Division of Engineering Research and Centre of Excellence for Advanced Sciences, National Research Centre, Tahrir Street, 12311 Cairo, Egypt.
J Biomed Inform. 2012 Jun;45(3):528-34. doi: 10.1016/j.jbi.2012.02.005. Epub 2012 Feb 25.
The investigation of small interfering RNA (siRNA) and its posttranscriptional gene-regulation has become an extremely important research topic, both for fundamental reasons and for potential longer-term therapeutic benefits. Several factors affect the functionality of siRNA including positional preferences, target accessibility and other thermodynamic features. State of the art tools aim to optimize the selection of target siRNAs by identifying those that may have high experimental inhibition. Such tools implement artificial neural network models as Biopredsi and ThermoComposition21, and linear regression models as DSIR, i-Score and Scales, among others. However, all these models have limitations in performance. In this work, a neural-network trained new siRNA scoring/efficacy prediction model was developed based on combining two existing scoring algorithms (ThermoComposition21 and i-Score), together with the whole stacking energy (ΔG), in a multi-layer artificial neural network. These three parameters were chosen after a comparative combinatorial study between five well known tools. Our developed model, 'MysiRNA' was trained on 2431 siRNA records and tested using three further datasets. MysiRNA was compared with 11 alternative existing scoring tools in an evaluation study to assess the predicted and experimental siRNA efficiency where it achieved the highest performance both in terms of correlation coefficient (R(2)=0.600) and receiver operating characteristics analysis (AUC=0.808), improving the prediction accuracy by up to 18% with respect to sensitivity and specificity of the best available tools. MysiRNA is a novel, freely accessible model capable of predicting siRNA inhibition efficiency with improved specificity and sensitivity. This multiclassifier approach could help improve the performance of prediction in several bioinformatics areas. MysiRNA model, part of MysiRNA-Designer package [1], is expected to play a key role in siRNA selection and evaluation.
小干扰 RNA（siRNA）及其转录后基因调控的研究已经成为一个极其重要的研究课题，无论是出于基础原因还是出于潜在的长期治疗益处。有几个因素会影响 siRNA 的功能，包括位置偏好、靶标可及性和其他热力学特征。最先进的工具旨在通过识别那些可能具有高实验抑制作用的 siRNA，来优化靶 siRNA 的选择。这些工具采用人工神经网络模型，如 Biopredsi 和 ThermoComposition21，以及线性回归模型，如 DSIR、i-Score 和 Scales 等。然而，所有这些模型在性能上都存在局限性。在这项工作中，我们开发了一种基于组合两个现有评分算法（ThermoComposition21 和 i-Score）的新 siRNA 评分/功效预测神经网络模型，以及全堆叠能（ΔG），应用于多层人工神经网络。这三个参数是在对五个著名工具进行综合比较组合研究后选择的。我们开发的模型“MysiRNA”在 2431 个 siRNA 记录上进行了训练，并使用另外三个数据集进行了测试。在评估研究中，MysiRNA 与 11 种替代的现有评分工具进行了比较，以评估预测和实验 siRNA 效率，它在相关性系数（R（2）=0.600）和接收者操作特征分析（AUC=0.808）方面都取得了最高的性能，相对于最佳可用工具的敏感性和特异性，预测准确性提高了高达 18%。MysiRNA 是一种新型的、免费可访问的模型，能够以提高的特异性和敏感性预测 siRNA 抑制效率。这种多分类器方法可以帮助提高几个生物信息学领域的预测性能。MysiRNA 模型是 MysiRNA-Designer 软件包[1]的一部分，预计将在 siRNA 选择和评估中发挥关键作用。
