Abstract
Long non-coding RNAs (lncRNAs) regulate gene expression through various mechanisms, including the specific formation of RNA:DNA:DNA triple helices (triplexes) via Hoogsteen hydrogen bonding in the DNA major groove. While computational prediction of these sites typically relies on sequence compatibility, current tools often yield high false-positive rates. In this study, we investigated whether incorporating DNA 3D structural features - specifically Helix Twist, Minor Groove Width, Propeller Twist, and Roll - can enhance the identification of functional triplex-forming sites. Analyzing 25 ChIRP-seq datasets from human and mouse, including seven lncRNAs with experimentally validated triplex activity, we identified distinct 3D shape profiles in target regions, characterized by lower helix twist and higher propeller twist. By employing nested logistic regression models to compare shape-augmented predictions against the state-of-the-art sequence-based tool (3plex), we demonstrate that integrating DNA shape features increases the Area Under the Curve (AUC) by 17.5 %. Crucially, this predictive improvement remains significant even when controlling for chromatin accessibility, indicating that DNA shape provides structural information independent of chromatin state. These findings suggest that sequence complementarity alone is insufficient for effective binding; rather, the DNA helix must adopt a "receptive" conformation to accommodate the third RNA strand. Our approach acts as a structural filter, distinguishing between thermodynamically plausible and structurally compatible sites, thereby refining the prioritization of lncRNA targets for experimental validation.