Adapt reasoning to deal with transitive relations

[gglusman]

Upon re-re-testing what drugs may treat Familial Pityriasis Rubra Pilaris (on ci), and comparing to the previous test, I see a new result with score 100:

Multiple issues with this result:

• The main logic is 'these two things are both inputs to the same binding activity, and since one of them treats the disease, the other might too'. This is extremely weak logic.
• The 'is input of' and 'has input' edges have no provenance attached.
• The 'retinol (aka vitamin A) treats PRP' triple cites as provenance just one paper about acne. The actual relevant content in the paper:

In 1953, Sidney Q. Cohlan (1953) observed that high doses of vitamin A had teratogenic effects on pregnant rats. G. L. Peck was initially researching vitamin A for the treatment of Darier’s disease, ichthyosis, and pityriasis rubra pilaris and discovered its therapeutic effects on acne vulgaris. Peck et al. (1979) also counseled on contraception, [...]

This clearly does not provide confident support for the 'retinol treats PRP' assertion on which the 9-cis-Retinal one depends. Yet it gets score 100. The direct 'treats' path cites Aragorn as source. The 'retinol treats PRP' triple cites BTE and TMKP.

[sandrine-m]

Thank you @gglusman for pointing this out. Those are excellent points in my opinion. Here is my analysis of your ticket, please tell me if I understood properly (I would split this ticket accordingly) and feel free to correct any misunderstanding :) :

1. scoring: I am unclear on the scoring piece as my understanding is that the scoring is still in progress.
2. reasoning: more rules are needed when is input of is present
3. EPC: add provenance on edges is input of and has input
4. SEMMED-related: filter text-mined or SEMMED edges treats if supported by only 1 publication

[gglusman]

@sandrine-m I'm thinking about this a bit differently. The main points:

1. Text mining: a 1979 paper mentions in passing that someone was initially researching using X for the treatment of Y, and this is interpreted to mean that X treats Y. This is not necessarily a semmed filtering issue. Of note, this is significantly weaker evidence of X actually treating Y than if there is a clinical trial testing the effectiveness of X in the treatment of Y... and we were told we cannot derive 'X treats Y' from the presence of such a clinical trial.

2. Reasoning: it's not just the 'input of' that is suspect in reasoning. There are several predicates that are hardly transitive in nature but are being interpreted as such. Some intuitive examples:

   • drug_A binds albumin, drug_B binds albumin, drug_B treats disease_X ... can we hypothesize that drug_A treats disease_X too? Tons of stuff binds albumin, it's just not specific!
   • key_A is_in key_ring, key_ring has_key key_B, key_B opens my_house ... it's rather weak to derive that key_A opens my_house (or even that it 'may open')
   • handkerchief is_in my_pocket, my_pocket has_content key_B, key_B opens my_house ... perhaps the handkerchief will open my_house too?

Yet somehow these two extremely weak things (a text misparsing and a weak inference path) become a top-scoring result. Agreed on the missing EPC bit and the not-clear-how-scoring-works. :)

[sandrine-m]

Thank you! I'll split the ticket accordingly to make sure we treat all points.

[sandrine-m]

I am keeping this original ticket for the reasoning part as I find the discussion in this thread important to keep for this topic. Changing title accordingly.

[gglusman]

Another example from the ASM case:

A number of genes increase secretion of dinoprostone and their corresponding things that are claimed to treat ASM. Guilt by association can only go so far...

[gglusman]

In #484 Sarah exposed another example of this, with with two drugs being 'part of Animals'... and somehow this being sufficient support to transfer the 'treats' assertion from one to the other.

[sandrine-m]

Thanks @gglusman for associating both issues, that is very helpful for triage. I have created a "transitive" sublabel for all these issues to help go back to them post relay.

[gglusman]

Noting #465 as affected by weak transitive logic.

[mbrush]

@gglusman do you happen to know what ARA prediction tools reported these problematic support paths? I suspect that is not a manually crafted or automated rule from an ARA like BTE, Unsecret, or ARAGORN, but more likely an arbitrary path through a larger explanation graph like those created by the ARAX reasoner - which are really not meant to be considered independently, outside of the context of this larger graph,

If it is indeed the case that these types of paths are from ARAX - then we will have perhaps isolated the source of the problem, and be better able to find a solution. This would also be interesting to know in upcoming EPC planning discussions about how to weight the strength of evidence provided by support paths with different topologies.

[gglusman]

@mbrush We collected a number of issues with the transitive label, and more than one ARA seems to be implicated.

[gglusman]

Testing on 2024/01/19, what drugs may treat Alzheimer Disease?

One answer is "Iron, dietary". The explaining path is:

This would be a correct redrawing of the graph:

That is, SLC11A2 causes decreased uptake of both dietary iron and zinc; zinc is claimed to treat AD. There is no support for iron->AD.

[sierra-moxon]

from TAQA:

• lots of results swamping the "answer" - these paths are non-transitive
• these are weak reasoning paths, they should be scored lower than good stuff.
• in progress on Aragorn side: score for each reasoning type (paths returned match rules -- we want a scheme that prioritizes good rules -- one that is correct more often -- vs. another path from a less-good-returning-rule) that can influence the rank -- which will end up down weighting lots of these answers. This is in progress this month, might not fully be completed.
• other ARAs might also have this issue
  • going a further step - direction matters in > 1 hop paths.
  • some predicates get more credibility than others - A x B more strength than A y B