Prediction of potent shRNAs with a sequential classification algorithm

Abstract

We present SplashRNA, a sequential classifier to predict potent microRNA-based short hairpin RNAs (shRNAs). Trained on published and novel data sets, SplashRNA outperforms previous algorithms and reliably predicts the most efficient shRNAs for a given gene. Combined with an optimized miR-E backbone, >90% of high-scoring SplashRNA predictions trigger >85% protein knockdown when expressed from a single genomic integration. SplashRNA can significantly improve the accuracy of loss-of-function genetics studies and facilitates the generation of compact shRNA libraries.

Figure 1

Computational modeling of advancements in shRNA technology. (a) Sequential advances in shRNA dataset development. The schematic shows diverse biological shRNA potency datasets and their feature and class label distribution biases. Unbiased large-scale sets include a comprehensive representation of negatives but contain few positives (left panel). Sets selected using prediction tools show higher rates of positives, leading to a more complete representation of this class, at the cost of changing the feature distribution of the negatives (middle panel). Use of the optimized miR-E backbone that boosts primary microRNA processing changes the requirements for potent RNAi, altering the target prediction rule (right panel). (b) Concept and equation of SplashRNA. We model the advancement in shRNA technology as a sequential support vector machine (SVM) classifier. The first classifier is trained on miR-30 data to remove non-functional sequences and the second classifier is trained on miR-E data to increase prediction performance of the remaining shRNAs. The final output is a weighted combination of the scores from both classifiers.

Figure 2

Benchmarking SplashRNA prediction performance. (a) Precision-recall curves of SplashRNA performance on the external shERWOOD miR-30 dataset. The first classifier alone, Splash_miR-30 (area under the precision-recall curve, auPR: 0.615), shows the best performance. SplashRNA (area under the precision-recall curve, auPR: 0.506) compromises slightly on miR-30 data to increase prediction accuracy on miR-E shRNAs (b), while still outperforming three other si/shRNA prediction tools (DSIR, seqScore, miR_Scan). (b) SplashRNA performance on miR-E data. SplashRNA (auPR: 0.611) clearly outperforms the miR-30 classifier alone (auPR: 0.572) as well as three other prediction tools. (c) Identification of “gold-standard” essential genes. The hit detection accuracy of top SplashRNA predictions was compared to larger sets of shRNAs and to CRISPR tools. Prediction performance as measured by the area under the receiver operating characteristic (auROC) curve indicates that the accuracy of the top 10 SplashRNA predictions is no different than the performance obtained by testing 25 shRNAs (the entire library). Additionally, the 10 top scoring shRNAs are significantly better predictors of the gold-standard genes set than the 10 bottom scoring shRNAs by SplashRNA (p < 0.001, empirical permutation test) and the bottom 5 SplashRNA predictions have lesser predictive value than the bottom 10 (auROC: 0.747 vs. 0.819, respectively). The dashed line represents the 10% false positive rate (FPR) threshold. (d-e) Western blot validation of de novo SplashRNA predictions. All shRNAs were expressed using LEPG at single-copy conditions. β-Actin (Actb) was used for normalization. Short (top) and long (bottom) exposures are shown. Immunoblotting of (d) Pten (median knockdown 96%, median score 1.60) and (e) Bap1 (median knockdown 93%, median score 1.05) in NIH/3T3s (Sup Figure 6i). C, miR-30 and miR-E control shRNAs. (f) Score distribution of fifth highest SplashRNA predictions for all human and mouse genes, indicating the proportion of genes with 5 predictions above a given score. Predictions were run only on the intersection of all transcript variants per gene and after shortening of transcripts due to ApA. The inset shows the score distribution of all human and mouse SplashRNA predictions. The kink in the curves represents the transition from Splash_miR-30 to combined SplashRNA scores. At least 80% of genes have five shRNAs with prediction scores above 1.