{"id":{"repo_id":"msu","oai_identifier":"oai:d.lib.msu.edu:etd_52891"},"canonical_url":"https://search.ndltd.org/etd/msu/oai:d.lib.msu.edu:etd_52891","repository":{"repo_id":"msu","name":"Michigan State University","base_url":"https://d.lib.msu.edu/oai"},"display":{"title":"Deep-Learning-Assisted Spectral Graph Algorithms and Persistent Sheaf Laplacians for Molecular Property Prediction","abstract":"Thesis (Ph.D.)--Michigan State University. Applied Mathematics - Doctor of Philosophy, 2025","abstract_html":"Thesis (Ph.D.)--Michigan State University. Applied Mathematics - Doctor of Philosophy, 2025","abstract_has_math":false,"creators":["Hayes, Nicole"],"institution":null,"degree_name":null,"degree_level":null,"degree_discipline":null,"degree_department":null,"school":null,"contributors":["Rapinchuk, Ekaterina","Wei, Guo-Wei","Huang, Longxiu","Tang, Moxun"],"advisors":[],"committee_chairs":[],"committee_members":[],"year":2025,"date_issued":"2025","date_published":"2025","updated_at":"2026-11-18T18:59:31Z","subjects":["Applied mathematics","Mathematics"],"languages":["eng"],"rights":["In Copyright"],"rights_urls":[],"identifier_entries":[{"key":"dc:identifier","label":"Identifier","values":["etd:52891","local:Hayes_grad.msu_0128D_21883"],"render_values":[{"text":"etd:52891","href":null,"code":true},{"text":"local:Hayes_grad.msu_0128D_21883","href":null,"code":true}]}],"person_identifier_entries":[]},"links":{"outbound_url":"https://doi.org/doi:10.25335/yxsr-6910","outbound_label":"DOI","outbound_source":"dc:identifier","record_url":"https://doi.org/doi:10.25335/yxsr-6910","record_label":"Repository record","record_kind":"doi","record_source":"dc:identifier","download_url":null,"download_label":null,"download_source":null,"access_status":null,"embargo_until":null,"access_note":null},"source_record":{"url":"https://d.lib.msu.edu/oai?verb=GetRecord&metadataPrefix=oai_dc&identifier=oai%3Ad.lib.msu.edu%3Aetd_52891","prefix":"oai_dc"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor","label":"Contributor","values":["Rapinchuk, Ekaterina","Wei, Guo-Wei","Huang, Longxiu","Tang, Moxun"]},{"key":"dc:creator","label":"Author","values":["Hayes, Nicole"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date","label":"Dc Date","values":["2025"]},{"key":"dc:relation","label":"Dc Relation","values":["Electronic Theses & Dissertations"]},{"key":"dc:type","label":"Dc Type","values":["text","Electronic dissertations"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["Applied mathematics","Mathematics"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:language","label":"Dc Language","values":["eng"]},{"key":"dc:rights","label":"Dc Rights","values":["In Copyright"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier","label":"Identifier","values":["etd:52891","local:Hayes_grad.msu_0128D_21883","https://doi.org/doi:10.25335/yxsr-6910"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description","label":"Description","values":["Thesis (Ph.D.)--Michigan State University. Applied Mathematics - Doctor of Philosophy, 2025","Knowledge of molecular properties is vital in applications such as drug discovery, design, and engineering. Drug discovery is a long and expensive process, and determining relevant molecular properties experimentally can be costly and may be subject to ethical constraints. Consequently, computational methods for predicting molecular properties, which may streamline critical stages of the drug discovery pipeline, have become popular. In order for such methods to be useful, they must be able to accurately predict desirable molecular properties, including toxicity, permeability, binding affinity, and flexibility. To that end, this dissertation introduces new computational models for predicting a variety of molecular properties in diverse settings.Due to the experimental challenges in molecular property determination, it is often difficult to obtain large amounts of labeled molecular data. However, many computational methods, in particular machine learning and deep learning models, require a significant amount of labeled data in order to accurately predict properties for unlabeled data. Existing algorithms such as the graph-based Merriman-Bence-Osher (MBO) technique have demonstrated success in data classification tasks with very limited labeled data. As such, we have developed methods integrating the MBO scheme with deep learning models that can encode molecular data in a latent space representation, leveraging the strengths of both techniques and enabling molecular classification. Our integrative deep-learning-MBO models perform very well compared to state-of-the-art methods even when as little as 1% of the data set is labeled.Another common challenge in molecular sciences is that of class-imbalanced data, where one class size is underrepresented with respect to the other class(es). Often, the minority class is the class of most interest, and the cost of misclassifying minority data points may be very high. Many models trained on imbalanced data demonstrate a bias toward the majority class(es), failing to identify data elements in the minority class. To address this challenge, we have adapted one of our highest-performing deep-learning-MBO models to better classify imbalanced molecular data. Our new model integrates a decision threshold adjustment step in the MBO classification algorithm and also implements distance correlation as a weight function for the MBO scheme's similarity graph-based framework. These advances result in better predictive performance on highly imbalanced data compared to existing approaches.The flexibility of proteins and protein-nucleic acid complexes is a crucial property that heavily impacts their various biological functions. For instance, protein flexibility plays a role in enzyme activity, cellular communication, and molecular transport. Nucleic acid flexibility is significant for processes like DNA packing and interactions between nucleic acids and proteins. Thus, understanding and accurately predicting this flexibility, often measured by atomic B-factors, is essential in contexts like drug design or protein engineering. In this dissertation, we introduce a persistent sheaf Laplacian (PSL) model for protein and protein-nucleic acid flexibility analysis. The PSL framework, a powerful tool in topological data analysis, captures the local topology and geometry of atoms in the harmonic spectra and non-harmonic spectra of PSLs, respectively. The construction of atom-specific features using these spectra enables accurate and robust B-factor prediction through downstream machine learning tasks. Additionally, the multiscale nature of the PSL framework makes it especially suitable for flexibility analysis of structures with interactions across a wide range of length scales, including protein-nucleic acid complexes. In particular, the PSL model achieves better predictive accuracy on protein and protein-nucleic acid data sets than traditional flexibility analysis methods, suggesting its potential utility for other complex biomolecules.","Description based on online resource. Title from PDF t.p. (Michigan State University Fedora Repository, viewed ).","Includes bibliographical references."]},{"key":"dc:format","label":"Dc Format","values":["application/pdf","1 online resource (135 p.) : ill."]},{"key":"dc:title","label":"Title","values":["Deep-Learning-Assisted Spectral Graph Algorithms and Persistent Sheaf Laplacians for Molecular Property Prediction"]}]}],"canonical_facts":{"dc:contributor":["Rapinchuk, Ekaterina","Wei, Guo-Wei","Huang, Longxiu","Tang, Moxun"],"dc:creator":["Hayes, Nicole"],"dc:date":["2025"],"dc:description":["Thesis (Ph.D.)--Michigan State University. Applied Mathematics - Doctor of Philosophy, 2025","Knowledge of molecular properties is vital in applications such as drug discovery, design, and engineering. Drug discovery is a long and expensive process, and determining relevant molecular properties experimentally can be costly and may be subject to ethical constraints. Consequently, computational methods for predicting molecular properties, which may streamline critical stages of the drug discovery pipeline, have become popular. In order for such methods to be useful, they must be able to accurately predict desirable molecular properties, including toxicity, permeability, binding affinity, and flexibility. To that end, this dissertation introduces new computational models for predicting a variety of molecular properties in diverse settings.Due to the experimental challenges in molecular property determination, it is often difficult to obtain large amounts of labeled molecular data. However, many computational methods, in particular machine learning and deep learning models, require a significant amount of labeled data in order to accurately predict properties for unlabeled data. Existing algorithms such as the graph-based Merriman-Bence-Osher (MBO) technique have demonstrated success in data classification tasks with very limited labeled data. As such, we have developed methods integrating the MBO scheme with deep learning models that can encode molecular data in a latent space representation, leveraging the strengths of both techniques and enabling molecular classification. Our integrative deep-learning-MBO models perform very well compared to state-of-the-art methods even when as little as 1% of the data set is labeled.Another common challenge in molecular sciences is that of class-imbalanced data, where one class size is underrepresented with respect to the other class(es). Often, the minority class is the class of most interest, and the cost of misclassifying minority data points may be very high. Many models trained on imbalanced data demonstrate a bias toward the majority class(es), failing to identify data elements in the minority class. To address this challenge, we have adapted one of our highest-performing deep-learning-MBO models to better classify imbalanced molecular data. Our new model integrates a decision threshold adjustment step in the MBO classification algorithm and also implements distance correlation as a weight function for the MBO scheme's similarity graph-based framework. These advances result in better predictive performance on highly imbalanced data compared to existing approaches.The flexibility of proteins and protein-nucleic acid complexes is a crucial property that heavily impacts their various biological functions. For instance, protein flexibility plays a role in enzyme activity, cellular communication, and molecular transport. Nucleic acid flexibility is significant for processes like DNA packing and interactions between nucleic acids and proteins. Thus, understanding and accurately predicting this flexibility, often measured by atomic B-factors, is essential in contexts like drug design or protein engineering. In this dissertation, we introduce a persistent sheaf Laplacian (PSL) model for protein and protein-nucleic acid flexibility analysis. The PSL framework, a powerful tool in topological data analysis, captures the local topology and geometry of atoms in the harmonic spectra and non-harmonic spectra of PSLs, respectively. The construction of atom-specific features using these spectra enables accurate and robust B-factor prediction through downstream machine learning tasks. Additionally, the multiscale nature of the PSL framework makes it especially suitable for flexibility analysis of structures with interactions across a wide range of length scales, including protein-nucleic acid complexes. In particular, the PSL model achieves better predictive accuracy on protein and protein-nucleic acid data sets than traditional flexibility analysis methods, suggesting its potential utility for other complex biomolecules.","Description based on online resource. Title from PDF t.p. (Michigan State University Fedora Repository, viewed ).","Includes bibliographical references."],"dc:format":["application/pdf","1 online resource (135 p.) : ill."],"dc:identifier":["etd:52891","local:Hayes_grad.msu_0128D_21883","https://doi.org/doi:10.25335/yxsr-6910"],"dc:language":["eng"],"dc:relation":["Electronic Theses & Dissertations"],"dc:rights":["In Copyright"],"dc:subject":["Applied mathematics","Mathematics"],"dc:title":["Deep-Learning-Assisted Spectral Graph Algorithms and Persistent Sheaf Laplacians for Molecular Property Prediction"],"dc:type":["text","Electronic dissertations"]},"updated_at":"2026-11-18T18:59:31Z"}