A new class of gene-prediction algorithms that recently been reported has shown the power of comparative genomics. The existing genefinding algorithms focus on locating exons in genomic sequence which have limitation on the input sequences as well as lack of statistical confidence. Another algorithm designed to detect conserved structural RNAs along with detecting coding regions is computationally heavy and is focused on structural RNAs. We use sequence similarity between human and mouse to classify alignments into coding regions and non-coding regions. Based on the aligned sequences of human and mouse, we propose a log-odds ratio score that based on conservation measurements and use the distribution of log-odds ratio scores of a fixed window size of a gapless alignment to separate alignments that contain coding regions from alignments that do not contain coding regions. The confidence level of our predictions of new coding regions is given by a multiple hypotheses testing that controls false discovery rate. The correctness of our prediction is validated by the 1M alignments of ancient repeats and 90,000 exons from refSeq mRNA and 932 pseudogenes produced by Sanger Institute.