First chat with Daniel Wegmann

[oolonek]

Discussion with Daniel Wegmann (https://www.unifr.ch/bio/en/groups/wegmann/) and Daniele.

A bit too high level for my poor statistical background. I will try my best to resume and Daniele will amend (and Daniel eventually if he is in !)

So in a nutshell:

Same concern as Daniele regarding pseudo absences.
It changes things drastically if we can or not assess trueness of the absence in our species per chemical matrices. And we can't (at least not completely).
Daniel argued that is such case he could basically have the simplest (and dumb model) with the matrix fully filled with 1 and that their would be no way to say that any more sophisticated model (taking into account phylogeny, research efforts or whatsoever ) would be better. Daniele disagreed on this point.

Things changed when considering the possibility to make posterior measurement (for example using the already profiled data we have in the PF dataset for which we could output a chemical composition for species present and also absent from the Lotus matrix).

We should thus :

1. establish an initial function predicting the full matrix (not a function for individual data points) equal to an intercept summed to the weighted constraints and a matrix of errors which would be a function of both biological and chemical taxo). See second picture.
2. predict chemical vectors per species from the latent variables (the Lotus matrix)
3. model the errors for each chemical taxa between the predicted vectors and the measured LCMS profiles (here we could use CANOPUS, keeping in mind its build on ClassyFire so maybe also use the ClassyFire results as inputs). See first pic

Voila for the drafty notes.

[Adafede]

Great to read!

I think that we cannot verify point 3 with single C18 runs coming from PF.
It would be way too brutal...

I would like to propose a better solution instead of "just saying no" but can't think of any better at the moment...

Thinking about such LC profiles... are we closer to a confirmation of absence rather than a confirmation of predicted presence.

For example, first, validate we are able to measure the presence of a given compound in our LC run (on a sample from species X). Then, fill in the data. In the predicted data, take species Y, where even after imputation, the compound is still not present and see if effectively the compound is not measured.

Would that bring something?

[Adafede]

Side note: Davy Guillarme just told me that filtering compounds with a logP -1 < x < 6 should be fine.

[Adafede]

Ok, so I did a quick and dirty function to filter LOTUS structures for C18 runs: anticipated-chemistry-of-life/elotus-trees@84ebcc5

With a filter of -1<xlogp<6 and 100<mass<1500 we end up with 228,030/293,590 structures.

[oolonek]

Great if we validate with PF profile these beeing EtOAc extracts I guess we could even be meaner with the logp filter.
Say elution between 1min and 6 min on a 8min linear gradient starting at 5%ACN finishing at 100%.
How did he give you the logp range ? Does this comes from his mighty HPLC excel sheet ?

[Adafede]

From an email exchange I had with him. I wanted to know if there was "less brutal" than rough logp and this is what he advised me for C18 runs...

[oolonek]

Maybe we could ask him for an unprotected version of https://ispso.unige.ch/labs/fanal/hplc_teaching:en ?
It looks like the gradient simulator takes logP of several compounds as inputs but couldn't find how to modify theses ...

[oolonek]

Follow up by Daniel
Metabolites.pdf

[oolonek]

and the afternoon scribbles