MapTracker is a massive graph database - over a terabyte on disk, 1.2B nodes, 2.0B edges, 3.5B metadata assignments. It's used at BMS to "resolve X to Y" - that is, given an object of "type X" find - in a qualitative way - all "related" objects of "type Y". This is done using an aggresively normalized triple store and a large set of rules that dictate what kinds of edges are reasonable to traverse when going from X to Y.
MapTracker is generally not used "on its own", but is rather a component in other tools. Examples avaiable here are:
- Chem-Bio Hopper - "Hop" from biology to chemistry, or vice-versa, using published chemical activities
- Hypergeometric Affy - Given a set of "interesting" (generally overexpressed) Affymetrix probesets, run Fisher's Exact Test to identify ontologies that appear overrepresented in the set.
- Standardize Gene - Given a set of gene identifiers (eg symbols), attempt to determine what they "really are" (ie, given messy gene symbols, convert to rigorous gene accessions)
The schema (tables) is relatively simple. What has made MapTracker particularly powerful is:
- Careful normalization of loaded data
- Segregation of nodes into namespaces. Ameliorates collisions, particularly with identifiers like gene symbols
- Exhaustive logic defining valid connections between
X-to-Y. Example, RNA to probeset - Generic transitive logic that lets
X-to-Ybe automatically merged withY-to-QandQ-to-Win order to findX-to-W. Such "chains" allow only fundamental connections to be defined yet allow the network to be (safely, rationally) explored far beyone its expected "neighbors"
The image below is an auto-generated network, created by sampling 20,000 random edges from the database (created by exploreSelf.pl). It represents, at a high level, the common node-edge-node triples held by the database.
All edges are part of a controlled vocabulary. Most (though not all) are directional. The edges in the above sample include:
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent