1. Definitive Screening Design on IDBV
Team: E. Romeu, M. Karagueuzian
March 24, 2015
Discovery Summit JMP

2. Agenda Background Danaher and Beckman-Coulter quick overview
Introduction
Objective
Format of the assay
Results
Comparison of lots A and B
Comparison of Design of Experiments  selection of DSD
DOE analysis on lots A and B: comparison of models
Confirmatory test on lot A
Conclusion

3. Background Danaher is a group of 30 OpCos (operating companies)
associates
17 billions dollars revenue
Beckman-Coulter is part of the Life Science and Diagnostic branch of Danaher representing
2 OpCos (operating companies)
associates
4 billions dollars revenue
Which customers are
Hospital, Physician Labs
Reference and Government Labs
Biopharma
Research institutes

4. Objective
This work will emphasize the use of DSD for selecting the best protocol during Development of an immunoassay
It will highlight the benefit in terms of time and material
Objective:
Select the best lot of protein for further testing
Understand why we’ve seen unexpected differences between lots
Remove confusion (give sense to weird results)
Restore confidence between stakeholders
Constraints:
Limited amount of material (protein)
Limited time/resources for testing (multiple proteins and peptides to evaluate)

5. Y Y Y Y Y Y Y Format of the assay Factors of the DSD
Coating of Protein
pH of buffer
[ Glycerol ]
200µL [ Protein ] Y Y
Assay
Step 1
Step 2
Step 3
3x wash Y
30, 60 or 90min
37°C
200µL serum (blood sample) Y
3x wash Y
1 hour
37°C
200µL [Biotinylated - Peptide] Y
substrate
30 min
37°C
3x wash Y
200µL [Streptavidin-ALP]
Read at
405 and 450 nm

6. Coating and assay steps
Intensive wet work:
Coating and assay steps
3 x Wash
Reading

7. Comparison of the 2 lots of protein
Signal / Noise ratio
During the assay developpment:
Selection of protein/peptide
Format of the assay (incubation time, volumes, quantities, …)
Formulation, …
R&D faced an issue in reproducibility of experiments while using 2 lots of the same recombinant Protein.
Lot B
Lot A
It was decided to perfom additional experiments with the 2 lots to understand if the issue came from
true difference between lots (unexpected)
human error
environment

8. Design of Experiments: usual approaches
At the time of selecting the appropriate design of experiment we had the choice between classical one-factor-at-a-time, which requires 13 conditions (three levels of each factor, the centre point being the same for all factors) and DOEs (32 to 64 conditions)

9. Design of Experiments: usual approaches vs. DSD
# conditions
Main factor
2nd order
higher order
Quadratic
OFAT 13
Yes No
Full Factorial 64
Fractional Factorial 32
Response Surface Design (CCD orthogonal or Box-Behnken)
53-54
DSD
DSD # trials:
Even #factors: 2m+1
Odd #factors : 2m+3
Compared to OFAT, with the same number of conditions we get the interactions and quadratic terms.
Compared to RSD, with 4 times less conditions we get the « same » level of information.

10. DSD: outputs and analysis methodology
Design
# factors
# conditions
Samples tested
Measurement
DSD 6 13
* Positive sample: undiluted and diluted
* Negative sample
OD at 405 and 450 nm
Lot
Outcome
Outputs
Model
R² adj
Lot A
Variability
Pooled variance component on √CV%
Stepwise
0.96
Signal/Noise
Positive Undiluted / Negative (450 nm)
Stepwise (simplified)
0.997
Positive Diluted / Negative (450 nm)
Sequential
0.94
Lot B
0.87
0.998
0.996
We mostly used the Stepwise methodology.
We validated our models on Signal/Noise with the RMSE.

11. CV on Signal: example on lot A
CV needs transformation (√CV) to get a nearly normal distribution for further ANOVA analysis.

12. √CV on Signal: ANOVA on lot A
ANOVA shows that √CV is not statistically different whatever the
Sample (Positive diluted-undiluted, negative)
Wavelenght of measurement (405 or 450 nm)
Intensity of signal (OD)
Consequently we decided to pool the variance component (3 samples, 2 wavelenghts) for further DSD analysis.

13. √CV on Signal: DSD on lot A on pooled variance
The DOE analysis preformed through a Stepwise process gave a model with
3 main factors
three 2nd order interactions
1 higher order interaction
We’ll go in deeper analysis later in the presentation.

14. Positive Undiluted / Negative: DSD on lot A
The complexicity of the model is efficiently captured by the Stepwise methodology.
The model is « validated » by the evaluation of residual error.
RMSE/Mean = 2.6%
while mean CV = 2.23% on previous analysis

15. Summary of all Outputs on Lot A
Focusing on Maximizing Signal / Noise ratio we are able to select the optimum condition for further testing.
Note: we did not use « Maximize Desirability » function as we take into account other informations.

16. Summary of all Outputs on Lot B
For Lot B the main driver is increasing Glycerol.
However we need to reduce its concentration for technical issues (viscosity/pipeting) which impair the optimization.

17. Comparison of the 2 lots of protein: Signal / Noise (undiluted)
The main differences between the 2 lots are due to different drivers
Lot A: Coating rate and SA-ALP play the main role, with an unexpected quadratic effect on Coating rate
Lot B: Glycerol plays the main role

18. Comparison of the 2 lots of protein: Signal / Noise (undiluted)
Lot A
Lot B
The main differences between the 2 lots are due to different interactions:
Lot A: Coating Rate x SA-ALP
Lot B: Glycerol x Coating rate

19. Lot A: confirmatory test
√CV
CV%
DSD predicted value: 1.83% (1.46%-2.24%)
Confirmatory test: 1.91%
Signal
DSD predicted value: 1.09 ( )
Confirmatory test: 0.95
Noise
DSD predicted value: ( )
Confirmatory test: 0.042
DSD predicted value: 28.6 ( )
Confirmatory test: 22.67
Signal/Noise

20. Conclusion The use of DSD helps us
Understanding true differences between the 2 lots of proteins
Save time and money
Restore confidence between stakeholders