Comparing artificial neural networks, general linear models and support vector machines in building predictive models...

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning in Genomics

Background:

• Exogenous short interfering RNAs (siRNAs) are crucial for gene knockdown but exhibit variable efficacy.
• Machine learning models are increasingly used to predict siRNA effectiveness, yet model comparison methods lack consensus.
• The impact of learning techniques, feature sets, and cross-validation on predictive model performance remains underexplored.

Purpose of the Study:

• To establish a robust statistical framework for comparing predictive models of siRNA efficacy.
• To evaluate the influence of different learning techniques and feature mapping methods on model performance.
• To identify optimal modeling approaches for predicting effective siRNA sequences.

Main Methods:

• Developed predictive models using Artificial Neural Networks (ANNs), General Linear Models (GLMs), and Support Vector Machines (SVMs).
• Employed five distinct feature mapping methods to generate models of siRNA activity.
• Utilized a 3x5 factorial Analysis of Variance (ANOVA) to assess the impact of learning technique and feature mapping.

Main Results:

• Both learning techniques and feature mapping significantly influenced predictive model variance, affecting precision and accuracy.
• Artificial Neural Networks (ANNs) and General Linear Models (GLMs) showed sensitivity to noisy features.
• Support Vector Machines (SVMs) demonstrated greater robustness with a larger number of features for precision and accuracy metrics.

Conclusions:

• The study provides a statistical framework for critically evaluating and comparing siRNA predictive models.
• Support Vector Machines (SVMs) with specific feature combinations significantly outperformed ANNs and GLMs in predictive accuracy.
• Feature relevance interpretation requires caution due to inconsistencies across different learning techniques.