First discussion with Daniele Silvestro

[oolonek]

Discussed today with Daniele (https://www.unifr.ch/bio/en/groups/silvestro/) and here are some notes of the chat :

• Daniele made a parallel with his research field (fossils) were pseudo absences are common: it is not because you dont find a fossil that it doesn't exists. This is pretty close to what we have in the lotus dataset: absence of molecules cannot be distinguished between real absence (the molecule is not present in the organisms) and pseudo absence (the molecule has not been discovered yet).
• Following the previous point he suggested that such dataset are not ideally suited for supervised learning because their is no ground truth. Other approaches such as bayesian stats could be better suited in such case.
• The probability of occurence of a given chemical class (i.e of a pseudo absence being indeed pseudo an not a real absence) could be formulated as the sum of probability of this occurence given the : chemical research effort context, phylogenetic context, covariance of other chemical classes.
• This last term was discussed: study of the covariance of chemical classes within a given taxa could be informative.
• Suggestion to study these stats and imputation possibilities first at the lotus original data level. That is predict occurrences of chemical classes for already studied species rather than for non studied ones.
• Reflexion and question with @Adafede : if we go for a bayesian statistical model rather than supervised learning does this mean that we can get rid of the threshold we taught about before (i. e. limit the dataset to plants having at least 50 molecules/paper associated) ?

[Adafede]

I knew we started discussing the molecules/threshold story somewhere...had to search for a while between the discussions to avoid creating a new one. This might become a limitation of GitHub discussions to organize our thoughts...

I wanted to start the following discussion:

• Do we need to restrict the dataset? If so, how:
• Restrict organisms to at least n publications?
• Restrict organisms to at least m compounds?
• Restrict organisms to less than o compounds?
• Restrict compounds to at least p publications?
• Restrict compounds to at least q organisms?
• Restrict compounds to less than o organisms?
• Restrict based on some percentile/quantile?
• A combination of what is mentioned above?