.png){kind=icon}
返回
Ensemble learning-based comparative study of feature representations for protein structural class classification
.png){kind=icon}
OA
AI
DOI：10.1007/s42452-026-09334-1.png)
.png){kind=icon}
原文求助.png){kind=icon}
分享.png){kind=icon}
收藏
摘要
En 中文
This study presents a feature-centric comparative analysis of protein structural class classification through ensemble modeling. We extract three complementary categories of features: (i) embedding-based features, derived from W2V-CBOW, W2V-SkipGram, fastText-CBOW, and fastText-SkipXGram models, where protein k-mers are converted into dense vectors and aggregated to represent full sequences; (ii) SkipXGram features, which leverage the skip-gram principle on structural sequence forms to capture contextual dependencies of tertiary elements such as helices, strands, and coils; and (iii) statistical descriptors, summarizing structural properties including frequency distributions, longest consecutive runs of H/E/C elements, and overall sequence length. Classification is performed using an ensemble VotingClassifier that integrates LinearSVC, Random Forest, and Gradient Boosting, enabling a systematic evaluation of individual feature types as well as their combinations. Experimental findings show that fusing embedding, structural-contextual, and statistical representations enhances classification accuracy, underscoring the effectiveness of integrated feature-driven ensemble learning for protein structure prediction.
Keyword:
Protein structure classification
Feature fusion
Embedding models
SkipXGram
Ensemble learning
