You could take a metagenomic approach first - assign each sequence as either more plant-like or more fungus-like using something like BLAST against a database of representative protein coding sequences from relevant organisms.

Then once you have the reads from the two organisms more or less separated into two bins, you could make a separate de novo assembly of the transcriptome of each organism using Trinity. You would then need to annotate these assemblies by aligning the assembled sequences to relevant reference genomes, or using something like PFAM.

Then you could do an analysis of expression by aligning each set of reads from the two bins in each sample against your assembled transcriptomes, and counting numbers of aligned reads against each transcript sequence, and assessing differential expression using standard methods.

Not for the faint hearted, but certainly possible.

I agree with Jonathan. This approach should work. I recently did something similar with (mostly) bacterial metagenomic data and a different assembler. It worked so well that we assembled a draft genome for one of the target organisms. Of course eukaryotes are a lot more complicated, but you ought to be ok.

One possible variation. If you have lots of reads (e.g. tens of millions), BLASTing all of them could be a problem. If this is the case, you can assemble all your reads first, then BLAST the assembled contigs in order to bin them. The assembler should tell you which reads went into which contigs. Note that the assembly will probably be terrible; it's only purpose is to let you BLAST and bin a few thousand contigs rather than a few million reads. After you bin the contigs, determine bins for the contributing reads, throw away the contigs, and if you want to assemble, assemble the bins of reads one at a time. 

Hope this helps.

Thanks! We will try it out. I was a bit worried about the task as I haven't worked with organisms where you have no full genome data to map the reads. It will be an interesting challenge. Looking forward.

@Philip Heller You are right about BLAST being too slow.  The authors of the MEGAN software (which uses BLAST-format as an input) recommend PAUDA or their own DIAMOND as alternatives, up to 5 orders of magnitude faster than BLAST, but maintaining the BLAST alignment format.