Proteomic models to predict pre-analytical variation

Background

In biomarker discovery, it is critical to assess any pre-analytical variation (PAV) in order to avoid artificial bias in the intended measurements. PAV may arise from both avoidable and unavoidable factors, resulting in misleading data and incorrect conclusions. Proteins, in particular, are vulnerable to variation in collection methods, storage temperatures, and processing protocols. It is vitally important to understand this PAV when analyzing samples using protein assays.

Methods

Human EDTA plasma and serum samples, subjected to standardized sample processing methods, with distinct excursions from ideal collection, were assayed on the SomaScan® Platform measuring ~7,000 analytes. Using machine-learning methods, these quantitative protein measurements were compared to sample processing truth standards (eg, time-to-spin) to create predictive models. These models, termed SomaSignal^® tests (SSTs), were developed to enable the assessment of PAV related to processing methods.

Results

SomaSignal tests (SSTs) have been developed to predict time-to-spin, time-to-decant and time-to-freeze, reported in the number of hours, for both plasma and serum. Models that predict the number for freeze/thaws a sample has been subjected to, have also been developed. All eight models had Lin’s CCC and R² values greater than 0.90 in hold-out validation datasets. In addition to these sample handling predictions, effect size calculations for all ~7,000 measurements have been determined for multiple time points, or freeze-thaw cycles, for each model.

Conclusions

SomaLogic has developed a unique class of PAV models that are able to assess variation related to processing methods. Results from these predictions can be used during biomarker evaluations to exclude samples due to apparent excessive delay in processing, identify collection site bias for current and future analysis and identify sample groupings that may impact analysis. Further, knowing the effect size metrics for all measurements could also enable the removal of specific analytes from modeling and/or be used as covariates in model development.

Authors

David Astling
Dan Drolet
Joe Gogain
Yolanda Hagar
Laura Sampson
Kaitlin Soucie
Kinsey Trinder
Ira von Carlowitz
Matthew Westacott

SomaLogic Operating Co., Inc., Boulder, CO USA