Feature Engineering for Biological Sequence and Protein Structure Prediction

Feature engineering remains central to predictive modeling in computational biology, governing how DNA, RNAand protein sequences are converted into numerical forms amenable to machine learning. The field hasprogressed from compact composition based descriptors such as amino acid composition (AAC) and dipeptidecomposition (DPC) to correlation based representations (pseudo amino acid composition, PseAAC, pseudoK-tuple nucleotide composition, PseKNC), evolutionary profiles (position specific scoring matrices, PSSM), and,more recently, to deep representation learning and biological foundation models. This narrative review focuseson biological sequence prediction, with particular emphasis on proteins and structure derived features. Webegin with Chou's five step framework for predictor construction, benchmark curation and redundancyreduction using CD-HIT and MMseqs2, and the specific challenges of multi label and hierarchical prediction. Wethen survey handcrafted descriptors, feature selection and dimensionality reduction, and the adaptation ofnatural language processing methods to biological sequences, including k-mer embeddings, convolutional andrecurrent architectures, and protein and genomic foundation models, including Transformer models such asESM-2 and DNABERT-2, the ESMFold structure predictor, and Hyena based genomic models. We distinguishthree common adaptation strategies: frozen embedding extraction, full fine tuning and parameter efficient finetuning (PEFT) - and discuss hybrid fusion of handcrafted and learned features, structure founded encodings(DSSP, Foldseek 3Di, ProstT5, SES-Adapter), and the role of explainable AI (SHAP, LIME, Integrated Gradients).Throughout we emphasize homology evaluation, nested validation to avoid selection leakage, confidencebased use of predicted structures, and cautious interpretation of attribution. We close with practical guidelinesfor choosing a representation under different data and compute regimes and outline open challenges.