BACKGROUND: Current RNA secondary structure prediction approaches predict prevalent pseudoknots such as the H-pseudoknot and kissing hairpin. The number of possible structures increases drastically when more complex pseudoknots are considered, thus leading to computational limitations. On the other hand, the enormous population of possible structures means not all of them appear in real RNA molecules. Therefore, it is of interest to understand how many of them really exist and the reasons for their preferred existence over the others, as any new findings revealed by this study might enhance the capability of future structure prediction algorithms for more accurate prediction of complex pseudoknots. METHODOLOGY/PRINCIPAL FINDINGS: A novel algorithm was devised to estimate the exact number of structural possibilities for a pseudoknot constructed with a specified number of base pair stems. Then, topological classification was applied to classify RNA pseudoknotted structures from data in the RNA STRAND database. By showing the vast possibilities and the real population, it is clear that most of these plausible complex pseudoknots are not observed. Moreover, from these classified motifs that exist in nature, some features were identified for further investigation. It was found that some features are related to helical stacking. Other features are still left open to discover underlying tertiary interactions. CONCLUSIONS: Results from topological classification suggest that complex pseudoknots are usually some well-known motifs that are themselves complex or the interaction results of some special motifs. Heuristics can be proposed to predict the essential parts of these complex motifs, even if the required thermodynamic parameters are currently unknown.