Provenance Levels for Biomedical In-vitro Experiments

Provenance data are an essential information source of scientific experiments and thus for maintaining the credibility of research findings [1]. There are several approaches for capturing provenance information, that either focus on the process (workflow provenance) or the actual data transformations (data provenance) [2]. The first is mainly used for describing activities including agents and their interdependencies by means of ontologies, such as PROV-O [3], [4], to answer the W7-questions [5]. The latter is grounded in database theory and employs a tuple based reconstruction to answer the questions why, how or where a data item/query result was created [6]. Thus, workflow provenance fails to answer questions about particular atomic data tuples, such as "Which parts of the concrete dataset were analyzed statistically?". On the other hand, the data provenance fails at incorporating the procedural view, such as who was involved in the analysis or relation between several datasets. We propose to combine these two provenance types into a unified framework enabling an integrated view on different granularity levels of the scientific experiments. In particular, we propose to include the modeling of data provenance similar to the approach of Ives et al. [7] in the PROV-O workflow provenance approach. This unified framework enables the reasoning over W7+1-questions [8] and their combinations in the particular level of detail needed, e.g. to identify the necessary parts of data sets for a particular analysis and the involved workflow including responsibilities and experimental settings. In order to illustrate the conceptual framework, we employ the use case of in-vitro experiments from the biomedical domain. We concentrate on in-vitro experiments, i.e. performing lab experiments including measuring characteristics followed by analyzing the data. In the biomedical domain these results are often used in simulations and the simulation findings than proven in lab experiments again. An excerpt of our concrete use case is shown in Figure 1. The upper part encodes the analyzation step of the raw data in terms of workflow provenance. While the raw data (prov:entity) are provided by the lab scientist (prov:agent), the analysis is performed (prov:activity) by the data analyst (prov:agent). The resulting data (prov:entity) are also attributed to the data analyst. The lower part of Figure 1 encodes the data provenances graph using our proposed modeling mechanism. In particular, the activities relate to database actions such as join (*) and union (+) of tuples. Additional comments specify the corresponding witnesses [6], which answers typical why-questions in data provenance. For the modeling of the different levels of details [8] including also workflow provenance and data provenance, we propose to employ the BFO [9] relations bfo:hasPart and bfo:partOf. These provenance levels are linked by dotted edges. In the example, the five data tuples on the data provenance level are bfo:partOf the verified data entity on the workflow level. In addition, we propose to employ bfo:hasPart between the levels of activities such as ∗ and the activity analysis. Summarizing this conceptual framework, it provides a unified view. Thus, it allows to reason over all W7+1-questions [8] including all necessary details without limiting provenance on either to process or tuple parts.