Artificial neural network prediction of antisense oligodeoxynucleotide activity

Abstract

An mRNA transcript contains many potential antisense oligodeoxynucleotide target sites. Identification of the most efficacious targets remains an important and challenging problem. Building on separate work that revealed a strong correlation between the inclusion of short sequence motifs and the activity level of an oligo, we have developed a predictive artificial neural network system for mapping tetranucleotide motif content to antisense oligo activity. Trained for high-specificity prediction, the system has been cross-validated against a database of 348 oligos from the literature and a larger proprietary database of 908 oligos. In cross- validation tests the system identified effective oligos (i.e. oligos capable of reducing target mRNA expression to <25% that of the control) with 53% accuracy, in contrast to the <10% success rates commonly reported for trial-and-error oligo selection, suggesting a possible 5-fold reduction in the in vivo screening required to find an active oligo. We have implemented a web interface to a trained neural network. Given an RNA transcript as input, the system identifies the most likely oligo targets and provides estimates of the probabilities that oligos targeted against these sites will be effective.

Prediction of Antisense Oligonucleotide Efficacy Using Local and Global Structure Information with Support Vector Machines

“…Among all 9 variations considered in Table 1, the best performance (ROC = 0.973) was achieved when the GS and Triplet profiling are used in conjunction on mRNA. This performance is apparently better than that of the previous methods on the same data set [5,6]. Profiling on parenthesis orientation sensitive triplets and quadruplets was also tested, but did not improve the performance, instead the performance decreased slightly (results not reported here).…”

“…Several computational methods have thus been developed to this end [3,4,5,6]. Essentially, these methods rely on various sequential and structural features to differentiate antisense oligonucleotides with high efficacy from those with low efficacy.…”

“…For example, in a paper by Matveeva et al, the occurrence of some motifs of 3 or 4 bases long, such as CCAC (GGGG), are reported to be correlated with high (low) activity efficacy. In the work by Giddings et al, 2002, an artificial neural network was developed to scan the AO sequences of all 256 tetranucleotide motifs, and achieved an accuracy of 53% for high efficacy AOs. In [5] by Camps-Valls et al, 2004, a support vector regressor was developed to select among a variety of features the most discriminating ones in identifying high efficacy AOs, and a success rate of 83.3% was reported for predicting AOs with activity efficacy higher than 0.75.…”

Ezrin mRNA target site selection for DNAzymes using secondary structure and hybridization thermodynamics

“…There has also been much interest in computational approaches to the to the secondary stuctures prediction of mRNA have advantages over experimental methods in terms of throughput, cost, and efficiency. Some approaches to efficacy prediction have been proposed for rational selection of target sites [29][30][31][32]. Thierry et al [33] indicated that single-stranded loop regions in mRNA secondary structure were the best target sites for accessibility of antisense reagents.…”

Profiled support vector machines for antisense oligonucleotide efficacy prediction

“…Another method is to look for motifs that occur more often in effective AOs. Ten sequence motifs have been identified with a correlation to AO efficacy in [20], and recently, motifs have been used as the input to neural network models [21,22] with reasonable success.…”