The approach to annotating a genome critically affects the number and accuracy of genes identified in the genome sequence. Genome annotation based on stringent gene identification is prone to underestimate the complement of genes encoded in a genome. In contrast, over-prediction of putative genes followed by exhaustive computational sequence, motif and structural homology search will find rarely expressed, possibly unique, new genes at the risk of including non-functional genes. We developed a two-stage approach that combines the merits of stringent genome annotation with the benefits of over-prediction. First we identify plausible genes regardless of matches with EST, cDNA or protein sequences from the organism (stage 1). In the second stage, proteins predicted from the plausible genes are compared at the protein level with EST, cDNA and protein sequences, and protein structures from other organisms (stage 2). Remote but biologically meaningful protein sequence or structure homologies provide supporting evidence for genuine genes. The method, applied to the Drosophila melanogaster genome, validated 1,042 novel candidate genes after filtering 19,410 plausible genes, of which 12,124 matched the original 13,601 annotated genes. This annotation strategy is applicable to genomes of all organisms, including human.