Design of a genome-wide siRNA library using an artificial neural network.

The largest gene knock-down experiments performed to date have used multiple short interfering/short hairpin (si/sh)RNAs per gene. To overcome this burden for design of a genome-wide siRNA library, we used the Stuttgart Neural Net Simulator to train algorithms on a data set of 2,182 randomly selected siRNAs targeted to 34 mRNA species, assayed through a high-throughput fluorescent reporter gene system. The algorithm, (BIOPREDsi), reliably predicted activity of 249 siRNAs of an independent test set (Pearson coefficient r = 0.66) and siRNAs targeting endogenous genes at mRNA and protein levels. Neural networks trained on a complementary 21-nucleotide (nt) guide sequence were superior to those trained on a 19-nt sequence. BIOPREDsi was used in the design of a genome-wide siRNA collection with two potent siRNAs per gene. When this collection of 50,000 siRNAs was used to identify genes involved in the cellular response to hypoxia, two of the most potent hits were the key hypoxia transcription factors HIF1A and ARNT.

An oligonucleotide microchip for genome-wide microRNA profiling in human and mouse tissues.

MicroRNAs (miRNAs) are a class of small noncoding RNA genes recently found to be abnormally expressed in several types of cancer. Here, we describe a recently developed methodology for miRNA gene expression profiling based on the development of a microchip containing oligonucleotides corresponding to 245 miRNAs from human and mouse genomes. We used these microarrays to obtain highly reproducible results that revealed tissue-specific miRNA expression signatures, data that were confirmed by assessment of expression by Northern blots, real-time RT-PCR, and literature search. The microchip oligolibrary can be expanded to include an increasing number of miRNAs discovered in various species and is useful for the analysis of normal and disease states.

Bridging expressed sequence alignments through targeted cDNA sequencing.

One of the major challenges in genome research is the identification of the complete set of genes in a genome. Alignments of expressed sequences (RNA and EST) with genomic sequences have been used to characterize genes. However, the number of alignments far exceeds the likely number of genes in a genome, suggesting that, for many genes, two or more alignments can be joined through overlapping sequences to yield accurate gene structures. High-throughput EST sequencing becomes less efficient in closing those alignment gaps due to its nonselective nature. We sought to bridge these alignments through a novel approach: targeted cDNA sequencing. Human expressed sequences from GenBank version 124 were aligned with the genomic sequence from NCBI build 24 using LEADS, Compugen's EST and RNA clustering and assembly software system. Nine hundred forty-eight pairs of alignments were selected based on EST clone information and/or their homology to the same known proteins. Reverse transcriptase PCR and sequencing yielded sequences for 363 of those pairs. These sequences helped characterize over 60 novel or otherwise incomplete genes in the recent UniGene build 153, which included over 1 million additional ESTs. These results indicate that this integrated and targeted strategy, combining computational prediction and experimental cDNA sequencing, can efficiently generate the overlapping sequences and enable the full characterization of genomes. Additional information about the contig pairs, the resultant overlapping sequences, tissue sources, and tissue profiles are available in a supplemental file.