Excerpt from the review:
Does PTGS by dsRNA Involve an RNase III–Like Enzyme?
Although the identity of the RNAi nuclease has not been determined, the characteristics of the short 21–25 nucleotide RNA pieces suggest they were generated by RNase III or a highly related enzyme (see  and , and references therein). RNase III is the only characterized nuclease known to cleave dsRNA at specific sites to generate dsRNA fragments of discrete sizes. For RNase III to stably bind a dsRNA, it must be at least two helical-turns in length, consistent with the observation that RNAi and transgene-induced silencing yield stable fragments of 22 base pairs. RNase III can produce fragments <22 class="apple-converted-space"> Figure 1, fragments less than 21–23 base pairs would not have been observed in the recent experiments because they would not remain stably bound to the enzyme and thus would be more accessible to degradation by other cellular nucleases.
Given the similarities between the cleavage products of RNase III and the RNAi nuclease, I have incorporated properties of the RNase III enzymes into the model of Figure 1. For example, RNase III makes staggered cuts that leave 3′ overhangs of two base pairs, as shown for the 23-mers of Figure 1. If RNAi involves an RNase III-like enzyme, it might explain why the small RNAs observed by Zamore and Tuschl range from 21–23 nucleotides. The initial cleavage might produce dsRNAs comprised of sense and antisense 23-mers, but the 3′ overhangs would be more accessible to single-strand–specific nucleases present in the extract, and trimmed to 21 and 22 nucleotide pieces. Zamore and Tuschl observe that cleavage of the dsRNA, unlike mRNA cleavage, does not absolutely require ATP. However, dsRNA cleavage is faster in the presence of ATP, and without ATP the pieces are predominantly the longer 23-mers. Certainly this is a clue to the role of ATP in this in vitro reaction, but at present its meaning is unclear.