HDRs 2024

Le logiciel présente le paradoxe d’ouvrir un champ immense de possibilités et de rester assez empirique et artisanal dans son développement. Nul doute qu’il soit le seul bien-service dont la maintenance génère des affaires aussi juteuses et les échecs si coûteux. Les travaux abordés dans ce document concernent le génie logiciel et son application. Nous nous inscrivons ici dans une vision modèle avec le cycle vertueux de construction (modélisaton, vérification de propriétés, exploitation) pour produire qualitativement du logiciel de qualité. La vérification de propriétés, telle que la sûreté de fonctionnement ou la sécurité, est fondamentale à l’établissement du contrat de service logiciel. Le logiciel étant un objet vivant dans un contexte mouvant, sa construction doit suivre et s’adapter en permanence. Nous proposons un processus pour cela, fait d’adaptation du contexte technologique et d’évolution du besoin. D’un point de vue applicatif, nous nous intéressons aux systèmes d’information d’entreprise, un écosystème complexe qui entremêlent pléthore de logiciels qu’il faut faire fonctionner ensemble, et aux systèmes de production industriels, qui y ajoutent la complexité des systèmes cyberphysique. Dans de tels cas, on comprend qu’une vision systémique s’impose pour gérer la complexité. Le document se termine sur une réflexion des compétences nécessaire pour diriger des recherches.

In this manuscript it is proposed to explore two representations of a biological system using computational modeling. These representations both gave birth to several methodological publications, and in some cases research projects in close collaboration with biologists. One is done through the sign-consistency modeling. In this approach a regulatory network (signed directed graph) is combined with a dataset of gene expression observations, using a logic program. This logic program, written in Answer Set Programming, expresses a rule that has to be valid for each species in the network, which relates the sign of a network species with its direct predecessors influences and signs. This rule is tested in a global way, through all the network species by using an efficient solver, clasp. The sign-consistency modeling framework we proposed is named Iggy. Iggy performs as well automatic and optimal correction of sign inconsistencies. The sign-consistency modeling framework has been applied to different biological case studies. For example, the signaling pathway of Hepatocyte Growth Factor, where some of the computational predictions of our model were validated experimentally. A case-study well described in this manuscript is the modeling of Multiple Myeloma patients gene expression data. Our main results on this system was to propose Multiple Myeloma markers, that is, species in the network, coupled with our computational predictions, that allow to identify patients having a better survival. Iggy has inspired MajS, our last sign-consistency modeling framework contribution. In this on-going research project we plan to integrate gene regulatory and metabolic network modeling. A second modeling approach is learning Boolean network families. In this framework, given a regulatory network (also called Prior Knowledge Network, PKN) and a set of network species observations upon multiple perturbations over the system, our framework learns a family of Boolean Networks (BNs), compatible with the PKN topology, that fits the perturbation data with minimal error. The first system we conceived is named caspo. It is also implemented using Answer Set Programming. An extension of caspo was implemented, so that new experimental designs (i.e. new experimental perturbations) can be proposed to decrease the number of learned BNs. Later, we proposed a system named caspo-ts, which deals with perturbation time-series data, and the output of this system is a family of dynamic BNs. caspo-ts has been applied to the data of HPN-DREAM challenge, concerning Breast cancer cell lines. Our objective was to identify the different BNs underlying the four Breast Cancer cell lines considered. In parallel, since multiple perturbation data, essential for \emph{caspo} or \emph{caspo-ts}, is sometimes hard to obtain in Human systems because of ethical reasons; we have begun a research subject towards the extraction of multiple pseudo-perturbations from non perturbed datasets, such as proteomics or RNA-Seq datasets. This method has been applied to discriminate Acute Myeloid Leukemia patients having different treatment prognosis. Currently, we are exploring to extract multiple pseudo-perturbations from single cell data, in the study of Human embryo development.