Short interfering RNAs (siRNAs) are widely used to silence the expression of specific genes. Current practice for designing effective siRNAs is to use algorithms based on sequence-efficacy correlations; however, there are many highly effective sequences that these algorithms do not anticipate. To ensure that the best siRNAs are identified, all possible gene-specific siRNA sequences of appropriate lengths should be screened in cell culture. Synthesizing and testing all such sequences individually is costly. A potentially much easier alternative is to prepare a mixture of all these sequences (a gene-specific library), express them in cells, select cells having the desired phenotype, and identify the siRNA contained within the selected cells. Here we describe two new methods for preparing and expressing such libraries. The first uses cloned Dicer or RNase III to digest gene-specific RNA duplexes to siRNAs, which are then converted to the corresponding DNA sequences by attaching RNA primers and performing reverse transcription-PCR. The second method involves partial DNase I digestion of gene-specific DNA, purification of a 20-30-bp fraction, and amplification by attaching DNA adapters followed by PCR. DNA libraries specific for TNF-alpha, DsRed, and part of the hepatitis C virus genome, generated by methods, were inserted into siRNA expression vectors between convergent human U6 and H1 promoters. Randomly selected clones from each library together with vectors expressing the corresponding target genes were cotransfected into 293FT cells and assayed for target gene inhibition. About 10%-20% of siRNAs represented in these libraries show significant inhibition of their target genes. Most of these inhibitory sequences are not predicted by existing algorithms.