Tran and Baral have proposed an action language (BioSigNet-RR) that is specific for the modeling of signalling networks from Biology and for answering queries relative to the expected response to a stimulus. Translation of their action language to logic programs under Answer Set semantics yields a reasoning mechanisms that gracefully handles incomplete/partial information, updates etc. Those features are extremely important since existing regulatory networks often contain missing or suspected interaction links, or proven interactions whose outputs are uncertain. We present our application experience in developing a BioSigNet-RR formalization of the Signalling network for Arabidopsis Brassinosteroid, a complex interaction that is at the base of growth in some plant species. Such modeling exercise has involved 'filling the gaps' between the terse graphical language of signalling networks literature and the precise specification of the triggering conditions required by BioSigNet-RR. This application experience leads us to propose a new formalization style for action theories representing signaling networks that allows for the description of non-immediate effects of actions. Empirical evaluation of our declarative model has involved formulating and testing several 'what if' queries and checking the quality of the answer with domain experts.
In Biology, signalling networks (also referred to as signalling pathways) are specific collections of interactions with a common theme. They are used to provide a summary, working model of the complex interactions that explain how a living cell receives and responds to signals from its environment. Modeling signalling networks is sometimes essential for understanding how cells function and it may lead to effective therapeutic strategies that correct or alter abnormal cell behaviors.
From the point of view of Artificial Intelligence signalling networks represent an interesting form of semi-formal knowledge representation: relevant cellular interactions are to be explicitly described, in a simple graphical language. However, two main issues make modeling signalling networks with action languages challenging: 1. inhibition, which we see as a special form of constraint, needs to be explicitly represented and reasoned about, 2. several unspoken assumptions lie in the background as they are assumed to be known by the (expert) reader. For instance, the time element, i.e., a description of the delay between stimulus and reaction, is not explicitly represented yet it is an essential element in reasoning about the long-term evolution of the cell.
This extended abstract reports the results of our work on formalization and automated reasoning with a signalling network that is considered3representative of the size, level of detail and complexity of such networks in the Biology literature. The chosen pathway, which is depicted in Figure 1, represents the current knowledge on the interaction that makes the arabdiposis of Brassinosteroids (represented by the lone diamondshaped node in the pathway) stimulate growth in plants.
Chory et al. [
From the point of view of knowledge representation and reasoning, the following scientific hypotheses motivate our work.
First, there is a question about the adequacy of BioSigNet-RR [
Second, there is a question of adequacy of our action languages with respect to reasoning about pathway interactions. We consider the following informal test: can we apply BioSigNet-RR to formalize the type of ’what if’ questions that an examiner would ask to check a student’s assimilation of the material. Hence, we proceed to formulate some easy sample questions and see how the query language part of BioSigNet-RR allows to formulate it. 3 Paccanaro and Bogre, personal communications.
Finally, our overall hypothesis is that, in the middle term, we should be able to design and implement a vertical solution by which the domain knowledge synthesized in a signalling pathway can be accessed and reasoned about automatically. 2
Our modeling effort has adopted the BioSigNet-RR [
When the initial situation is only partially defined, or actions are unknown or even
may have non-deterministic effects, alternative answer sets account for the alternative
evolutions of the domain. The translation from action languages, including
BioSigNetRR, to logic programs is modular, in the sense that it can be done line-by-line by a
parser and generator. The translation, which is described in detail in [] has been adopted
as is and implemented by a Python-language program derived from Gregory Gelfond’s
Al2ASP project [
Signalling pathways are a graphical, synthetic representation of knowledge. However, to fully grasp the dynamics represented by the pathway one often needs to read attentively a natural-language background description that comes with the network.
In our project we have spent a great amount of time to understand and organize the
background description of the Arabidopsis Brassinosteroid process given by Chory et
al. [
BKI1 from the plasma membrane. 3. It has been proposed that the physical interaction between BRI1 and BAK1 leads to the formation of a signalling-competent hetero-oligomer. 4. The signals transmitted from the plasma membrane-localized BRI1-BAK1 heterooligomer negatively regulate the activity of a glycogen synthase kinase 3 (GSK-3), called BIN2.
It relatively easy to associate each phrase to one of the arcs represented by the signalling pathway. Such association, however, is not always straightforward and will be further commented upon.
It must be pointed out that in our analysis we have discovered that the interaction which is represented by arc:
+=
BR1 =) BAK1
is not found in the pathway depicted in Chory et al. [
The formalization of the pathway proceeds as follows. For each named cellular component, e.g., BR, we introduce two fluents5: high(br) e low(br). Then, we introduce two activation actions: activate(br) and inactivate(br), where the latter is an inhibition action that, in some sense, depresses BR. 4 http://www.sciencemag.org 5 The labels used in the signalling pathway are in lowercase, since they are constant names in the domain description.
high(bki1) inhibits activate(bri1) “Binding of BR to preformed BRI1 homo-oligomers lead to the dissociation of BKI1 from the plasma membrane.” binding(br; bki1) causes dissociated(bki1) if high(bri1) (3) Even though it looks like the description of a local, direct interaction, this formalization may be the most effective in capturing the rationale of the pathway. An alternative formulation, which has hitherto not been tested is the following:
high(br) high(bri1) triggers dissociated(bki1) “A suppressor screen using a weak allele of BRI1 identified a secreted and active carboxypeptidase, called BRS1, although its molecular target in the BR signaling pathway is unknown. It has been proposed that the physical interaction between BRI1 and BRS1 leads to the formation of a signaling-competent hetero-oligomer.”
Next, the remaining arcs are formalized, by coupling each arc to the illustrative phrase found in the description. For instance, the arc connecting BRI1 to BAK1 is connected to the phrase: “BRI1 interacts directly with BAK1 [through a phosphorylation process].” activate(bak1) causes up(bri1) “BKI1 interacts directly with the kinase domain of BRI1 to negatively regulate the signalling pathway.” (1) (2) (4) (5) (6) (7) (8) high(brs1) triggers activate(bri1) “In vitro and in vivo, BRI1 associates with TTL, transthyretin-like protein. Overexpression of TTL causes slight dwarfing, suggesting that it may play a negative role early in the BR signaling pathway.”
high(bri1) triggers downregulate(ttl) “Biochemical studies identified TRIP-1 as a BRI1 interactor.”
high(bri1) triggers activate(trip1) “The signals transmitted from the plasma membrane-localized BRI1-BAK1 heterooligomer negatively regulate the activity of a glycogen synthase kinase 3 (GSK-3), called BIN2.”
high(bri1); high(bak1) inhibits activate(bin2) “Although the mechanism is as yet uncharacterized, inactivation of BIN2 leads to the dephosphorylation of BES1 and BZR1, members of a new family of plant-specific transcription factors.” inactivate(bin2) causes low(bzr1) inactivate(bin2) causes low(bes1) (9) “Dephosphorylated BZR1 binds to a novel element in the promoters of BR biosynthetic genes to repress their expression.” high(bzr1) triggers activate(br) (10) “Current data suggest that dephosphorylated BES1 is then able to form homo-or hetero-dimers with the basic helix-loop-helix (bHLH) transcription factor BIM1, to bind to E-box elements in the promoters of br-regulated genes.” (11) (12) (13) (14) (15) (16) (17) (18) high(bes1) triggers activate(bim1) high(bsu1) triggers activate(bes1) low(bsu1) inhibits activate(bes1) high(bri1) triggers activate(bsu1) high(serk1) triggers activate(bri1)
It should be noticed again that in Chory et al. [
high(bri1) inhibits (bin2) 3.2
For each fluent we have had to introduce a couple of actions that represent the upregulation and downregulation of the fluent itself. Hence, we need to introduce the following two schematic rules, to be instantiated to each fluent:
activate(C) causes high(C) downregulate(C) causes low(C) In a sense, these axiom schemata naturally complement the action theory by capturing the essence of the += labeling of the arcs. However, they introduce an extra level of complexity in the representation, since new conditions must be devised to disallow these definitions for specific values of C, e.g., activate(bak1), that do not have a direct Biological interpretation.
To assess the adequacy of the representation language and of our specific action theory we have considered the following classroom scenario: questions about the Arabdiposis brassinosteroid process that a teacher would use to check whether her students have properly learn the material and are now able to reason about Brassinosteroids and their effects. Such questions were formulated with the goal of stressing the connections between the several cellular components of the cell.
To illustrate how natural BioSigNet-RR queries are, we now list, for each question, the expected answer in English, paired with BioSigNet-RR that captures, to some extent, the question itself.
Q: Looking at the pathway, how does BR affect the cell? A: BR causes the activation of BRI1 and BAK1, which, in turn, inhibit the activation of BIN2.
Our formalization is based on three distinct queries: high(bri1) after activate(br) high(bak1) after activate(br) low(bin2) after activate(br) Q: What are the effects of activation of BAK1 ? A: BAK1 brings about activation of BR1 ; subsequently, BR1 shall affect the whole cell network.
This question can be translated directly in the following formula (query): ? ? ? ? ? (19) (20) (21) (22) (23) Q: What are the effects of inactivation of BIN2 ? A: inactivation of BIN2 shall cause the inhibition of BZR1 and BES1. Again, me must resort to two separated queries:
The formalization and deployment project described in this article can be considered successful from the point of view of assessing what can be done with action languages (and Logic Programming in general) in the context of Biological knowledge representation and automated reasoning. The overall Artificial intelligence goal, i.e. to have computers process the meaning synthesized in a signalling pathway without human intervention, is yet to be achieved, as our formalization had to deal with a time-consuming human analysis of the accompanying textual explanation, often a heavy-going technical explanation.
BioSigNet-RR has shown to be an ideal platform for formalization in this domain. However, we believe that more research is needed in order to have the action theory match the pathway. One problem that was, in our opinion, only partially solved here is that action is understood in two ways. The first understanding, which is probably what Gelfond-Lifschitz meant, is that of external intervention, i.e., alteration of the state of affairs. The second understanding, which we suspect is more frequent in Biology, is that an action may also be an alteration, called upregulation or downregulation, of a component of the cell. For this second understanding, a specific action sort should be devised and introduced to the formalization language.
We are grateful to Alberto Paccanaro and Laszlo Bogre at Royal Holloway, University of London for providing us careful advice and guidance in the Biological and Bioinformatics aspects of this work. Many thanks to Gregory Gelfond for providing us constant advice and support on our work to extend his AL2ASP program.