It is stating the obvious that we live on a planet consisting in continuous landscapes. Yet, there exists genuine barriers to developing sound statistical models that accommodate for continuous spatial information and genetic data in a satisfactory way. In fact, dominant spatial models in population genetics rely on the crude assumption that populations are divided in discrete demes. Other approaches make predictions about the spatial distribution of individuals that are generally not supported by biological evidence. Current limitations in the models and the inference techniques available hamper our understanding of biodiversity in space and time. They are thus the main focus of our project. Recent advances in theoretical population genetics have produced a new model, the spatial Lambda-Fleming-Viot model, that alleviates the limitations of current methods. This model considers the habitat as a truly continuous area and allows for a stationary distribution of individuals in time and space. A straightforward probabilistic description of the ancestral locations and genealogical relationships between sampled individuals is also available, thereby defining a simple way to calculate the likelihood of this model (the probability of the data given the model parameters). Yet, this likelihood involves a lot of latent variables, i.e., parameters that are not of utmost biological interest but are mandatory in order to proceed with the evaluation of the function of interest. It is therefore not clear whether the spatial Lambda-Fleming-Viot model is amenable to parameter inference. We have implemented and tested a prototype of a Bayesian sampler that estimates the posterior distribution of this model parameters from the analysis of geo-referenced genetic data. Preliminary results indicate that, when harnessed to state of-the-art statistical inference techniques, this new model indeed provides accurate estimates of the population densities and the dispersal range, two parameters that cannot be estimated separately with most traditional approaches. These promising results suggest that the spatial Lambda-Fleming-Viot model can indeed serve as a sound basis to tackle important biological questions. In particular, we will assess the impact of non-homogeneous landscapes on migration of individuals in this project. We will also investigate the detection of variability of a population density in space and during the course of evolution. Alongside these extensions of the original model, mathematical simplifications of the likelihood function will be examined. We have in fact identified analytical "shortcuts" that should considerably simplify, and therefore speed up, the calculations. Extensions and improvements of the models and inference techniques developed in this project will be applied to the analysis of large population genomics datasets from two flagship species of considerable economic importance: the "harlequin ladybird" and the spotted wing drosophila. We will quantify levels of gene flow and population densities throughout their respective habitats, thereby gaining some insight into the biology of these organisms. Software applications will be produced that implement the most relevant approaches developed in this project. These applications will be thoroughly tested through extensive simulations and then made available to a wide scientific audience.

AI4Sci develops hybrid AI methods at the intersection of machine learning, distributional semantics and knowledge representation in order to analyze online discourse and in particular scientific controversies taking place on the web with an applications related to the COVID-19 pandemic. Scientific insights form a central part of public discourse, in particular in the context of the COVID-19 pandemic. However, due to the inherent complexity of scientific claims as well as the mechanisms of online platforms, where controversial topics are shown to generate more user interaction, retention and virality, scientific findings tend to be represented in a simplified, decontextualized and often misleading way. In this context, AI4Sci addresses the challenge of providing hybrid AI methods for tracing and interpreting scientific claims in online discourse, as a means to tackle and understand misinformation in society. Progress in areas such as transfer learning and neural NLP have opened up new possibilities for the AI-based interpretation of online discourse. On the other hand, it has been shown that structured knowledge can improve transparency and performance of neural models, while neural language models themselves carry relational knowledge. Building on these insights, the project will develop hybrid AI methods, able to classify and disambiguate online discourse about scientific findings as observable in online news media and the social Web. AI4Sci will build on recent advances in AI at the intersection of neural NLP, distributional semantics and symbolic knowledged to develop methods geared towards the particular problem of extracting and classifying scientific claims about controversial topics together with related contextual information from online discourse and matching them to their respective scientific context. The hybrid methodology of AI4Sci will also contribute to widely recognised issues such as transparency and reproducibility of neural models. Given the very discipline-specific contexts of scientific claims, both in science as well as online discourse, AI4Sci will evaluate methods in two use-cases centered around the COVID-19 pandemic, involving the life sciences as well as the social sciences. The joint expertise of LIRMM (France) and GESIS (Germany) combines backgrounds in symbolic AI and knowledge graphs , with expertise in NLP/NLU. In particular with respect to mining and understanding online discourse on the (social) Web, the two partners complement each other with applications in the context of computational social science (GESIS) and life science (LIRMM). This will advance the AI-related agendas of both organisations and contribute to the AI strategies at the national and international level. The project will build on joint work and GESIS- and LIRMM-hosted corpora, such as knowledge graphs about online discourse, unique Web and social Web crawls as well as scientific data and bibliographic archives, which will accelerate and facilitate the AI4Sci work programme. AI4Sci brings together a highly diverse team of 4 established and 3 young researchers from LIRMM and GESIS, which will be enhanced by two PhD projects.

The coming decade will see the generalization of dematerialized computation and communication according to which computations and data will be « mobile » and served transparently to users. This will result from the convergence of current numerical technologies, including embedded systems via Internet of Things (IoT), high-performance computing (HPC) and cloud computing, towards pervasive large-scale distributed virtualized systems. This trend will have a heavy impact on the way to address scientific problems (e.g., high energy physics, chemistry, material science), industrial sectors (e.g., automotive, aeronautics, energy, complex system optimization) and societal challenges (e.g., climatology, medicine, energy, social networking, smart cities management). This upcoming mutation of the computing ecosystem will come with a number of major technological challenges amongst which one can underline security, reliability and energy-efficiency. The CONTINUUM project mainly focuses on the last challenge by investigating a low power multicore design solution to answer the energy-efficiency demand in foreseen future systems. It explores a design continuum for compute nodes, which seamlessly goes from software to technology levels via hardware architecture. Power saving opportunities exist at each of these levels, but real measurable gains will come from the synergistic focus on all these levels as considered in this project. Then, a cross-disciplinary collaboration is promoted between computer science and microelectronics, to achieve two main breakthroughs: i) combination of state-of-the-art heterogeneous adaptive embedded multicore architectures with emerging communication and memory technologies and, ii) power-aware dynamic compilation techniques that suitably match such a platform. Involved partners are two very renown academic institutions carrying out cutting-edge research on compilation (Inria Rennes – Bretagne Atlantique), integrated system architectures and technologies (LIRMM / CNRS-UM), and a low power core technology leader company (Cortus S.A.S.).

Pharmacogenomics (PGx) studies how individual gene variations cause variability in drug responses. A state of the art of PGx is available and constitutes a basis for implementing personalized medicine, i.e., a medicine tailored to each patient by considering in particular her/his genomic context. However, most of the state of the art of this domain is not yet validated, consequently not yet applicable to medicine. Indeed, most of it results from assays that do not fulfill statistics validation standards and are difficult to reproduce because of the rarity of gene variations studied (making hard to recruit sufficiently large cohorts) and of the multifactorial aspect of drug responses. Beside, the generalizing use of Electronic Health Records (or EHRs) generates large repositories that offer new opportunities such as composing patient cohorts for the study of clinical hypotheses hard to test experimentally. Typically, EHR repositories make possible to assemble cohorts of patients to study, on the basis of practice-based data, the impact of gene variations on drug responses. The goal of the PractiKPharma project (Practice-based evidences for actioning Knowledge in Pharmacogenomics) is to validate or moderate PGx state-of-the-art (SOTA) knowledge on the basis of practice-based evidences, i.e., knowledge extracted from EHRs. Units of knowledge in PGx typically have the form of ternary relationships gene variant–drug–adverse event, and can be formalized to different extents using biomedical ontologies. To achieve our goal, we propose: (1) to extract SOTA knowledge from PGx databases and literature, (2) to extract observational knowledge (i.e., knowledge extracted from observational data) from EHRs, (3) to compare knowledge units extracted from these two origins, to confirm or moderate SOTA knowledge, with the ultimate goal of enabling personalized medicine. (4) Finally, we intend to emphasize confirmed knowledge by investigating omics databases for molecular mechanisms that underlie and explain drug adverse events. For this investigation will use and contribute to the biomedical Linked Open Data. The PractiKPharma project involves a multidisciplinary consortium of 4 academic partners: 2 informatics experts, the LORIA of Nancy (coordinator), the LIRMM of Montpellier; and 2 biomedical experts, the HEGP (Paris) specialized in EHRs management and PGx, and the SSPIM (from CHU Saint-Etienne) specialized in medical informatics and pharmacovigilance. The expected impacts of PractiKPharma both in computer science and in its application domain are: novel methods for knowledge extraction from text and EHRs; multilingual semantic annotation of EHRs; methods for representing and comparing SOTA and observational knowledge; a database that maps genotypes to quantitative traits to facilitate the study of PGx with EHRs; the completion and connection of Linked Open Data related to PGx; methods for hypothesizing on mechanisms of adverse events; validated PGx knowledge units. Consequently, our ultimate goal is to provide clinicians with actionable PGx knowledge to establish guidelines that, implemented in personalized medicine, will reduce drug adverse events, and then improve the quality of clinical care.