東華大學圖書館 |

Deep Learning and Physics-Based Methods for Macromolecular Structure Prediction and Design.

紀錄類型:	書目-電子資源 : Monograph/item
正題名/作者:	Deep Learning and Physics-Based Methods for Macromolecular Structure Prediction and Design./
作者:	Pearce, Robin.
面頁冊數:	1 online resource (227 pages)
附註:	Source: Dissertations Abstracts International, Volume: 84-12, Section: B.
Contained By:	Dissertations Abstracts International84-12B.
標題:	Bioinformatics. -
電子資源:	http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=30548564click for full text (PQDT)
ISBN:	9798379566654

Deep Learning and Physics-Based Methods for Macromolecular Structure Prediction and Design.
Pearce, Robin.

Deep Learning and Physics-Based Methods for Macromolecular Structure Prediction and Design. - 1 online resource (227 pages)

Source: Dissertations Abstracts International, Volume: 84-12, Section: B.

Thesis (Ph.D.)--University of Michigan, 2023.

Includes bibliographical references

Proteins and non-coding RNA are the macromolecules responsible for performing the vast majority of biological functions in living organisms. These functions are mediated by the diverse structures adopted by different macromolecules, which in turn are determined by their primary sequences. Understanding the principles that govern this sequence-structure-function paradigm has become a hallmark of structural biology. The work presented in this thesis focuses on elucidating these principles by developing state-of-the-art deep learning and physical models for computational protein/RNA structure prediction and protein design. Despite the immense progress witnessed in protein structure prediction through the use of deep neural networks to predict spatial restraints, the modeling accuracy for proteins that lacked sequence and/or structure homologs remained to be improved. Thus, we developed an open-source program, DeepFold, which integrates spatial restraints predicted by multi-task deep residual neural networks along with a physics-based potential to guide rapid gradient-descent folding simulations. The results on large-scale benchmark tests showed that DeepFold created full-length models with accuracies significantly beyond classical folding approaches and other leading, contemporaneous deep learning methods. Of particular interest was the modeling performance on targets with very few homologous sequences, where DeepFold achieved an average TM-score that was ~40-45% higher than deep learning methods such as trRosetta and DMPfold, while being 262 times faster than traditional folding simulations. Inspired by the revolutionary advances in self-attention-based structure prediction, we developed DeepFoldRNA, which is an extension of the DeepFold pipeline that predicts RNA structures from sequence by coupling deep self-attention neural networks with gradient-based folding simulations. The method was tested on two independent benchmark datasets, including the RNA-Puzzles experiment, where DeepFoldRNA constructed models with an average RMSD of 2.72 A, which was significantly better than the best models submitted by the community (RMSD=6.90 A). Overall, these findings illustrate the major advantage of advanced deep learning techniques at capturing detailed structural information over human-engineered potentials. The second area of research that will be covered in the proceeding chapters is protein design, which is often regarded as the conceptual inverse of protein structure prediction. Protein design generally consists of two sub-problems, namely sequence design and structure design. For the first sub-problem, we developed an online server system, EvoDesign, which uses evolutionary profiles alongside a physical potential to guide the sequence search simulations. EvoDesign demonstrated advantages over pure physics-based approaches in terms of more accurately designing proteins that adopt desired target folds. Furthermore, as one of the essential difficulties in computer-based protein design is the expensive cost of experimental validation, the server aims to provide various transparent intermediate data to allow for a detailed annotation and analysis of the confidence of the designed sequences. Lastly, for the second design sub-problem, we developed FoldDesign to create novel protein folds from specific secondary structure (SS) assignments through sequence-independent replica-exchange Monte Carlo simulations. The method was tested on a large-scale dataset of non-idealized, SS topologies, where FoldDesign outperformed other state-of-the-art methods and consistently created stable structural folds with local characteristics that closely matched native structures. Notably, while sharing similar local characteristics, a large portion of the designed scaffolds possessed novel global folds that were completely different from natural proteins in the PDB. This highlights FoldDesign's ability to explore areas of protein fold space through computational simulations that have not been explored by nature.

Electronic reproduction.
Ann Arbor, Mich. :
ProQuest,
2023

Mode of access: World Wide Web

ISBN: 9798379566654Subjects--Topical Terms:

553671
Bioinformatics.
Subjects--Index Terms:

Protein structure predictionIndex Terms--Genre/Form:

542853
Electronic books.

Deep Learning and Physics-Based Methods for Macromolecular Structure Prediction and Design.
LDR:05470nmm a2200397K 4500 001 2363547
005 20231127093438.5
006 m o d
007 cr mn ---uuuuu
008 241011s2023 xx obm 000 0 eng d
020 $a 9798379566654
035 $a (MiAaPQ)AAI30548564
035 $a (MiAaPQ)umichrackham004885
035 $a AAI30548564
040 $a MiAaPQ $b eng $c MiAaPQ $d NTU
100 1 $a Pearce, Robin. $3 3704315
245 1 0 $a Deep Learning and Physics-Based Methods for Macromolecular Structure Prediction and Design.
264 0 $c 2023
300 $a 1 online resource (227 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: 84-12, Section: B.
500 $a Advisor: Omenn, Gilbert S.
502 $a Thesis (Ph.D.)--University of Michigan, 2023.
504 $a Includes bibliographical references
520 $a Proteins and non-coding RNA are the macromolecules responsible for performing the vast majority of biological functions in living organisms. These functions are mediated by the diverse structures adopted by different macromolecules, which in turn are determined by their primary sequences. Understanding the principles that govern this sequence-structure-function paradigm has become a hallmark of structural biology. The work presented in this thesis focuses on elucidating these principles by developing state-of-the-art deep learning and physical models for computational protein/RNA structure prediction and protein design. Despite the immense progress witnessed in protein structure prediction through the use of deep neural networks to predict spatial restraints, the modeling accuracy for proteins that lacked sequence and/or structure homologs remained to be improved. Thus, we developed an open-source program, DeepFold, which integrates spatial restraints predicted by multi-task deep residual neural networks along with a physics-based potential to guide rapid gradient-descent folding simulations. The results on large-scale benchmark tests showed that DeepFold created full-length models with accuracies significantly beyond classical folding approaches and other leading, contemporaneous deep learning methods. Of particular interest was the modeling performance on targets with very few homologous sequences, where DeepFold achieved an average TM-score that was ~40-45% higher than deep learning methods such as trRosetta and DMPfold, while being 262 times faster than traditional folding simulations. Inspired by the revolutionary advances in self-attention-based structure prediction, we developed DeepFoldRNA, which is an extension of the DeepFold pipeline that predicts RNA structures from sequence by coupling deep self-attention neural networks with gradient-based folding simulations. The method was tested on two independent benchmark datasets, including the RNA-Puzzles experiment, where DeepFoldRNA constructed models with an average RMSD of 2.72 A, which was significantly better than the best models submitted by the community (RMSD=6.90 A). Overall, these findings illustrate the major advantage of advanced deep learning techniques at capturing detailed structural information over human-engineered potentials. The second area of research that will be covered in the proceeding chapters is protein design, which is often regarded as the conceptual inverse of protein structure prediction. Protein design generally consists of two sub-problems, namely sequence design and structure design. For the first sub-problem, we developed an online server system, EvoDesign, which uses evolutionary profiles alongside a physical potential to guide the sequence search simulations. EvoDesign demonstrated advantages over pure physics-based approaches in terms of more accurately designing proteins that adopt desired target folds. Furthermore, as one of the essential difficulties in computer-based protein design is the expensive cost of experimental validation, the server aims to provide various transparent intermediate data to allow for a detailed annotation and analysis of the confidence of the designed sequences. Lastly, for the second design sub-problem, we developed FoldDesign to create novel protein folds from specific secondary structure (SS) assignments through sequence-independent replica-exchange Monte Carlo simulations. The method was tested on a large-scale dataset of non-idealized, SS topologies, where FoldDesign outperformed other state-of-the-art methods and consistently created stable structural folds with local characteristics that closely matched native structures. Notably, while sharing similar local characteristics, a large portion of the designed scaffolds possessed novel global folds that were completely different from natural proteins in the PDB. This highlights FoldDesign's ability to explore areas of protein fold space through computational simulations that have not been explored by nature.
533 $a Electronic reproduction. $b Ann Arbor, Mich. : $c ProQuest, $d 2023
538 $a Mode of access: World Wide Web
650 4 $a Bioinformatics. $3 553671
650 4 $a Biostatistics. $3 1002712
650 4 $a Molecular biology. $3 517296
653 $a Protein structure prediction
653 $a RNA structure prediction
653 $a Protein design
653 $a Deep learning
655 7 $a Electronic books. $2 lcsh $3 542853
690 $a 0715
690 $a 0307
690 $a 0308
710 2 $a ProQuest Information and Learning Co. $3 783688
710 2 $a University of Michigan. $b Bioinformatics. $3 3560866
773 0 $t Dissertations Abstracts International $g 84-12B.
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=30548564 $z click for full text (PQDT)