Abstract
Multi-phenotype deep mutational scanning (DMS) experiments provide a powerful means to dissect how protein variants affect different layers of molecular function, such as abundance, surface expression, and ligand binding. When these phenotypes are connected through a molecular pathway, interpreting variant effects becomes challenging because downstream phenotypes often reflect both direct and indirect consequences of mutation. We introduce Cosmos, a Bayesian framework for residue-level causal inference in multi-phenotype DMS data. Cosmos addresses three key questions: (1) whether a causal relationship exists between two phenotypes; (2) the strength of that relationship; and (3) the expected downstream phenotype if the upstream phenotype were normalized, enabling counterfactual interpretation. The framework uses position-level aggregation and Bayesian model selection to infer interpretable causal structures, without requiring phenotype-specific biophysical assumptions. We apply Cosmos to three datasets-Kir2.1 (abundance and surface expression), PSD95-PDZ3 (abundance and CRIPT binding), and KRAS (abundance and RAF1-RBD binding) and show that it effectively distinguishes direct from indirect functional effects. Across these applications, Cosmos provides a generalizable and interpretable approach to disentangle causal relationships in high-throughput protein functional screens.