Development and performance evaluation of robust methods for protein secondary structure prediction

Abstract

newline In the post-genomic era, the study of the sequence to structure relationship and newlinefunctional annotation have an extensive role in molecular biology. The structural newlineclass of proteins has an imperative role in rational drug design, pharmacology newlineand furnishes useful insight towards protein structure determination. Notwith- newlinestanding,theexponentialdevelopmentofnewlydiscoveredproteinsequencesby newlinemainstream scientific communities has moved a wide gap between the quantity newlineofsequence-knownandstructure-knownproteins. Henceforththereexistsabasic newlineneed to establish computerized strategies for quick and exact assurance of pro- newlineteinstructure. Proteinsecondarystructurepredictionisatwo-stepprocess. Inthe newlinefirstphase,therawaminoacidsequencesaresubstitutedinahighlyinterpretable newlineway to embed conformational, physiochemical properties of the molecule. In the newlinesecond phase, the converted sequences are fed to a classifier for further predic- newlinetion. In conventional methods for representing raw amino acid sequences, the newlineimpact of noise has been severely overlooked, which can critically hamper the newlineoverall prediction accuracy of the subsequently developed classifier. In biomed- newlineical signals, noise can creep into the samples at various stages of data curation, newlinehence can t be averted altogether. Like most machine learning methods, contem- newlineporary approaches are deprived of generalization ability. Models trained with newlinesome set of data consuming a significant amount of time, fail to predict when newlinesubjected to a different dataset. newlineHere, we address three concerns in protein structural class prediction. newline1. Development and evaluation of noise-resistant representation for protein newlinesequences. newline2. Development of classifiers which can be trained faster with advanced op- newlinetimization techniques like stochastic gradient descent (SGD). newlinevi newline3. Development of deep learning based models with better generalization