Assessment of Variability in the Plasma 7k SomaScan Proteomics Assay

Key results

1. 01

   ShownNormalization that ended with median signal normalization of every sample type, applied after calibrator-based inter-plate correction, gave the lowest discordance between technical duplicates on every duplicate metric and on grid-search CV estimates, compared with the less processed levels.

   Fully normalized data, median ± MAD across 7,288 human-protein SOMAmers from 102 inter-plate duplicate pairs: RMSV 5.42 ± 2.13%, MADV 4.79 ± 1.80%, PV 4.20 ± 1.56%, in the order theory predicts. In PCA without buffer wells, PC1 variance after normalization was less than half that of raw data, and PC2 separated experimental from control samples. · Results; Discussion; Figure legends

2. 02

   ShownPer-SOMAmer coefficient of variation (CV) was estimated by a grid search that fits log-normal fold-change models to the duplicate data using one-sample Kolmogorov-Smirnov tests. CV was low overall, differed by dilution group and correlated with signal-to-background ratio (SBR).

   CV median ± MAD: 4.5 ± 1.5% overall; 4.5 ± 1.5% (20% dilution, n = 6001), 4.0 ± 1.5% (0.5%, n = 1100), 6.5 ± 2.2% (0.005%, n = 187); pairwise Wilcoxon adjusted p < 0.05. CV vs SBR: Spearman r = 0.35 (p = 4 × 10^−210), driven mainly by the 20% group (r = 0.43); no significant correlation in the 0.005% group. · Results; Discussion; Figure legends

3. 03

   ShownA five-step normalization built only on references derived from the data themselves matched SomaLogic's external-reference output. The steps are hybridization, calibrator median-signal, plate-scale, inter-plate calibration and all-sample median-signal normalization.

   For every human-protein SOMAmer, the Spearman RFU correlation across experimental samples between fully normalized internal-reference and external-reference data was r ≥ 0.95 (p ≈ 0). Because Spearman correlation is rank-based, this holds under monotonic transforms such as log. · Data Normalization Procedures; 5. Median signal normalization on all sample…; Discussion

4. 04

   ShownThe 7k assay was highly sensitive compared with buffer-only wells. Nearly all human-protein SOMAmers sat well above background, and very few fell below an estimated limit of detection (eLoD) derived from buffer wells.

   99.1% of human-protein SOMAmers were >2-fold brighter in experimental samples than in buffer wells; 68.8% were 5-fold and 43.8% were 10-fold or more brighter. Median SBR was 8.1 ± 7.2 (median ± MAD). Only seven SOMAmers (e.g., Caspase-14, DC-SIGNR) had median experimental RFU below eLoD. · Results; Discussion

5. 05

   SuggestedNegative controls expected at buffer level, chiefly non-biotin SOMAmers and spuriomers, gave excess signal in a significant fraction of plasma samples. The authors infer some low-range RFUs may be spurious and that the vendor's 10-order dynamic range claim is open to further scrutiny.

   Hybridization controls (n = 12) matched buffer, as expected. All non-cleavable controls (n = 4) and most spuriomers (n = 20) had SBR < 2, some spuriomers reached SBR 2-3, and non-biotin controls (n = 10) were higher still. Measured range was 1.2 × 10^9-fold over 1,799 samples; if low RFUs are unreliable, the authors argue the effective range would shrink by a further 1-2 orders of magnitude. · Results; Discussion; Figure legends