Hi Everyone!
I am realtively new to JMP and would like to ask if I could get some feedback / answers to a few question. I am quite sure some of them are based in a lack of understanding on my side, so please be patient.... I have of course tried to find answers first in documentation and training videos but I'm still left with this.
At work, I recently took over the development of growth media for our cells (biotech). The medium consist of about 80 different components as well as some additional factor (for example pH setting). There are about 14 different stock solutions / groups of ingrendients with are mixed at different ratios. We currently have an issue with this formulation to keep the cells alive as long as possible which means viability on the last day (D10_viability) should be as high as possible.
To figure out which component components might influence this without too much bias from previous obsevations, my idea was to perform a definitiive screening design with 12 of these stock solutions as factor. I would assume that quadratics will be present for almost all. Obviously it would be nicer to do this for each individual ingredient in a way I can also asses the (certainly present) interactions, but that's not feasible by hand.
So what I did was using the current formulation as center point and go from there. Since I assume that for some ingredients (for example trace elements) a much bigger change is expected to see something, I chose different "intervals". What I did however is that I changed the values from the setpoint using different factors which is supposed to mean for exmaple: for ingredient 1 (let's just imagine this was glucose), the factor was 5, so the center point (0) is for example 5 g/L, the low point (-1) is 1 g/L and the high point (+1) is 25 g/L. In this particular case 0; 5 and 10 would probably also have worked, but in case of trace elements, just doubling for +1 probably wouldn't do much. Also for pH I need to use the "pH expected" and not what it turned out be in reality (which was along those lines, but of course not 100% the same).
I attached the file also containing some responses and the real values and the ones used in the DoE as well as some responses.
So my questions(s):
Does this approach make any sense / is this correct? If I use the real values, I cannot assess / fit a definitive screening and the expected performance of the DoE is terrible.
If I perform a fit anyways (not using the definitve screening fitting tool) using mains and quadratics I get something that looks like a decent model but I have no idea if this makes any real sense....
Can I put in interactions that I believe exist despite the "evaluate model" dialog telling me these can't be estimated?
It turned out that one of the ingredients (nr 6) was extremely deadly for the cells on already by D03. I can identify this quite nicely (although it's also obvious by eye) using the definitive screening tool.
When I use the same platform to look at viability on D10_viability as response I get some weird results including very strange residual by row plots. When I kick out all the conditions with the high ingredient_6 (marked in red in the file) which is about 40% of the conditions, I cannot use the definitive screening platform anymore, but can still fit models that look good to me. Obviously the evaluate model platform tells me this this is no longer great.
Does this still make any sense??
Also, this only works if I remove the original high 100 from the recoding for factor 6...
Sorry for all these somewhat messy and long questions, but I would really appreciate any kind of input!
