REVEAL: RNA Ensemble Validation and Exploration with AI–guided Landscapes

Capturing the sparsely populated but functionally important, excited states of RNA remains a central challenge for molecular simulation, because the underlying transitions are rare events that unbiased molecular dynamics cannot reach from a single starting structure. Here we introduce REVEAL (RNA Ensemble Validation and Exploration with AI–guided Landscapes), a workflow that couples generative secondary structure sampling, secondary structure conditioned tertiary structure prediction with trRosettaRNA2, unbiased molecular dynamics and Markov state modeling. Using only the sequence of the engineered HIV–1 TAR mutant, in which the excited state ES2 is the dominant conformation, we show that REVEAL captures the reversible ES2 to GS register shift from 6.8 μs of aggregate unbiased sampling across 34 seeds, recovering both basins, the intermediate states, and correctly identifies ES2 as the dominant state, consistent with the design of the construct. In contrast, molecular dynamics seeded from the ES2-dominant HIV–1 TAR mutant NMR ensemble remains trapped in ES2 despite substantially longer aggregate simulation time, and sequence–only ensemble predictors (AlphaFold3, Boltz-2, ESMFold2, OpenFold3 and Protenix–v2) place every structure near the ground state and invert the experimentally observed population ordering. The information content of the seed ensemble, rather than the total simulation time, determines whether the transition is sampled. We further introduce ″population-weighted consensus score″ (PWCS) that ranks conformers by how representative they are of the equilibrium ensemble directly from the simulation data, without experimental restraints and independent of the choice of collective variables. PWCS identifies the dominant state even when low–dimensional representations fail to separate the states. Built entirely from open–source tools, REVEAL provides a transparent and extensible route to RNA conformational ensembles beyond the single–structure paradigm.