From Data to Phenotype: Integrating Data Structure and Prior Knowledge to Model Biological Systems

This manuscript addresses the methodological challenges of inferring phenotypes from biological data by framing the task as the approximation of a fonction P  = f( G, E, 0), where Gand E represent genotype and environment, and O denotes ad­ ditional structured or latent influen ces. A central question underlies the manuscript: how can one define an appropriate form for the mapping f, and optimize its parameters, in the presence of dataset dependencies and prior biological knowledge?

The first chapter introduces this conceptual framework and highlights how dependencies-between individuals, between features, or through latent biological structure-complicate inference but also offer structural eues for more accurate mod­ eling. It argues that appropriate choices of variables, models, and optimization procedures must be informed by the data structure and available biological knowl­ edge.

Chapter 2 focuses on statistical dependencies across individuals, such as repeated measures, paired experiments, or shared ancestry. It presents contributions based on mixed-effect models and Bayesian approaches, showing how they allow proper estimation of phenotypes and variance components when individuals are not inde­ pendent.

Chapter 3 addresses explicit dependencies between features, for instance intime­ series or dose-response experiments. It proposes models that account for these structures, including differential equation systems and nonparametric projection tech­ niques. The idea of representing phenotypes as fonctional trajectories is developed through case studies in aging and cellular dynamics, such as spindle elongation.

Chapter 4 considers more complex dependencies, particularly those that require external biological knowledge to be revealed or modeled. It includes contributions on predicting metabolic fonctions from microbiome data and on improving the interpretability of machine learning models using prior knowledge. These examples illustrate how structured biological information can act as a constraint to guide inference and interpretation.

The final chapter reflects on the need to evaluate the robustness and influence of prior knowledge in machine learning models. Rather than designing new models, the focus is shifted toward developing meta-analytical tools to assess reproducibility, sensitivity to assumptions, and interpretability. Two ongoing projects are presented as illustrations: one on comparing representations of genetic sequences, and one on evaluating ontology-based aggregation methods.

Across the manuscript, three themes emerge: dependencies as informative features of the dataset; prior knowledge as a source of structure in modeling-including dataset dependencies, which can themselves be seen as implicit or protocol-derived prior knowledge; and the need for interpretability tools tailored to knowledge­ informed models. The overall goal is to better understand the interplay between data-driven learning and structured biological insight in phenotype inference.