WEBVTT Kind: captions Language: en 00:00:00.440 --> 00:00:01.990 Ok, thank you, Lauren. 00:00:01.990 --> 00:00:05.560 And thank you, Florence for inviting me. 00:00:05.560 --> 00:00:09.660 Ok, I'm going to share my screen. 00:00:09.660 --> 00:00:15.440 So, thank you for inviting me to participate in this webinar. 00:00:15.440 --> 00:00:22.230 I am Carlos Badenos-Olmedo, a researcher from the Ontology Engineering Group and I am also 00:00:22.230 --> 00:00:25.289 an assistant profit professor in the Universidad Politecnica de Madrid. 00:00:25.289 --> 00:00:26.940 I am going to present Drugs4COVID project. 00:00:26.940 --> 00:00:32.759 This is a project that we propose to create an knowledge graph. 00:00:32.759 --> 00:00:39.840 This is trying to represent the drugs that we used during the the pandemic to define 00:00:39.840 --> 00:00:43.000 the clinical control of the coronavirus. 00:00:43.000 --> 00:00:51.989 The idea is that during the COVID-19 pandemic, worldwide institutions tried to identify or 00:00:51.989 --> 00:01:00.300 to create data sets with scientific research articles that are related to coronavirus. 00:01:00.300 --> 00:01:07.570 This kind of information could be useful for hospital pharmacies, for clinical practitioners. 00:01:07.570 --> 00:01:14.210 Our research group, the Ontology Engineering Group, was trying to promote how can we retrieve 00:01:14.210 --> 00:01:18.189 information - retrieve knowledge from this kind of data. 00:01:18.189 --> 00:01:25.330 The first step was to identify what is the most important dataset that we can use. 00:01:25.330 --> 00:01:32.960 In this case, from the European Union was provided the COVID-19 European data portal, 00:01:32.960 --> 00:01:35.120 also the [inaudible] COVID data set. 00:01:35.120 --> 00:01:42.909 Mainly, the COVID-19 open presets data set was the most important data set that combines 00:01:42.909 --> 00:01:47.979 information from PubMed BioRxiv, MedRxiv, and arXiv information. 00:01:47.979 --> 00:01:53.210 It also combines information from the World Health Organization and provides more than 00:01:53.210 --> 00:02:03.920 400,000 articles that we need - we thought that we can leverage to provide knowledge. 00:02:03.920 --> 00:02:13.670 Combining this kind of knowledge with the self-service from Madrid, the [inaudible]. 00:02:13.670 --> 00:02:20.890 We provide the mechanisms to create knowledge graphs that we can exploit and provide knowledge 00:02:20.890 --> 00:02:22.690 from this kind of of people. 00:02:22.690 --> 00:02:31.170 So, first of all, we define a workflow with different steps that we provide not only the 00:02:31.170 --> 00:02:38.970 step but also the recommendations the steps that you can follow to create a final analysis 00:02:38.970 --> 00:02:42.550 graph and also facilitate the exploitation. 00:02:42.550 --> 00:02:46.360 We define a six step workflow. 00:02:46.360 --> 00:02:47.780 The first step is the harvesting. 00:02:47.780 --> 00:02:54.430 In this step the idea is that you need to identify the data set that is related to coronavirus, 00:02:54.430 --> 00:02:59.459 and also to evaluate if the data is completely available or not. 00:02:59.459 --> 00:03:05.799 The second step is the pre-processing because you need to organize the data that is provided, 00:03:05.799 --> 00:03:13.140 in this case a data set, because this data is not total because we are working with tests. 00:03:13.140 --> 00:03:17.180 We need to modelize a certain way of [evaluating] the data. 00:03:17.180 --> 00:03:21.560 Then, we move to the third step, which is information extraction. 00:03:21.560 --> 00:03:28.569 In this step we need to discover the main elements of this kind of data that we can 00:03:28.569 --> 00:03:30.709 use to create the analysis graph. 00:03:30.709 --> 00:03:37.599 In our case, the main elements are the biomedical concepts (for example, the drugs, the diseases, 00:03:37.599 --> 00:03:39.340 and the genetic information). 00:03:39.340 --> 00:03:47.231 Then, we need to define a formal description of the domain, which is the fourth step, which 00:03:47.231 --> 00:03:48.849 is the semantification. 00:03:48.849 --> 00:03:55.110 In this case, we need to create an ontology to define the relations between the biomedical 00:03:55.110 --> 00:04:01.020 concepts and to define all the concepts, all the elements, that finally appear in the knowledge 00:04:01.020 --> 00:04:02.020 graph. 00:04:02.020 --> 00:04:04.590 The next step is the knowledge graph generation. 00:04:04.590 --> 00:04:09.580 In this step we need to define the rules to create the instances in the knowledge graph. 00:04:09.580 --> 00:04:16.120 And finally we can provide the mechanisms to exploit - to facilitate the use of the 00:04:16.120 --> 00:04:19.560 information that the knowledge graph contains. 00:04:19.560 --> 00:04:21.780 So we focus on the first step. 00:04:21.780 --> 00:04:29.020 The objective is to identify the relevant data sources and also to evaluate the availability 00:04:29.020 --> 00:04:30.590 of the data. 00:04:30.590 --> 00:04:36.479 Our proposal is that you need to perform a systematic literature review, taking into 00:04:36.479 --> 00:04:41.460 account the main concepts of the coronavirus, and then define the data set. 00:04:41.460 --> 00:04:46.380 For example, from digital repositories, PubMed, BioRxiv, and so on, but also combining with 00:04:46.380 --> 00:04:51.470 other sources, for example, clinical collections from the coordinating corpus, the lead COVID 00:04:51.470 --> 00:04:53.830 dataset and also additional resources. 00:04:53.830 --> 00:04:58.310 For example, patents, encyclopedic articles from Wikipedia. 00:04:58.310 --> 00:05:03.240 And all this data is organized in this first step. 00:05:03.240 --> 00:05:10.280 In the next step, we need to transfer this unstructured data, which are text, into tables, 00:05:10.280 --> 00:05:12.120 which are structured data. 00:05:12.120 --> 00:05:17.540 Then, the methodology that we propose is to identify the minimal information you need. 00:05:17.540 --> 00:05:23.510 The most easy way to transform text into a structured way is to define the full text 00:05:23.510 --> 00:05:26.630 of the article as the data. 00:05:26.630 --> 00:05:32.400 In our opinion, this is not the best way to do that and our proposal is to define as minimal 00:05:32.400 --> 00:05:38.120 information you as need: the paragraph of the articles. 00:05:38.120 --> 00:05:43.780 In that area you can discover all the references or the relation between the biomedical concepts. 00:05:43.780 --> 00:05:47.180 Then the next step is the informational structure. 00:05:47.180 --> 00:05:54.360 In this case, the idea is to create annotations based on that paragraphs that discover drugs, 00:05:54.360 --> 00:05:57.440 diseases, and genetic information. 00:05:57.440 --> 00:06:03.590 In our experience, we fine-tune different language models for each different biomedical 00:06:03.590 --> 00:06:04.590 concept. 00:06:04.590 --> 00:06:11.380 The idea is that you need to define a specific language model to identify drugs and also 00:06:11.380 --> 00:06:16.030 to normalize the drugs according to different vocabularies because in different countries 00:06:16.030 --> 00:06:20.740 we use different standard codes. 00:06:20.740 --> 00:06:29.490 Once we have the annotations with the entities and also the codes we can define the formal 00:06:29.490 --> 00:06:33.139 space to describe all this information. 00:06:33.139 --> 00:06:36.849 This is the step that we need to create an ontology. 00:06:36.849 --> 00:06:42.360 In the biomedical domain exist a lot of ontologies so the idea is not to create from scratch 00:06:42.360 --> 00:06:43.440 an ontology. 00:06:43.440 --> 00:06:48.970 The idea is to reduce system ontologies and to provide the missing information in the 00:06:48.970 --> 00:06:49.970 new ontology. 00:06:49.970 --> 00:06:59.300 In our case the ontology was EBOCA and the missing information was to provide the evidence 00:06:59.300 --> 00:07:05.960 that supports the relationships between drugs, between diseases, and between genetic information. 00:07:05.960 --> 00:07:12.599 In our case we used the unified medical language system and also the DISNET platform. 00:07:12.599 --> 00:07:14.550 All this information is combined. 00:07:14.550 --> 00:07:19.889 We also provide, this is the purple area, the information about the evidence. 00:07:19.889 --> 00:07:21.599 Which is the evidence? 00:07:21.599 --> 00:07:29.289 The evidence is the paragraph where the relation between these element is reported in the scientific 00:07:29.289 --> 00:07:30.289 article. 00:07:30.289 --> 00:07:39.169 Once we have defined the formal domain, the ontology, we need to identify the instances, 00:07:39.169 --> 00:07:45.449 the claims, the statements retrieved from the scientific articles to create instances 00:07:45.449 --> 00:07:46.449 in the knowledge graph. 00:07:46.449 --> 00:07:54.470 This is the knowledge graph generation step so ours our methodology is proposed to create 00:07:54.470 --> 00:08:01.169 rules to identify using the previous model language the entities and also the relations 00:08:01.169 --> 00:08:03.300 between them. 00:08:03.300 --> 00:08:10.289 Finally, once we have the ontology, we have the instances, we are able to identify the 00:08:10.289 --> 00:08:11.690 nodes in the graph. 00:08:11.690 --> 00:08:17.530 For example, the blue ones are the elements, the orange ones are the relations between 00:08:17.530 --> 00:08:23.259 them, and the purple ones are the evidence that support these kind of relations. 00:08:23.259 --> 00:08:29.789 The evidence is the minimal information unit which is the paragraph and also the articles. 00:08:29.789 --> 00:08:36.370 Once we have the knowledge graph, finally, we are available to facilitate the exploitation 00:08:36.370 --> 00:08:38.149 of the information. 00:08:38.149 --> 00:08:45.290 The best - the first step is of course we can use SPARQL queries - this is a specific 00:08:45.290 --> 00:08:51.670 language model that you can create queries to exploit the language, the analysis graph. 00:08:51.670 --> 00:08:56.300 This requires an expert in this kind of domain. 00:08:56.300 --> 00:09:03.720 Our second methodology is to create a question-answer interface that provides the information not 00:09:03.720 --> 00:09:09.480 only from the knowledge graph but also combining from external sources, from others and knowledge 00:09:09.480 --> 00:09:15.140 graphs from others, document connections and then support questions in natural language 00:09:15.140 --> 00:09:18.860 to provide answers, also in natural language. 00:09:18.860 --> 00:09:24.339 So our platform is for a COVID platform. 00:09:24.339 --> 00:09:29.959 All this information is policy available: the knowledge graphs, the models, the data 00:09:29.959 --> 00:09:36.250 sets, and also the services are - these resources are completely free, are completely public. 00:09:36.250 --> 00:09:40.160 These are available from these URLs. 00:09:40.160 --> 00:09:43.769 Thanks for your attention and I'm able to to answer any questions that you have.