Argudas: lessons for argumentation in biology based on a gene expression use case

Gene expression information, inconsistency, and incompleteness

Gene expression information describes whether or not a gene is expressed (active) in a location. Broadly speaking there are two types of gene expression information: one that concentrates on where the gene is expressed, and a second whose primary concern is the strength of expression. Throughout, this work focuses on the former category, in particular a technology called in situ hybridisation gene expression.

Information on gene expression is often given in relation to a structure - or it may be given for a point in 3D space - in a particular model organism. Here the model organism of interest is the mouse. This organism is studied from conception until adulthood. The time window is split into twenty-eight Theiler Stages. Each stage has its own anatomy, and corresponding anatomy ontology called EMAP [1]. The first twenty-six stages cover the developmental mouse: the mouse from conception until birth. Stage 27 is the new born mouse, and 28 the adult.

The result of an in situ experiment is represented as an image displaying a slice (2D cross section) of a mouse (from a particular Theiler Stage) in which some subsections of the mouse are highly coloured. Areas of colour indicate that the gene is expressed in that location. In addition to showing where the gene is expressed, the image provides some indication of the strength (or level) of expression. The more intense the colour, the stronger the expression.

Result images are analysed manually. A human expert determines in which structures the gene is expressed, and at what level of expression. As strength information is not the main focus of the experiment, its description is often vague using loose natural language terms such as strong, moderate, weak or present. For example, the gene bmp4 is strongly expressed in the future brain at Theiler Stage 15.

Once an in situ gene expression experiment is completed the experiment may be published in a traditional journal. Regardless of whether or not this is true, the experiment usually will be published by one (or more) online resources. Two of the main resources in the current domain of interest are EMAGE [2] and GXD [3] - both use the EMAP anatomy ontology. These online databases publish so-called annotations that contain particular types of information: provenance, details of the technique used, analysis of the result, and perhaps some indication of how reliable the resource believes the experiment to be. It is possible to supply information directly to the resources, and thus omit the traditional journal publication. Often such a route is favoured by large-scale projects that conduct a great number of experiments.

Although EMAGE and GXD are substantial resources they cannot be considered complete [4]. There is a range of reasons for this phenomenon including: some large scale projects publish their own results in a proprietary database, some experiments are deemed of insufficient quality by the resource curators, and others simply 'slip under the radar'. Consequently, in order to build as complete a picture of the domain as possible, it is necessary to consult multiple resources.

In addition to being incomplete, online biological resources are often inconsistent [4]. In terms of gene expression, this means that the same resource publishes one annotation suggesting the gene is expressed in a particular structure, and a second annotation suggesting it is not. As biologists treat absent and expressed as mutually exclusive, this implies an inconsistency. Due to the complexity of the underlying experiments there are a number of possible reasons for the different results, including: differences in the interpretation of results, unrecognised differences in the experiments, and human error (by either the research team or the resource's curators). As discussed previously, it is necessary to use synchronously multiple resources, but doing so raises the prospect of inconsistency between those resources, in addition to the inconsistency inside each resource.

Although there are many different methods to tackle the issues described, the use of argumentation [5] is considered in this work.

Argumentation

Argumentation [5] is a multidisciplinary field that studies arguments and arguing.

An argument is a reason to believe something is true. This may be a formal proof, or a piece of natural language: for example, a reason to carry an umbrella. The crucial attribute of an argument is its defeasibility: an argument may provide a reason to believe something is true, but it does not prove it is definitely true.

Defeasibility is important because arguments can be in conflict: one argument may suggest a conclusion is true whilst a second argument intimates that the same conclusion is false. Often it is desirable to make a decision about the validity of a conclusion. As the evidence and corresponding arguments change over time, it may be necessary to revise that decision. Defeasibility allows this to happen.

Arguing (commonly called argumentation) is the process of using arguments to justify a point of view. This process may take place between multiple agents (human or software) inside a debate, or it may be carried out by a single agent: e.g. a political speech justifying the government's decision to increase taxes.

The subdomain of computational argumentation involves the use of computers for constructing and using arguments. There is a wide range of domains in which argumentation has been applied including: Artificial Intelligence & Law [6], ontology matching [7], medical decision support systems [8], and agent communication [9].

Argumentation in relation to biology is surprisingly rare. Jefferys et al. [10] use argumentation to analyse the output of a protein prediction tool. However, most work involves pedagogical efforts to improve the construction of natural language scientific arguments by students, e.g. [11].

Although argumentation is rare in biology, it is common within the medical world and is employed in a wide range of decision support systems, for example [8]. Additionally, argumentation can be used to persuade patients to change their behaviour [12], and generate pamphlets to explain complex issues to patients [13]. Although clear differences exist between medicine and biology, there are many parallels. If argumentation has been shown to work well in medicine, it should be relatively successful in similar domains, e.g. biology.

Arguing over gene expression information

Initial attempts to tackle the twin problems of inconsistency and incompleteness are documented in length in previous work [14–16]; a brief reprise is given here.

McLeod et al. [14] describe an earlier system designed to tackle the above problems. In effect, this system allows a user to enquire if a gene is expressed in a particular anatomical structure from an individual Theiler Stage. Doing so causes the system to generate a number of arguments, evaluate those arguments, and present the results (argument(s) and associated evaluation) to the user.

For the arguments to be meaningful, they have to be based on expert knowledge - the expert employed for this task was the senior editor of EMAGE. Expert knowledge is captured in a series of natural language inference rules, so-called argumentation schemes [17]. Essentially, these schemes provide a natural language if-then (modus ponens) rule and an associated series of questions that can be asked to ensure the rule's application is suitable in the current context. Additionally the expert assigns a degree of confidence to each scheme.

The natural language schemes are converted into a logical form using the method described by Verheij [18]. The logic in question is a PROLOG-like logic employed by the ASPIC argumentation engine [8]. This tool provides a means to generate arguments, and conduct a virtual debate between two agents in order to determine which argument is strongest. Arguments are created by the ASPIC argumentation engine using the rules, and biological facts (information pulled dynamically at runtime from EMAGE and GXD).

The rules inherit the confidence values of their schemes, and likewise the arguments take the confidence values from their rules. When two arguments are in conflict, the engine compares the confidence values, and declares the argument with the higher value to be valid.

In McLeod et al. [14] the arguments are converted back into natural language and presented to the user, the presentation mechanism experts deemed most suitable. Figure 1 part A shows a screen shot of the results page: two arguments are displayed using one of ASPIC's in-built presentation mechanisms. Both arguments are undefeated, which means the system believes them to be true. Unfortunately, the default presentation style uses a mixture of natural language and logic rendering it unsuitable for use by the expected user group.

Subsequent work concentrated on the development, and evaluation of an improved interface (see Figure 1 part B for the updated results page). This time the arguments are presented entirely in natural language, and preceded by an image that summarises the argumentation, which is below a single line conclusion (the gene is expressed).

Sutherland et al. [16] discuss the inclusion of this work in a semantic web browser for the life sciences.

Full details of the evaluation can be found in Ferguson et al. [19]; though, McLeod et al. [15] publish the key findings. In particular, the following issues are prominent:

subjectivity - this affects the final conclusion (gene is/is not expressed), the schemes, and their associated degrees of belief;

breadth of resources - there is a need to include other gene expression resources for the developmental mouse, preferably featuring different techniques for measuring gene expression.

Tackling subjectivity

One of the problems with the previous implementations is the notion of subjectivity. During the evaluation it was clear that each subject had their own approach to interpreting the information contained within EMAGE and GXD. Accordingly, users disagreed with the system in three key areas: the argumentation schemes, the confidence values assigned to those schemes, and the final conclusion the system presented. Each area of conflict will now be addressed in turn.

Extending Argudas for richer argumentation

Argudas aims to improve on previous work with the integration of further resources - more resources means extra information and potentially richer arguments. Initially the microarray data contained in the ArrayExpress [34] resource was targeted. This highlights a number of integration issues that have yet to be resolved.

Firstly, the ArrayExpress resource does not use the EMAP anatomy ontology. Secondly, accessing the data held by ArrayExpress was difficult as they did not provide a direct programmatic access to their database. Instead access was via a RESTFUL web service, which provided limited functionality and did not allow access to the data required by this work. Initially it was impossible to ask for all the genes expressed in a healthy mouse's pancreas at stage 24 because ArrayExpress did not compute multi-factor statistics. That is, they computed which genes were expressed in the pancreas and which genes were expressed in stage 24 separately and there was no way of presenting the intersection at that time. The team behind the resource were working on improving this interface and claimed that such functionality would become available in the future; however, the delay was problematic because of the time constraints associated with Argudas. Finally, ArrayExpress had less data for the developmental mouse than expected: only three stages were covered. Comparing the costs and benefits it was decided not to pursue this integration further.

As work on ArrayExpress stopped an investigation of the Allen Brain Atlas (ABA) [35] and Gene Expression Nervous System Atlas (GENSAT) [36] began. Both of these resources are databases of in situ experiments focusing predominately on the adult mouse's nervous system, i.e. brain. The latter project makes available a full database dump. The former supplies an extensive range of RESTFUL interfaces that provide access to the desired information.

However, bringing the data from these two new resources into Argudas is problematic. Neither resource uses the EMAP anatomy - as these resources focus on the brain they have a far finer granularity for the brain structures than EMAP. Hence it is necessary to attempt some form of mapping from their respective anatomies to EMAP. Secondly, these resources use their own measures to describe the level of expression, GENSAT natural language terms and ABA floating point numbers, which also must be mapped across to the corresponding EMAGE/GXD terminology. EMAGE and GXD use very similar labels, for example when a gene is not expressed in a particular structure GXD describe the gene as absent whereas EMAGE use the term not detected.

Mapping between the different anatomy ontologies employed by the resources is based on a series of alignments produced by Jiménez-Lozano et al. [37]. As both GENSAT and ABA have a finer granularity than EMAP, mapping from those resources to EMAGE/GXD commonly results in a loss of precision.

The second task is straightforward for GENSAT as their choice of labels is similar to EMAGE's, and in turn GXD's. Whereas EMAGE has not detected, detected, weak, moderate, and strong GENSAT has not done, undetectable, weak signal, and moderate to strong signal.

Mapping EMAGE/GXD expression levels to ABA is a more complex task. There are three different measures of expression level published by ABA. Firstly there is the raw experimental information, secondly there is the average information (across all the experiments for a particular gene and structure), and finally there is a mathematical aggregation of the expression level and expression density. For current purposes, the first class of information is most suitable. Subsequently, the ABA generated expression level mappings must be applied. These mappings are a series of cut-offs that determine whether the expression level is not expressed, weak, moderate or strong. There are different limits for different parts of the brain. In order for the limits to be applied to the structures lower down in the anatomy hierarchy, the limits need to be propagated through the brain in a similar manner to the gene expression information.

When this work has been done it is necessary to determine what level of integration is appropriate for these resources. At the most basic level it would be possible to merely report the results contained in ABA and GENSAT. If either of these resources agreed with an annotation from EMAGE/GXD, it would increase the confidence in that annotation. Fully integrating ABA and GENSAT would require the generation of schemes for these resources, which is substantially more work and would necessitate involvement from a resource expert. In the case of ABA such an approach may not be fruitful; ABA does not publish all the data it collects, accordingly some of the attributes provided by an expert may be hidden from the public, and thereby Argudas. The restricted resources of Argudas meant that only the former option was realistic.

Although an interested biologist may raise a number of concerns regarding the anatomy and expression level mappings described above, currently there is no better way of aggregating data between the four resources of interest.

Further issues identified during the development of Argudas

The first Argudas prototype tests using the revised schemes and confidence values. For the reasons discussed before, it removes the final conclusion and instead presents a list of textual arguments. When this initial system was demonstrated to a group of end users they raised concerns over the number of arguments shown and the style of their presentation. These topics shall be explored in greater depth before a possible conclusion is presented.

The notion of argument reconsidered

Previous work, and the initial version of Argudas, use the ASPIC argumentation engine to generate and evaluate arguments inside a virtual debate. These arguments are presented to the user as a natural language paragraph - this display mechanism is chosen as it is the preference of the original expert. However, feedback suggests this choice is subjective [15]. Furthermore, the potential for a large number of arguments to be generated appears to imply that the original approach is sub-optimal. There is a clear need to find an alternative method for displaying arguments.

During internal discussions it was proposed that the argumentation mechanism should be reconsidered. This approach was based on the belief that users wanted quick access to certain key attributes of the annotation. Theoretically there is no need to employ the argumentation engine to create and evaluate arguments. Instead, the most important schemes (as identified by the scheme revision process), should be the basis for a range of key attributes that describe the annotation. The schemes indicate whether or not the information stored in EMAGE/GXD for a particular annotation should increase or decrease a user's confidence in that annotation. As such, Argudas should extract information from the resources and present it, with associated key highlights, to the user. It then becomes the duty of the user to evaluate the information.

In order to test this hypothesis two mock interfaces were created and evaluated. There are three steps to each interface. The first two steps are the same: select a gene and/or structure of interest; report on the available annotations and allow the user to ask for more information if desired. Figure 2 shows both of these: initially the query is bmp4 - future brain in all stages; the query causes all combinations of the gene and structure to be displayed in a table. The table presents all relevant annotations, summarises what each annotation shows, and provides a link to the resource's web page for that annotation.

In some situations the table in Figure 2 will be enough to resolve a biologist's question; i.e. it is clear that bmp4 is expressed in the future brain in stage 14. On the occasions when the table is not helpful, or does not provide enough information, clicking the argue button provides a range of arguments.

In the first mock interface a number of textual arguments are displayed - in a similar manner to Figure 1 part B. The second interface can be seen in Figure 3 - the 'arguments' are now a list of key attributes such as multiple annotations agree. Whether or not an attribute should strengthen a user's confidence in the annotation is indicated with a tick or cross. The attributes are divided into two layers - firstly by expression level, and then by annotation. For each level of expression there are three attributes that indicate how likely that level of expression is. Asking for more information causes the second layer of attributes to appear. This allows the user to evaluate the annotations individually, and collectively as a group that promotes a specific expression level.

Argument presentation reconsidered

The mock interfaces, shown in Figures 2 and 3, were explored with the help of two expert users. During this small informal exercise (described later in the Methods section), the test users were kept apart; however, they reached identical conclusions:

1. 1.

   the revised interface, using key attributes rather than textual arguments, is an improvement;

    
2. 2.

   a further improvement could be made by placing the content of Figure 3 into a table.

    

Separately, the experts both had the idea contained in Figure 4. Effectively, the important schemes become columns, with individual annotations represented as rows. Ticks indicate positive aspects that suggest the annotation is more likely to be correct, with crosses having the opposite semantics. Blue dashes inform the user that a piece of information is unavailable.

Enacting these recommendations has a major side effect: there is no need to perform computational argumentation. As such, there is no need for the argumentation engine. Instead the user now performs the argumentation themselves using the aggregated and curated information presented by Argudas.

This change helps address the issue of subjectivity, because the user is now able to apply his/her own confidence values and criteria to the decision making process. Furthermore, because the user is able to build his/her own arguments, Argudas becomes a more flexible tool. For example, it is now possible to use Argudas to decide where, and at what level, a gene is expressed.

Implementing these changes results in the second, and current Argudas system (the user interface can be seen in Figure 4). One addition to the advice offered by the expert users is the inclusion of mouse-overs to present extra information. Discrete dots underneath a tick/cross, or column title, indicate that more information is available if the user hovers his/her mouse over the dots. For example, hovering the mouse over the dots in the cell for 'strong - multiple annotations agree' (Figure 4, top table) reveals the experiment(s) that suggest the gene is strongly expressed.

The evaluation of this system, described later, demonstrates that the changes are both appropriate and effective because they allow users to quickly scan the important information.

Future work

This work set out to model expert knowledge and use it to reason with information sources available through the Internet. In the current use case, this appears to be beyond the scope of what a typical end user wishes. However, the current use case is relatively constrained, and thus contained: it focuses on one kind of gene expression information for one model organism. Extending the use case to include different types of biological information, for example gene regulatory networks, makes the use case considerably more complex. As the intricacy of the biological investigation increases, the need for user support likewise increases. Argumentation is one possible support mechanism.

Another avenue for future work relates to CUBIST, Combing and Uniting Business Intelligence with Semantic Technologies, an EU FP7 project that aims to combine the essential features of Semantic Technologies, Business Intelligence, and Visual Analytics. Data from both unstructured and structured sources will be federated within a Business Intelligence enabled triplet store, before visual analysis techniques such as Formal Concept Analysis [38] are applied. One of the project's three use cases involves the gene expression data described in this paper. Although it is early in the life of the CUBIST project, a semantic Extract Transform Load [39] process that includes computational argumentation may be envisioned. The inclusion of argumentation may provide an intelligent transformation of data, and a user-friendly explanation of the transformation.

Revising the argumentation schemes and confidence assignments with multiple experts

As discussed above, when the previous work was evaluated [15] a number of evaluation subjects disagreed with the expert's schemes and his assignment of degrees of confidence to those schemes. Argudas did not have the resources to create a new set of schemes as this was a sizeable task; nevertheless, it was possible to review the schemes and their confidence values. To this end, two experts were asked to appraise the list of previously generated schemes. Separately, each expert awarded each scheme a score:

0 disagree with the scheme, i.e. reject;

? don't know - scheme is very weak and is on the border between being rejected and being classified as a weak scheme;

1 weak scheme, i.e. low confidence;

2 moderate scheme, i.e. medium confidence;

3 good scheme, i.e. high confidence.

In total, the two experts were asked to assign a score to sixty-eight schemes. The experts completely agreed upon - that is they gave exactly the same score to - sixteen schemes. A further thirty-three schemes were assigned a similar score. The notion of similar being defined as an adjacent score, i.e. if one expert assigned 2, then either a 1 or a 3 would be classified as similar. If the two experts assigned scores that were neither adjacent nor exact matches, they were deemed to disagree - this happened with nineteen schemes.

In conclusion, the experts broadly agreed on 72% of the schemes. This left 28% of the schemes for which the disagreement was marked. Regrettably, one of the experts emigrated shortly after this exercise was completed and was no longer available to assist in the development of Argudas. Therefore this disagreement was never resolved, nor was its root cause investigated.

Potentially the source of the disagreement was very interesting, as it was not clear whether the conflict between the experts was caused by a genuine difference of opinion or a difference of interpretation. As the schemes were written in natural language, the latter is a distinct possibility.

Reconsidering the presentation of arguments

Originally arguments were presented entirely in natural language, e.g. Figure 1 part B. An alternative was to summarise the information by displaying only key attributes, e.g. Figure 3. This small informal exercise attempted to decide which mechanism was more appropriate for the potential users.

The two presentation styles of argument were evaluated with the assistance of two expert users from the Medical Research Council's Human Genetic Unit [41]. One expert had participated in the previous revision of argumentation schemes and confidence values. The second expert had attended a presentation on Argudas, but had not previously had a role in Argudas' development. The expert evaluations were undertaken independently with no discussion between the experts prior to the evaluation.

Each expert user was presented with a description of the planned evaluation, then a structured walkthrough was conducted. Using a protocol, the user was guided through each interface using the same search example: bmp4 - future brain - stage 15. They were asked to raise any issues or aspects they liked or disliked while undertaking the interface evaluations. The experts were then asked to score the interfaces out of ten in terms of their usability. Ultimately, a limited set of questions was asked to determine the user's opinions on the requirements for refining aspects of the interface and argument presentations.

In particular, one question directly asked which style of presentation the subject preferred. A second question asked how their favoured approach could be improved further. As discussed above, the users provided very similar answers, with a clear preference for displaying only key attributes. Additionally, both test subjects proffered the idea of using tabulation to improve the presentation of the attributes.

Evaluation of second Argudas implementation

An evaluation of the new interface and latest version of Argudas, represented by Figure 4, was undertaken with eight users (consisting of biologists, bioinformaticians, and computational biologists) working at the Medical Research Council's Human Genetics Unit. Work pressures necessitated the restriction of each evaluation session to twenty minutes.

The evaluation comprised a structured walkthrough with the user being guided through a scenario of a typical Argudas query. This was followed by a usability evaluation questionnaire and a number of open-ended questions to gather the user's opinions on the presentation of results. Due to the small number of users, results were largely qualitative rather than quantitative.

Users were asked to rate their opinion of a number of aspects of the Argudas system using a five point scale. Considering the response time of the system: seven users rated it as "excellent" and one as "good". Focusing on the appearance of Argudas: one user described it as "excellent" and seven as "good".

The user ratings for how easy the three results tables were to understand favoured "easy" and neutral, with only two users finding the second and third results tables (e.g. Figure 4) "difficult" to understand.

Whilst conducting the evaluation, the observer noted that five of the users appeared to not see the key explaining the tables, suggesting that the key should be placed in a position on the page that makes it more prominent.

Failure to spot the key meant that the same users did not read the explanation of the mouse-overs that it contained. This was believed to be one contributing factor to the apparent inability of several users to discover the mouse-overs. Another relevant factor may be the deliberately discrete nature of the underlining used to indicate the presence of a mouse-over. Attempts to ensure the underlining did not intrude too significantly, appear to have made the visual clue too difficult to see.

A number of general conclusions were drawn from user comments made during the evaluation. For example, the idea of triggering help information by mouse-over. Additionally, several users felt the results tables were too big, with the information too widely spaced. These users recommended a reduction in white space and cell size. When a user asks for more information from the top results table (e.g. Figure 2), the top table is hidden to allow more space for the presentation of the two tables from Figure 4. To see the first table again, the user must click on a hyper-link ("Show/hide gene expression results" in Figure 2). It transpires that this hyperlink should be more visible.

Overall, the conclusion drawn from the evaluation is that the evolution of the interface to a tabular display has achieved the desired aim of making the results easier to "eyeball". Accordingly, user satisfaction with the tables is largely positive. However, the mechanisms that provide the user with extra information about, and context for, the results need to be improved.