Long non-coding RNAs (lncRNAs) play diverse roles across cellular compartments, yet predicting subcellular localization from the RNA sequence remains challenging. We present a new framework that casts each lncRNA sequence as a graph and trains graph neural networks (GNNs) to perform binary localization for a given lncRNA and cell line. We consider five approaches for graph construction: (1) normalized Pointwise Mutual Information (nPMI) with gapped co-occurrence and reversecomplement (RC) pairing, (2) positive PMI (PPMI) withinwindow, (3) directed de Bruijn path with RC links, (4) traditional directed de Bruijn graph without RC link, and (5) positional bin profile graphs. For the graph neural network models, we considered four GNN families, namely: GCN, GAT, GraphSAGE, and GIN. Nodes are$k$-mers, while edges encode statistical or positional relations between nodes, with normalization that controls for sequence length. Combined, these resulted in 20 different graph structure-GNN model combinations. We then integrated these 20 models to build a stacking classifier ensemble for subcellular lncRNA localization. Further, we evaluate inexact (Hamming-1)$k$-mer bucketing under this framework using graph-level SMOTE to handle data imbalance problems. On held-out tests, our best configurations improve overall accuracy compared to single-model baselines, while avoiding problem of fold-leakage by constructing graphs on per sequence basis.

In this work, we studied the question of whether the two ends of long non-coding ribonucleic acids (lncRNAs) (i.e., the 5′ end and 3′end) carry similar information about subcellular localization of lncRNAs. We considered this problem from three viewpoints using machine learning models: (1) consideration of the classification performance of the machine learning models using features from defined regions (or segments) along the sequence, (2) correlation-based analysis using models built on regions/segments along the LncRNA sequence, and (3) analysis of the relative positions of predicted lncRNA localization motifs along the LncRNA sequence. Our results and observations suggest that the 5′ region of the lncRNA sequences (the prefixes) tend to carry more localization signals when compared with the 3′ region (the suffixes) of the sequences. These could have implications on how we use machine learning models for improved analysis of lncRNA subcellular localization.

BACKGROUND: Metagenomic sequencing is becoming a powerful technology for exploring micro-ogranisms from various environments, such as human body, without isolation and cultivation. Accurately identifying genes from metagenomic fragments is one of the most fundamental issues. RESULTS: In this article, we present a novel gene prediction method named MetaGUN for metagenomic fragments based on a machine learning approach of SVM. It implements in a three-stage strategy to predict genes. Firstly, it classifies input fragments into phylogenetic groups by a k-mer based sequence binning method. Then, protein-coding sequences are identified for each group independently with SVM classifiers that integrate entropy density profiles (EDP) of codon usage, translation initiation site (TIS) scores and open reading frame (ORF) length as input patterns. Finally, the TISs are adjusted by employing a modified version of MetaTISA. To identify protein-coding sequences, MetaGun builds the universal module and the novel module. The former is based on a set of representative species, while the latter is designed to find potential functionary DNA sequences with conserved domains. CONCLUSIONS: Comparisons on artificial shotgun fragments with multiple current metagenomic gene finders show that MetaGUN predicts better results on both 3' and 5' ends of genes with fragments of various lengths. Especially, it makes the most reliable predictions among these methods. As an application, MetaGUN was used to predict genes for two samples of human gut microbiome. It identifies thousands of additional genes with significant evidences. Further analysis indicates that MetaGUN tends to predict more potential novel genes than other current metagenomic gene finders.

SUMMARY: We proposed a tool named MetaTISA with an aim to improve TIS prediction of current gene-finders for metagenomes. The method employs a two-step strategy to predict translation initiation sites (TISs) by first clustering metagenomic fragments into phylogenetic groups and then predicting TISs independently for each group in an unsupervised manner. As evaluated on experimentally verified TISs, MetaTISA greatly improves the accuracies of TIS prediction of current gene-finders. AVAILABILITY: The C++ source code is freely available under the GNU GPL license via http://mech.ctb.pku.edu.cn/MetaTISA/.

Abstract Summary: We report a new and simple method, TriTISA, for accurate prediction of translation initiation site (TIS) of microbial genomes. TriTISA classifies all candidate TISs into three categories based on evolutionary properties, and characterizes them in terms of Markov models. Then, it employs a Bayesian methodology for the selection of true TIS with a non-supervised, iterative procedure. Assessment on experimentally verified TIS data shows that TriTISA is overall better than all other methods of the state-of-the-art for microbial genome TIS prediction. In particular, TriTISA is shown to have a robust accuracy independent of the quality of initial annotation. Availability: The C++ source code is freely available under the GNU GPL license viahttp://mech.ctb.pku.edu.cn/protisa/TriTISA. Contact: [email protected] Supplementary information: Full documentation of the program, containing installation instructions and other operational details, is available on our website. Supplementary data are available at Bioinformatics online.

BACKGROUND: Accurate annotation of translation initiation sites (TISs) is essential for understanding the translation initiation mechanism. However, the reliability of TIS annotation in widely used databases such as RefSeq is uncertain due to the lack of experimental benchmarks. RESULTS: Based on a homogeneity assumption that gene translation-related signals are uniformly distributed across a genome, we have established a computational method for a large-scale quantitative assessment of the reliability of TIS annotations for any prokaryotic genome. The method consists of modeling a positional weight matrix (PWM) of aligned sequences around predicted TISs in terms of a linear combination of three elementary PWMs, one for true TIS and the two others for false TISs. The three elementary PWMs are obtained using a reference set with highly reliable TIS predictions. A generalized least square estimator determines the weighting of the true TIS in the observed PWM, from which the accuracy of the prediction is derived. The validity of the method and the extent of the limitation of the assumptions are explicitly addressed by testing on experimentally verified TISs with variable accuracy of the reference sets. The method is applied to estimate the accuracy of TIS annotations that are provided on public databases such as RefSeq and ProTISA and by programs such as EasyGene, GeneMarkS, Glimmer 3 and TiCo. It is shown that RefSeq's TIS prediction is significantly less accurate than two recent predictors, Tico and ProTISA. With convincing proofs, we show two general preferential biases in the RefSeq annotation, i.e. over-annotating the longest open reading frame (LORF) and under-annotating ATG start codon. Finally, we have established a new TIS database, SupTISA, based on the best prediction of all the predictors; SupTISA has achieved an average accuracy of 92% over all 532 complete genomes. CONCLUSION: Large-scale computational evaluation of TIS annotation has been achieved. A new TIS database much better than RefSeq has been constructed, and it provides a valuable resource for further TIS studies.

BACKGROUND: Despite a remarkable success in the computational prediction of genes in Bacteria and Archaea, a lack of comprehensive understanding of prokaryotic gene structures prevents from further elucidation of differences among genomes. It continues to be interesting to develop new ab initio algorithms which not only accurately predict genes, but also facilitate comparative studies of prokaryotic genomes. RESULTS: This paper describes a new prokaryotic genefinding algorithm based on a comprehensive statistical model of protein coding Open Reading Frames (ORFs) and Translation Initiation Sites (TISs). The former is based on a linguistic "Entropy Density Profile" (EDP) model of coding DNA sequence and the latter comprises several relevant features related to the translation initiation. They are combined to form a so-called Multivariate Entropy Distance (MED) algorithm, MED 2.0, that incorporates several strategies in the iterative program. The iterations enable us to develop a non-supervised learning process and to obtain a set of genome-specific parameters for the gene structure, before making the prediction of genes. CONCLUSION: Results of extensive tests show that MED 2.0 achieves a competitive high performance in the gene prediction for both 5' and 3' end matches, compared to the current best prokaryotic gene finders. The advantage of the MED 2.0 is particularly evident for GC-rich genomes and archaeal genomes. Furthermore, the genome-specific parameters given by MED 2.0 match with the current understanding of prokaryotic genomes and may serve as tools for comparative genomic studies. In particular, MED 2.0 is shown to reveal divergent translation initiation mechanisms in archaeal genomes while making a more accurate prediction of TISs compared to the existing gene finders and the current GenBank annotation.

MOTIVATION: At present the computational gene identification methods in microbial genomes have a high prediction accuracy of verified translation termination site (3' end), but a much lower accuracy of the translation initiation site (TIS, 5' end). The latter is important to the analysis and the understanding of the putative protein of a gene and the regulatory machinery of the translation. Improving the accuracy of prediction of TIS is one of the remaining open problems. RESULTS: In this paper, we develop a four-component statistical model to describe the TIS of prokaryotic genes. The model incorporates several features with biological meanings, including the correlation between translation termination site and TIS of genes, the sequence content around the start codon; the sequence content of the consensus signal related to ribosomal binding sites (RBSs), and the correlation between TIS and the upstream consensus signal. An entirely non-supervised training system is constructed, which takes as input a set of annotated coding open reading frames (ORFs) by any gene finder, and gives as output a set of organism-specific parameters (without any prior knowledge or empirical constants and formulas). The novel algorithm is tested on a set of reliable datasets of genes from Escherichia coli and Bacillus subtillis. MED-Start may correctly predict 95.4% of the start sites of 195 experimentally confirmed E.coli genes, 96.6% of 58 reliable B.subtillis genes. Moreover, the test results indicate that the algorithm gives higher accuracy for more reliable datasets, and is robust to the variation of gene length. MED-Start may be used as a postprocessor for a gene finder. After processing by our program, the improvement of gene start prediction of gene finder system is remarkable, e.g. the accuracy of TIS predicted by MED 1.0 increases from 61.7 to 91.5% for 854 E.coli verified genes, while that by GLIMMER 2.02 increases from 63.2 to 92.0% for the same dataset. These results show that our algorithm is one of the most accurate methods to identify TIS of prokaryotic genomes. AVAILABILITY: The program MED-Start can be accessed through the website of CTB at Peking University: http://ctb.pku.edu.cn/main/SheGroup/MED_Start.htm.