GPML

GPML is an XML markup language for biological pathway diagrams. It solves an awkward problem: a pathway drawing is simultaneously a picture that a biologist reads and a formal model that software needs to compute over. Most formats pick a side. Vector formats like SVG keep the picture and lose the biology; systems-biology formats like SBML keep the biology and lose the layout. GPML deliberately keeps both, storing the coordinates, colours and line styles of every object next to the database identifiers, literature citations and interaction semantics that make the drawing analysable.

The acronym has been expanded two ways over its life. Early schemas and most of the literature call it the GenMAPP Pathway Markup Language, after the desktop application it came from. The current documentation and the GPML2021 schema call it the Graphical Pathway Markup Language, reflecting a format that long ago outgrew its original host program.

History and origins

GPML’s ancestry runs through GenMAPP, the Gene Map Annotator and Pathway Profiler. GenMAPP began around 2000 as a prototype in Bruce Conklin’s laboratory at the J. David Gladstone Institutes in San Francisco, with a first public release, GenMAPP 1.0, in 2002. Its purpose was to let biologists paint microarray expression data onto hand-drawn pathway diagrams and see which parts of a pathway had changed. Diagrams lived in a proprietary binary format with the extension .mapp, readable only by GenMAPP itself on Windows.

That was tenable while GenMAPP was the only consumer of the data and fatal as soon as it was not. GenMAPP 2.0 followed in 2004, and its pathway schema is the direct source of GPML. The provenance is still legible in the schema files: the earliest version kept in the official repository, GPML2007, carries the documentation string “GenMAPP version 2.0 Pathway Schema version 1.0”, along with the note that all graphics measurements are in centimeters - a unit choice inherited from a tool built for printing pathway diagrams rather than for the web.

GPML was designed to be backwards compatible with MAPP, in the specific sense that anything expressible in a MAPP file can be expressed in GPML and existing pathways can be converted. That compatibility requirement is why the format is so unapologetically graphical. It was not designed from first principles as an ontology of biochemistry; it was designed to losslessly carry an existing archive of drawings into an open, parseable, cross-platform world, and to add annotation on top.

Two 2008 projects turned that into an ecosystem. PathVisio, an open-source Java reimplementation of the GenMAPP idea developed at Maastricht University and the Gladstone Institutes, adopted GPML as its native format and was described in BMC Bioinformatics in 2008. WikiPathways launched in the same year as a wiki where the research community curates pathways collaboratively, with a GPML document behind every page. From that point GPML stopped being an application file format and became an interchange standard.

Design philosophy

Three commitments shape the language.

The drawing is the model. A GPML file is a flat list of graphical objects with explicit coordinates. There is no separate abstract layer that a renderer then lays out. When a curator moves a box, the file changes. This is the opposite of the SBML approach, and it is why GPML pathways look like the figures in a review article rather than like automatically generated graphs.

Identity comes from external databases. A node labelled TP53 means nothing on its own. GPML attaches an Xref to it, giving a database name and an identifier - Ensembl, Entrez Gene, HGNC, ChEBI, HMDB and so on - so that downstream tools can map the node to whatever identifier space their data uses. This indirection, supported in practice by the BridgeDb identifier-mapping framework, is what makes a GPML pathway usable as an analysis target rather than only as a picture.

Provenance travels with the content. Literature citations, comments and curation metadata are part of the document. In GPML2013a and earlier this is done by embedding BioPAX PublicationXref records inside a <Biopax> element and pointing at them with BiopaxRef; GPML2021 replaces this with native Citations, Annotations and Evidences collections. Either way, a pathway carries the evidence for its own claims.

Anatomy of a GPML document

The root element is Pathway, carrying the name, organism and schema namespace. Inside it, GPML2013a defines a compact vocabulary:

A minimal fragment in the GPML2013a dialect looks like this:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22

<?xml version="1.0" encoding="UTF-8"?>
<Pathway xmlns="http://pathvisio.org/GPML/2013a"
         Name="Example signalling" Organism="Homo sapiens">
  <DataNode TextLabel="TP53" GraphId="a1b2c" Type="GeneProduct">
    <Graphics CenterX="220.0" CenterY="140.0" Width="80.0" Height="20.0"
              ZOrder="32768" FontSize="10" Valign="Middle"/>
    <Xref Database="Ensembl" ID="ENSG00000141510"/>
  </DataNode>
  <DataNode TextLabel="MDM2" GraphId="d4e5f" Type="GeneProduct">
    <Graphics CenterX="220.0" CenterY="240.0" Width="80.0" Height="20.0"
              ZOrder="32768" FontSize="10" Valign="Middle"/>
    <Xref Database="Ensembl" ID="ENSG00000135679"/>
  </DataNode>
  <Interaction GraphId="idff01">
    <Graphics ZOrder="12288" LineThickness="1.0">
      <Point X="220.0" Y="150.0" GraphRef="a1b2c" RelX="0.0" RelY="1.0"/>
      <Point X="220.0" Y="230.0" GraphRef="d4e5f" RelX="0.0" RelY="-1.0"
             ArrowHead="Arrow"/>
    </Graphics>
    <Xref Database="" ID=""/>
  </Interaction>
</Pathway>

Everything characteristic of the language is visible here. Coordinates are absolute and mandatory. Identity is delegated to Ensembl. Endpoints are anchored to nodes by GraphRef with relative offsets, so a connection survives its endpoints being dragged around. And an Xref with empty attributes is perfectly legal - GPML tolerates unannotated content, because a half-curated pathway is better than none, and curation is expected to be incremental.

Evolution

The schema versions are named for their years, and each is a distinct XML namespace, which makes version detection trivial for parsers:

• GPML2007 - http://genmapp.org/GPML/2007, the original published schema
• GPML2008a - http://genmapp.org/GPML/2008a
• GPML2010a - http://genmapp.org/GPML/2010a
• GPML2013a - http://pathvisio.org/GPML/2013a, introduced on WikiPathways in July 2013 and the long-lived workhorse version. The repository also ships a GPML2013a-fuller variant, which its maintainers describe as reflecting the GPML actually used between 2013 and 2017 more accurately than the nominal schema does
• GPML2017 - a draft version that appears in the schema repository but was reportedly never adopted as a release
• GPML2021 - http://pathvisio.org/GPML/2021, the current version, released around July 2021

The 2013 move from the genmapp.org namespace to pathvisio.org marks the point where the format’s centre of gravity shifted from the original Gladstone application to the PathVisio and WikiPathways community. GPML2021 is the largest structural change in the language’s history: it introduces container elements (DataNodes, Interactions, Labels, Shapes, Groups), promotes Author, Annotation, Citation and Evidence to first-class schema types, adds a Description element on the pathway, and retires the embedded BioPAX block. The accompanying LibGPML library exists precisely because the two dialects differ enough that hand-rolled parsers are a liability; it handles reading, writing and conversion between versions.

The tooling followed the same arc. PathVisio 3, published in PLOS Computational Biology in February 2015, added a plugin system so that importers, exporters and analysis modules could be written independently of the editor. PathVisio 4 is built on the 3.x architecture plus LibGPML, and targets a workflow in which pathway models are submitted directly as pull requests to the wikipathways-database GitHub repository.

Current relevance

GPML is not dormant, though it is easy to see why it might look that way from outside bioinformatics: it has no package manager, no conference, no benchmarks, and its schema has changed only a handful of times - four revisions between 2007 and 2013, then nothing until 2021. What it has is a live corpus. The WikiPathways 2024 database paper, published in November 2023, reports approximately 1,900 human-curated and reviewed pathways spanning roughly 27 species, all stored as GPML, with homology mapping generating a much larger secondary collection for vertebrate model organisms. Monthly snapshots are published in GPML alongside GMT, SVG and RDF derivatives.

The infrastructure around it has been modernised rather than replaced. WikiPathways migrated from MediaWiki to Jekyll and GitHub Pages, putting pathway curation under ordinary version control with continuous integration - GPML files are now reviewed in pull requests like source code, which is arguably the most natural home a diagram-as-text format could have found. The classic WikiPathways site has been made read-only, with its retirement announced for 1 May 2026.

Where GPML sits in the wider standards landscape is worth stating plainly. SBML and BioPAX are richer and more formal about biochemistry; SBGN specifies a rigorous visual grammar. GPML is looser than all three, and that looseness is the point - it is a curation format, designed so that a bench biologist with a review article and an afternoon can produce something structured enough to compute over. Converters between GPML and the stricter formats exist, and GPML’s role in practice is usually the human-facing end of that pipeline.

A note on the name

Three unrelated things share the acronym, and search engines mix them freely:

• GPML, the pathway markup language described here
• GPML, the GPlates Markup Language, a GML application for plate-tectonic and geological feature data
• GPML, the Gaussian Processes for Machine Learning toolbox for MATLAB and Octave, which accompanies the Rasmussen and Williams textbook published by MIT Press in 2006 - software, not a markup language

A fourth, reportedly an XML standard for interchanging genetic programming trees, has also been proposed in the genetic-programming literature. Encyclopedia entries that file GPML under machine learning have almost certainly collided with one of the latter two; the markup language with a 2000s origin and an XML schema is the pathway one.

Why it matters

GPML is a good case study in what a domain-specific markup language is actually for. It was not designed by a standards committee reasoning about the ontology of biology. It was extracted from a working application, under a hard backwards-compatibility constraint, to get an existing body of curated knowledge out of a proprietary binary blob and into text that anyone could parse. That origin explains its shape - the centimetre measurements, the absolute coordinates, the tolerance for empty annotations - and it also explains its longevity. Two decades on, the drawings still open, the identifiers still resolve, and the corpus has moved from a Windows desktop tool to a wiki to a Git repository without ever being re-authored.

For anyone building a format for a specialised community, the lesson is that the interesting design work happens at the boundary between what humans want to draw and what machines need to read. GPML chose to serve both, accepted the redundancy that entails, and got a durable archive out of the bargain.