Biology subjects

Xia, L. C.

Publications and source records attributed to Xia, L. C..

9 recordsLinked to original sources

bioRxiv2026

sxRaep: A Rapid and Accurate Enzyme Predictor for high-throughput mining of enzymatic sequences

MotivationMetagenomic sequencing generates petabyte-scale sequence datasets that strain both deep learning and alignment based enzyme annotation tools. A lightweight rapid and accurate filter tool is needed to filter and identify enzymatic sequences prior to resource-intensive functional prediction. ResultsWe present sxLaep (Lightweight and Accurate Enzyme Predictor), a resource-efficient framework using lightweight physicochemical features for enzyme pre-screening. On the external validation set, sxLaep completed prediction in only 0.002 s/sequence, which is 22.9-fold faster than Diamond (0.0457 s/sequence). It used 372.16 MB peak memory, corresponding to a 54.4% memory reduction relative to Diamond (815.64 MB). sxLaep achieved an accuracy of 99.34% and the highest recall in remote homology detection, including enzyme candidates missed by alignment-based methods. We further successfully applied sxLaep to a marine metagenomic enzyme-mining workflow, demonstrating its utility for high-throughput discovery from large-scale metagenomic sequences. Availability and ImplementationsxLaep is available as a Python package at https://pypi.org/project/sxlaep and is maintained as an open-source software repository at https://github.com/labxscut/sxLaep. Detailed installation, usage, and Docker deployment instructions are provided in the GitHub repository to support reproducible enzyme prediction and model execution. Contactlcxia@scut.edu.cn

bioinformatics↗

bioRxiv2024

Predicting Enzyme Functions Using Contrastive Learning with Hierarchical Enzyme Structure Information

Enzyme functional annotation is a fundamental challenge in biology, and many computational tools have been developed. Accurate function prediction of enzymes relies heavily on sequence and structural information, providing critical insights into enzyme activity and specificity. However, for less studied proteins or proteins with previously uncharacterized functions or multiple activities, most of these tools cannot accurately predict functional annotations, such as enzyme commission (EC) numbers. At the same time, functional hierarchical information between enzyme species categorized based on EC numbers has not been sufficiently investigated. To address these challenges, we propose a machine learning algorithm named EnzHier, which assigns EC numbers to enzymes with better accuracy and reliability compared to state-of-the-art tools. EnzHier cleverly learns the functional hierarchy of enzymes by optimizing triplet loss, enabling it to annotate understudied enzymes confidently and identify confounding enzymes with two or more EC numbers. By incorporating both sequence and structural information, EnzHier enhances its predictive capabilities. We experimentally demonstrate its excellent performance. We anticipate that this tool will be widely used to predict the function of uncharacterized enzymes, thereby advancing many fields such as drug design and discovery and medical diagnostics.

bioinformatics↗

bioRxiv2023

Building a genetic and epigenetic predictive model of breast cancer intrinsic subtypes using large-scale data and hierarchical structure learning

Breast cancer subtyping is a difficult clinical and scientific challenge. The prevalent Prediction Analysis of Microarray of 50 genes (PAM50) system and its Immunohistochemistry (IHC) surrogate showed significant inconsistencies. This is because of the limited training samples, highly variable molecular features and in-efficient strategies used in these classifiers. The rapid development of early screening technologies, especially in the field of circulating tumor DNA, has also challenged the subtyping of breast cancer at the DNA level. By integrating large-scale DNA-level data and using a hierarchical structure learning algorithm, we developed Unified Genetic and Epigenetic Subtyping (UGES), a new intrinsic subtype classifier. The benchmarks showed that the use of all classes of DNA alterations worked much better than single classes, and that the multi-step hierarchical learning is crucial, which improves the overall AUC score by 0.074 compared to the one-step multi-classification method. Based on these insights, the ultimate UGES was trained as a three-step classifier on 50831 DNA features of 2065 samples, including mutations, copy number aberrations, and methylations. UGES achieved overall AUC score 0.963, and greatly improved the clinical stratification of patients, as each stratas survival difference became statistically more significant p-value=9.7e-55 (UGES) vs 2.2e-47 (PAM50). Finally, UGES identified 52 subtype-level DNA biomarkers that can be targeted in early screening technology to significantly expand the time window for precision care. The analysis code is freely available at https://github.com/labxscut/UGES.

bioinformatics↗

bioRxiv2023

Identifying gastric cancer molecular subtypes by integrating DNA-based hierarchical classification strategy and clinical stratification

BackgroundMolecular subtyping has been introduced to better understand the genetic landscape of gastric cancer (GC), but current subtyping methods only had limited success because of the mixed use of molecular features, a lack of strategy optimization, and the limited availability of GC samples. The community urgently calls for a precise, and easily adoptable subtyping method to enable DNA-based early screening and treatment. MethodsBased on TCGA subtypes, we developed a novel classifier, termed Hierarchical DNA-based Classifier for Gastric Cancer Molecular Subtyping (HCG), leveraging all DNA-level alterations as predictors, including gene mutations, copy number aberrations and methylations. By adding the closely related esophageal adenocarcinomas (EA) dataset, we expanded the TCGA GC dataset for training and testing HCG (n=453). We optimized HCG with three hierarchical strategies evaluated by their overall accuracy using Lasso-Logistic regression, and by their clinical stratification capacity using multivariate survival analysis. We used difference tests to identify subtype-specific DNA alteration biomarkers based on HCG defined subtypes. ResultsOur HCG classifier achieved an overall AUC score of 0.95 and significantly improved the clinical stratification of patients (overall p-value=0.032). 25 subtype-specific DNA alterations were identified by difference tests, including high level of mutations in SYNE1, ITGB4 and COL22A1 genes for the MSI subtype, high level of methylations of ALS2CL, KIAA0406 and RPRD1B genes for the EBV subtype. ConclusionsHCG is an accurate and robust classifier for DNA-based GC molecular subtyping with high-performing clinical stratification capacity. The training and test datasets and analysis programs of HCG are available at https://github.com/labxscut/HCG.

bioinformatics↗

bioRxiv2022

A comparative study of structural variant calling strategies using the Alzheimer's Disease Sequencing Project's whole genome family data

BackgroundReliable detection and accurate genotyping of structural variants (SVs) and insertion/deletions (indels) from whole-genome sequence (WGS) data is a significant challenge. We present a protocol for variant calling, quality control, call merging, sensitivity analysis, in silico genotyping, and laboratory validation protocols for generating a high-quality deletion call set from whole genome sequences as part of the Alzheimers Disease Sequencing Project (ADSP). This dataset contains 578 individuals from 111 families. MethodsWe applied two complementary pipelines (Scalpel and Parliament) for SV/indel calling, break-point refinement, genotyping, and local reassembly to produce a high-quality annotated call set. Sensitivity was measured in sample replicates (N=9) for all callers using in silico variant spike-in for a wide range of event sizes. We focused on deletions because these events were more reliably called. To evaluate caller specificity, we developed a novel metric called the D-score that leverages deletion sharing frequencies within and outside of families to rank recurring deletions. Assessment of overall quality across size bins was measured with the kinship coefficient. Individual callers were evaluated for computational cost, performance, sensitivity, and specificity. Quality of calls were evaluated by Sanger sequencing of predicted loss-of-function (LOF) variants, variants near AD candidate genes, and randomly selected genome-wide deletions ranging from 2 to 17,000 bp. ResultsWe generated a high-quality deletion call set across a wide range of event sizes consisting of 152,301 deletions with an average of 263 per genome. A total of 114 of 146 predicted deletions (78.1%) were validated by Sanger sequencing. Scalpel was more accurate in calling deletions [≤]100 bp, whereas for Parliament, sensitivity was improved for deletions > 900 bp. We validated 83.0% (88/106) and 72.5% (37/51) of calls made by Scalpel and Parliament, respectively. Eleven deletions called by both Parliament and Scalpel in the 101-900 bin were tested and all were confirmed by Sanger sequencing. ConclusionsWe developed a flexible protocol to assess the quality of deletion detection across a wide range of sizes. We also generated a truth set of Sanger sequencing validated deletions with precise breakpoints covering a wide spectrum of sizes between 1 and 17,000 bp.

genetics↗

bioRxiv2021

SeqDistK: a Novel Tool for Alignment-free Phylogenetic Analysis

Algorithms for constructing phylogenetic trees are fundamental to study the evolution of viruses, bacteria, and other microbes. Established multiple alignment-based algorithms are inefficient for large scale metagenomic sequence data because of their high requirement of inter-sequence correlation and high computational complexity. In this paper, we present SeqDistK, a novel tool for alignment-free phylogenetic analysis. SeqDistK computes the dissimilarity matrix for phylogenetic analysis, incorporating seven k-mer based dissimilarity measures, namely d2, d2S, d2star, Euclidean, Manhattan, CVTree, and Chebyshev. Based on these dissimilarities, SeqDistK constructs phylogenetic tree using the Unweighted Pair Group Method with Arithmetic Mean algorithm. Using a golden standard dataset of 16S rRNA and its associated phylogenetic tree, we compared SeqDistK to Muscle - a multi sequence aligner. We found SeqDistK was not only 38 times faster than Muscle in computational efficiency but also more accurate. SeqDistK achieved the smallest symmetric difference between the inferred and ground truth trees with a range between 13 to 18, while that of Muscle was 62. When measures d2, d2star, d2S, Euclidean, and k-mer size k=5 were used, SeqDistK consistently inferred phylogenetic tree almost identical to the ground truth tree. We also performed clustering of 16S rRNA sequences using SeqDistK and found the clustering was highly consistent with known biological taxonomy. Among all the measures, d2S (k=5, M=2) showed the best accuracy as it correctly clustered and classified all sample sequences. In summary, SeqDistK is a novel, fast and accurate alignment-free tool for large-scale phylogenetic analysis. SeqDistK software is freely available at https://github.com/htczero/SeqDistK.

bioinformatics↗

bioRxiv2019

Whole genome analysis identifies the association of TP53 genomic deletions with lower survival in Stage III colorectal cancer

DNA copy number aberrations (CNA) were frequently observed in colorectal cancers (CRC). There is an urgent call for CNA-based biomarkers in clinics, in particular for Stage III CRC, if combined with imaging or pathologic evidence, promise more precise care at the timing. We conducted this Stage III specific biomarker discovery with a cohort of 134 CRCs, and with a newly developed high-efficiency CNA profiling protocol. Specifically, we developed the profiling protocol for tumor-normal matched tissue samples based on low-coverage clinical whole-genome sequencing (WGS). We demonstrated the protocols accuracy and robustness by a systematic benchmark with microarray, high-coverage whole-exome and -genome approaches, where the low-coverage WGS-derived CNA segments were highly accordant (PCC>0.95) with those derived from microarray, and they were substantially less variable if compared to exome-derived segments. A lasso-based model and multivariate cox regression analysis identified a chromosome 17p loss, containing the TP53 tumor suppressor gene, that was significantly associated with reduced survival (P=0.0139, HR=1.688, 95% CI = [1.112-2.562]), which was validated by an independent cohort of 187 Stage III CRCs. In summary, the new low-coverage WGS protocol has high sensitivity, high resolution and low cost and the identified 17p-loss is an effective poor prognosis marker for Stage III patients.

genomics↗

bioRxiv2018

Assembly of Mb-size genome segments from linked read sequencing of CRISPR DNA targets

We developed a targeted sequencing method for intact high molecular weight (HMW) DNA targets as large as 0.2 Mb. This process uses HMW DNA isolated from intact cells, custom designed Cas9-guide RNA complexes to generate 0.1 - 0.2 Mb DNA targets, electrophoretic isolation of the DNA targets and sequencing with barcode linked reads. We used alignment methods as well as local assembly of the target regions to identify haplotypes and structural variants (SVs) across multi-Megabase genomic regions. To demonstrate the performance of this approach, we designed three assays that covered a 0.2 Mb region surrounding the BRCA1 gene, a set of 40 overlapping 0.2 Mb targets covering the entire 4-Mb MHC locus, and 18 well-characterized structural variants. Using the highly characterized NA12878 genome, we achieved on-target coverage of more than 50X, while overall whole genome coverage was approximately 4X. We generated haplotypes that completely covered each targeted locus, with a maximum size of 4 Mb (for the MHC region). This method detected structural variants such as deletions and inversions with determination of the exact breakpoints and genotypes. Even breakpoints inside highly homologous segmental duplications are precisely determined with our high-quality assemblies. Overall, this is a new method to sequence large DNA segments.

genomics↗

bioRxiv2018

SVEngine: an efficient and versatile simulator of genome structural variations with features of cancer clonal evolution

BackgroundSimulating genome sequence data with features can facilitate the development and benchmarking of structural variant analysis programs. However, there are a limited number of data simulators that provide structural variants in silico. Moreover, there are a paucity of programs that generate structural variants with different allelic fraction and haplotypes.\n\nFindingsWe developed SVEngine, an open source tool to address this need. SVEngine simulates next generation sequencing data with embedded structural variations. As input, SVEngine takes template haploid sequences (FASTA) and an external variant file, a variant distribution file and/or a clonal phylogeny tree file (NEWICK) as input. Subsequently, it simulates and outputs sequence contigs (FASTAs), sequence reads (FASTQs) and/or post-alignment files (BAMs). All of the files contain the desired variants, along with BED files containing the ground truth. SVEngines flexible design process enables one to specify size, position, and allelic fraction for deletion, insertion, duplication, inversion and translocation variants. Finally, SVEngine simulates sequence data that replicates the characteristics of a sequencing library with mixed sizes of DNA insert molecules. To improve the compute speed, SVEngine is highly parallelized to reduce the simulation time.\n\nConclusionsWe demonstrated the versatile features of SVEngine and its improved runtime comparisons with other available simulators. SVEngines features include the simulation of locus-specific variant frequency designed to mimic the phylogeny of cancer clonal evolution. We validated the accuracy of the simulations. Our evaluation included checking various sequencing mapping features such as coverage change, read clipping, insert size shift and neighbouring hanging read pairs for representative variant types. SVEngine is implemented as a standard Python package and is freely available for academic use at: https://bitbucket.org/charade/svengine.

bioinformatics↗