Knowledge of regulation relations is widely applied by biomedical researchers in for example experiment design on regulatory pathways and in systems biology. In the work presented here, we analyze in total 28 verbs - and in dept 6 frequently used verbs denoting the regulation relations regulates, positively regulates and negatively regulates through corpus analysis. We propose a formal representation of the acquired knowledge as domain specific semantic frames and the semantic types of the relata of the resulting relationships. We suggest that the acquired knowledge patterns can be used to identify and reason over knowledge represented in texts from the biomedical domain.
Relations representing positive and negative
regulations are widely used in the biomedical domain in
systems biology for representation of pathway
relations, e.g. [
For a more precise retrieval and representation of regulation relations, however, a deeper analysis of the semantics of the different verbs denoting regulation relations can be useful. The knowledge that is acquired from such an analysis can be translated into textual knowledge patterns, or semantic frames, similar to those in the lexical resources FrameNet and VerbNet.
BioFrameNet [
In another related work focusing on regulation,
[
In [
Our aim is to develop a formal semantics of the
regulates relation developed from [
Additionally, the focus on agent-patient roles in
regulation can support reasoning over additional
extracted events as suggested by [
In this section, we describe the corpus analysis that is a means of identifying the lexico-syntactic patterns that exist for regulatory verbs and their arguments in biomedical texts. These analyses are the basis for a bio-extension of the semantic frames as presented in section 3.
In order to categorize the different types of regulation relations other than through the isa-relation hierarchy in which positive and negative regulations are specializations of regulation, cf. figure 1, we analyze a corpus compiled of biomedical texts or, more specifically, a collection of PubMed abstracts. Through this analysis, we have identified four general types of regulations patterns as outlined below.
Verbs denoting regulation, negative regulation and positive regulation have a somewhat different usage in biomedical texts than they do in general language texts. In order to identify this usage, we created a concordance of all occurrences for a selection of regulatory verbs in a corpus consisting of 40.000 PubMed abstracts. Our search covered the specific verb forms “regulates” (323 occurences) (denoting regulation), “inhibits” (781 occurences) (denoting negative regulation), “reduces” (699 occurences) (denoting negative regulation), “decreases” (1119 occurences) (denoting negative regulation), “increases” (3171 occurences) (denoting positive regulation), and “stimulates” (372 occurences)(denoting positive regulation).
By examining the concordances for these verb forms, we can classify the usage of the examined verbs with respect to types of arguments into four general frame types or patterns (analyzed further in section 3.1). In the patterns presented below, arguments may be of the type processes or substances. Substances can for example be gene products (e.g. proteins and functional rna) or small molecules and processes can for example be glucagon release or glucose transport.
A majority of the identified frame parts are
overlapping with the ones identified in [
Action-NN] V-active caused by
roughly 10% of the occurences of the examined
verbs. For this pattern, the regulated substances
are most frequently enzymes. Example:“...lithium
inhibits the enzyme glycogen synthase kinase-3...”, .
In terms of [
ers roughly 10% of the occurences of the examined
verbs. Example: “...nitric oxide pathway regulates
pulmonary vascular tone...” . In terms of [
a minor part of the occurences of the examined verbs. Very few examples of this pattern were found, only in the analysis of the verb regulates. Example: “...Proximal tubular dopamine production regulates basolateral Na-K-ATPase ...”.
Thus, not many textual instances of regulations
have a process on their left hand side, i.e. present
the patterns processes regulate substances or
processes regulate processes, but the vast majority of
our examples present a pattern where a substance
regulates a substance/process. Normally, when
regulation relations are represented in biochemical
interaction webs such as KEGG [
The over-representation of the pattern substances
regulate processes is also reflected in the number of
text patterns found. For example, in [
In an extended corpus analysis as well as framecomparison from the resources of FrameNet, VerbNet and WordNet, we notice that some of the verbs representing regulatory relations exhibit a deviant behaviour compared to the identified frame parts. For example, the verbs “increase” and “decrease” often appear in a passive or nominal form, and in these cases, they do not have an expressed agent. We must thus add frame parts such as [ Patient Action-NN] V-passive and NN ’in’ [ Patient Action-NN]. This type of linguistic knowledge is important for the outcome of the semantic annotation and eventually for a reasoning over the extracted knowledge. 3
The results of the corpus analysis as presented in section 2, can be viewed as an extensional definition of regulation relations. However, to be able to perform a reliable semantic annotation of text, there is a need for an understanding of the intensional side of the relations.
Here, we discuss the connection between the
patterns presented in section 2 and the intensional
descriptions of the relations as it is described in [
Though the proposed transformations are purely formal, they can be useful for a reasoning process as well as for a foundation for a semantic annotation.
In line with [
This means that the abovementioned knowledge patterns can be processed such that the semantic constraint, Substances inhibit processes, can be incorporated into the knowledge patterns using concepts from the Semantic Network. For example, “Substance(T167)” is a type, having subtypes such as “Amino Acid Peptide or Protein” and “Chemical”. Additionally, “Phenomenon or Process(T067)” is type representing “process” having subevents like “Physiologic Function” and “Cell-function”.
Since all concepts in the individual UMLSresources have a direct link into the Semantic Network, this method will make it possible to capture the ontological types of a large number of domain specific terms. The understanding of the main types are as follows:
Substances are, like for instance “continuants” in
BFO [
Substances can regulate other substances or processes, but processes can also regulate other processes or substances. Focusing on for example the relation regulates there are therefore four possible relations among individuals combining the two types of relata substance and process, corresponding to the four general patterns given in section 2. We name these relations regulatesss, regulatessp, regulatesps, and regulatespp, where the subscript “ss” means that it is a relation that can only exist between two substances, “sp” means that it is a relation that can only exist between a substance and a process,“ps” means that it is a relation that can Regulation FN.Objective_inuence,BFN.Regulation regulate a€ect inuence Positive Regulation Negative Regulation Activation Stimulation FN.Encoding,FN.cause_to_start,FN.Initially_create FN.Cause_change_on_position_of_a_scale encode express produce generate activate start (Increase)
Increase amplify stimulate promote elevate induce facilitate
Irreversible inhibition FN.Removing eliminate remove delete abolish (Decrease) FN.Change_on_position_of_a_scale
Reversible inhibition FN.Hindering,FN.Eclipse, FN.Cause_change_on_position_of_a_scale reduce inhibits decrease block limit suppress attenuate Inactivate only exist between a process and a substance, and finally, “pp” means that it is a relation that can only exist between two processes given in section 2.
We can formalize the four general types of patterns discussed in section 2. We use s, s1, s2, . . . to range over substances, and p, p1, p2, . . . to range over processes.
Substances regulate substances
⇒ s1 regulatesss s2 Substances regulate processes
⇒ s regulatessp p Processes regulate substances
⇒ p regulatesps s Processes regulate processes
⇒ p1 regulatespp p2
However, introducing a “production of” and a
“output of” operator as proposed in [
The “production of” operator works on a substance s by transforming it to the process that produces s. Similarly the “output of” operator transforms a process p to the substance that is the output of p. With these operators, the instance relations regulatesss, regulatesps, regulatespp and regulatessp can be transformed into one, namely the regulatessp relation. These transformations are given below s1 regulatesss s2
⇒ s1 regulatessp production of (s2) s regulatessp p
⇒ s regulatessp p p regulatesps s
⇒ output of (p) regulatessp production of (s) p1 regulatespp p2
⇒ output of (p1) regulatessp p2 Additionally, we note that the pattern s1 regulatessp f unction of (s2) denoting a slightly different meaning, namely a regulation by a substance of the function of an enzyme (or another substance), frequently occurs. The specific lexical patterns representing this relation are for example the expressions: [Agent]V − active[Patientactivity/function] [Agent]V − active[activity/function of Patient] Similarily, the production of -operator in s1 regulatessp production of (s2), has the expressions: [Agent]V − active[Patientproduction/secretion/ transcription/expression/synthesis/release] [Agent]V − active[(the) synthesis/production/ secretion/expression/transcription/release of Patient] These identified specific lexical patterns expressing the patient, can be of use for extraction of knowledge regarding the pattern for later reasoning as presented briefly earlier in this section.
Reasoning over Knowledge Patterns These transformations reflect the underlying semantics of verbs denoting regulates relations in biomedical texts. For example, the verb stimulate has a usage where it denotes the relation positively regulates1: “Insulin stimulatesss glycogen”, “insulin stimulatessp the glycogenesis”, and “insulin stimulatessp the production of glycogen (through the glygoneonesis)” where the process glycogenesis is equal to “the production of glycogen”. Likewise, we can construct the sentences: “beta cell secretion stimulatespp glycogenes” that can be transformed to “outout of beta cell secretion stimulatessp production of glycogen”, where the output of beta cell secretion is insulin.
It may be argued that the relations regulatesps and regulatespp are not genuine relations since we can question whether it is at all possible for processes to stimulate other processes or substances directly or whether it is rather through their outputs they stimulates.
However, as part of our aim is to be able to identify and annotate instances of regulations in texts, it is important to include the patterns that cover the forms as they actually occur in biomedical texts, and not only as we know them to function. In this cross field between form and meaning, it may be possible both to grab meaningful contents from texts through the semantic roles (e.g. agent and patient) of the relata.
In this work we have performed a preliminary
investigation of a subset of English verbs denoting
regu1This example is also used in [
In order to achieve a deeper knowledge about the semantics of the studied relations, we initially mapped the relata of the relations into the UMLS Semantic Network. Second we grouped the different verbs denoting regulates relations and their corresponding text-patterns into four different types corresponding to the types of their relata.
Finally, we gave a formal description of the four observed types of regulation relations and transformed these into one. Using the operators “production of”, “output of”, and “function of”, the four observed types of regulation relations, regulatesss, regulatessp, regulatesps, and regulatespp, can be transformed into one, namely the regulatessp relation.
We stress the importance of identifying textual forms of the relations as knowlege patterns as they actually occur in biomedical texts, even if we know that a given pattern does not reflect how regulations function in reality as we know them to function.
The semantic patterns that are identified
through our corpus analysis, can form a background
for further knowledge extraction, where for
example text is automatically annotated by use of the
patterns and subsequently fed to a machine
learning algorithm for identification of new patterns (in
line with [
Additionally, through a deeper linguistic
analysis, these parts can be part of the basis for a domain
specific FrameNet describing regulatory events (an
extended BioFrameNet in line with the vision of [