlation coefficients between SomaScan and prior immunoassay results of the select analytes (S4 Fig). This was particularly evident in the injury group, where the correlation between the two measures for 4 out of 9 proteins (IL6, IL8, TGFß1, TIMP1) changed from strongly correlated to weakly or not correlated (S4B Fig). Interestingly, some of the measured cytokines (which had been selected because of their putative disease relevance) such as MCP1, IL-8 and TGFß1, correlated with the intracellular protein score (S6 Table). Batch correction for plate and bimodal signal status (as part of our optimised standardisation) was seen to have minimal impact on immunoassay agreement (S4 Fig).

Description and reduction of pre-defined technical confounding by protein filtering

In addition to identifying badly performing SOMAmers and samples for filtering according to assay performance (see methods), we also identified filters based on pre-defined technical confounders (S2 Table).

Confounding technical factors were dealt with in different ways. Samples showed systematic biases in signal intensity plate position (S7 Table, ‘Plate position’), but this was also deemed unlikely to confound downstream analyses because samples were randomised across and within plates, so its effect was not adjusted for. Blood staining (S7 Table, ‘Visual blood staining’) and sample volume were both drivers of IPS, but we felt that both could contain biological signals of relevance, so they were not adjusted during QC but were considered covariates in the downstream analyses. Sample age (S7 Table, ‘Sample age’) could introduce technical variation, therefore significantly associated proteins were removed by filtering (S4 Table). Freeze-thawing was also shown to be a potential technical confounder (S7 Table, ‘Sample freeze thaw cycles’). This was investigated further.

For freeze-thawing, we had repeatedly freeze-thawed (five times per sample) one aliquot each of the pooled OA and the pooled injury samples. This had an effect, particularly in the injury samples. However, the majority (77%) of proteins retained a good %CV (<20%) even after five freeze-thaws (S3 Fig), suggesting that such samples remained usable. Technical variation brought about by freeze-thaw was nonetheless adjusted for by filtering out significantly associated proteins following Bonferroni correction (S4 Table).

Assessment of centrifugation effect on protein measurements

Although most of the samples had been centrifuged prior to initial storage as per our eligibility criteria, 240 samples included in the replication analysis were unspun. In anticipation of this we assessed further the impact of centrifugation on the 18 pairs of samples that had either been spun or left unspun at time of collection. We compared the SomaScan data to identify proteins that changed upon centrifugation (S5 Fig). The effect of centrifugation on the data depended on whether the data were adjusted for IPS. The unadjusted data showed that centrifugation status was a major driver of variation across the paired samples, with a significant correlation with PC1 (S5A Fig, paired t-test p = 0.0066), but after adjusting for IPS, the top PCs were no longer driven by spun status (S5B Fig, paired t-test p = 0.2089). Centrifugation impacted the concentration of a large number of individual proteins in both IPS-unadjusted (n = 5638, 74%, at BH adjusted p<0.05) and, to a lesser extent, IPS-adjusted data (n = 3731, 49%, at BH adjusted p<0.05), although the majority of proteins were significantly correlated between the paired spun and unspun samples (n = 6402, 85% in unadjusted data and n = 4558, 60% in IPS adjusted data, S5C and S5D Fig respectively).

We concluded that spun and unspun samples were comparable (in that they captured similar information), but that any analysis that included both types together would need to adjust for systematic shifts in abundance and the small numbers of uncorrelated proteins. In our discovery and replication analysis plans relating to our primary analysis, only spun samples are therefore considered, with unspun samples used for secondary sensitivity analyses.

Effect of blood staining on protein measurements

A subset of samples had information on blood staining, graded by visual inspection at the time of joint aspiration, prior to centrifugation. As shown in S7 Table, the presence of blood measured in this way was a significant driver of protein variation. It was also a potential biological driver as haemarthrosis is common after significant joint injury and is known to be pro-inflammatory and associated with persisting knee symptoms [60–62]. Visual blood staining could reflect presence of either intact or lysed red blood cells. The analyte haemoglobin A (HBA) correlated reasonably well with visual blood staining grade prior to adjustment for IPS (S6A Fig), and less so after adjustment for IPS (S6B Fig). The log concentration of HBA relative abundance level (without IPS adjustment) was subsequently used as a measure of blood content, as a covariate in downstream analyses.

Validation of data quality after QC

Following application of filters, 1720 samples and 6290 SOMAmers (features) remained. The total numbers of samples and proteins filtered out are shown in S4 Table. After filtering, median %CV of pooled OA and injury samples remained relatively unchanged at 11.25 and 12.42 respectively (S7 Fig). An overview of the end-to-end data processing and quality control pipeline, from raw data to final filtered data, is shown in Fig 3.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Overview of the final data processing and quality control pipeline for synovial fluid SOMAscan data used by the STEpUP OA consortium, broken down into three stages: Standardisation (yellow box), technical confounder correction (blue box) and filtering (green box).

More details on filtering thresholds, and the number removed by each filter, can be found in S4 Table.

https://doi.org/10.1371/journal.pone.0309677.g003

The association of all these variables with the top 10 PCs of the standardised, bimodal signal corrected data after filtering is shown in S7 Table. The strongest associations in the IPS adjusted, filtered data are shown in Fig 4. IPS adjusted non-filtered, and non-IPS adjusted data are shown in S8 Fig.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Visualisation of selected predefined confounders against select principal components of the batch corrected, filtered, IPS adjusted data.

(A) The average value of PC9 (most strongly associated with plate position) by sample well position, (B-F) visualisation of the two PCs most strongly associated with each confounder, coloured by confounder value. Pre-defined confounders shown are (B) blood staining grade of sample after aspiration assessed by visual inspection, (C) volume of sample taken during aspiration, (D) age of the sample in years, measured from aspiration to sample processing at Oxford, (E) the number of times the sample was thawed and re-frozen before sample processing at Oxford, (F) the disease group of the sample (osteoarthritis [OA], acute knee injury [Injury], healthy control, inflammatory arthritis control).

https://doi.org/10.1371/journal.pone.0309677.g004

Finally, we visualised the different diagnostic subgroups (OA, joint injury, inflammatory control, disease-free control) on UMAPs of the standardised, corrected and filtered data, with and without IPS adjustment (Fig 5A and 5B respectively). Both datasets showed clustering of knee injury and OA cases in overlapping but distinguishable areas of high-dimensional proteomic space, though the smaller groups (disease-free controls and inflammatory controls) were more evenly distributed. Inflammatory controls tended to segregate with acute knee injury samples. These patterns were also reflected at the PC level (Fig 4F and S9 Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5.

UMAP visualisation of two reduced dimensions (D1 and D2) of the top PCs of the log abundance data with (A) and without (B) IPS adjustment followed by filtering, coloured by disease group. These groups were osteoarthritis (OA, acute knee injury (injury), healthy controls, inflammatory arthritis controls. UMAP, Uniform Manifold Approximation and Projection.

https://doi.org/10.1371/journal.pone.0309677.g005

Patient and public involvement statement

People with lived experience of osteoarthritis have been involved in the design of this project. A patient research panel was involved in discussing and inputting on the STEpUP OA project in February 2020 (invited to the Centre for Osteoarthritis Pathogenesis Versus Arthritis in Oxford, as part of its involvement activities). Aspects relevant to the development of the project were further discussed with the panel in July 2022. The working groups for the consortium include one focused on patient involvement and engagement. A lay summary is included in the appendix of our publicly available analysis plan. A short video about the project was produced and is available on our website: https://www.kennedy.ox.ac.uk/oacentre/stepup-oa/stepup-oa. In addition, the various constituent cohorts contributing to STEpUP OA also typically have lay or patient members on their steering committees.