GoPubMed is a powerful adjunct to PubMed proper. It enables users to search from the get-go in a straightforward what, who, where and when fashion but then enables them to drill down to ever increasing levels of specificity. GoPubMed facilities contacts between users and the authors of the papers they are reading about in a rudimentary but useful social networking cum search engine fashion. It is most useful for power searchers in the health sciences. I had the opportunity to talk recently with Dr. Michael Alvers, CEO of Transinsight, the company behind GoPubMed.

The Interview

Before we get started Michael, I would like our readers to give a bit of background. I work in a medical library and spend most of my time in PubMed. Medical librarians are very attached to PubMed proper. So my first question is a tough one. Why would a sophisticated searcher looking for medical information use GoPubMed instead of PubMed? Or do you see medical librarians, physicians and clinical researchers automatically repeating their searches in GoPubMed as a way of ensuring the best results?

The semantic approach of GoPubMed is fundamentally different to that of a key word search engine like PubMed. GoPubMed uses background knowledge in the form of semantic networks – also called ontologies – to make search intelligent. For example, a search in PubMed for “Heart Diseases” yields only 52,738 results. GoPubMed delivers 834,687 results. Why do we get so many more results? It’s because we include all 544 ontological sub-concepts from MeSH and their synonyms found under the concept “Heart Diseases.” This is the big difference between semantic search and key word search and it guarantees completeness of results – of course only as complete as the background knowledge used is.

Then, GoPubMed goes further by providing a great navigational instrument. The tree on the left bar, allows users to narrow down huge result sets with only a few mouse clicks. It also works as a facet filter; it shows you something like a statistical distribution of your results. The Top Categories list shows the most important concepts in the semantic network. In PubMed you just get a list sorted by date.

Based on this semantic analysis, we also provide some very useful statistical information on top authors, journals and cities, and networks of co-authorship. None of these features are available with PubMed. Using GoPubMed makes searching more efficient and saves time, which is important to all scientists.

Here is a real world example. Say, I want to learn about the possible use of stem cells in the treatment of amyotrophic lateral sclerosis (ALS). Can you provide instances of where GoPubMed would surpass PubMed vis-a-vis results?

Type “Amyotrophic Lateral Sclerosis” in GoPubMed (the auto completion feature in GoPubMed should suggest the full term after typing “Amyo” on the second position). Then navigate the tree on the left bar to anatomy > all of anatomy > cells > stem cells (as shown in the figure below at left). The children of stem cells are fibroblasts, embryonic stem cells, mesenchymal stem cells, etc. Mousing over shows a tooltip (see the figure below at right) with descriptions of the different types of stem cells. A click on stem cells brings up 133 relevant articles. Since you are interested in treatments, type “treatment” in the search box (auto-completion suggests Therapeutics (Treatment) as concept). The Find It button delivers 47 highly relevant articles.

One feature I particularly like about GoPubMed is the “Contact” buttons. That is a pretty exciting, powerful feature. For instance, I did a search on ALS and found that a certain PN Leigh is a leading expert on the condition. I clicked on the contact button and got a box that said, “To contact Leigh, PN fill in the form. We will forward your message.” Can you tell us about how you set up relationships for this mediated form of communication? Do you have relationships with universities and labs worldwide, with scientific societies and/or individual researchers?

This is indeed an exciting and powerful feature. We decided to link people through us rather than allow “spam emails.” We filter for that, but I must say that no one has abused the service so far.

Originally, we had planned to expand GoPubMed to include a Facebook-like feature for biomedical people. But based on feedback we received from users, we realized that they were not yet ready for a mixed search engine/community platform. We decided to put that feature on hold and just left the top-author feature online, but we might change that back again.

What are the underlying relationships and technologies of this feature?

Again, semantics is the key technology used to provide this service. Author disambiguation is an especially tough task. Keep in mind that, for example, the author name Lee, S. exists 17,835 times and Smith, J. even 20,062 times in all almost 19 Million PubMed abstracts. Scientists change places, research topics and collaborators, which makes it extremely difficult to tell who is who. We reach about 85 to 90% accuracy. I think achieving more accuracy based only on PubMed data is also hard for human curators.

Have you any evidence that this sort of mediated social networking is leading to fruitful ongoing relationships? Have the researchers who are thus contacted been pleased with the results? Who is contacting them and with what sorts of inquires?

For now, we see very good traffic – which means a lot of work for us – but we can’t say whether this leads to fruitful relationships. However, our customers do use this feature extensively to find the right experts. We do not comment on the content of requests, and most often we don’t know what the content is. We just forward the contact request to the author and amazingly 100% agree to be contacted.

Do any of your competitors offer this service? It seems unique to me and could lead to the sort of advance in patient empowerment and productive interactions between those afflicted with certain conditions and those researching those conditions that is promised by firms like Private Access, which connects researchers with those interested in enrolling in clinical trials.

No, the service is unique in this semantic form. We have many requests from platforms for medical doctors and see a strong need there. I assume we’ll OEM our technology to those companies and I expect it’ll cause a big impact as a highly efficient tool for finding the right specialists.

Please discuss how you see search engines and social networking coalescing. Heretofore, they have been pretty distinct entities.

I personally believe in altruism. Wikipedia is a good example that it can work. I see a future in which the community creates semantic networks and ontologies, and since these enhance searches so drastically, there is a big enough benefit for creators to contribute. Of course, vanity also plays a role; people enjoy ranking high on contributor lists. And why not take advantage of that.

Prerequisite is an all-in-one system with search engine, semi-automated ontology creator, and integrated ontology editor and curator module. GoPubMed PRO [Transinsight's semantic federated enterprise search product] provides all these features and customers appreciate this a lot.

Tell us, more, please about the social computing aspect of GoPubMed. For instance we read, “As a curator you can curate our search engine.” Tell us about that. What are the advantages to all GoPubMed users of this feature and how does it work in practice vis-a-vis quality control and user loyalty and buy-in?

As stated above, the Wiki approach is a good example of free contributions to help all. Our curation-mode allows curating our non-perfect text mining. We now reach about 90–95% fmeasure but not 100%. Manual curation helps to understand the missing 10–5% and to fine tune our algorithms. The response rate here is also fantastic! 98% of the people asked contributed to curating their own publications. The community profits from this effort and we improve our free service. Our customers benefit indirectly from better text mining algorithms. It’s a win-win situation for all.

One of the biggest problems in marketing search engines is simply explaining persuasively and understandably to potential users why a search tool is so obviously superior by far to its competitors. Can you give a brief description of the following terms with specific examples of how they play out in the world of GoPubMed and why that should matter to those who have never used GoPubMed?

We’ll let computers do the hard work. After “OG,” you will find the definitions that our GoPubMed ontology editor and the semi-automated ontology generator offers, followed by my remarks.

Taxonomies and ontologies

OG: Taxonomy is the science of classification according to a pre-determined system, whose resulting catalogue is used to provide a conceptual framework for discussion or analysis.

OG: Ontology is an explicit, formal representation of the concepts, objects or other entities in a particular domain and the relationships among them.

Semantic networks of concepts would have been a better and more understandable term for what today we often call taxonomies and/or ontologies. I’ll give you a simple example to explain why such a network improves the search process. Consider searching for investments in biotechnology in Germany, in 2008. With traditional technologies, you would not find this article: “GlaxoSmithKline (GSK) invests 94 Million Euro in its Sächsisches Serumwerk Dresden”. Why not? Because the keywords “Germany” and “Biotechnology” are not mentioned at all. The knowledge that GSK is a pharmaceutical company and Dresden is a German city (besides 16 others in the U.S.) helps to classify this article as (semantically) important. We, as humans, have such networks in our heads and make those connections; computers should use existing knowledge as well.

Controlled vocabulary

OG: Controlled vocabulary is an established list of words and phrases (generally referred to as subject headings or descriptors) that provide a standard vocabulary used in a database.

The GeneOntology is a good example.

Text mining

OG: Text mining is an emerging field at the intersection of several research areas, including data mining, natural language processing, and information retrieval.

Text mining is often referred to the process of finding facts in texts. For example, ” … pifithrin-alpha, a P53 inhibitor …” (PMID 19360335) gives a “link” between P53 and pifithrin-alpha in this context. In GoPubMed, text-mining is referred to semantically matching ontologies with concepts in texts. A semantic search engine has to be smart enough to distinguish between Jaguar, the animal, or a car, watch, boat, scissor, harvester, guitar, operating system, caddy and more. Ken and Barbie, in most biological contexts, are proteins. Yes and no are also proteins and have to be identified as such by the machine context sensitive. The task seems to be simple since humans are very good at combing knowledge and reasoning, but bringing all this to the level of computation is not trivial. Still, I’m very confident that for many applications it’ll happen in the next ten years or so.

Semantic Web applications

OG: Semantic Web application is an application using semantic and distributed technology.

The idea of Tim Berners-Lee was to get more use out of the/his internet, which literally translated form Latin means “the net(work) between”. But as we know, today the internet is a connector for hyperlinks and emails, but not for applications on content. In short, the semantic web is the idea – not the reality – of bringing more meaning to a huge collection of texts. The next logical and doable step is making online texts more useful by extracting information and maybe turning this information into actionable knowledge. GoPubMed is a great tool, even though we’re still in the early stages. I think it’s comparable to the Wright brothers about 100 years ago. The first flight was only 13 seconds … We are just at the beginning of the semantic age and I’m excited thinking about the developments to come.

Please give an example of an ontology-based literature search. Say, I am a nurse researcher trying to determine whether any studies have been done on nutritional issues in post-partum depression. Would I want to do an ontology-based literature search? What questions would I ask? And would GoPubMed be the place to ask them? If so, why?

Let’s start by typing the query post partum depression. We see that 3,143 articles have been retrieved. The same results are found in Pubmed.

Now, you asked, “Would I want to do an ontology-based literature search?” Absolutely! Because now you can navigate the ontologies looking for “nutrition” and all aspects of nutrition, and filter results according to your needs, without typing one more keyword.

If you are familiar with the GO and MeSH ontologies you can open the tree and click on it. If you don’t know their structure, then use the option “find related concepts …” at the top of the tree. Type “nutrition”.

All facets of nutrition appear on the tree: nutrition therapy, processes, disorders, metabolic diseases, phenomena, overnutrition, etc. You can mouse over a term to read its definition. By clicking on the concepts, GoPubMed will filter the results.

In terms of what questions to ask and why GoPubMed would be the place to ask them, let’s look at some questions and I’ll show you why GoPubMed is the place to go.

Show evidences between maternal depression and malnutrition => filter with “malnutrition” => 14 results are retrieved

Which nutritional disorders are related to postpartum depression? => filter with nutrition disorders => 34 of 3,143 articles related to “Nutrition Disorders”

Even if you add more keywords, like nutrition, you cannot not reach the results that you do using ontology-based search. You miss many, many relevant papers because you miss all synonyms, natural language, spelling variations, and related terms to this concept, known as “children.” For example, obesity is a child of nutrition disorders. In keyword search, if you type “nutrition disorder” you will not find any of the papers where obesity appears.

Ontology-based search is more complete than keyword search.

Compare the results using keyword search (left) and ontology-based search (right):

Results using ontology-based search: type one keyword and navigate the results (nutrition is used in this example). Lots of papers were found by means of synonyms and related concepts (left).

How does knowledge-based search differ from search in general? Can you give a specific example? Please use juvenile diabetes as an example.

Here is a request by one of our OG users:

Email from a user:

[I] Searched the term “diabetes” and then filtered results for the protein “insulin.” That resulted in 311 of the last 10,000 from 326,718 articles related to “insulin,” which seem incredibly low for the term insulin in general. However, insulin as a MeSH term under “Chemicals and Drugs” results in 2,319 articles. How is it possible that you can extract the term insulin specifically as a protein from 311 of the abstracts?

Our answer:

The Ontology term “insulin” contains synonyms and children with synonyms. By searching, it produces an expansion of terms and approximately 2,300 papers (of the last 10,000 from 326,773 related to diabetes) are found. In order to identify the GoPubMed term “insulin” as a protein, the context of the paper is semantically interpreted. Papers where insulin is referred to as a protein are shown and in this case, approximately 300 papers are found.

How does GoPubMed differ from the natural language processing health search engine of Cognition Technologies vis-a-vis underlying technologies and results?

Cognition’s white papers show the company is seriously trying to understand the semantic problem. But we think that trying to understand the problem from a theoretical point of view is not enough. At the moment, the demonstrator (in biology) shows rather intransparent results for the user. Semantic approaches should show the user the complexity of semantics, meaning, and disambiguation. We have to “understand and interpret” what the user needs and show this in a rather simple way.

Examples that demonstrate the problems:

1. Cognition uses synonyms and definition, but the source they come from is not transparent.
2. The query syntax is different from PubMed; users are familiar with PubMed grammar for searching.
3. The option “Use drop down menus to change the meanings” is unclear to us. Especially if only one “meaning” is shown. For example, in the case of the query “heart diseases,” the meaning is “1) n: a circulatory system part condition.” By consulting the definition of heart diseases in MeSH you find: “Pathological conditions involving the heart including its structural and functional abnormalities.”
4. In the help section, we find this example: “genetic correlates of diabetes.” The results state:
   The following word meanings were selected. Use dropdown menus to change the meanings.

   Genetic: 1) adj: related to genetics: the deformation was a genetic defect
   Correlate: 1) n, v, tr: find connection: the study correlates smoking with cancer
   Diabetes: 1) n: a gland disease

   For me this is an example of user-unfriendly feedback. In GoPubMed, you can easily find the genetic correlates of diabetes by searching “Genetic Predisposition to Disease” [mesh] and the user can filter or navigate all types of diabetes, i.e. diabetes mellitus, type 1, 2, etc.

5. The ranking algorithm is intransparent. It’s not clear to me, for example, why for the heart diseases query, a paper from 1993 is ranked in first place.
6. Only a long list of papers is shown. No faceted result semantic exploration is possible.

In general, this effort also shows that semantics is really in its very early stages.

The homepage of GoPubMed features the slogan, “Searching is now sorted!” What do you mean by that? I mean, in PubMed I can arrange my results by date of publication, first author, last author, date of publication, journal, title. What can I sort by in GoPubMed and what would be the benefit of those additional categories?

As I explained before with my example of a search for heart diseases, with PubMed you can sort what today we refer to as meta-data, but not content. We sort on content level too! And we do it intelligently. Using the tree to navigate the concept “Heart Diseases” gives you all abstracts belonging to this category, not just articles containing the keywords “Heart” and “Diseases”. This, again, is the fundamental difference. Moreover, in GoPubMed you can arrange the results by date of publication, first author, last author, journal and title. We use PubMed’s syntax/grammar, but extend its expressivity.

What is the difference between, “find it” and “get all” in GoPubMed? I ask because I just typed into the search box “amyotrophic lateral sclerosis” and I wasn’t sure which option to choose. Such hesitancy and confusion on the part of first time might be a problem, right?

Yes, that’s true. We will actually change that in the next version. The idea was to clearly show the user that we do semantic analysis on almost all 19 million documents, not just a small subset. If you click on the “Get all” button, all PubMed documents are retrieved and analyzed. Technologically, this is an extremely difficult task and we thought it would be a good idea to promote it.

We see that we can, “Export to: RDF Xml BibTex Endnote PlainText” – any other options in the offing?

We already have export for 5 formats but we also welcome requests and consider their implementation.

More and more medical information is starting to reside in such formats as video. For example, in 2008 we saw the Journal of Visualized Experiments (JoVE) become the first online video journal accepted in PubMed. Are you positioning yourself for the non-text-based world of medical literature?

Videos are fantastic and help a lot in understanding problems. However, we are focused on our customers’ needs and the biggest bottleneck is still textual material. Information managers often have to deal with hundreds of millions of documents. Keeping track of content and changes is impossible with today’s key word based document and/or knowledge management systems. The only solution is a semantic approach. The machine must better understand the content so users can rely on classifications.

Allow me to mention that we will participate as one of only 12 selected groups in the biggest German IT-Research Project called “Theseus” (total volume 100 Million Euro) where we will focus on hard semantic problems in the area of antibodies and gene-interaction networks. Images already play a role in Theseus. Video will be coming soon.

As someone who spends a huge amount of time in PubMed, I would like to ask if it is a disadvantage is for you that “PubMed” is in your title. After all, if I want PubMed features can’t I just go to PubMed? If I want to supplement what I could get in PubMed wouldn’t I then try other science search engines and databases that offer material that may not be immediately available in PubMed, such as ScienceDirect, Science.gov, Mednar, ScienceRoll and Vadlo?

GoPubMed is a free, non-commercial service that demonstrates our semantic capabilities. We see GoPubMed as a PubMed add-on. We use PubMed because we think it is an outstanding service and we highly appreciate the effort of Americans! But we also add our contribution with semantics and our users appreciate that. To pick one from your list: Mednar is a very good federated search engine, but it is not semantic to my knowledge. If users of GoPubMed ask for more sources, we will consider integrating them. GoPubMed PRO – the product – does already include many sources (and many ontologies).

PubMed offers RSS feeds and email alerts of new entries. Mednar offers email alerts. You don’t seem to offer, either. Could you address that? After all, if I spend several hours tweaking my search in GoPubMed, I would like to be able to receive notification of new entries on that topic as they are generated.

Yes, it’s an often requested feature. We are working on a Semantic Alerts feature that we hope we’ll be able to provide soon.

Please tell us about GoGene and GoWeb. Who are your competitors in both realms?

GoGene is a fantastic new tool when it comes to the elucidation of gene-interaction networks. It is brand new and not public yet. GoWeb we created to explore how a general semantic search engine could look like and how typical Google users react to the semantic approach.

Go3R should be better known, as it is an intriguing, rather touching concept, ” … semantic search to avoid animal experiments.” Shouldn’t you have added “duplicative” or “redundant?” After all, some animal experimentation is necessary in the biological sciences, right? The point of that search engine is to prevent animal experiments that have already been performed, not to prevent animal experimentation entirely, right? Can you discuss the origins of Go3R and the meaning of the name?

Well, we could have included terms like “unnecessary animal trials” but our slogan reflects the ideal situation and touches on a more specific case. In regards to research with sentient animals, the EU Directive 86/609/EEC for the protection of laboratory animals requires scientists to consider whether any planned animal experiment can be substituted by other scientifically satisfactory method not entailing the use of animals, or entailing fewer animals or less animal suffering, before performing the experiment. Thus, a collection of relevant information is indispensable in order to meet this legal obligation. The 3Rconcept developed by Russell and Birtch in 1959 addresses this directive and is a “gold standard” in the community:

1. Refinement of scientific techniques?
2. Reduction in the numbers of animals used?
3. Replacement of animal procedures with non-animal procedures

Go3R helps scientists find all relevant information when planning to conduct animal trials. That’s why we called it Go – traditionally for GeneOntology – and 3R for the 3Rconcept.

Can you give us the names of notable medical librarians and/or researchers who adore GoPubMed?

We compiled a collection of blog comments about GoPubMed (and Transinsight) on our web site (http://www.transinsight.com/peopleAbout). But as a rule we never use “user data” like domain names for marketing. The only thing I can tell is users come from literally all over the world. The biggest user is Stanford University.

As we know, we are in a very challenging economic climate. How is Transinsight doing in these tough times? Have you had to scale back in any way or is it full steam ahead?

We are doing quite well. We are not planning an IPO “next month,” but we are making good progress as a company. I think people are starting to recognize the power of semantic search and seeing clear benefits. There is often a 90% reduction in the amount of time spent searching. And this translates into increased productivity and financial returns.

More and more people are realizing how bad and time-consuming traditional search is and how much more they can accomplish with our semantic systems.

You are based in Germany, but your products are in English. Can you discuss any cultural or differences in business practices and scientific mores between the US and Germany you have encountered over the years?

Most of our largest clients are in the U.K. and the U.S. and we are constantly working to fulfill – and anticipate – their needs. We never really had problems in terms of cultural differences with companies from the Western world. Business in Asia was and is still a challenge for us. We tried to start selling there, but it never took off. South America is also an interesting (Next) market.

Who are your personal heroes in science, medicine, technology or in any other respect?

As a Geophysicist, I admire Einstein because he thought the impossible. That’s why Transinsight’s slogan is “Think the impossible.” I also admire Gerd Binnig, the 1986 Nobel Laureate in Physics, with whom I had the pleasure to work for 5 years and who thinks beyond the visible. What he and our team developed nine years ago is still (although not very well known) leading edge: Self-Similar, Self-Organizing Semantic networks [1].

Final thoughts?

Semantics will revolutionize the world … again. I would say that even more than the internet did. The chasm is still to come, but “smarter machines” will impress us sooner rather than later. For us, it’s exciting to participate in such an amazing development.

Thank you for your time.

Thank you for your interest in GoPubMed and Transinsight.

Are you a Twitter user? Tweet this!

References

1. Binnig et al. Will machines start to think like humans? Europhysics News 33(2), 44-47, 2002.