AI4scMED

Cell-based precision medicine has the potential to revolutionize healthcare. To reach its full potential, we need to better understand and integrate the variability and the multiscale facets of most diseases. Single-cell (sc) multi-omics has the unique capacity to provide high-throughput molecular portraits of individual cells, and to identify predictive biomarkers of a disease trajectory. Understanding this intrinsic complexity and its translation into better treatment requires the exploitation of heterogeneous information, which can be achieved only through new AI-based breakthroughs that will unravel the multiscale processes governing cell trajectories towards disease.

Our consortium will tackle new methodological challenges to bridge the gap between sc data and the advent of personalized targeted treatments. Sc-data can resolve the confounding effects of distinct cell types by providing measures of genome-wide features at the level of individual cells. When integrated, sc-multi-omics data have already demonstrated the multifacets and heterogeneity of diseases. In addition, combined with image analysis, sc-molecular readouts can be integrated at the cellular scale to investigate the compartmentation and spatialization of signaling events. To address the complexity of the human body and to combine genomic information with other assays, we will develop new AI-based methods to handle, integrate, analyze, and visualize this multiscale complexity that characterizes diseases. These new developments will be based on cutting-edge advances in AI-based data analysis to decipher the complex geometry of sc-genomic datasets.

Contributions of sc-genomics to health requires multiscale models to infer causal mechanisms governing the evolution towards disease at different levels. Our developments will account for the specific challenges presented by sc multi-omics (dimensionality, scarcity). Moreover, they typically consist of observational datawithout systematic perturbations, which makes the identification of causal effects challenging. To meet this challenge, we will combine causal/logical/stochastic modeling for the integration of heterogeneous data.

Our models will integrate the different temporal scales of the dynamical processes that govern gene activity, and will account for biophysical priors to reduce the complexity of model inference. We will produce network inference methods for the characterization of the molecular mechanisms underlying the phenotypic behaviors of a clinical sample. Our objective is to couple symbolic AI with machine learning for an efficient characterization of bestpredictive executable models. Our methods will allow the identification of molecular targets for controlling the so-called pathological state. The inferred networks will then be investigated to identify genes having a key role in the analyzed clinical samples or to predict the impact of therapeutic interventions in a clinical sample.

Finally, a challenge for precision medicine consists in integrating the different levels of variability that govern cell decision making, a multiscale dynamical process that associates signaling networks with cell trajectory and a tissue-level organization. A paradigm shift must be realized to consider disease as a stochastic state that could be predicted, perturbed, and controlled. Predictive capability to describe any trajectory towards disease requires the coupling of intracellular and biochemical dynamics with cell population dynamics.

Meeting this challenge requires reconciling different mathematical formalisms and integrating heterogeneous biological knowledge to represent in a common framework biological processes described on very contrasting spatial and temporal scales. Our ambition is to build predictive executable models, ultimately digital twins, to provide data-driven solutions for the implementation of targeted personalized treatments based on the control of cell fate decision.