Speaker
Description
Inferring signed and directed interaction structure from multivariate time-series data is a central problem in regulatory network reconstruction and dynamical systems biology. Albeit many biological systems are modeled as $\dot{x}=f(x)$, identifying the structural architecture underlying $f$ from observational data remains challenging: equation fitting requires strong parametric assumptions, correlation-based methods lack directionality, and many causal frameworks rely on interventions or restrictive identifiability conditions. We propose the Classification-Based Inference of Regulatory Networks (CIRN), a classification-theoretic reformulation of structural inference for continuous-time dynamical systems. Instead of modeling state magnitudes, CIRN models the direction of temporal change by encoding $sgn(dx_j/dt)$ as a binary outcome and relating it to lagged states and derivatives via Bayesian logistic regression. Directed edges are inferred using posterior credible intervals, yielding an uncertainty-aware signed interaction structure. CIRN does not estimate functional forms of $f$; it identifies sign-consistent interaction constraints, producing a directed topology compatible with deterministic and stochastic models. By inferring regulatory structure independently of specified functional forms, CIRN provides a general framework for reconstructing regulatory architecture across gene regulatory, epidemiological, ecological, environmental, and other complex dynamical systems.
