LSTrAP-denovo: Automated Generation of Transcriptome Atlases for Eukaryotic Species Without Genomes

Abstract

Structured AbstractMotivationDespite the abundance of species with transcriptomic data, a significant number of the species still lack genomes, making it difficult to study gene function and expression in these organisms. Whilede novotranscriptome assembly can be used to assemble protein-coding transcripts from RNA-sequencing (RNA-seq) data, the datasets used often only feature samples of arbitrarily-selected or similar experimental conditions which might fail to capture condition-specific transcripts.ResultsWe developed the Large-Scale Transcriptome Assembly Pipeline forde novoassembled transcripts (LSTrAP-denovo) to automatically generate transcriptome atlases of eukaryotic species. Specifically, given an NCBI TaxID, LSTrAP-denovocan (1) filter undesirable RNA-seq accessions based on read data, (2) select RNA-seq accessions via unsupervised machine learning to construct a sample-balanced dataset for download, (3) assemble transcripts via over-assembly, (4) functionally annotate coding sequences (CDS) from assembled transcripts and (5) generate transcriptome atlases in the form of expression matrices for downstream transcriptomic analyses.Availability and ImplementationLSTrAP-denovois easy to implement, written in python, and is freely available athttps://github.com/pengkenlim/LSTrAP-denovo/.Supplementary InformationSupplementary data are available in the forms of supplementary figures, supplementary tables, and supplementary methods.

Constructing Ensemble Gene Functional Networks Capturing Tissue/condition-specific Co-expression from Unlabled Transcriptomic Data with TEA-GCN

“…Transcripts per million (TPM) values of primary transcripts of RNA-seq samples estimated by Kallisto were used as gene expression abundances to construct gene expression matrices for the different species. RNA-seq samples with poor read alignment were identified based its %_pseudoaligned statistic (reported by Kallisto) below a certain threshold (5% for H. sapiens; 20% for S. cerevisiae and A. thaliana) as described [90][91][92] (Table S10 to 12). After removal RNA-seq samples with poor read alignment, gene expression matrices comprising of gene expression abundances from 150,096 (79.8%), 57,759 (87.5%), and 95,553 (66.9%) retained RNA-seq samples were designated as the public transcriptomic datasets for S. cerevisiae, A. thaliana and H. sapiens, respectively.…”

“…TPM values of each RNA-seq sample in the gene expression matrices were standardized to unit variance (within-sample standardization) to control for batch effects at the sample level as described 33,90 . Next, the standardized gene expression values were subjected to dimension reduction using Principal Component Analysis (PCA; decomposition.IncrementalPCA function from scikit-learn version 1.3.0) where only the first 1,000 Principal components (PCs) were used as high-level features to embed RNA-seq samples in PC space (PC embeddings) that describes their relative transcriptome similarity between each other.…”