Parses textual descriptions of reactions into pathways and OWL expressions
(NOTE: we will soon have docker which will be easier)
Install SWI-Prolog from http://www.swi-prolog.org
Run directly using bin/reactioner
- data/go-rhea-newsyns.tsv - suggested syns from RHEA
- data/go-rhea-check.tsv - checking RHEA xrefs
This code parses structured text definitions of reactions (in particular, those used by GO).
For example, the def for sucrose synthase activity is
Catalysis of the reaction: UDP-glucose + D-fructose = UDP + sucrose.
We employ a 3-step Parse-Match-Balance on each. After this we translate the expression to an OWL expression.
The reaction string is parsed according to a grammar:
Reaction ::= Participant[ + Participant]* = Participant[ + Participant]*
The grammar is encoded by a prolog DCG. The grammar is semi-deterministic, so each string has zero or one parses
TODO: co-efficients/stoichiometry
See also: Jupe et al. (https://doi.org/10.1093/database/bau060)
for generative approach in Reactome
Each of the participants is matched using CHEBI. Because strings do not uniquely and unambiguously identify a chemical entity, we produce all combinations of possible reactions. We score each reaction according to the strengths of each match.
We prioritize matches to name/label above synonym matches. We treat all synonym scopes as equivalent, as CHEBI uses related vs exact arbitrarily.
TODO: use rhea-suggested synonyms
We prioritize balanced reactions above unbalanced ones, or ones in which balance can not be determined. The assumption is that false CHEBI matches will be less likely to balance.
A subset of CHEBI classes have a property chebi:formula that maps to
a string such as C15H24N2O17P2. We use this to count the number of
each element on each side. A penalty is introduced for every
unabalanced element.
For example:
Catalysis of the reaction: UDP-glucose + D-fructose = UDP + sucrose.
Parses and matches to the prolog term:
['http://purl.obolibrary.org/obo/CHEBI_18066'/'UDP-D-glucose',
'http://purl.obolibrary.org/obo/CHEBI_15824'/'D-fructose'] =
['http://purl.obolibrary.org/obo/CHEBI_58223'/'UDP(3-)',
'http://purl.obolibrary.org/obo/CHEBI_17992'/sucrose],21,['N'-0,'P'-0,'C'-0,'H'- -3,'O'-0]).
(note the above is an actual prolog term, the predicate = is written in infix form
Goal is to infer definitions of grouping classes
For any grouping class C, we take all asserted ancestors, and determine the reaction for each one.
We want to compare this against all other reactions.
We treated this as a standard classification problem and use tree learning in scikit-learn, where the features/predictors are properties of the reaction and the target is a binary variable for member of class / not member
Each reaction is decomposed to chemical entities to derive base features, and then these generate features for the overall reaction of the form
The BaseFeature is derived from the CHEBI class
We use the "Ygevny transform" by default.
Exact match can be performed. For true equivalence or subsumption test use an OWL reasoner.
Often RHEA will use a CHEBI ID that does not seem justified by the label they use.
We can query their RDF, RHEA uses their own labels for each compound, and map this to CHEBI. We can query these pairs:
- data/go-rhea-newsyns.tsv - suggested syns from RHEA
We note whether the label they choose is a new string or corresponds to a different ID (ambiguous)
We don't always expect exact matches.
We can do standard semantic similarity matching, using CHEBI graph. However, we may want to pull in other features of chemical entities rather than rely on pre-asserted hierarchy.
Chym See: https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1000937