To create this resource I would:

• get the table knownGene from UCSC: http://hgdownload.cse.ucsc.edu/goldenPath/hg18/database/knownGene.txt.gz
• build the cDNA and the translated protein using 'knownGene' and the reference sequences: http://hgdownload.cse.ucsc.edu/goldenPath/hg18/chromosomes
• fetch the swissprot entry for each gene
• align the swissprot entry and your reconstituted protein
• map the annotation from the swissprot entry to your reconstituted protein
• map the position of the domains back to the reference genome.

Few weeks back, I was also looking for such a resource for my analysis and realised exactly what you figured out: you won't be able to get this information from BioMart. I contacted the Ensembl help desk and they suggested me to integrate data using Ensembl resources (some of the data via Biomart and rest via the Ensembl Core/Variation API). So you have two options now, you may explore the Ensembl API path or proceed as described by Pierre using UCSC resources. Also remember, it will get a bit more complex due to alternate transcripts and alternate exons; This can change final protein product and exact genomic location of the domains, because of this complexity you may not be able to get a perfect one to one-mapping.

What you're describing is a coordinate conversion problem.

It's possible of course; if you know all of the required coordinates in both coordinate systems (i.e. exons and domains, in amino acid and nucleotide coordinates), but it is quite technically challenging.

One solution, if you're comfortable in Perl/Bioperl might be the Bioperl module Bio::Coordinate::GeneMapper, which was written for just this purpose. There may be similar libraries available for other languages.

As Pierre mentioned, you may also be able to use the UCSC tables, many of which have positional information.

You say that you are using Biomart? Does that mean your genome of interest is in Ensembl? If so, the work may already done for the annotated protein domains; these are stored as Bio::EnsEMBL::ProteinFeatures which have a location both on the protein (in protein coordinates) and on the genome (in chromosome coordinates).

To find these you would need to obtain genes, then transcripts and from those the translations. Given a translation, you can get the protein features and then filter these to include only those whose analysis type you require e.g. Pfam.

While this is possible according to the API docs, I don't know whether these data are present for your organism. It's probably worth checking, though because it will only take a short script to find out.

For COGs there is Genome ProtMap:

http://www.ncbi.nlm.nih.gov/sutils/protmap.cgi?cluster=COG4690E&result=map

The (very) hard way would be to map selected Pfam domains back to genome of interest using genewise.

PAL2NAL is a tool that can project a protein alignment onto nucleotide sequences. It's not exactly meant for what you want to do, but might be usable if you use the domain sequence and the nucleotide sequence of the gene.