I really appreciate when a vendor takes the time to publish a work that is both education and informative.  The folks at GE did a really nice job with this article that describes how their Capto adhere resin could be used to clear aggregates from the process stream.  What I love about their work is that they include the data so the folks like me can generate our own analysis.

They begin with a screening design looking at load amount, load conductivity, and load pH.  The experimental design was to run the column in weak partitioning mode such that the impurities are have a higher affinity to the column than the antibody.  The column was then run in flow-through mode with loads ranging from 93-312 mg/mL.  While some antibody might be lost due to binding on the column, the fraction is small relative to the amount passing through.

One of the lovely elements of JMP is the ability to code values within the table.  For example, if my values in a column are across a range, then I can choose to have the lower and upper limit represent -1 and 1, respectively.

The authors then looked at recovery, CHO HCP, leached Protein A and aggregate (as a percentage).  The Protein A was always below LOQ and thus omitted from the analysis.  CHO HCP, aggregate, and recovery could all be analyzed.  Starting with the recovery, the Analysis of Variance (ANOVA) for the main effects and pair-wise interactions was

Source	DF	Sum of Squares	Mean Square	F Ratio
Model	5	0.22177718	0.044355	16.2981
Error	5	0.01360754	0.002722	Prob > F
C. Total	10	0.23538473		0.0041*

and the parameter estimates were

Term		Estimate	Std Error	t Ratio	Prob>|t|
Intercept		0.921899	0.027004	34.14	<.0001*
Load(93,312)		0.1192675	0.02355	5.06	0.0039*
Load cond(10,30)	Biased	0.1853178	0.04106	4.51	0.0063*
Load pH(6,7.5)	Biased	-0.221614	0.043222	-5.13	0.0037*
Load*Load cond	Biased	0.0251376	0.024798	1.01	0.3572
Load*Load pH	Zeroed	0	0	.	.
Load cond*Load pH		-0.069163	0.036368	-1.90	0.1156

Clearly the pair-wise interactions should be neglected (p>0.05).  When only the main effects are taken into account, then the ANOVA became

Source	DF	Sum of Squares	Mean Square	F Ratio
Model	3	0.20926078	0.069754	18.6907
Error	7	0.02612395	0.003732	Prob > F
C. Total	10	0.23538473		0.0010*

and the parameter estimates were

Term		Estimate	Std Error	t Ratio	Prob>|t|
Intercept		0.8820829	0.019367	45.55	<.0001*
Load(93,312)		0.1042488	0.025469	4.09	0.0046*
Load cond(10,30)		0.1298189	0.037178	3.49	0.0101*
Load pH(6,7.5)		-0.158015	0.036643	-4.31	0.0035*

With these, we can now develop a model of how the recovery will be affected by the three factors using the Prediction Profiler in JMP to assemble a linear function of the three factors.  The dashed blue lines represent the 95% confidence interval around the model obtained using the defined estimates.

It's a bit late for moi, so I'm gonna pick this up tomorrow.