myExperiment: A MySpace for the self-serving bioinformatician (Carole Goble)
I hurt. I have a hang-over in my eyebrows, nose, ears, toes,... Don't drink sour beer. Just say NO.
Carole certiainly knows the art of swearing in the right places during talks. She's talking about workflows, why scientists share, and so on. This talk doesn't seem as polished as her normal stuff. However, ...
Cut'n'paste vs workflow - bla, bla, bla. Boxes with data-flow arrows - bla, bla, bla. Services, SOAP, blast, interpro, coordination, change, complexity, heterogeneity. Feel better yet? At least Taverna at least automates this stuff.
'In The Cloud' - somebody uses somebody elses stuff, without support. It's just out there.
mm - publishers are starting to ask for the workflow that goes with the paper. Interesting...
Peer reviewed workflows. Workflows as patterns.
Now talking about mySpace, facebook, ... and then the equivalent for e-science. myExperiment as social-networking arround workflows. Tagging vs structure. Meta-data, controlled-vocabulary to structure this.
"Social-science lterature tells you the bleeding obviouis" - people are selfish, will share when it is of obvious benefit to them.
I'm not quite sure what this myExperiment stuff is, other than propper web community support for workflows. All seems good and sensible. Finally, manchester is starting to pay attention to actual people rather than technologies.
OXL - part of ondex, the 'native' xml format for post-integration data.
Do xml dumbly - glue all links in-place --note: how does it handle cycles then? Seems like a thinly-veiled excuse to give an ondex talk, showing pictures it draws.
Emperor's new cloths? Probably usefull if you are already in the Ondex world.
Analysis of Integrated biomolecular somethings - CABiNET
Same old slide about integrating networks - different ones, different types, with external kinds of data.
CABiNET - semi-automated network processing pipeline e.g. to identify funcitonal information about proteins.
Identification of sub-structures in the network graphs. Then go thorugh them using biologists. Lots of by-hand triage on the networks.
Uses a manually curated 'alias db' containning db-specific IDs for shared entities.
All seems propperly engineered. J2EE, EJBs, WS, DB, the usual suspects. Nice layered architecture. Seems to have decent through-put. Seems to have a workable web front-end for people. Unfortunately, in our simple during-talk tests, it doesn't work too good.
VINEdb: a data warehouse for integration and interactive exploration of life science data
Is he going to show us some screenies? Not sure ...
Event-driven import. Llistens to OMIM, KEGG, UniPROT, IntAct and GO. When these change, the monitor triggers a re-integration. Grabs data from: flat files, go sql dump. Exports as data marts, warehouses + web interface.
Ah, screenies finally. Not sure what they tell me though.
Possibly worth looking at. Didn't address the effort needed to set one of these guys up from scratch. No idea how brittle it is to source schemas. No idea how much schema integration is done.
Exploring PSI-MI XML Collections using describeX
XML is widely used by bioinformatics community. They claim the flexibility of xml introduces heterogenity. I'd prefer to think that it allowed latent hetergenity to be exposed.
DescribeX builds xml structural summaries - labelled graphs of paths (xpath roots?) - basically, suffix-tree compression on the document structure. MM - is popularity of a node (element path type) correlated with importance of a node?
XSD schema doesn't constrain how you use the format - can often encode the same stuff in different ways while being complient.
All human-driven. It is basically semantic zooming, data mining of document structures. Doesn't tackle document content at all. I'm sure it's interesting to people managing xml formats (e.g. PSI) but is of no discernable use to me.
Data Linkage Graph: computation, querying and knowledfe discovery of life science data.
Databases in sciences are structured. Databases are organised by domains. --note: of course, this is wrong on both counts - they are badly structured and have very overlapping, messy domains, although a single protein may have a 'core' concept of interest that doesn't overlap others much.
Life-science databases are already heavily linked via xrefs - 87% of lrelevant databases have some sort of interlinks.
Database links are explicit - html, dbms integral reference integrity, web-service parameters; implicit - text mining, sequence alignment, pattern matching, ... --note: I'd call these schematic and data-hapenstance but who's counting?
Data warehouses - model unified schema, import (crappy - format handling etc.), make complex queries 9SQL, XQuery, ...), expose through canned queries on user interface
Virtual integrated db - the first 2 steps become lazy and distributed. Requires a lot of knowledge.
Query by navigation: data linkage graph - building a semantic network. E.g. make all xrefs in html forms links. Make all seqquence renderings have a link to allignments. This, of course, is only useful to people.
They make a distinction between inter-entity relationships (links) and inner-entity relationships (functions). The resulting schema is awesom - 318 nodes, 375 relations!!! And those are the ones they thought where interesting from just 5 data sources!!!
This is all about the data. It fails to at all address the meaning of the data.
Having said that, it looks like it's been done right. I'd almost go as far as to say Ondex done right.
BioXMash: XML data summaries for user-driven data integration.
Long-term goal: data integration for end-users.
XML-based. Selection/projection on trees. Interactive querying. Indexing pseudo-off-line (e.g. index while at lunch, then have interactive session). Working with multi-gig xml files. Can work with thousands of files (presumably they are small files).
Fun statements like "we know that in genecards, each file is a gene" that say a lot about how they are thinking about data-integration. From indexes, find possibly relevant xml files/fragments, slurp into memory, blat it out (with select/project) as the query answer.
No semantic integration. Very focussed upon structural-metrics for source-document extraction. Could be helped by full-text indexing. These guys had to go ask for an xml dump of databases. Could do a RESTFUL xml export of source data, but this is firmly future work.
Quite fun, but seems a bit 'good idea' and not so 'just works'. In pilot phaze it seems.
VisGenome.
Community-based linking of biological network resources: Database, Formats and Tools
To understand your data, you do need to view it in context. Use the kind of network that will be relevant to your question.
How to find Networks? Want structured, dynamic, user-driven content. Got monolithic, unstructured, static content.
The solution? A colaborative web site - WiKi. Sounds a lot like you basically punt all the hard work of dealing with the data by harnessing 'lots of monkeys' in the form of domain experts to deal with it all for you. Use on-line tools to make it as easy as possible for the domain experts.
Their model has Objects and Connections. Objects include databases, files, formats, tools. Connections describe relations between these e.g. Kegg can export in KGML and this can be read by VANTED. seems like lots of stuff on user-management, access control etc. to give a feeling of data quality by restricted/audited authoring.
How does it deal with conflicting information? What support is there for reasoning? Is it all just human-readable? Are the semantics all in our heads, or actually in the types of connections and objects?
Suports 'semantic-drill-down' of data.
It's very small right now. Probably has a worth that scales non-linearly with the size and activity of the user-base. Very human-facing.
A methodology for comparative functinoal genomics. (e-Fungi, Mike Cornell)
Sequence analysis of 36 (to be 48) fungal genomes. Blasts, alignments, clusters, phylogeny, cellular localization, ...
This seems to be old-school bioinformatics done with no e-science techniques, on reasonably sized datasets, everything hand-rolled and hand-analyzed. Very dull. Trees, sequences, bla, bla, bla. Let's play on facebook untill this travesty has gone away.
Keynote 1: Understanding interactomes by data integration
Soren Brunak
Protein-protein interaction networks - topology, temporal, phenotypic aspects. Text mining, predict 'disease genes'.
Starting with the cell cycle (an intrinsically dynamic system).
Data-cleaning is a pain - quality, formats, ...
Data following linear growth (?not in exponential phaze yet?)
Different databases provide very different data (did he mean knowledge?) about interactions.
Don't trust other people's 'components list' - they are wrong, missguided - re-analize raw data e.g. they found that 10% of yeast proteins are periodic (cell-cycle) - found 600 proteins.
Most cell-cycle control is a mix of non-cyclic, constant-expression components and up/down components - just-in-time activation of nearly-assembled complexes. Periodically present proteins have more degredation signals, more phosphorylation, post-translational modification, ...
Very little conservation of periodic behavior between human, pombe, cerevisiae, arabadopsis. However, complexes have the same members in each organisms, but have per-linneage choice of periodic vs stable expression. The important thing is that the complex is periodically active, not which component gives rise to this. Complexes and their behavior is conserved, not the details of the complex members.
This is very important for drug testing - need to make sure that drugs are tested in model organisms with compattible regulation to humans e.g. chimp responds to that immunosupressent like human does, macack doesn't (where the guys swelled up and had massive anafalactic shocks) - be scared wherever there is phosphorylation and tight regulation - it may be one of these stable/variable complexes.
Similar symptoms from diseases with genetic components may be due to them arrising from impairment with different proteins, but both interacting with the same complex. Combine phenotype data with linkage data with interactions.
Project this knowledge into a vector space - apply standard kernel magick (in their case possibly just dot product) to ccreate a scoring function.
Lots of good, solid work. The bioinformatics really does seem to be tracking the biology in this case.
