HI Xavier, thanks a lot for your answer :) !

This indicative number of ~100 libraries per run is exactly what I needed; I could not find this information in the Illumina website. I was searching this information for a classical RAD-seq protocol, actually just to get a glance at the cost of applying RAD-seq for a very large number of individuals. Mean expected number of reads per library is not the only limitation I notice, as you could potentially pool more individuals (and still get a correct coverage, depending on what you want to do with the data of course) than allowed by ready-to-use Illumina kits. But I might be wrong !

Thank you very much again, all the best !!

Since you are using RADseq, when calculating coverage don't count on getting the full output of the lane, as the bases used for cluster delimitation (I think the first 11 bases?) are low diversity because this overlaps with the restriction site overhang (e.g. TGCAG for PstI). So in order to aid the system in discriminating between clusters on the flow cell the core facility will often add a phiX spike-in (I think our facility usually uses 15-50% phiX spike in). For our facility, the HiSeq 4000 usually yields ~280M reads for low-diversity libraries and 350-400M for 'balanced' libraries. You can add a variable length insert in your P5 adaptors to increase diversity and squeeze some more yield out, but make sure you tell the facility about any custom adapter changes you make so they can figure out how best to load your libraries. See Tin et al (2015). 

Although if you are going to mess with the adapter sequences, you may also consider adding a degenerate region for detection of PCR duplicates, although this isn't necessary for RAD methods with length polymorphism (like when doing paired-end sequencing following the Baird et al 2008 protocol). Attached some papers on adapter modifications. We use the Peterson et al (2012) dual-indexing strategy, which has a P5-end 'barcode' and a P7 TruSeq index. 

For calculating coverage and allowable level of multiplexing you need the yield and also target number of loci and desired average coverage. We do ddRAD and predict loci yield using in silico digest of reference genomes, and usually target around 40,000 loci at a minimum coverage of 30X reads per locus. So for 280M reads on the HiSeq 4000: 280M reads / 40K loci / 30X coverage = 233 individuals could be multiplexed. We use 48 P5 barcodes per P7-end TruSeq index, and usually pool 96 dual-index individuals (=2 TruSeq indices @ 48 P5 barcodes each) per lane, because we are used to the yield of the HiSeq 2500 and have only switched to the HiSeq 4000 very recently. Soon we will likely move to 144 or 192 individuals per lane. 

Another note is that using the dual-indexing strategy of Peterson et al you don't need to sequence paired-end, as the TruSeq indices are supported by the Illumina platform and have their own sequencing primer, and will be sequenced in a special short 'index' read cycle. But with classical RADseq (Baird et al 2008) you probably want to go paired-end so you can build longer contigs. 

Additionally if you are sequencing a VERY large number of samples (into the thousands) you can save a significant amount by combining RAD with a target-capture approach (e.g. Rapture or RAD-cap). The idea here is that many loci may be too randomly recovered, or have a systematic allele dropout caused by mutation in a restriction site, or be repetitive, highly conserved... Many reasons to not necessarily be informative for whatever downstream analyses you have. So you do RAD to scan for 'useful' loci, design probes (see Mycroarray MYbaits) for those, and then enrich your libraries for just those loci. Then, rather than wasted sequencing effort on 'bad' loci, you could target a subset and sequence many more individuals per lane. I'm working on a project now that is expected to scale to a large size and we plan to use probes to target ~1/4 of the loci, meaning we can multiplex 4 times as many per lane without sacrificing coverage. So rather than (I think, approximate) a sequencing cost now of about $16 per sample, we would pay more like $4 per sample (with a negligible change in library prep per sample, other than up-front cost of the baits). Attached some papers on that as well. 

Probably way more information than you were looking for, but if not helpful for you than maybe for someone else. Good luck!

Article Degenerate adaptor sequences for detecting PCR duplicates in...

Article Detection and Removal of PCR Duplicates in Population Genomi...

Article RAD Capture (Rapture): Flexible and Efficient Sequence-Based...

Article RADcap: Sequence Capture of Dual-digest RADseq Libraries wit...

Article Double digest RADseq: An inexpensive method for de Novo SNP ...

Thank you very much Tyler ! This was very helpful and interesting. I have heard about the Rapture but I did not notice someone already used it :-) That's good to know !

All the best !

Anyone having experience with sequencing RAD libraries on the Hiseq 4000 would like to share? Due to the unresolved issues with this platform, I am hesitating between sticking to Hiseq2500. How much did you spike in the libraries on the Hiseq4000? Thanks for any input