Triplexis: Prediction of DNA Triplex Formation Using Deep Learning on Synthetic Libraries

Abstract: A triplex forms when a single RNA strand joins the DNA double helix through Hoogsteen base pairing, leaving the Watson–Crick pairs between the two DNA strands intact. Such triplexes have been shown to regulate gene expression in cells and could enable programmable molecular targeting. Predicting which RNA sequences bind to which DNA sites is challenging: experiments cover little of the vast interaction space, and their measurements confound molecular recognition with sequence abundance and experimental biases. We generate datasets using TripClick-seq, a novel assay that chemically links short strands to DNA sites so pairs can be counted. For RNA, deep learning models predict observed pair abundance (held-out Pearson r ≈ 0.40–0.52), performing comparably to nucleotide-composition baselines. Attribution analysis identifies a strong preference for RNA rich in uridine (U), while comparisons of sequences with identical composition reveal additional sensitivity to base order. We then computationally screen all 16.8 million possible 12-base RNA sequences against a GAA repeat target associated with Friedreich ataxia, an inherited neurological disease. Highly ranked candidates contain U-rich stretches interrupted by other bases, motivating experimental tests of whether these interruptions help RNA align with repeated DNA sites. These candidates provide a starting point for developing synthetic RNA-based regulators of gene expression at disease-associated DNA repeats.