Biologists frequently refer to the functions and activities of gene products and the complexes they form, for example, asserting that a particular gene product has protein kinase activity, or that another funcations as a receptor for glutamate. In building an ontology to represent these functions and activities we need to decide what general class or classes of thing terms like protein kinase activity and glutamate receptor activity refer to. This question is intimately connected to the question of how we relate these activities and functions to gene products and to the biolgical processes such as signal transduction, glycolysis or gastrulation.
One account of functions and activities is that they are 'realizables'. By saying that a gene product has 'protein kinase activity' I am stating that it has the potential, given the right environmental conditions and the presence of a suitable substrate, to be the active participant (the catalyst) in the process of protein phosphorylation. In ontology jargon we say the function or activity is realized by participation in the process. This fits nicely with the everyday use of the terms function and activity by biologists: they'd be happy to say that the molecules of a restriction enzyme sitting in glycerol in a -20'C freezer have site-specific endonuclease activity even though they are not currently involved in cutting DNA.
Despite this, the GO asserts that its molecular_function terms refer not to the 'realizables' but to the processes in which gene products are the active participants. More precisely, a GO molecular_function is a process, carried out by a single gene product or complex. This has the advantage that it is simple to relate molecular functions to biological processes involving the activities of many gene products: An EGFR receptor activity is part of cell-cell signaling; pyruvate kinase activity is part of glycolysis. Within the GO, part of is used to relate molecular functions to biological processes which they are always part of. The newly developed LEGO/nocuta annotation system for GO annotation uses part of to capture the component molecular functions of particular instances of biological processes. *(TBA - broader argument - almost all of the direct relationships which are useful to make are relations to the process not the realizable: what its inputs and output are, what it regulates...)
Defining GO molecular_function requires one further clause. The GO and its associated annotations aim to record the cannonical (non-pathalogical) function of gene products. To make this clear, we define a GO molecular_function as any process carried out by a single gene product or complex, which that gene product or complex has evolved. By this definition, virus receptor activity is not a GO molecular function: CD4 did not evolve to bind HIV, HIV evolved to bind CD4. It would be correct, however, to record that HIV:gp120 has the GO molecular function 'protein binding' with input CD4).
Some molecular functions are simple biochemical process such as an enzyme catalysed reaction or ion binding. Other molecular functions involve multiple, causally linked processes. For example: a receptor activity has a ligand binding component that regulates and effector component, such as a kinase activity; Ca2+ binding regulates protein binding by calmodulin; in an active transporter activity, ATPase activity provides energy for ion transporter activity. Some MF classes are defined, in part, by process context. For example, chemoattractant activity involves a single gene product binding and activating signaling receptor in context context of positive chemotaxis.
(See ticket: Decide on new upper level classification in MF for further discussion.)
Molecular function is very poorly formalised. Only 16% of terms have logical definitions (compared to 70% for biological_process) and only 10% of classifications are inferred (compared to 64% in biological_process).
Multiple inheritance classification summary:
| No. Terms | No. parents |
|---|---|
| 7672 | 1 |
| 2197 | 2 |
| 626 | 3 |
| 193 | 4 |
| 89 | > 4 |
| (tested on 16:02:2017) |
Given that manually maintained multiple inheritance heirarchies are typically incomplete, the true level of multiple inheritance is almost certainly higher. The propsed new classifications in this refactoring will increase this further. This situation is not sustainable without the use of design patterns to drive inferred classification. In the absence of this, errors and ommissions in the class heirarchy will inevitably accumulate.
Some branches of molecular_function lack potentially useful intermediate classes for grouping functions. For example, we have no general classes for nuclear receptors or for activities that modify the tertiary sructure of DNA. Adding such classes will improve the usefulness of enrichment analysis results and
Compound molecular functions are under-respresented in the current heirarchy. For example there is no single term in GO that can be used to record that calmodulin binds calcium and that the calcium binding regulates binding of calmodulin to other proteins. Instead we have separate terms for calcium binding and calcium-regulated protein-binding. Where compound functions are represented, they are not represented consistently. For example, kinase receptor activities are classified under kinase activity, and have their ligand recorded using either has_input or a has_part relationship to a binding term for the ligand. In contrast, G protein-coupled receptor activity classes in GO currently have no link to their effector function (GEF activity).
LEGO annotation allows users to construct networks describing particular instances of biology via their component molecular functions and biological processes, recording a variety of relationships between these components including causal (regulatory) relationships and part relationships. The power and flexibility of LEGO allows for great expressiveness, but also makes annotation consistency challenging. It also allows curators to break down functions and processes that are formally defined in the ontology using the same set of relations and classes creating yet more consistency problems.
To get around these problems, we need to develop templates for LEGO curation that promote annotation consistency and reduce the burden on curators. These need to be compatible with ontology design patterns for equivalent/related classes. With these in place, we can use standard reasoning software in combination with the LEGO models and the ontology to infer additional annotations to LEGO models, further improving consistency as well as enhancing annotation.
- Clarify the definition of molecular function and its major subdivisions
- Add new intermediate nodes that provide biologically meaningful and useful ways for biologists to group functions
- Allow curators to better capture the compound nature of functions
- Automate classification via the implementation of design patterns
- Add error checking in the form of disjointeness axioms
- Provide templates and automated inference to simplify and standardise LEGO curation
Starting point: Paul Thomas' manual refactoring proposal in WebProtege
Document usefulness and practicality of proposed new terms via checklist, recording whether the proposed new term:
- is useful grouping of molecular functions not provided by an existing biological process?
- is useful for error checking (via disjointness axioms)?
- is useful for construction of design patterns?
- is useful for curators - e.g. used in templates or guiding template usage?
- requires multiple inheritance classification
- if so, can we use design patterns & reasoning to populate subclasses?
Many tickets on the MF refactoring tracker follow this pattern.
Work up from existing classes under Paul's new classes to try to find common design patterns & LEGO templates to automate classification.
-
Edits are taking place on a branched version of the GO.
-
Design patterns and templates are specified in the GO design pattern repository. Files here are automatically checked for validity using a Travis job.
-
Documentation of standard practises for LEGO annotation to ensure correct and complete inference. For example, should binding between two gene products be recorded as one node or two?.
Combined design-pattern / template docs will be specified using DOS-DPs, which now support specification of LEGO templates.
The components of compound molecular functions are causally linked. Co-reference issues prevent these causal links from being represented on the class level, but we can represent them directly in LEGO templates. We can also use conventions within class design patterns to specify regulatory input and effector/output.
In addition to the standard OWL-EL tools for inference that we use in the ontology, we can also use SRWL rules to drive inference within the LEGO model. This turns out to be particularly useful for reducing the burden on curators and maximising inference.
Inference on the individual level can also take advantage of inverse relations. We cannot safely conclude from that statment (all) head has_part some hair that (all) hair part_of some head. But if we record in the ontology that a necessary condition of being a 'kinase receptor activity' is that it has part some kinase activity, then this is fulfilled by a part of relationship between a kinase activity and a 'receptor activity' in a LEGO model.
As far as possible, we will leverage existing classifications such as EC and Rhea. For details see ticket: geneontology#22
In network models, we need to be able to refer to processes and their
Refering to activities and functions as processes can be linguistically awkward given that everyday usage of the terms is consistent with them being realizables.
Where two gene products contribute to a molecular_function, as is the case with binding, there are two realizables (the binding activities of each gene product) that are realized in the same process (see ticket).