While reading about how SPARQL computes its solutions I considered that we can get much better efficiency by choosing when and how to do joins based on our own needs rather than letting RDFLib's sparql plugin handle it. I think it took me so long to make this change in thinking since I'm used to doing things in relational databases using static queries with a set number of joins. I've started writing some demo programs that build paths in the RDF graph to query. The solution uses generators, so porting to other languages will require a slightly different approach that breaks up the code in the generators for other languages, like Java, that do not support them. Happily, other facilities, like being able to iterate over an RDF graph given a triple pattern is available in Java through Jena with list_statements and in C through Redland with librdf_model_find_statements.

Reading

I am currently doing some reading on how SPARQL queries are evaluated in order to improve the inefficiency of complicated queries created by PyOpenWorm.

PyOpenWorm R1

We (myself, Stephen Larson, Anton Shekhov, Travis Jacobs) have continued working on PyOpenWorm in order to get it release-ready. The work that has been done has focused on documentation of previously existing work, addition of data-integrity tests, and inclusion of data in the PyOpenWorm database from existing spreadsheets and the sqlite3 database contributed by Timothy Busbice.

End of GSOC

We've reached the end of Google Summer of Code (it was weeks ago now, but I'm writing this blog post anyway).

Although I've written a lot of code and made a lot of plans, I can't really back off of the project until I see it being used in a way that people (outside and inside of OpenWorm) are actually seeing it. When I came into the project, my long-term goal was an "annotated simulation". This is a system for adding annotations to data, carrying them through to model files, and finally exposing the annotations to end-users upon inspection into simulation results. This is to answer our long standing criticism that, in some way or another, our simulations don't line up with the actual science. (Similar concerns are addressed from a computational perspective in this thread) Over this passed summer, I've gotten the first step to a place where I almost like it, but the second step has stalled significantly, and the third was never started.

Cleaning up documentation...

As I've been using PyOpenWorm, I've encountered some cracks in the design. The primary issue is the interaction between configuration, object properties, and database set-up. Some examples of problems that have resulted:

1. Run-time configuration has been set-up such that each object has to receive its configuration anew from the object that creates it or else use a globally set default configuration. This is a very brittle framework that falls apart any time the appropriate configuration is not passed to Configureable objects contained within other Configureable objects which can result in configuration values hanging around between connect() and disconnect(). This isn't a big deal for simple one-off scripts, but can be annoying when testing or when designing more complex scripts.
2. Registration of a Property doesn't take place until the first time its owner is initialized. This prevents us from resolving a Property to an object directly from the graph without having made one of its owners first. This results in properties which are identical having different names in sub-classes due to the dynamic nature of property-name creation.
3. Each DataObject sub-class has to be registered as a separate step from defining the class. This is an minor annoyance when creating new classes.
4. It's necessary to either know the structure of generated namespace URIs (e.g., http://openworm.org/entities/Neuron/) and recreate them or to create an object of the desired type when referencing objects in another namespace. The first approach is sub-optimal because we might have reason to change the structure of the URIs in the future which is already a problem for some auxiliary methods designed to extract information from a URI. The second approach is also problematic since it requires the creation of objects which are never used otherwise, creating overhead for the garbage collector as well as the programmer

My proposed solution to these problems is to push the configuration and class-dependent initialization to the class-definiton phase through the use of Python's decorators which run on library-load. This takes care of the fourth point by having all of the namespaces and namespace sturcture made static by the time any object should be created.

The third point is covered by folding the current registration into the class decoration.

The second point is covered by attaching each Property subclass to the owner's class while retaining the initialization in the __init__() call for accessing owner-instances.

Finally, the first point can be addressed without the proposed solution by not having configuration passed through __init__. Initially I conceived of a tree structure for configuration which could be augmented by internal nodes and passed on to children with specific configuration needs that didn't need communication to higher level users. This structure is exemplified by Data which is Configureable, but also a Configure object suitable for configuring other objects. Unfortunately, this is the only case where the feature gets any use and isn't required for passing augmented configuration to the module as a whole, so it can simply be removed.

I hope to address some of these issues soon after outstanding issues are closed and features for the next release are well-defined.

UPDATE: Most of these issues have been addressed in the new-classes branch. Not yet merged into master.

It seems I have more to learn about rdflib's SPARQL query engine. A persistent issue was the mis-use of indexing for what (I thought) should be simple queries. My error was in including triples of the form ?x rdf:type openworm:type in the query patterns. For reasons I don't yet understand, the engine evaluates such statements before others even if there's a more reductive join it could perform from other patterns. Hopefully I can understand this eventually.

In any case, standard object->property->value queries are effectively O(1) time again--yay!

small update:

The issue from the previous post was one of indexing. The Property object identifiers were being generated differently for every Neuron object which obviously made minimal use of the index. Basing identifiers on the identity of the owner object and the value of the property takes the query time to get a single property back to constant time. For example:

for x in range(1000):
    Neuron().save() # creates and saves a randomly generated neuron object
Neuron().name()

Takes about 10 seconds for the call to name() on rdflib's memory store as well as the SleepyCat store (closer to 11 seconds) bundled with rdflib. That's about .01 seconds for getting the name of a specific neuron regardless of the database size.

I adjusted the query according to the previous post and also added patterns for the cell the property is actually on (oops):

prefix openworm: <http://openworm.org/entities/>
prefix neuron: <http://openworm.org/entities/Neuron/>
prefix sp: <http://openworm.org/entities/SimpleProperty/>
select distinct ?type where
{ 
?Neuron rdf:type openworm:Neuron .

?Neuron neuron:lineageName ?Neuron_lineageName .
?Neuron_lineageName rdf:type openworm:Neuron_lineageName .
?Neuron_lineageName sp:value ?lineageName .

?Neuron neuron:name ?Neuron_name .
?Neuron_name rdf:type openworm:Neuron_name .
?Neuron_name sp:value "AVAL" .

?Neuron neuron:type ?Neuron_type .
?Neuron_type rdf:type openworm:Neuron_type .
?Neuron_type sp:value ?type .

?Neuron neuron:receptor ?Neuron_receptor .
?Neuron_receptor rdf:type openworm:Neuron_receptor .
?Neuron_receptor sp:value ?receptor 
}

For the data stores we've been using, such queries don't seem to execute very efficiently.

I would hope for it to be quick since

1. there should be exactly one statement matching (a, sp:value, "AVAL") to look up in the 'pos' index,
2. and then only one matching (b, neuron:name, a) looked up again in the 'osp' index,
3. and an index lookup on the 'spo' index for the types of 'b' to check it has the correct type should be efficient.
4. The neuron should then have only one neuron:name,
5. and one of all of the other properties which can be looked up in the 'spo' index.

For queries against Sesame, I have to try using the Java API to be more sure these indexes are actually there, otherwise I'm making a lot of assumptions about how well the indexes work. Besides that, I don't have an easy way of understanding how the queries are getting executed without measuring often very long query times. Maybe I can get some sense of how these queries are being handled with some test data.

I've started reading a little about ZODB, a python object storage database. The core of what I've been making for PyOpenWorm is an Object<->RDF mapper. The choice of RDF storage solutions to back PyOpenWorm was based on the work that had already been done, my experience in working with some RDF tools, and an expectation that joining with existing datasets may be easier with RDF than other storage options. It may be useful to re-evaluate going forward.

I've been trying out the KyotoCabinet store for rdflib. Uploading the whole worm under the current parameters takes about 17 minutes.

Timing any of the queries on the full data set is difficult since even 'simple' ones tend to take several minutes. The queries I'm running look something like this for getting the type ('interneuron', 'sensory', 'motor') :

select distinct ?type where { ?Neuron_type0056fa3c098f873e <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://openworm.org/entities/SimpleProperty> .
?Neuron_type0056fa3c098f873e <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://openworm.org/entities/Property> .
?Neuron_type0056fa3c098f873e <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://openworm.org/entities/DataObject> .
?Neuron_type0056fa3c098f873e <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://openworm.org/entities/Neuron_type> .
?Neuron98fc353f03788d3e <http://openworm.org/entities/Neuron/type> ?Neuron_type0056fa3c098f873e .
?Neuron_type0056fa3c098f873e <http://openworm.org/entities/SimpleProperty/value> ?type }

I've adopted a pattern of mirroring the triples into the upload query and the select query. The objective with this is to make it easier to change the data structures without having to update both the upload format and the select with every change. This is the reason for the query above containing 3 rdf:type patterns -- in lieu of RDF inference (intending to include later), I want to be able to utilize the class hierarchy to, for instance, get all of the Neuron objects when querying for Cell objects. To have this, I insert, for each object, all of its super-types. This would let me query for objects of type Cell and get all those of type Neuron (and Muscle). Any object should have all of the super-type tags in addition to its most specific form, so there's no problem, theoretically, querying for the right object.

The main issue is that these queries get executed very slowly. I suspect it's something to do with scanning the table. Interrupting a query (because who has time for that?) we find ourselves here:

  File "build/bdist.linux-x86_64/egg/rdflib/plugins/sparql/evaluate.py", line 47, in evalBGP
  File "build/bdist.linux-x86_64/egg/rdflib/graph.py", line 1373, in triples
  File "build/bdist.linux-x86_64/egg/rdflib_kyotocabinet/KyotoCabinet.py", line 315, in triples

There is an index on the predicates-objects which should make getting the objects of a given type quick:

%ls -lh worm.db
total 193M
-rw-r--r-- 1 markw markw 6.1M Jul 15 13:35 contexts.kch
-rw-r--r-- 1 markw markw  55M Jul 15 13:35 c^o^s^p^.kch

-rw-r--r-- 1 markw markw  55M Jul 15 13:35 c^p^o^s^.kch

-rw-r--r-- 1 markw markw  55M Jul 15 13:52 c^s^p^o^.kch
-rw-r--r-- 1 markw markw  14M Jul 15 13:35 i2k.kch
-rw-r--r-- 1 markw markw  14M Jul 15 13:35 k2i.kch
-rw-r--r-- 1 markw markw 6.1M Jul 15 13:35 namespace.kch
-rw-r--r-- 1 markw markw 6.1M Jul 15 13:35 prefix.kch

I'm thinking that for my query above though, the database engine is doing a join between the results from the index lookup. Adding in an option to toggle between the triples generated for querying versus uploading should make these queries more reasonable while still allowing us to get subtypes from a super-type query. Some of these issues I've run into (there are others...) suggest that the shared query/upload pattern may be inappropriate. I'll have to think more about it.

I haven't posted in a couple weeks!

Since the last update, our design has moved away from a remote database to one locally hosted. So each user of PyOpenWorm would have his/her own copy of the database on their machine together with a serialization of the database which can be version controlled under git. The serialization isn't too large (40MB for trix, 20MB for n-quads, 13MB for trig).

One of my long standing issues is attributing data to data sources -- labs, articles, curated data. This aspect of attribution, assigning metadata, can be done before the data model is fully specified, and has been in various places. We would only (!) need to translate into our data model after that.

For now I'm implementing the API.

Lately, I've been attempting to generate morphology files from the database to use in RegenerateConnectome.py. Generating them is fine, but obviously pulling from the database for every one of the 302 neurons is pretty slow. It may not be worth it to speed up that loop more directly rather than caching the database. Need to test this.

EDIT: Comparing to the full database copied to a local Sleepycat store, the queries are actually slower. Same with OpenRDF Sesame...probably due to my laptop being way underpowered compared to the server...or resource sharing.

EDIT2: Think I'll look into HTTP caching.

PyOpenWorm feature freeze

I have declared a feature freeze for PyOpenWorm. For this week we're just writing code to pass the tests we've already written and those obviously derived from the API.

I am trying to figure out how ReadTheDocs can bring in documentation from the python source files.

• update (10:33): Sphinx installed
• update (10:44): Got Network's documentation printing out. Let's see about the others.
• update (10:57): aaand, now the docs don't update from the source files. Not even deleting the build directory changes it.
• update (11:19): Started using sphinx-apidoc. Has really noisy output.

Citations, DOIs, Pubmed IDs

Since we've started discussing the API for PyOpenWorm, we've started talking about putting up citations of data sources. What we're looking at doing is making a declarative API that allows us to write things like:

n1 = Neuron('AVAL')
n2 = Neuron('ADER')
r1 = n1.connectsWith(n2) # r1 is now a representation of this 'connectsWith' relationship
s = Chemical('X')
r2 = s.modulates(r) # r2 is the representation of the 'modulates' relationship
e = Evidence(doi="http://dx.doi.org/10.1007%2Fs00454-010-9273-0")
e.asserts(r2)

In English, "10.1007%2Fs00454-010-9273-0 suggests that chemical X modulates the connection between AVAL and ADER". This 'relationship' API (a sub-API to the PyOpenWorm API), has the primary feature of allowing statements made as to be part of higher order statements.

Lineage data

I finally dug into the daughter_of tree (unfortunately, a forest. see below) cell division data.

After cleaning it up a little I put it into OpenRDF: http://107.170.133.175:8080/openrdf-workbench/repositories/OpenWorm2

There are some discontinuities which make the graph unconnected. The roots in our forest of division trees can be shown with this query:

PREFIX ns1:<http://openworm.org/entities/>

select distinct ?z where { ?y ns1:daughter_of ?z . filter(not exists { ?z ns1:daughter_of ?p }) }

These discontinuities have nothing to do with the input errors mentioned in the github issue as all of those in the daughter_of table were easily fixed.
To look at the path-to-root for 'AB plppaaaapa' :

PREFIX ns1:<http://openworm.org/entities/>

select distinct ?z where { ns1:AB_plppaaaapa ns1:daughter_of+ ?z }

The code I used to extract and upload: http://pastebin.com/NVnAfD8D

It's just the cell divisions that we've had for some time. We've discussed bringing in volume data for each of the cells in development in order to model differentiation waves discussed by Richard Gordon and colleagues. As of yet, we don't have this data readily available.

OpenRDF problems

I was having difficulty with setting up OpenRDF on a Digital Ocean machine, but figured it out. A permissions problem was preventing sesame from creating its log files. Changing the ownership of /usr/share/tomcat7 to user tomcat7 fixed it.

I've switched from using a local SQLite database to an OpenRDF SPARQL service which supports SPARQL 1.1 updates. The update statements should include mandatory information about the updater as well as other checks specific to the kind of data being modified. I'm looking into an inference engine that stands between the DB and PyOpenWorm, but in the meantime I'll be encoding the rules in PyOpenWorm itself.

I have an OWL file with WormBase data in it. And a sparql endpoint that has the same data.

Wormbase has it going on (with respect to anatomical tree displays).

Also:

• WormBase data modeling.
• The database for WormBase
• REST interface for WormBase entities

Work has started on the requirements writeup.

Cell lineage data

Just links:

Relevant meeting notes

Whitepaper

Github issue