Hey,

You can use a nice package by a guy who is on here - Brian Bushnell (if I remember correctly). It's called BBmerge and if you google it, you can find out the syntax for calculating paired end insert size for reads - to my mind it doesn't need a reference genome. I had to do this a few weeks ago when I was trying out SOAPtrans which asked for an insert size.

Hope that helps you.

1. There is no exact size, it's a distribution. This distribution is useful to know how many N's to insert when scaffolding.
2. The de-novo assembly is usually done without the knowledge of paired-end, each side is treated as if it's a single end to generate contigs. You should get long enough contigs to be able to map both ends of a fragment to estimate insert size (even with mate-pair).

1) Does one have to know the exact length of the insert for the paired end or mate-pair sequences to be useful?

It is important to have an estimate, so that the aligner can distinguish between _"proper"_ pairs that are likely to truely represent the molecule they originate from, and the artefacts where one mate is misaligned, usually very far from the other mate. How much "far" means depends on the method. For instance, in transcriptome sequencing, it is expected that some proper pairs will align hundreds of kilobases apart, and short read aligners such as BWA need to know that.

2) ... do you also have some knowledge of the approximate length of the insert ?

First, as explained above, the distribution observe lengths after alignment will differ according to the kind of sequencing method (transcriptome, genome, ...). In addition, for genome sequencing, the sequencing templates can be prepared in such a way that the distance after alignment should be within a given range.

3) How is this done in de novo sequencing...

De novo assembly typically takes advantage of the prior information on what the distance between the mates should be, in order to sort the contigs, predict gap size, etc.