In this paper we present SMART Protocols, a semantic and NLP -based infrastructure for processing and enacting experimental protocols. Our contribution is twofold; on the one hand, SMART Protocols delivers a semantic layer that represents the knowledge encoded in experimental protocols. On the other hand, it builds the groundwork for making use of such semantics within an NLP framework. We emphasize the semantic and NLP components, namely the SMART Protocols (SP) Ontology, the Sample Instrument Reagent Objective (SIRO) model and the text mining integrative architecture GATE. The SIRO model defines an extended layer of metadata for experimental protocols; SIRO is also a Minimal Information (MI) model conceived in the same realm as the Patient Intervention Comparison Outcome (PICO) model that supports search, retrieval and classification purposes. By combining comprehensive vocabularies with NLP rules and gazetteers, we identify meaningful parts of speech in experimental protocols. Moreover, in cases for which SIRO is not available, our NLP automatically extracts it; also, searching for queries such as: What bacteria have been used in protocols for persister cells isolation is possible.
Experimental protocols are fundamental information structures that support the
description of the processes by means of which results are generated in experimental
research [
In this paper, we present the semantic and Natural Language Processing (NLP)
components for SMART Protocols (SP); we want to facilitate the semantic
representation and natural language processing for these documents. Our NLP layer makes use
of various ontologies as well as of the Sample Instrument Reagent Objective (SIRO)
model for minimal information (MI) that we have defined. Our work is based on an
exhaustive analysis of over 400 published experimental protocols1 (molecular biology,
cell and developmental biology, biochemistry) and guidelines for authors from
repositories and journals, e.g. Nature Protocols2, Plant Methods (Methodology)3, Cold
Spring Harbor Protocols4. Moreover, the SP ontology and SIRO are built upon
experiences such as those under the BioSharing umbrella, e.g. the MIBBI project [
We have carefully considered and reused several ontologies, for instance: i) The
Ontology for Biomedical Investigations (OBI) aims to provide a representation of
biomedical investigations. OBI builds upon BFO and structures the ontology by
using ”occurrences” (processes) and ”continuants” (materials, instruments, qualities,
roles, functions) relevant to the biomedical domain [
Unlike other approaches, the SP ontology is an application ontology designed to support NLP over experimental protocols, publish protocols as Linked Open Data (LOD) and annotate and classify these documents according to particularities in their workflows. The SP-document delivers an structured vocabulary for representing a specific type of document, the protocol. We extend the IAO to provide a structured vocabulary of concepts to represent the information that is necessary and sufficient for describing and experimental protocol as a document. In addition, the SP ontology also considers the protocol as an executable element to be carried out and maintained by humans it may also be transformed to workflow languages used by enactors such as robots or machines per se, the SP ontology is not a workflow language but a model.
For the representation of instructions in the SP-workflow module, we expand the
class ”experiment action” from EXACT and reuse classes from OBI, the BioAssay
Ontology (BAO) [
SIRO has been conceived in a way similar to that of the Patient Intervention
Comparison Outcome (PICO) model; it supports information retrieval and provides
an anchor for the records [
For convenience, in this paper we use the protocol ”Extraction of total RNA from
fresh/frozen tissue (FT)” [
The development of the SP ontology, was the first step [
The SMART Protocols (SP) ontology is an application ontology designed to
facilitate the representation of experimental protocols in two ways, as a document and
as a workflow [
The workflow module9 represents the ”steps/instructions”, ”actions” and experimental inputs such as ”reagents”, ”instruments”, and ”samples/speciments” for enacting the workflow. The representation of executable elements of a protocol (instructions or steps), is modeled by reusing and expanding ”experimental actions” from EXACT; in addition, we reused and expanded terminology related to ”instruments”, ”reagents/chemical compounds”, ”organisms” and ”sample/specimen” from ontologies such as: OBI, BAO, EFO, ERO, NCBI Taxonomy and ChEBI. The representation of steps/instructions is modeled with the class sp:protocol instruction, The property sp:has instruction, is used to define instructions involved in protocols. The order in which these instructions should be executed is captured by the property ”is preceded by” and ”precedes” from BFO.
Our running example, ”Extraction of total RNA from fresh/frozen tissue (FT)”
[
the SP-Document module is presented in table 1 and Fig. 1; metadata elements are organized in SP-Document as ”textual entities”. Modeling the workflow aspects, SPWorkflow module is presented in table 2 and Fig. 2; the first column in table 2 includes frequently used instructions that should be executed in protocols for the extraction of nucleic acids. The second column includes instructions extracted from the protocol. Olga Giraldo domain expert and author of this papersemi automatically extracted the information for each metadata element and identified instructions frequently used in different types of protocolsmainly in molecular biology.
Extraction of total RNA from fresh/frozen tissue (FT) Kim M. Linton, Yvonne Hey, Sian Dibben, Crispin J. Miller, Anthony J. Freemont, John A. Radford, and Stuart D. Pepper
DOI:10.2144/000113260 sp:title of the protocol sp:author name sp:protocol identifier Descriptive metadata sp:application of the protocol ”Methods comparison for high-resolution transcriptional analysis of archival material on Affymetrix Plus 2.0 and Exon 1.0 microarrays” ”The extraction method (steps 221) is taken from sp:provenance of the protocol the method supplied with TRIzol reagent (Invitrogen, Paisley, UK).” Metadata about the materials used sp:specimen name ”tumor tissue” sp:reagent name ”TRIzol” sp:manufacturer ”Invitrogen” name sp:reagent name ”Chloroform” sp:manufacturer ”Sigmaname Aldrich” sp:reagent name ”Ethyl alcohol” sp:manufacturer ”Sigmaname Aldrich” sp:reagent name ”Isopropyl alcohol” sp:manufacturer ”Sigmaname Aldrich” sp:equipment or supplies name ””FToisrscueepss”t,or”aSgcealcpoenl”ta,i”nSecra”l,p”eHlohmolodgeern”izer blades”,
SIRO represents the minimal information for describing an experimental protocol. In doing so, it serves two purposes. Firstly, it extends and structures available metadata for experimental protocols, for instance, author, title, date, journal, abstract, and other properties that are available for published experimental protocols usually as PDFs. SIRO extends this layer of metadata by aggregating information about Sample, Instrument, Reagent and Objective hence the name. If this information is part of the abstract or the full content, SIRO extracts and structures it as Linked Open Data (LOD). Secondly, SIRO, in combination with NLP and semantics, provides an anchor and structure for the minimal common data elements in experimental protocols. This makes it possible to find specific information about the protocol; if the owner of the protocol chooses not to expose the full content, as in the case of publishers and/or laboratories, SIRO may be exposed without compromising the full content of the document.
SIRO was developed after the SP ontology; Fig. 3 (step 2) illustrates the
development process. The identification of common elements involved the following activities.
Our ”kick-off ” phase started by redefining the use cases focusing on the identification
of commonalities; it also entailed preparing the material to be used, e.g., ontologies,
protocols and planning. Our main input was the SP Ontology and our corpus of
documents. We then started to manually identify commonalities across protocols and,
map these to the SP ontology as well as to ontologies in Bioportal10, and OntoBee11.
This Domain Analysis and Knowledge Acquisition (DAKA) phase allowed us to
gather common terminology with a raw classification. Our Linguistic and Semantic
Analysis (LISA) was carried out in parallel with DAKA. LISA allowed us to
automatically classify and identify the terminology we were gathering; LISA was extensively
supported by GATE [
The SMART Protocols ontology models the execution of the workflow and relates reagents and instruments to steps in the workflow. Ontologies provide the gazetteers with the necessary knowledge for annotating beyond entity recognition. The gazetteers 10 http://bioportal.bioontology.org/ 11 http://www.ontobee.org/index.php make it possible to identify SIRO elements in the narrative; they are structured with information such as definition, URIs, provenance, synonyms, etc. Gazetteers based on ontologies have context; rules making use of these gazetteers find meaningful parts of speech in the text. GATE uses the gazetteers and the rules for annotating the documents. In this way, it is possible to differentiate between ”centrifuge” as an action and ”centrifuge” as an instrument. Furthermore, the rule engine in GATE, in combination with the gazetteers, make it possible to find statements related to, for example, ”precipitation instructions” that include words related to an action such as ”centrifuge” and a reagent like ”isopropanol ” (used to facilitate the precipitation of DNA) see Fig. 6 for a more complete example.
The development of the SEmantic GAzetteers (SEGA) was a very complex and domain knowledge intensive activity. Developing the gazetteers entailed the identification of ontologies with terminology related to ”sample/specimen” (including organisms), ”instruments” and ”reagents”. Ontology repositories and their corresponding Application Programming Interfaces (APIs) were reviewed so that the process could be automated see Fig. 3, step 3 ”Review of Ontologies”. The ontologies identified during the development of SP ontology were then more carefully inspected; overlaps were identified, and availability of metadata for each term, object properties and complexity in the classification were addressed. For ”organisms” (related to sample/specimen), the NCBI Taxonomy was chosen. For ”instrument ”, the choices included EFO, ERO, OBI and SP ontology. For ”reagents” and ”chemical compounds”, ChEBI and SP were selected Fig. 3, step 3 ”Selection of ontologies”. During the stage ”Extraction of Terms” see Fig. 3, step 3, we focused on enriching the terminology; depending on the limitations of the SPARQL endpoints for OntoBee and Bioportal, we were using either SPARQL queries or locally parsing the ontologies. The terminology was gathered with the corresponding annotation properties. Axioms and annotation properties were used to, for instance, discriminate if a term is synonymous with another term due to a case of acronyms or common name.
At this point, we had some gazetteers with over half a million terms. For quality
control, we then started the depuration of the terminology. We removed the terms
that had comments from the curators about the suitability of the terms in specific
sub-domains. For instance, the class ”cell harvester ” (OBI 0001119) has a specific
comment ”A device that is used to harvest cells from microplates and deposit samples
on a filter mat. NOT AN INSTRUMENT ”. We also removed terminology that was
reused across ontologies. For instance, the OBI class ”thermal cicler ” is reused by SP
and EFO. In this particular case, we use the term only once and from the original
source -OBI. Classes with the same label represented in several ontologies with
different axioms were conserved. For instance, SP reuses from the Sequence Ontology (SO)
[
For Testing the Gazetteers, we followed a manual process against our corpus of documents. Documents were loaded into GATE, and words related to SIRO elements were identified and annotated. We evaluated the following aspects, i) execution time, ii) correctness in the annotation of the terms and their synonyms, iii) failures in the recognition of terms in the texts, and iv) identification of terms incorrectly annotated, namely a word with different meaning For example, the word ”cat ” is a term from NCBI Taxonomy used to represent the common name of ”Felis catus”, but cat (or cat., Cat, CAT) also represents the short word for ”catalog ”. From the gazetteers, linguistic patterns were identified so that The Iterative Rule Writing (see Fig. 5) step could commence.
We are using JAPE (Java Annotation Patterns Engine) to code the rules. In this stage, we are designing rules to automate the identification of meaningful elements in the narrative. This step runs iteratively with previous stages; as linguistic structures and meaningful PoS, e.g. instructions, are characterized, then rules are written, tested and improved. Ontologies and domain terminology is mapped to the corresponding vocabularies.
We have presented our approach to the Semantics for representing experimental
protocols, the SP ontology and the SIRO model. The SP ontology is composed of two
modules, namely SP-document and SP-workflow. In this way, we represent the
workflow, document and domain knowledge implicit in experimental protocols. Actions,
as presented by [
The SIRO model for minimal information breaks down the protocol in key elements that we have found to be common to our corpus of experimental protocols: i) Sample/ Specimen (S), ii) Instruments (I), iii) Reagents (R) and iv) Objective (O). For the sample, it is considered the strain, line or genotype, developmental stage, organism part, growth conditions, pre-treatment of the sample and volume/mass of sample. For the instruments, it is considered the commercial name, manufacturer and identification number. For the reagents, it is considered the commercial name, manufacturer and identification number; it is also important to know the storage conditions for the reagents in the protocol. Identifying the objective or goal of the protocol helps readers to make a decision about the suitability of the protocol for their experimental problem. SIRO and the SP Ontology facilitate the generation of a self-describing document with structured annotation.
Our NLP layer makes use of the semantics we have defined. We currently have six gazetteers with over 1.400.000 terms in all; these terms will be further refined and then added to the SP ontology. The gazetteers are currently reusing terminology from EFO, ERO, OBI, NCBI Taxonomy and ChEBI; we will continue adding terminology from other ontologies and also adding more documents to our corpus. We are making use of existing infrastructure provided by BioPortal and OntoBee. For managing large ontologies, we are not using their respective SPARQL endpoints but locally parsing them, e.g. NCBI Taxonomy and ChEBI. Our Semantics plus NLP infrastructure makes it possible to retrieve information where specifics from the protocols are used to construct the query. Our NLP layer is able to extract SIRO automatically; we have encountered issues with the free narrative often used for describing the objectives.
Experimental protocols are meant to capture a complex and nested set of roles actions, derivations of original plans, actions executed by personnel, robots taking care of some specific steps in the workflow, computational workflows often used in support of laboratory work, data being produced at every step of the workflow, etc. Representing and enacting all of these is not a simple task; laboratories require flexibility in their conceptual models so that parameterizing their own workflows won’t become an overwhelming task. The laboratories only carry out a limited set of actions over a limited set of samples; high-level abstractions for general process models are needed. These could be made more concrete as workflow constructs, samples, roles, actions, reagents, instruments, etc. are aggregated. Representing the execution requires the confluence of metadata that allows tracking of everything that has occurred, who has done it, how, where, etc. Our ontology model may easily be extended and adapted to these realities. The metadata schemata to represent laboratory protocols should be kept independent from the workflow enactors; robots will surely have their own procedural languages. The descriptive schemata should interoperate with the workflow enactors. The SP ontology was conceived considering all of these; our use cases are incrementally becoming more complex as we are moving from protocols published in journals to those registered in laboratory notebooks needless to say, that gaining access to laboratory notebooks is not easy.