249 research outputs found
Digital signal processing techniques for audio applications
No full textThis thesis proposes two effective digital signal processing techniques for the audio applications. The full-band technique performs the audio equalization at the full frequency band and the subband technique at the low-frequency band. Both full-band and subband techniques have been implemented on the Motorola DSP56300 digital signal processor. Objective and subjective methods have been applied to evaluate the performance of the proposed techniques. The modeling of room transfer functions is discussed and the experiments on a pipe and two rooms are conducted for the proposed techniques. The most suitable model for the audio equalization is the all-pole modeling. In addition, two common acoustic pole techniques, least squares technique and averaging technique, have been presented for the proposed full-band and subband techniques. The least squares technique requires much higher computational power than the averaging technique
Development of structure-based machine learning approaches for protein interfaces
No full textProtein interfaces play vital roles as essential hubs in executing diverse protein functions within various biological processes, including proteolytic cleavage, enzymatic activities, viral attachment and entry, signal transduction, and transport. Machine learning (ML) has rapidly grown into one of the most popular and promising areas for predicting interface specificities efficiently and accurately. This dissertation aimed to use interface structures in machine learning for specific interface function prediction and designs. It focused on developing interpretable and generalizable models, with broad applications for various interface systems. The study also explored the importance of data quality, variety, and quantity for learning.
To understand protease specificities, we need to quickly grasp the full substrate landscape of proteolytic cleavage and comprehend the underlying biological basis for functionally distant proteases. Herein, we introduced Protein Graph Convolutional Network (PGCN), a novel approach that leverages a physically-grounded, structure-based molecular interaction graph to capture molecular topology and interaction energies for the prediction of protease specificity. We showed PGCN’s high accuracy in predicting specificity landscapes for various variants of two model proteases, namely the NS3/4 protease from the Hepatitis C virus (HCV) and the Tobacco Etch Virus (TEV) protease. Through node and edge ablation tests, we identified key elements within the graph that contribute significantly to specificity prediction, aligning with known biochemical constraints governing protease:substrate recognition. We further employed a pre-trained PGCN model to guide the design of TEV protease libraries for cleaving non-canonical substrates, achieving excellent agreement with experimental cleavage results. Most importantly, we found the model can accurately assess designs featuring diversity at positions not present in the training data.
We expanded PGCN's use with CleavEX, a user-friendly pipeline for preprocessing deep sequencing data, enabling efficient categorization of substrates for various enzymes. It successfully prepared data for six 3C-like proteases, achieving state-of-the-art accuracies with PGCN models. We also explored PGCN's potential in predicting binding interactions, specifically between the RBD of the SARS-COV-2 spike protein and ACE2, as well as PDZ domain binding.
To further explore data quality, we analyzed small molecule ligand data from the Protein Data Bank, identifying distribution and dependencies of ligand structure quality metrics. This led to composite ligand quality indicators for simplified ligand selection in drug design. Additionally, we introduced a novel all-and-subsample clustering framework to handle large-scale clustering problems, requiring human review for biological significance verification. The process can be extended to analyze various types of large datasets.
Our research in this dissertation presents promising solutions for interface prediction, design, and structure quality assessment using structure-aware machine learning approaches. These findings offer possibilities for developing customized protein editors capable of selectively and irreversibly modifying specific targets, opening new avenues for future therapeutic applications.Ph.D.Includes bibliographical reference
BIAN‐NHC Ligands in Transition‐Metal‐Catalysis: A Perfect Union of Sterically Encumbered, Electronically Tunable N‐Heterocyclic Carbenes?
No full textEmerg Infect Dis
No full textA 13-valent pneumococcal conjugate vaccine against invasive pneumococcal disease (IPD) was introduced in China in April 2017. We describe 105 children <5 years of age who were hospitalized for IPD at Soochow University Affiliated Children's Hospital in Suzhou, China, during January 2010-December 2017. We calculated the incidence of hospitalization for IPD as 14.55/100,000 children in Suzhou. We identified 8 different capsular serotypes: 6B (28.4% of cases), 14 (18.9% of cases), 19A (18.9% of cases), 19F (12.2% of cases), 23F (10.8% of cases), 20 (4.1% of cases), 9V (4.1% of cases), and 15B/C (2.7% of cases). These results provide baseline data of IPD before the introduction of this vaccine in China, enabling researchers to better understand its effects on IPD incidence
Hybrid Model for Network Traffic Anomaly Detection Based on Parallel Two-Stage Feature Fusion
No full textThe increasing frequency and complexity of network attacks underscore the need for robust detection models capable of accurately identifying malicious activities in network traffic. In this paper, we propose GRU-ResCBANet, a novel deep learning-based hybrid model for network traffic anomaly detection. The proposed model utilizes a parallel two-stage feature extraction and fusion technique. In the first stage, we enhance the convolutional neural network by introducing residual units and convolutional block attention module to improve spatial and channel feature extraction. In the second stage, gated recurrent units are used to capture temporal dependencies in the network traffic data. The features extracted from both stages are then fused to form a comprehensive representation of the data. Finally, abnormal traffic detection is performed through a fully connected layer, which classifies the network traffic. We evaluate GRU-ResCBANet’s performance on the NSL-KDD, UNSW-NB15, and CICIDS2017 datasets, achieving binary classification accuracies of 99.70%, 99.16%, and 99.83%, and multi-class classification accuracies of 99.69%, 98.23%, and 99.79%, respectively. Comparative analysis with eight other models, as well as models reported in the literature, demonstrates that GRU-ResCBANet offers superior detection performance in both binary and multi-class classification tasks
Vaccine
Full text linkBackground:Data are limited on the economic burden of seasonal influenza in China. We estimated the cost due to influenza illness among children < 5-year-old in Suzhou, China.Methods:This study adopted a societal perspective to estimate direct medical cost, direct non-medical cost, and indirect cost related to lost productivity. Data to calculate costs and rates of three influenza illness outcomes (non-medically attended, outpatient and hospitalization) were collected from prospective community-based cohort studies and hospital-based enhanced laboratory-confirmed influenza surveillance in Suzhou during the 2011/12 to 2016/17 influenza seasons. We used mean cost-per-episode, annual incidence rates of episodes of each outcome, and annual population size to estimate the total annual economic burden of influenza illnesses among children < 5-year-old for Suzhou. All costs were reported in 2017 U.S. dollars.Results:The mean cost-per-episode (standard deviation) was 161.05 (176.98) for influenza outpatient illnesses, and 7.37 (95% confidence interval, 6.9\u20137.8) million, with estimated costs for non-medically attended influenza of 3.5 (3.3\u20133.8) million, and influenza hospitalizations 1.3 million) of total economic burden, accounting for 21,994 days of lost productivity annually. Among inpatients, the indirect cost was 22.1% ($829,000), accounting for 18,431 days of lost productivity annually.Conclusions:Our findings show that influenza in children < 5-year-oldcauses substantial societal economic burden in Suzhou, China. Assessing the potential economic benefit of increasing influenza vaccination coverage in this population is warranted.CC999999/ImCDC/Intramural CDC HHSUnited States/U2G GH000018/GH/CGH CDC HHSUnited States
SPDK Vhost-NVMe: Accelerating I/Os in Virtual Machines on NVMe SSDs via User Space Vhost Target
No full text- …
