In theory you can: You could take your fastq and align it using bwa. And from the sam file you could filter out the aligned reads and re-construct a fastq file with the unaligned ones and supply that to tophat. However, as the first stage of mapping, tophat will use bowtie. But I'd guess that there shouldn't be many reads that are mapped using bowtie on unaligned reads.
But then, I also suggest (as Geparada) the approach of using tophat2 and bowtie2, although bowtie2 seems to be in beta version. I am testing this right now.