Fast and reliable identification of different species of the genus Candida is important to define adequate therapeutic decisions, because the different species have highly variable susceptibilities to antifungal drugs; azoles and amphothericin B. Accurate statistical records on case history and epidemiological studies also depend on effective identification. To address this problem we established a RAPD method that enabled direct identification of five very common species of Candida. Initially, reference band patterns were established for C. albicans, C. tropicalis, C. parapsilosis, C. glabrata and C. krusei. One of the primers, M2, showed remarkably conserved intra-specific patterns of approximately 10 bands each, ranging in size from 2.0 to 0.1 kb. These patterns were significantly different and species-specific. Few bands were conserved between different species of Candida, which was assumed to be consistent with their phylogenetic relatedness. In addition, band patterns were constant and reproducible and DNA isolated from single colonies yielded sufficient DNA for identification. The reference band patterns were then used, in blind experiments, to identify species of Candida in 50 randomly chosen samples, including clinical isolates and ATCC strains. RAPD results were 100% consistent with results obtained by conventional diagnostic methods and were achieved in one day instead of several days taken by conventional methods. Because ideal identification methods should be consistent with phylogeny and taxonomy we tested whether RAPD could be used to calculate genetic distances. Comparison of RAPD phylogenetic trees with 18S rRNA trees showed significant differences in tree topologies which indicated that RAPD data could not accurately measure the relative distances between different species. Also, computer simulations of RAPD random patterns were used to test whether the observed degree of RAPD band pattern similarities could occur at random. These simulations suggested that the level of inter-specific band pattern similarities observed in our data could be obtained at random, while intraspecific pattern similarities could not. RAPD would be helpful to discriminate between isolates but not to quantitate the differences. We suggest that the inaccurate estimate of genetic distances from RAPD is a general limitation of the technique and not a specific problem of our identification method. Because of the repetitive character of the target sequences, genetic distances calculated from RAPD could be affected by paralogy, namely, recombination and duplication events not parallel with speciation events.