At the atomic level, simulating transitions between metastable states of the free-energy landscape, often hindered by slow molecular processes, poses a formidable challenge. Overcoming the associated free-energy barriers and accelerating the underlying dynamics is commonly addressed by importance-sampling schemes, which necessitate adequately defined reaction-coordinate models within compact, low-dimensional sets of collective variables (CVs). While traditional approaches rely on human intuition for dimensionality reduction, recent machine-learning (ML) algorithms provide robust alternatives. We compare two variational data-driven ML methods, state-free reversible variational approach for Markov processes networks (SRVs) and variational committor-based neural networks (VCNs), to capture the slowest decorrelating CV and committor probability in a paradigmatic transition between metastable states. Examining simple model systems, both methods demonstrate the capacity to discover relevant descriptors and are adaptable to importance-sampling schemes using a reweighting algorithm that approximates the kinetic properties of the transition.