As the demand for accurately aligning gene sequences to the genome of a related species grows with the sequencing of new genomes, spaced seeds emerge as a promising vehicle for increasing alignment sensitivity. We extend the existing {0, 1} match-mismatch models for sensitivity evaluation to take into account the compositional structure of coding sequences and ultimately produce seeds better suited to this particular application. Designing seeds for alignment programs, however, needs to balance sensitivity and specificity. We assess the effects of seed variations on both sensitivity and specificity in an extended model that incorporates transitions and differentiates among the three codon positions, and show that spaced seeds with transitions offer a better sensitivity-specificity tradeoff. Furthermore, we propose a theoretical formulation for rigorously assessing seed specificity, starting from Bernoulli and Markov models of the mRNA and genomic sequences. Within this framework, we perform the first comprehensive analysis of seeds to serve as a blueprint for selecting sensitive and specific seeds for practical applications. Our analyses show that specificity is relatively constant for seeds of a given weight, while sensitivity varies widely, with the highest values attained by seeds allowing a small (2-6) number of transitions.A strategy for designing seeds, therefore, is to first select the weight of the seed by identifying the desired sensitivity-specificity tradeoff, then choose the most sensitive seed(s) within that weight group. We illustrate our methods with the alignment of chicken coding sequences against the human genome assembly version HG17.