Deep Learning in Biomedical Research and Statistical Inference on Time Warping Functions
Saved in:
| Title: | Deep Learning in Biomedical Research and Statistical Inference on Time Warping Functions |
|---|---|
| Authors: | N/A |
| Committee Members: | Lin, Menghan (author); Wu, Wei (professor directing dissertation); Yang, Wei (university representative); Barbu, Adrian G., 1971- (committee member); Zhang, Jinfeng (committee member); Florida State University (degree granting institution); College of Arts and Sciences (degree granting college); Department of Statistics (degree granting department) |
| Summary: | In this dissertation, I present two research projects that I have worked on during my doctoral study at Florida State University. I provide a brief summary of each of these projects below. Detailed studies are presented in the chapters of this dissertation. Deep Learning Application Part 1: Protein-Ligand DockingDeep learning has transformed protein-ligand docking by improving the accuracy and efficiency of interaction predictions, surpassing traditional molecular docking and free-energy simulation methods that often suffer from computational limitations and predefined heuristics. This study explores deep learning-based approaches for protein-ligand docking, emphasizing the role of reinforcement learning, generative models, and geometric deep learning in optimizing ligand placement within protein binding pockets. Our previous study formulated docking as a sequential decision-making problem and applied a reinforcement learning (RL) method—the Asynchronous Advantage Actor-Critic (A3C) model—to optimize docking trajectories. To further enhance protein-ligand docking, we introduce DisDock, a deep learning-based method specifically designed for metal ion redocking. DisDock employs a U-net architecture with axial-attention modules to process pairwise distance matrices, capturing spatial relationships between metal ions and protein atoms. Two postprocessing techniques—multidimensional scaling alignment and direct estimation via distance constraints—are implemented to refine metal ion localization. Evaluations on large-scale datasets reveal that DisDock significantly outperforms existing methods, including BioMetAll, AutoDock Vina, and RL-based docking models. To address docking challenges where binding pockets are unknown, we propose to adopt and fine-tune ESM-GearNet, a structure-aware docking residue prediction model that integrates sequence-based protein representations with geometric deep learning. Pretrained on large-scale protein databases, GearNet effectively captures residue-level spatial dependencies, facilitating the identification of potential ligand-binding sites. Part 2: Biomedical Relation ExtractionBiomedical relation extraction (RE) is essential for identifying relationships between biomedical entities such as genes, diseases, and chemicals in scientific literature. In this study, we enhance RE by leveraging pretrained language models, including PubMedBERT, RoBERTa, and BioLinkBERT, evaluating their performance to optimize the extraction pipeline. We investigate various pretraining strategies, such as warmup scheduling and multi-corpus training, to improve model effectiveness. Ablation studies confirm that data augmentation and warmup training enhance both robustness and accuracy. Statistical Inference on Time-Warping Functions Time-warping functions capture phase variability in functional data. However, statistical modeling and inference for time-warping functions remain challenging due to their non-linearity in conventional L² space. In this work, we propose a statistical inference framework for time-warping functions by employing a linear, generative, and stochastic model under the centered log-ratio (CLR) transformation. By establishing an isometric isomorphism between the time-warping function space and a subspace of L², we enable the direct application of statistical models. To demonstrate the utility of our framework, we introduce hypothesis tests for the mean and covariance structures of time-warping functions. The test for equality of means assesses whether two groups of time-warping functions have the same mean functions—i.e., whether they exhibit similar phase variability on average. Meanwhile, the test for equality of covariance structures evaluates differences in the covariance structure of phase variability across groups. Specifically, we implement an L²-norm-based test statistic, whose distribution is approximated using the eigenvalues of the covariance operator. Simulation studies confirm the effectiveness of these tests, demonstrating their power in detecting subtle variations in phase structure. |
| Database: | OpenDissertations |
| Abstract: | In this dissertation, I present two research projects that I have worked on during my doctoral study at Florida State University. I provide a brief summary of each of these projects below. Detailed studies are presented in the chapters of this dissertation. Deep Learning Application Part 1: Protein-Ligand DockingDeep learning has transformed protein-ligand docking by improving the accuracy and efficiency of interaction predictions, surpassing traditional molecular docking and free-energy simulation methods that often suffer from computational limitations and predefined heuristics. This study explores deep learning-based approaches for protein-ligand docking, emphasizing the role of reinforcement learning, generative models, and geometric deep learning in optimizing ligand placement within protein binding pockets. Our previous study formulated docking as a sequential decision-making problem and applied a reinforcement learning (RL) method—the Asynchronous Advantage Actor-Critic (A3C) model—to optimize docking trajectories. To further enhance protein-ligand docking, we introduce DisDock, a deep learning-based method specifically designed for metal ion redocking. DisDock employs a U-net architecture with axial-attention modules to process pairwise distance matrices, capturing spatial relationships between metal ions and protein atoms. Two postprocessing techniques—multidimensional scaling alignment and direct estimation via distance constraints—are implemented to refine metal ion localization. Evaluations on large-scale datasets reveal that DisDock significantly outperforms existing methods, including BioMetAll, AutoDock Vina, and RL-based docking models. To address docking challenges where binding pockets are unknown, we propose to adopt and fine-tune ESM-GearNet, a structure-aware docking residue prediction model that integrates sequence-based protein representations with geometric deep learning. Pretrained on large-scale protein databases, GearNet effectively captures residue-level spatial dependencies, facilitating the identification of potential ligand-binding sites. Part 2: Biomedical Relation ExtractionBiomedical relation extraction (RE) is essential for identifying relationships between biomedical entities such as genes, diseases, and chemicals in scientific literature. In this study, we enhance RE by leveraging pretrained language models, including PubMedBERT, RoBERTa, and BioLinkBERT, evaluating their performance to optimize the extraction pipeline. We investigate various pretraining strategies, such as warmup scheduling and multi-corpus training, to improve model effectiveness. Ablation studies confirm that data augmentation and warmup training enhance both robustness and accuracy. Statistical Inference on Time-Warping Functions Time-warping functions capture phase variability in functional data. However, statistical modeling and inference for time-warping functions remain challenging due to their non-linearity in conventional L² space. In this work, we propose a statistical inference framework for time-warping functions by employing a linear, generative, and stochastic model under the centered log-ratio (CLR) transformation. By establishing an isometric isomorphism between the time-warping function space and a subspace of L², we enable the direct application of statistical models. To demonstrate the utility of our framework, we introduce hypothesis tests for the mean and covariance structures of time-warping functions. The test for equality of means assesses whether two groups of time-warping functions have the same mean functions—i.e., whether they exhibit similar phase variability on average. Meanwhile, the test for equality of covariance structures evaluates differences in the covariance structure of phase variability across groups. Specifically, we implement an L²-norm-based test statistic, whose distribution is approximated using the eigenvalues of the covariance operator. Simulation studies confirm the effectiveness of these tests, demonstrating their power in detecting subtle variations in phase structure. |
|---|
