Protein Characterization using Deep Learning Models

Loading...
Thumbnail Image

Files

TR Number

Date

2026-05-26

Journal Title

Journal ISSN

Volume Title

Publisher

Virginia Tech

Abstract

Protein characterization is a fundamental problem in computational biology because protein function is shaped by sequence, evolutionary history, and three-dimensional structure. Recent advances in protein language models, large-scale reference databases, and structure prediction methods have created new opportunities to characterize proteins using multiple complementary data modalities. Although this dissertation explores specific biological applications, the computational strategies investigated here are broadly applicable to protein sequence and structure-based downstream classification tasks. However, the contribution of each modality is often task dependent, and existing machine-learning evaluations may over- estimate true performance when homologous sequences are not partitioned carefully across training and test sets. These challenges motivate the need for robust, interpretable, and generalizable frameworks that integrate sequence-based, alignment-based, and structure-based information while supporting fair and realistic benchmarking. This dissertation investigates how different computational representations of proteins can be used individually and in combination to improve protein characterization across diverse biological tasks. In Chapter 2, we study protein characterization using protein language models and alignment-based scoring, developing hybrid frameworks that combine learned sequence representations with homology-based evidence to improve classification and an- notation, particularly in low-similarity and low-data settings. While this chapter focuses on antimicrobial resistance gene characterization, the framework is relevant to broader pro- tein annotation problems where learned representations and evolutionary evidence provide complementary information. Chapter 3 extends this direction to subcellular localization prediction by examining how protein language models and structural information can be integrated to capture biologically meaningful patterns associated with protein localization. More broadly, this chapter evaluates how sequence-derived embeddings and structure-aware representations can support downstream classification tasks where protein function depends on spatial organization, residue interactions, or geometric constraints. Chapter 4 addresses a key methodological challenge in biological machine learning by introducing a homology- aware protein sequence partitioning framework for robust model evaluation. This framework is designed to reduce homology bias and data leakage, enabling more realistic benchmarking of machine-learning methods across protein sequence datasets. Finally, Chapter 5 summarizes the main findings of the dissertation and discusses future directions for multimodal protein characterization. Collectively, this work provides a systematic view of how sequence, homology, and structural information contribute to protein characterization. It demonstrates that the most informative modality depends on the biological task, highlights the value of hybrid and multimodal learning strategies, and emphasizes the importance of rigorous evaluation protocols for obtaining reliable estimates of model generalization. Together, these contributions support the development of more accurate, interpretable, and practically useful computational approaches for a broad range of protein sequence- and structure-based classification tasks.

Description

Keywords

Protein Language Models, Protein Structure, Alignment-Based Scoring, Graph Neural Network

Citation