Whether to discard row 1 depends purely upon your criterion. Commonly, one requires X number of samples with values at least Y. Having said that, don't use FPKMs in normal RNAseq for statistics.

Time to read count based methods for RNASeq DE analysis and get a hang of what kind of selection one should be applying to get more compact gene sets having read outs(raw read counts). The manual of egdeR and DESEQ2 should be having the different thresholds for consideration while before applying the normalization metrics, I do not remember if the latest one is having or not but older ones did cite some examples. Either you consider rowsum or you can consider 40% of the first quantile but these entirely depends on the experiments and the read count sets you receive. So you can see that yourself and then apply. FPKM is not used for making any DE analysis, it is normalized gene expression data which is not for any statistical metrics as @Devon pointed out.