A computational solution to improve biomarker reproducibility during long-term projects

Abstract

Biomarkers are fundamental to basic and clinical research outcomes by reporting host responses and providing insight into disease pathophysiology. Measuring biomarkers with research-use ELISA kits is universal, yet lack of kit standardization and unexpected lot-to-lot variability presents analytic challenges for long-term projects. During an ongoing two-year project measuring plasma biomarkers in cancer patients, control concentrations for one biomarker (PF) decreased significantly after changes in ELISA kit lots. A comprehensive operations review pointed to standard curve shifts with the new kits, an analytic variable that jeopardized data already collected on hundreds of patient samples. After excluding other reasonable contributors to data variability, a computational solution was developed to provide a uniform platform for data analysis across multiple ELISA kit lots. The solution (ELISAtools) was developed within open-access R software in which variability between kits is treated as a batch effect. A defined best-fit Reference standard curve is modelled, a unique Shift factor "S" is calculated for every standard curve and data adjusted accordingly. The averaged S factors for PF ELISA kit lots #1-5 ranged from -0.086 to 0.735, and reduced control inter-assay variability from 62.4% to <9%, within quality control limits. S factors calculated for four other biomarkers provided a quantitative metric to monitor ELISAs over the 10 month study period for quality control purposes. Reproducible biomarker measurements are essential, particularly for long-term projects with valuable patient samples. Use of research-use ELISA kits is ubiquitous and judicious use of this computational solution maximizes biomarker reproducibility.

Fig 1. Biomarker controls with time.

ELISAs were completed over ~10 months as described in Methods for biomarkers myeloperoxidase and PF. (A,B) Optical density (O.D. at 450nm) readings and biomarker concentrations calculated from each plate’s standard curve (C,D) for the BMC internal control samples are shown versus time. Two ELISA kit lots were used for myeloperoxidase, and five kit lots for PF. Two BMC control preparations were used for PF (C1 and C2), and one BMC control preparation for myeloperoxidase. OD readings for PF controls are reasonably constant with time, but unlike myeloperoxidase, PF concentrations decreased by 62.4% over the study period.

Fig 2. Lot-to-lot variability in PF ELISA.

(A) The averaged standard curves for each lot of PF ELISA as the mean optical density O.D. ± S.D. at each standard concentration. N = 28, 19, 8, 4, 9 curves for lots #1–5, respectively. (B) Variability of standard curve optical density at 450nm within and between PF ELISA kit lots #1 and #5.

Fig 3. Reference curve for calculation of S factor.

The best-fit, 4-parameter logistic (4pl) Reference curve (closed circle, R, solid line) from twenty-eight PF ELISA kit lot #1 standard curves is shown as derived from the indicated equation and D = -0.01; A = 3.20; C = 1300.00; B = -1.30. The averaged standard curves for kit lot #2 has an S factor = 0.0690 (open square; n = 19 curves) and S = 0.6994 (open triangle; n = 9) for curves from kit lot #5.

Fig 4. Calculated shift factors for PF ELISA standard curves.

A 4pl Reference curve was derived from PF kit lot #1 standard curves data and a Shift factor “S” was calculated to quantify the difference between each plate’s standard curve and the best-fit reference curve. The S factors for each curve in lots #1–5 is shown with the mean (horizontal line) for each lot. S factors for lots #1 and #2 curves are not different, but those for lots #3–5 differ significantly from lot #1 (***P<0.0001, ANOVA with Bonferroni post-test).

Fig 5. BMC control PF concentrations.

A BMC control sample from preparation C1 or C2 was included on each ELISA plate for the 10 month study period and assayed with kit lots #1–5. C1 was used primarily on plates from kit lots #1 and #2. C2 was used primarily on plates from kit lots #3–5. Standard curves from each plate were adjusted according to their calculated S factor and the control PF concentrations were re-calculated. The data shows the PF concentrations before and after correction with the S factors. ***P<0.0001; ns, not significant (Student’s t test).