Section: New Results

Knowledge Discovery in Healthcare and Life Sciences

Participants : Alexandre Bazin, Miguel Couceiro, Adrien Coulet, Sébastien Da Silva, Florence Le Ber, Jean-François Mari, Pierre Monnin, Amedeo Napoli, Abdelkader Ouali, Yannick Toussaint.

Ontology-based Clustering of Biological Data

Biomedical objects can be characterized by ontology annotations. For example, Gene Ontology annotations provide information on the functions of genes, while Human Phenotype Ontology (HPO) annotations provide information about phenotypes associated with diseases. It is usual to consider such annotations in the analysis of biomedical data, most of the time annotations from only one single ontology. However, complex objects such as diseases can be annotated at the same time w.r.t. different ontologies, making clear distinct dimensions. We are investigating how annotations from several ontologies may be cooperating in disease classification. In particular, we classified Genetic Intellectual Disabilities, on the basis of their HPO annotations and of Gene Ontology annotations of genes known for being responsible for these diseases [88]. We used clustering algorithms based on semantic similarities that enable us to compare sets of annotations. In particular, this experiment illustrates the fact that considering several ontologies provides better results in clustering, while selecting the best set of ontologies to combine is depending on the dataset and on the classification task. This study is still going on.

Validation of Pharmacogenomic Knowledge

State of the art knowledge in pharmacogenomics is heterogeneous w.r.t. validation. Some units of knowledge are well validated, observed on a large population and already used in clinical practice, while a large majority of this knowledge is lacking validation and reproducibility, mainly because of scarce observation. Accordingly, validating state of the art knowledge in pharmacogenomics by mining Electronic Health Records (EHRs) is one objective of the ANR project “PractiKPharma” initiated in 2016 (http://practikpharma.loria.fr/).

To carry out this validation, we define a minimal data schema for pharmacogenomic knowledge units (PGxO ontology), which is instantiated with data of different provenance (e.g. biomedical databases, literature and EHRs). The output of this instantiation is a (unique) knowledge graph called PGxLOD (https://pgxlod.loria.fr/). We defined and applied a set of so-called “reconciliation rules” that compare and align whenever possible knowledge units of different provenance [9]. The results of these rule applications are of particular interest since they highlight knowledge units defined in various data and knowledge sources. We are continuing this effort by studying how graph convolutional networks enable us to learn and then to compare the representation of $n$-ary relationships in the form of graph embeddings [39].

In addition, following our participation in the Biohackathon 2018 in Paris (https://2018.biohackathon-europe.org/), we firstly updated PGxLOD and improved its quality, completeness, and interconnection with other resources. Secondly we mined PGxLOD and searched for explanations about molecular mechanisms of adverse drug responses. Preliminary results where presented at the MedInfo Conference [59].

Mining Electronic Health Records

In the context of the Snowball Inria Associate Team, we studied the use of Electronic Health Records (EHRs) to predict at first prescription the need for a patient to be prescribed with a reduced drug dose [6]. We particularly focused on drugs whose dosage is known to be sensitive and variable. We used data from the Stanford Hospital to construct cohorts of patients that either did or did not need a dose change for each considered drug. After feature selection, we trained Random Forest models which successfully predict whether a new patient will or not require a dose change after being prescribed one of 23 drugs among 22 drug classes. Several of these drugs are related to clinical guidelines that recommend dose reduction exclusively in the case of adverse reaction. For these cases, a reduction in dosage may be considered as a surrogate for an adverse reaction, which our system could help to predict and to prevent.

In collaboration with Stanford University, we continued studying the development of predictive models from EHR data, in particular to evaluate the risk of atherosclerotic cardiovascular diseases (ASCVD). The evaluation of ASCVD risk is crucial for deciding upon the prescription of preventive therapies such as statins and others lipid lowering therapies. The prevalence of these diseases is depending on subgroups in a population, such as African-American and Asian people, which are indeed under-represented in cohorts that were used to fit the model currently used in clinics to evaluate the risk of ASCVD [25]. Due to such under-representation, biases are appearing in the evaluation of the risk when considering these different subgroups in the population. Then we proposed a method and a predictive model that controls, to some extent, the variability in the prediction of ASCVD when considering such “foreign” subgroups [40].