How to use the SMAFIRA search engine

Please refer to the Externer Link:Data Protection Declaration.

Currently, you can conduct a similarity search using a PMID (PubMed identifier). Please go to PubMed first and identify an article that represents your research goal and can serve as a reference for identifying similar contents. Then, you can enter the PMID of the reference into the input field of SMAFIRA. 

A valid PMID retrieves a citation from PubMed. This is validated by SMAFIRA. Furthermore, the citation must contain an abstract and must be linked with similar articles.

A valid PMID retrieves a citation from PubMed. This is validated by SMAFIRA. Furthermore, the citation must contain an abstract and must be linked with similar articles.

No, each search (session) must contain only one PMID.

SMAFIRA performs an automatic classification to predict the experimental models mentioned in the abstracts, followed by a re-ranking according to the similarity to the reference abstract. For more details about our methods for classification and re-ranking, see "Methods and data" below.

Yes. You can start from a detailed summary of your experimental project in a language other than English. Then, you should use translation tools to translate your summary into English, and the Externer Link:MeSH on Demand tool to identify an equivalent citation in PubMed. Use the PMID of the latter citation for your search.

SMAFIRA presents the following information:

• Rank (number) of the citation in the list, according to the chosen method.
• PMID, with a link to the side-by-side visualization of the reference and similar abstract (user opinion page).
• External link to the article in PubMed. ([1])
• SMAFIRA Rank (default ranking): rank of the citation according to the SMAFIRA re-ranking.
• PubMed Rank: rank of the citation according to PubMed.
• My Rank: an icon that indicates the similarity between the reference abstract and the current abstract as judged by the user. This icon is only available after a user has activated the respective checkbox on the side-by-side comparison of a single article with the abstract.
• Title of the cited article.
• Labels: one or more icons that indicate the assigned classes of experimental models that were automatically predicted or specified by the user (no cogwheel icon).
• If available, icons that indicate whether the article is a review or that no abstract is available .

[1] The icons in this application were made by Externer Link:Freepik, Externer Link:Good Ware, Externer Link:juicy_fish, Externer Link:Roundicons from Externer Link:www.flaticon.com.

This occurs when none of the articles in the list mentions the respective experimental model in the abstract text, i.e., it is not possible to filter the list by this label.

By default, the tool ranks the citations according to the SMAFIRA re-ranking. However, the user has two other options for ranking the articles: (a) the PubMed ranking, i.e., the one that is carried out by the PubMed ranking algorithm; and (b) the user-defined ranking (My Rank) based on the similarity validation, and in this order: Thumb up, thumb up/down, thumb down. 

For the sake of completeness, we do not remove them, but they are usually located at the end of the list. Since no abstract is available, we can neither predict the labels nor calculate its rank with the SMAFIRA re-ranking.

We give a short definition below, please refer to our Externer Link:publication for details.

• in vivo: living vertebrates (including cephalopodes)
• tissue/biopsy: isolated vertebrate organs and tissues
• primary cell: finite vertebrate primary cells or stem cells
• immortal cell: established vertebrate cancer or immortalized cell lines
• invertebrate: invertebrates or invertebrate material (excluding cephalopods)
• human: human/patients and or any kind of human material (including organs/tissues, primary cells and immortal cells)
• virtual model: computer simulations
• others:  other types of research, including observational studies 

We chose labels that indicated not only the level of invasiveness but also the species or the origin of material used. Please refer to our Externer Link:publication for details.

Yes, please save (or bookmark) the URL of the session that is specified at the top of the results page.

Yes, please click on the export icon on the top right of the results page. All articles in the list will then be exported, along with the feedback provided by the user for the labels and similarity.

We show a side-by-side visualization with the reference abstract (on the left) and the abstract from the currently chosen citation (on the right). On the top left is a link of results to go back to the list of citations, while on the top right are navigational links to the previous  (if available) and the next  (if available) citation. The page includes the possibility to give feedback on the similarity between the two abstracts (see details below). For each abstract, we present the following information:

• Information about the automatically predicted classes of experimental models.
• PMID of the citation and link to it in PubMed.
• Title of the corresponding article, as available in PubMed. 

Yes, the user can override the automatically predicted labels. Any changes to the labels are automatically saved within the context of the respective user. To suggest a global change of a prediction – affecting all users of SMAFIRA – the change has to be forwarded to the SMAFIRA team, by sending the session URL to the contact e-mail. Then, the suggestion will be considered during a follow up training of the prediction model. 

We consider two abstracts as similar, if the described scientific objectives agree. Please refer to our Externer Link:publication for a stringent specification of similarity (“equivalence”). 

Simply choose one of the three options (thumb up – similar, thumb up/down – uncertain, thumb down – not similar). The selected one is automatically saved. It is possible to clear your selection by clicking on the "clear" link next to the options.

First of all, users can use their own judgement of similarity to rank citations (“My rank”). Furthermore, user feedback will be shown in the other rankings (“SMAFIRA” rank, “PubMed rank”) as well. Thus, users can evaluate the performance of the ranking algorithms: the more citations on the top positions are judged as similar to the reference, the better. If users send us the URL of their session, including such feedback on similarity by e-mail, we can internally use the data to train or evaluate the prediction models of SMAFIRA.

We fine-tuned a machine (deep) learning model that is based on the Externer Link:Transformers framework. This is the current state-of-the-art approach for text classification tasks. The Externer Link:PubMedBERT language model was pretrained on PubMed® contents (MEDLINE®). For fine-tuning we produced a data set of annotated biomedical abstracts (see next question). The obtained model is used for a real-time prediction of the labels assigned to the reference and similar abstracts. Please refer to our Externer Link:publication for details.

As described above, the pre-trained language model was further fine-tuned with a collection of around 1,600 annotated biomedical abstracts (called the GoldHamster corpus). A team of more than 10 experts manually annotated the abstracts with labels according to the mentioned experimental model(s). Please refer to our Externer Link:publication for details. 

Users can override the automatic prediction on the user opinion page. If the URL of the respective session is sent to our contact address by e-mail, we can use this information for future fine-tuning of our model. 

Not yet, but we plan to implement such a functionality in the future.

We start by representing each article as a vector of words whose represention uses Externer Link:biomedical word embeddings provided by the NCBI. Then we calculate the similarity score between the two vectors using a Externer Link:dot similarity from Python numpy. The tool ranks the citations according to these scores.  

We relied on our Externer Link:SMAFIRA-c dataset. It contains four use cases for which the top 100 PubMed similar articles were annotated regarding the similarity (scientific objective) of test and reference abstract. Please see our Externer Link:publication for details on our experiments with various methods.