Fig 1.
RNA primary (left), secondary (middle), and tertiary structures (right). The RNA folding process is hierarchical, i.e., the RNA secondary structure forms rapidly from linear RNA (primary structure) with a large energy loss, while the formation of a complex tertiary structure is usually much slower.
Table 1.
Summary of the ML-based RNA secondary structure prediction methods.
Fig 2.
Framework for RNA secondary structure prediction methods with ML-based score schemes. Wet lab data, RNA sequence data, or RNA structure data can be employed to train an ML model to obtain a score scheme.
Using this score scheme, an RNA secondary structure can be predicted using a traditional score-based approach from a single RNA sequence.
Fig 3.
Framework for RNA secondary structure prediction methods with ML-based preprocessing or postprocessing.
In RNA secondary structure prediction, ML models (trained by sequence data, in green) can be also used in pretreatment for selecting an appropriate prediction method or a group of appropriate parameters; ML models (trained by structure data, in brown) also can provide a means of determining the most likely structures among the outcomes.
Fig 4.
Framework for the RNA secondary structure prediction methods with ML-based prediction process.
ML models (trained by wet lab, RNA sequence, or RNA structure data) are directly used to predict RNA secondary structures in an end-to-end way or followed by a filter or optimizer to obtain the optimal RNA secondary structure.