This answer may not reduce confusion: precise reconstruction genes is still an un-attained goal of genome informatics, with many approaches and differing interpretations of what are biologically accurate gene reconstructions. An expert curator can often look at evidence such as your gene map and decide, but none of the automated methods will return those expert decisions for all genes and alternates.

My approach in the EvidentialGene project is to use coding sequence measures for biological validation, and to include any and all the gene reconstruction methods you mention, then reduce that over-assembly of too many models by scoring their coding accuracy (homology to other species, plus other bits). If you accept that a consensus of genes across species is a good biological ruler, then this approach returns a more precise gene set than others I've tested, with or without reference chromosomes to guide gene reconstruction.

This is a recent summary of how well EvidentialGene works, and why, versus both chromosome-gene modelling, and de-novo gene assembly of animals and plants. This is a recent independent assessment of EvidentialGene with the spiny mouse.

With EvidentialGene, you should apply as many gene reconstruction/assembly methods as you can afford if they return some valid genes. I aim for around 100 models of each transcript, i.e. 10 million alternate transcript models, using different assemblers and different settings (many KMER steps), for complete and accurate gene set reconstructions. Your de-novo assembly methods are a good selection: Velvet/Oases is most accurate, when run properly; idba-tran second, and SOAP-trans thirdish, Trinity is fourth. The above evigene_bothgalmod1606iu documents this.

For the chromosome mapped methods, cufflinks makes more errors than de-novo assembly (ommission/commisson) last time I tested, but it does return some most accurate of the weakly expressed genes. I gave up on Scripture a while back as too error prone. I've not tested String-Tie.

EvidentialGene works with gene transcripts of any source (chromosome-based or not). It returns the subset of most accurate coding genes, classified into alternate transcripts/locus. The current version for transcripts doesn't use chromosome mapping to classify loci, but I'm working on that addition. The locus calls it makes by gene exon alignment are accurate, but for a small subset of high identity paralogs versus alternates. You can use one of various transcript x chromosome aligners to get back to your genome map views: GMAP, BLAT, even BLASTn.

Caveat: EvidentialGene documentation is still sparse, not in the easy to follow category. At the Galaxy meeting this summer I tried to interest those folks to help make this plug-in ready to use, so far no luck. But others with small informatics expertise do use this method with success (and griping at the docs :).

There is a trick to accurate de-novo assembly: use large KMER values (read shredding) up to size of your paired reads. Velvet/Oases uses this well: for complex genes, the most accurate subset are assembled with kmers in 60-90 bp range for 100 bp paired Illumina reads. And VelO runs much faster, with less memory, at these high KMER settings. You can use it for only a hi-K subset, other methods for low-K. Idba-tran also uses the kmer range well, and paper on it explains how/why this helps, esp. with alternate transcript reconstruction. This is a published trick, but one most have not noticed in this complex literature.

-- Don Gilbert