東華大學圖書館 |

Advancing Machine Learning Methods for Small and Large Biological Datasets: Applications in Redox Potential Prediction, Enzyme Interaction Prediction, and Spectra Annotation /

Record Type:	Electronic resources : Monograph/item
Title/Author:	Advancing Machine Learning Methods for Small and Large Biological Datasets: Applications in Redox Potential Prediction, Enzyme Interaction Prediction, and Spectra Annotation // Apurva Kalia.
Author:	Kalia, Apurva,
Description:	1 electronic resource (107 pages)
Notes:	Source: Dissertations Abstracts International, Volume: 86-08, Section: B.
Contained By:	Dissertations Abstracts International86-08B.
Subject:	Computer science. -
Online resource:	https://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31767567
ISBN:	9798304914369

Advancing Machine Learning Methods for Small and Large Biological Datasets: Applications in Redox Potential Prediction, Enzyme Interaction Prediction, and Spectra Annotation /
Kalia, Apurva,

Advancing Machine Learning Methods for Small and Large Biological Datasets: Applications in Redox Potential Prediction, Enzyme Interaction Prediction, and Spectra Annotation /Apurva Kalia. - 1 electronic resource (107 pages)

Source: Dissertations Abstracts International, Volume: 86-08, Section: B.

Advancements in computational methods, especially with deep learning and AI, have driven breakthroughs in areas such as computer vision and natural language processing. This progress has sparked efforts to apply these powerful techniques to complex biological problems, such as drug design, protein engineering, and metabolic engineering. However, biological data presents unique challenges since it is frequently multi-modal, limited in size due to costly experimentation, and often noisy. These factors add complexity to feature selection and learning. Consequently, it is essential to develop computational methods tailored to the characteristics and quality of data available for each specific biological problem. This thesis addresses three biological problems, where one problem has a data-scarce scenario requiring feature selection and evaluation, while the other two problems have data-abundant scenarios enabling the use of deep learning for representation learning.First, we address the challenge of predicting redox (reduction/oxidation) potential, which is a molecular property that is crucial for understanding electron transfer processes in chemical and biological systems. As there are currently no established "best" method for creating redox models for small datasets, we explore the use of existing molecular features, including structural properties, molecular energies, and drug-like attributes, to develop Gaussian Process Regression (GPR) models for predicting redox potentials across three datasets of organic molecules. We demonstrate that combining these features yields higher predictive accuracy than when using each feature independently, and validate the model's performance on an experimental quinone dataset. Our results highlight the importance of integrating structural, electronic, and pharmacokinetic properties to capture critical redox characteristics, offering a recommended method for accurate redox potential prediction when dealing with smaller datasets. Next, we explore the problem of molecule-enzyme interaction prediction, a critical factor in understanding enzyme promiscuity and its implications for biocatalysis and evolutionary biology. We develop Contrastive Stratification for Interaction Prediction, CSI, a novel model which partitions data by interaction features and applies contrastive learning to generate embeddings to maximize mutual information across congruent views. By incorporating multiple views on enzymatic reactions, CSI learns high-quality representations of enzymes and molecules, resulting in improved interaction predictions over methods that do not utilize such views. Finally, we address the problem of annotating spectra measured via untargeted mass spectrometry. Assigning chemical identities to the measured spectra is essential for identifying biomarkers and diagnosing diseases. Despite recent computational efforts, annotation rates remain low. We present Joint Embedding Space Technique for Ranking, JESTR, a deep-learning approach to embed both molecular structures and their spectra into a joint embedding space. Through contrastive learning, JESTR aligns molecular and spectral representations, ranking candidate molecules by cosine similarity to the query spectrum. JESTR improves annotation accuracy over state-of-the-art existing methods by avoiding the explicit reconstruction of molecules from spectra, or the construction of spectra from candidate molecules. Collectively, these projects illustrate the importance of exploiting data to derive high-quality features, achieved through either manual engineering or deep learning, in addressing critical biological challenges.

English

ISBN: 9798304914369Subjects--Topical Terms:

523869
Computer science.
Subjects--Index Terms:

Computational biology

Advancing Machine Learning Methods for Small and Large Biological Datasets: Applications in Redox Potential Prediction, Enzyme Interaction Prediction, and Spectra Annotation /
LDR:05239nmm a22004573i 4500 001 2436538
005 20251023121238.5
006 m o d
007 cr|nu||||||||
008 261011s2025 miu||||||m |||||||eng d
020 $a 9798304914369
035 $a (MiAaPQD)AAI31767567
035 $a AAI31767567
040 $a MiAaPQD $b eng $c MiAaPQD $e rda
100 1 $a Kalia, Apurva, $e author. $3 3827233
245 1 0 $a Advancing Machine Learning Methods for Small and Large Biological Datasets: Applications in Redox Potential Prediction, Enzyme Interaction Prediction, and Spectra Annotation / $c Apurva Kalia.
264 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2025
300 $a 1 electronic resource (107 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertations Abstracts International, Volume: 86-08, Section: B.
500 $a Advisors: Hassoun, Soha Committee members: Liu, Liping; Chang, Remco; Sonkusale, Sameer; Krishnan, Dilip.
502 $b Ph.D. $c Tufts University $d 2025.
520 $a Advancements in computational methods, especially with deep learning and AI, have driven breakthroughs in areas such as computer vision and natural language processing. This progress has sparked efforts to apply these powerful techniques to complex biological problems, such as drug design, protein engineering, and metabolic engineering. However, biological data presents unique challenges since it is frequently multi-modal, limited in size due to costly experimentation, and often noisy. These factors add complexity to feature selection and learning. Consequently, it is essential to develop computational methods tailored to the characteristics and quality of data available for each specific biological problem. This thesis addresses three biological problems, where one problem has a data-scarce scenario requiring feature selection and evaluation, while the other two problems have data-abundant scenarios enabling the use of deep learning for representation learning.First, we address the challenge of predicting redox (reduction/oxidation) potential, which is a molecular property that is crucial for understanding electron transfer processes in chemical and biological systems. As there are currently no established "best" method for creating redox models for small datasets, we explore the use of existing molecular features, including structural properties, molecular energies, and drug-like attributes, to develop Gaussian Process Regression (GPR) models for predicting redox potentials across three datasets of organic molecules. We demonstrate that combining these features yields higher predictive accuracy than when using each feature independently, and validate the model's performance on an experimental quinone dataset. Our results highlight the importance of integrating structural, electronic, and pharmacokinetic properties to capture critical redox characteristics, offering a recommended method for accurate redox potential prediction when dealing with smaller datasets. Next, we explore the problem of molecule-enzyme interaction prediction, a critical factor in understanding enzyme promiscuity and its implications for biocatalysis and evolutionary biology. We develop Contrastive Stratification for Interaction Prediction, CSI, a novel model which partitions data by interaction features and applies contrastive learning to generate embeddings to maximize mutual information across congruent views. By incorporating multiple views on enzymatic reactions, CSI learns high-quality representations of enzymes and molecules, resulting in improved interaction predictions over methods that do not utilize such views. Finally, we address the problem of annotating spectra measured via untargeted mass spectrometry. Assigning chemical identities to the measured spectra is essential for identifying biomarkers and diagnosing diseases. Despite recent computational efforts, annotation rates remain low. We present Joint Embedding Space Technique for Ranking, JESTR, a deep-learning approach to embed both molecular structures and their spectra into a joint embedding space. Through contrastive learning, JESTR aligns molecular and spectral representations, ranking candidate molecules by cosine similarity to the query spectrum. JESTR improves annotation accuracy over state-of-the-art existing methods by avoiding the explicit reconstruction of molecules from spectra, or the construction of spectra from candidate molecules. Collectively, these projects illustrate the importance of exploiting data to derive high-quality features, achieved through either manual engineering or deep learning, in addressing critical biological challenges.
546 $a English
590 $a School code: 0234
650 4 $a Computer science. $3 523869
650 4 $a Bioinformatics. $3 553671
650 4 $a Pharmacology. $3 634543
650 4 $a Biochemistry. $3 518028
653 $a Computational biology
653 $a Machine learning
653 $a Metabolomics
653 $a Proteins
653 $a Small molecules
690 $a 0984
690 $a 0800
690 $a 0715
690 $a 0487
690 $a 0419
710 2 $a Tufts University. $b Computer Science. $e degree granting institution. $3 3816160
720 1 $a Hassoun, Soha $e degree supervisor.
773 0 $t Dissertations Abstracts International $g 86-08B.
790 $a 0234
791 $a Ph.D.
792 $a 2025
856 4 0 $u https://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31767567