Assay Analytics

 

Delivers Higher Assay Precision and Better Data

  • Artificial intelligence software that learns to quantify assay result differences by individual well


  • The process identifies repeating patterns. Each well’s Gaussian Curve is used to calibrate results and measure precision.


In this experiment, the only variable is the brand of microplate. In theory, all the wells should report identical results.

There appears to be significant variation in cell growth between brands of microplates.

This variation exists between microplates of the same polymer and between types of polymers. Two plates are polystyrene; one is cyclic olefin copolymer.

Luminescence is a relative measurement. The plates were not run with standard curves nor were background measurements taken. In other runs with the same brands, cytotoxic-treated wells reported statistically significant differences in background.

All three brands were run as single batch to achieve as much process uniformity as possible.

Surface Chemical Differences of Microplates – Analysis by ToF-SIMS

Brand 3 - Green Lines 5, 6
Brand 2 - Red Lines 3, 4
Brand 1 - Blue Lines 1, 2
(Brand 3 had C3H8N/C3H7O/C4H8N/C8H10N/C8H20N >20X Brands 1+2)
Each brand has a unique blend of surface contaminants that source from polymer production and molding.

Variation in the MS results show that every spot analyzed is exposing different concentrations of contaminants, i.e. surfaces are inconsistent.

This concentration inconsistency effects each brand of plate’s statistical performance; small growth rate differences in wells lead to increasing assay SD.

Brand 1 Well Value Drift Measured in 10ths SD
(Plate 1:A1…)-(Plate 2:A1…)/(Standard Deviation)

MultiBrand Well Value Driftmeasured in 10ths SD
(B1, Plate 1:A1…)-(B3, Plate 2:A1…)/(Standard Deviation)

Conclusions from Above 2 Graphs


1. Comparing the two Brand 1 plates individual wells by position, 74% of the well pairs reported results within 1/2 SD. An SD value of 1.15 is normally associated with 75% confidence limits.

2. That would indicate a 57% reduction in SD by comparing by well position versus using whole plate statistics.

Three Brand Well Value Drift Grouped by .5 SD
(Plate 1:A1…)-(Plate 2:A1…)/(Standard Deviation)

Conclusions from Above Graph


1. This graph summarizes the comparison of well position between all three brands.
2. The data show the effects of consistency of plate performance.

Homogeneous Assay Well Value Drift, grouped by .5 SD
(Plate 1:A1…)-(Plate 2:A1…)/(Standard Deviation)

Conclusions from Above Graph


1. This graph summarizes the comparison of well position between a reference and 9 others from the same lot.
2. The reference and #6 correlate well. The other 8 correlate well. All 10 don’t correlate well as a group.

Learning Algorithm Results

How it Works


1. Plates are initially analyzed. The process measures plate to plate differences that occur due to the assay, plate uniformity and the analytical measurement technology being used.
2. Results are statistically evaluated by individual well position over several runs. Multidimensional matrices (technically tensors) are generated that include the adjusted average for each well and approximations of Gaussian statistics for each well.
3. Each well can be mulitplied by a linear factor or adjusted by a non-linear transformation process if the assay readout is not linear or another process that aligns results in a uniform and consistent manner. In the graphic examples, the homogeneous assay on the right was adjusted with a linear factor. The cell assay on the left was transformed by a process using center points of the three highest peaks of each graph for alignment.

Conclusions from Above Graphs


1. This cell assay showed a SD reduction of 60% using adjusted Brand 1 data from the Learning Algorithm. The SD reduction is high due to the large number of wells needing parallax correction.
2. Learning examples for Brand 1 were N=2. A larger number of learning examples is expected to improve the analytics process.
3. Results for other assays should be anticipated to be different. The reduction in the SD for the Homogeneous Assay was measured at about 30% with N=4 learning examples.