K-mer Representations Research Articles

Improved functions for nonlinear sequence comparison using SEEKR.

SEquence Evaluation through k-mer Representation (SEEKR) is a method of sequence comparison that uses sequence substrings called k-mers to quantify the nonlinear similarity between nucleic acid species. We describe the development of new functions within SEEKR that enable end-users to estimate P-values that ascribe statistical significance to SEEKR-derived similarities, as well as visualize different aspects of k-mer similarity. We apply the new functions to identify chromatin-enriched lncRNAs that contain XIST-like sequence features, and we demonstrate the utility of applying SEEKR on lncRNA fragments to identify potential RNA-protein interaction domains. We also highlight ways in which SEEKR can be applied to augment studies of lncRNA conservation, and we outline the best practice of visualizing RNA-seq read density to evaluate support for lncRNA annotations before their in-depth study in cell types of interest.

HostNet: improved sequence representation in deep neural networks for virus-host prediction

BackgroundThe escalation of viruses over the past decade has highlighted the need to determine their respective hosts, particularly for emerging ones that pose a potential menace to the welfare of both human and animal life. Yet, the traditional means of ascertaining the host range of viruses, which involves field surveillance and laboratory experiments, is a laborious and demanding undertaking. A computational tool with the capability to reliably predict host ranges for novel viruses can provide timely responses in the prevention and control of emerging infectious diseases. The intricate nature of viral-host prediction involves issues such as data imbalance and deficiency. Therefore, developing highly accurate computational tools capable of predicting virus-host associations is a challenging and pressing demand.ResultsTo overcome the challenges of virus-host prediction, we present HostNet, a deep learning framework that utilizes a Transformer-CNN-BiGRU architecture and two enhanced sequence representation modules. The first module, k-mer to vector, pre-trains a background vector representation of k-mers from a broad range of virus sequences to address the issue of data deficiency. The second module, an adaptive sliding window, truncates virus sequences of various lengths to create a uniform number of informative and distinct samples for each sequence to address the issue of data imbalance. We assess HostNet's performance on a benchmark dataset of “Rabies lyssavirus” and an in-house dataset of “Flavivirus”. Our results show that HostNet surpasses the state-of-the-art deep learning-based method in host-prediction accuracies and F1 score. The enhanced sequence representation modules, significantly improve HostNet's training generalization, performance in challenging classes, and stability.ConclusionHostNet is a promising framework for predicting virus hosts from genomic sequences, addressing challenges posed by sparse and varying-length virus sequence data. Our results demonstrate its potential as a valuable tool for virus-host prediction in various biological contexts. Virus-host prediction based on genomic sequences using deep neural networks is a promising approach to identifying their potential hosts accurately and efficiently, with significant impacts on public health, disease prevention, and vaccine development.

Classifying the bacterial taxonomy with its metagenomic data using the deep neural network model

Because of the two sequenced methods stated above, SG and AMP, are being used in different ways, present a deep learning methodology for taxonomic categorization of the metagenomic information which could be utilized for either. To place the suggested pipeline to a trial, 1000 16 S full-length genomes were used to generate either SG or AMP short-reads. Then, to map sequencing as matrices into such a number space, used a k-mer model. Our analysis of the existing approaches revealed several drawbacks, including limited ability to handle complex hierarchical representations of data and suboptimal feature extraction from grid-like structures. To overcome these limitations, we introduce DBNs for feature learning and dimensionality reduction, and CNNs for efficient processing of grid-like metagenomic data. Finally, a training set for every taxon was obtained by training two distinct deep learning constructions, specifically deep belief network (DBN) and convolutional neural network (CNN). This examined the proposed methodology to determine the best factor that determines and compared findings to the classification abilities offered by the RDP classifier, a standard classifier for bacterium identification. These designs outperform using RDP classifiers at every taxonomic level. So, at the genetic level, for example, both CNN and DBN achieved 91.4% accuracy using AMP short-reads, but the RDP classifier achieved 83.9% with the same information. This paper, suggested a classification method for 16 S short-read sequences created on k-mer representations and a deep learning structure, that every taxon creates a classification method. The experimental findings validate the suggested pipelines as a realistic strategy for classifying bacterium samples; as a result, the technique might be included in the most commonly used tools for the metagenomic research. According to the outcomes, it could be utilized to effectively classify either SG or AMP information.

Investigation of the BERT model on nucleotide sequences with non-standard pre-training and evaluation of different k-mer embeddings.

In recent years, pre-training with the transformer architecture has gained significant attention. While this approach has led to notable performance improvements across a variety of downstream tasks, the underlying mechanisms by which pre-training models influence these tasks, particularly in the context of biological data, are not yet fully elucidated. In this study, focusing on the pre-training on nucleotide sequences, we decompose a pre-training model of Bidirectional Encoder Representations from Transformers (BERT) into its embedding and encoding modules to analyze what a pre-trained model learns from nucleotide sequences. Through a comparative study of non-standard pre-training at both the data and model levels, we find that a typical BERT model learns to capture overlapping-consistent k-mer embeddings for its token representation within its embedding module. Interestingly, using the k-mer embeddings pre-trained on random data can yield similar performance in downstream tasks, when compared with those using the k-mer embeddings pre-trained on real biological sequences. We further compare the learned k-mer embeddings with other established k-mer representations in downstream tasks of sequence-based functional prediction. Our experimental results demonstrate that the dense representation of k-mers learned from pre-training can be used as a viable alternative to one-hot encoding for representing nucleotide sequences. Furthermore, integrating the pre-trained k-mer embeddings with simpler models can achieve competitive performance in two typical downstream tasks. The source code and associated data can be accessed at https://github.com/yaozhong/bert_investigation.

RegCloser: a robust regression approach to closing genome gaps

BackgroundClosing gaps in draft genomes leads to more complete and continuous genome assemblies. The ubiquitous genomic repeats are challenges to the existing gap-closing methods, based on either the k-mer representation by the de Bruijn graph or the overlap-layout-consensus paradigm. Besides, chimeric reads will cause erroneous k-mers in the former and false overlaps of reads in the latter.ResultsWe propose a novel local assembly approach to gap closing, called RegCloser. It represents read coordinates and their overlaps respectively by parameters and observations in a linear regression model. The optimal overlap is searched only in the restricted range consistent with insert sizes. Under this linear regression framework, the local DNA assembly becomes a robust parameter estimation problem. We solved the problem by a customized robust regression procedure that resists the influence of false overlaps by optimizing a convex global Huber loss function. The global optimum is obtained by iteratively solving the sparse system of linear equations. On both simulated and real datasets, RegCloser outperformed other popular methods in accurately resolving the copy number of tandem repeats, and achieved superior completeness and contiguity. Applying RegCloser to a plateau zokor draft genome that had been improved by long reads further increased contig N50 to 3-fold long. We also tested the robust regression approach on layout generation of long reads.ConclusionsRegCloser is a competitive gap-closing tool. The software is available at https://github.com/csh3/RegCloser. The robust regression approach has a prospect to be incorporated into the layout module of long read assemblers.

Deep Effective k-mer representation learning for polyadenylation signal prediction via co-occurrence embedding

The polyadenylation process begins as the transcription of a gene terminates, and polyadenylation signal prediction by the genomic sequence is the key to understanding the mechanism of mRNA metabolism. Computational algorithms are widely proposed for polyadenylation signal prediction, but they mainly rely on sequences’ statistical and compositional characteristics. In this work, we conduct adaptively context-aware feature learning for polyadenylation signal prediction via a deep model and co-occurrence embedding. Specifically, we devise an activation function with dynamic parameters to control the scale of features channel-wisely for the positive and negative part value of neuron outputs, and then use the statistical information of k-mer occurrences as features of polyadenylation data, which is pre-trained by an unsupervised learning algorithm. Next, by combining multiscale convolutions with adaptive coefficients, long short-term memory networks jointly extract spatial and temporal contextual patterns. The patterns are fused to obtain contextual information by an identity skip connection and addition strategy. Experimental results demonstrate that our proposed model achieves more remarkable performance than state-of-the-art methods across polyadenylation signal benchmarks, and an ablation study shows the effectiveness of the model, and k-mer embedding is a good feature representation for polyadenylation signals.

BERT contextual embeddings for taxonomic classification of bacterial DNA sequences

Biological taxonomic classification is an important task needed for the identification and discovery of organisms, as well as the inference of their evolutionary relationships. The order and structure of biological sequence components has an essential and primary role in what the sequence’s identity and function is. In order to be able to efficiently differentiate between different bacterial categories, interactions and positions of the biological components in sequences must be known — which is an essential challenge in biological sequence classification. In this light, a considerable amount of recent research has been made to explore efficient representations of biological sequences such as spectral k-mer representation, one-hot encoding, Hilbert space curves and classical word embeddings such as Word2Vec. This paper identifies the taxonomic classification of bacterial 16S rRNA genes at five resolutions mapping hierarchical taxonomic ranks. A Bidirectional Encoder Representations from Transformers (BERT) model is pretrained using biological sequences, which to the best of our knowledge is the first time BERT has been trained with such sequences. A complete prediction model is then proposed – BioSeqBERT-CNN – that initially extracts contextual embeddings representations of DNA sequences using the pretrained BERT model. Extracted representations are further used for taxonomic classification through a Convolutional Neural Network (CNN). For boosting the deep learning classification performance, a data augmentation step is applied. Classification with the original dataset on the most fine-grained rank produced an accuracy of 93.5%, which surpasses that of recent works by 1.5–24.3%. Using data augmentation, an accuracy of 99.9% is achieved, which exceeds values of recent works by a minimum and maximum of 7.9% and 30.7%, respectively on the most fine-grained taxonomic rank. This exhibits promising performance promoting the study of using contextual embeddings to represent biological sequences, with Deep Learning networks.

Regional and Single Nucleotide Correction of Sequence Bias in Chromatin Accessibility Data

Chromatin accessibility assays directly measure which regions of DNA are actively regulated in the genome. Genomic chromatin accessibility data provide the gold standard of spatial and temporal resolution of regulatory landscapes in various cell types and conditions. Accessibility assays depend on enzymatic digestion of DNA and these enzymes are biased with respect to the sequences they recognize and cut. Sequence biases can lead to incorrect conclusions about the nature of the chromatin landscape within an experiment. Previous work from our group has shown that we can use observed k‐mer cut frequencies and k‐mer representation in the genome to correct the majority of this bias. However, these k‐mer scaling approaches did not effectively correct ATAC‐seq data, which assesses chromatin accessibility using the Tn5 transposase enzyme. Here, we find that the oligomeric state of Tn5 and wide sequence bias window precludes effective k‐mer scaling to account for Tn5 sequence bias. Tn5 sequence specificity is influenced by sequence content up to ten bases away from the Tn5 insertion site in both directions. In comparison, Benzonase, Cyanase, MNase, and DNase are primarily influenced by the sequence within 3 bases of the cleavage site. We could not effectively apply k‐mer scaling to ATAC‐seq data because the majority of insertion sites happen at unique 20‐mers in the genome and the method requires many genomic instances per k‐mer. To overcome this limitation, we applied an interpretable machine learning approach, prediction rules ensemble, to correct Tn5‐digested naked DNA. Application of a prediction rules ensemble effectively corrects local (single nucleotide resolution) sequence bias as well as sequence biases influenced by regional GC content. Correction of Tn5 sequence biases is important as the field continues to generate ATAC‐seq data at accelerating rates and use these data to accurately infer regulatory element activity.

RNA transcription and degradation of Alu retrotransposons depends on sequence features and evolutionary history.

Alu elements are one of the most successful groups of RNA retrotransposons and make up 11% of the human genome with over 1 million individual loci. They are linked to genetic defects, increases in sequence diversity, and influence transcriptional activity. Still, their RNA metabolism is poorly understood yet. It is even unclear whether Alu elements are mostly transcribed by RNA Polymerase II or III. We have conducted a transcription shutoff experiment by α-amanitin and metabolic RNA labeling by 4-thiouridine combined with RNA fragmentation (TT-seq) and RNA-seq to shed further light on the origin and life cycle of Alu transcripts. We find that Alu RNAs are more stable than previously thought and seem to originate in part from RNA Polymerase II activity, as previous reports suggest. Their expression however seems to be independent of the transcriptional activity of adjacent genes. Furthermore, we have developed a novel statistical test for detecting the expression of quantitative trait loci in Alu elements that relies on the de Bruijn graph representation of all Alu sequences. It controls for both statistical significance and biological relevance using a tuned k-mer representation, discovering influential sequence features missed by regular motif search. In addition, we discover several point mutations using a generalized linear model, and motifs of interest, which also match transcription factor-binding motifs.

Metagenomic Geolocation Using Read Signatures.

We present a novel approach to the Metagenomic Geolocation Challenge based on random projection of the sample reads from each location. This approach explores the direct use of k-mer composition to characterise samples so that we can avoid the computationally demanding step of aligning reads to available microbial reference sequences. Each variable-length read is converted into a fixed-length, k-mer-based read signature. Read signatures are then clustered into location signatures which provide a more compact characterisation of the reads at each location. Classification is then treated as a problem in ranked retrieval of locations, where signature similarity is used as a measure of similarity in microbial composition. We evaluate our approach using the CAMDA 2020 Challenge dataset and obtain promising results based on nearest neighbour classification. The main findings of this study are that k-mer representations carry sufficient information to reveal the origin of many of the CAMDA 2020 Challenge metagenomic samples, and that this reference-free approach can be achieved with much less computation than methods that need reads to be assigned to operational taxonomic units—advantages which become clear through comparison to previously published work on the CAMDA 2019 Challenge data.

Set-Min Sketch: A Probabilistic Map for Power-Law Distributions with Application to k-Mer Annotation

k-mer counts are important features used by many bioinformatics pipelines. Existing k-mer counting methods focus on optimizing either time or memory usage, producing in output very large count tables explicitly representing k-mers together with their counts. Storing k-mers is not needed if the set of k-mers is known, making it possible to only keep counters and their association to k-mers. Solutions avoiding explicit representation of k-mers include Minimal Perfect Hash Functions (MPHFs) and Count-Min sketches. We introduce Set-Min sketch-a sketching technique for representing associative maps inspired from Count-Min-and apply it to the problem of representing k-mer count tables. Set-Min is provably more accurate than both Count-Min and Max-Min-an improved variant of Count-Min for static datasets that we define here. We show that Set-Min sketch provides a very low error rate, in terms of both the probability and the size of errors, at the expense of a very moderate memory increase. On the other hand, Set-Min sketches are shown to take up to an order of magnitude less space than MPHF-based solutions, for fully assembled genomes and large k. Space-efficiency of Set-Min in this case takes advantage of the power-law distribution of k-mer counts in genomic datasets.

Genome-Wide Mutation Scoring for Machine-Learning-Based Antimicrobial Resistance Prediction.

The prediction of antimicrobial resistance (AMR) based on genomic information can improve patient outcomes. Genetic mechanisms have been shown to explain AMR with accuracies in line with standard microbiology laboratory testing. To translate genetic mechanisms into phenotypic AMR, machine learning has been successfully applied. AMR machine learning models typically use nucleotide k-mer counts to represent genomic sequences. While k-mer representation efficiently captures sequence variation, it also results in high-dimensional and sparse data. With limited training data available, achieving acceptable model performance or model interpretability is challenging. In this study, we explore the utility of feature engineering with several biologically relevant signals. We propose to predict the functional impact of observed mutations with PROVEAN to use the predicted impact as a new feature for each protein in an organism’s proteome. The addition of the new features was tested on a total of 19,521 isolates across nine clinically relevant pathogens and 30 different antibiotics. The new features significantly improved the predictive performance of trained AMR models for Pseudomonas aeruginosa, Citrobacter freundii, and Escherichia coli. The balanced accuracy of the respective models of those three pathogens improved by 6.0% on average.

DeepD2V: A Novel Deep Learning-Based Framework for Predicting Transcription Factor Binding Sites from Combined DNA Sequence.

Predicting in vivo protein–DNA binding sites is a challenging but pressing task in a variety of fields like drug design and development. Most promoters contain a number of transcription factor (TF) binding sites, but only a small minority has been identified by biochemical experiments that are time-consuming and laborious. To tackle this challenge, many computational methods have been proposed to predict TF binding sites from DNA sequence. Although previous methods have achieved remarkable performance in the prediction of protein–DNA interactions, there is still considerable room for improvement. In this paper, we present a hybrid deep learning framework, termed DeepD2V, for transcription factor binding sites prediction. First, we construct the input matrix with an original DNA sequence and its three kinds of variant sequences, including its inverse, complementary, and complementary inverse sequence. A sliding window of size k with a specific stride is used to obtain its k-mer representation of input sequences. Next, we use word2vec to obtain a pre-trained k-mer word distributed representation model. Finally, the probability of protein–DNA binding is predicted by using the recurrent and convolutional neural network. The experiment results on 50 public ChIP-seq benchmark datasets demonstrate the superior performance and robustness of DeepD2V. Moreover, we verify that the performance of DeepD2V using word2vec-based k-mer distributed representation is better than one-hot encoding, and the integrated framework of both convolutional neural network (CNN) and bidirectional LSTM (bi-LSTM) outperforms CNN or the bi-LSTM model when used alone. The source code of DeepD2V is available at the github repository.

Classification of Precursor MicroRNAs from Different Species Based on K-mer Distance Features

MicroRNAs (miRNAs) are short RNA sequences that are actively involved in gene regulation. These regulators on the post-transcriptional level have been discovered in virtually all eukaryotic organisms. Additionally, miRNAs seem to exist in viruses and might also be produced in microbial pathogens. Initially, transcribed RNA is cleaved by Drosha, producing precursor miRNAs. We have previously shown that it is possible to distinguish between microRNA precursors of different clades by representing the sequences in a k-mer feature space. The k-mer representation considers the frequency of a k-mer in the given sequence. We further hypothesized that the relationship between k-mers (e.g., distance between k-mers) could be useful for classification. Three different distance-based features were created, tested, and compared. The three feature sets were entitled inter k-mer distance, k-mer location distance, and k-mer first–last distance. Here, we show that classification performance above 80% (depending on the evolutionary distance) is possible with a combination of distance-based and regular k-mer features. With these novel features, classification at closer evolutionary distances is better than using k-mers alone. Combining the features leads to accurate classification for larger evolutionary distances. For example, categorizing Homo sapiens versus Brassicaceae leads to an accuracy of 93%. When considering average accuracy, the novel distance-based features lead to an overall increase in effectiveness. On the contrary, secondary-structure-based features did not lead to any effective separation among clades in this study. With this line of research, we support the differentiation between true and false miRNAs detected from next-generation sequencing data, provide an additional viewpoint for confirming miRNAs when the species of origin is known, and open up a new strategy for analyzing miRNA evolution.

Taxonomic classification of metagenomic sequences from Relative Abundance Index profiles using deep learning

We propose a Convolutional Neural Network approach based on k-mer representation for metagenomic fragment classification problem. The proposed model consists of two steps; the first step is representation of DNA based on k-mer frequency with Relative Abundance Index (RAI) and the second step is classification metagenomic fragments with CNN. RAI scores, as DNA fragment representations are fed to CNN classifiers (CNN-RAI). RAI consist of the over- and under abundance statistics gathered from the taxon for each k-mer. In order to compare the performances of CNN-RAI and RAIphy, which classifies metagenomic fragments using the same input attributes with an expectation-maximization based approach, databases of different metagenomic scenarios were tested. Metagenomics data that were generated (or simulated) by different Next-Generation Sequencing platforms, respectively Illumina technology and Oxford Nanopore MinION were compiled into shotgun metagenomics or 16S rRNA datasets. RAI based method and CNN models were trained on represented data with read lengths ranging between 200 and 10,000 bp, also with distinct k-mer size (3≤k≤7) at genus level. RAI score was used for the first time in the deep learning algorithm as a spectral representation with improved performance thanks to the ability of deep learning on each dataset for a range of parameters. The proposed representation was compared to the current spectral methods and shown to be competitive for all datasets used in this study.

Systematic analysis of binding of transcription factors to noncoding variants.

Many sequence variants have been linked to complex human traits and diseases1, but deciphering their biological functions remains challenging, as most of them reside in noncoding DNA. Here we have systematically assessed the binding of 270 human transcription factors to 95,886 noncoding variants in the human genome using an ultra-high-throughput multiplex protein-DNA binding assay, termed single-nucleotide polymorphism evaluation by systematic evolution of ligands by exponential enrichment (SNP-SELEX). The resulting 828 million measurements of transcription factor-DNA interactions enable estimation of the relative affinity of these transcription factors to each variant in vitro and evaluation of the current methods to predict the effects of noncoding variants on transcription factor binding. We show that the position weight matrices of most transcription factors lack sufficient predictive power, whereas the support vector machine combined with the gapped k-mer representation show much improved performance, when assessed on results from independent SNP-SELEX experiments involving a new set of 61,020 sequence variants. We report highly predictive models for 94 human transcription factors and demonstrate their utility in genome-wide association studies and understanding of the molecular pathways involved in diverse human traits and diseases.

Word2vec based deep learning network for DNA N4-methylcytosine sites identification

As an important epigenetic modification, N4-methylcytosine not only controls DNA replication and cell cycle, but also participates in regulating cell differentiation and gene expression. However, the understanding of its biological function is far from enough. In order to further reveal the function and regulatory mechanism of 4mC, it is important to accurately identify the 4mC site and detect its distribution in the genome. In this study, we propose 4mcDeep-W2VC, a general and efficient deep neural network to identify 4mC sites. Different with other methods, our proposed predictor can automatically extract features based on DNA sequences. We use the word2vec algorithm to learn the distributed representation of k-mers instead of one-hot encoding. Compared with the traditional k-mers method, the distributed representation we obtained considers the potential relationship between k-mers. Next, we input the distributed representation of the DNA sequence into convolutional neural network to extract hidden high-level and more biological features. The experimental results show that our predictor can achieve better performance in identifying 4mC sites compared to the state-of-art predictors.

KmcEx: memory-frugal and retrieval-efficient encoding of counted k-mers.

K-mers along with their frequency have served as an elementary building block for error correction, repeat detection, multiple sequence alignment, genome assembly, etc., attracting intensive studies in k-mer counting. However, the output of k-mer counters itself is large; very often, it is too large to fit into main memory, leading to highly narrowed usability. We introduce a novel idea of encoding k-mers as well as their frequency, achieving good memory saving and retrieval efficiency. Specifically, we propose a Bloom filter-like data structure to encode counted k-mers by coupled-bit arrays-one for k-mer representation and the other for frequency encoding. Experiments on five real datasets show that the average memory-saving ratio on all 31-mers is as high as 13.81 as compared with raw input, with 7 hash functions. At the same time, the retrieval time complexity is well controlled (effectively constant), and the false-positive rate is decreased by two orders of magnitude. The source codes of our algorithm are available at github.com/lzhLab/kmcEx. Supplementary data are available at Bioinformatics online.

MicroPheno: predicting environments and host phenotypes from 16S rRNA gene sequencing using a k-mer based representation of shallow sub-samples.

Author(s): Asgari, Ehsaneddin; Garakani, Kiavash; McHardy, Alice C; Mofrad, Mohammad RK | Abstract: MotivationMicrobial communities play important roles in the function and maintenance of various biosystems, ranging from the human body to the environment. A major challenge in microbiome research is the classification of microbial communities of different environments or host phenotypes. The most common and cost-effective approach for such studies to date is 16S rRNA gene sequencing. Recent falls in sequencing costs have increased the demand for simple, efficient and accurate methods for rapid detection or diagnosis with proved applications in medicine, agriculture and forensic science. We describe a reference- and alignment-free approach for predicting environments and host phenotypes from 16S rRNA gene sequencing based on k-mer representations that benefits from a bootstrapping framework for investigating the sufficiency of shallow sub-samples. Deep learning methods as well as classical approaches were explored for predicting environments and host phenotypes.ResultsA k-mer distribution of shallow sub-samples outperformed Operational Taxonomic Unit (OTU) features in the tasks of body-site identification and Crohn's disease prediction. Aside from being more accurate, using k-mer features in shallow sub-samples allows (i) skipping computationally costly sequence alignments required in OTU-picking and (ii) provided a proof of concept for the sufficiency of shallow and short-length 16S rRNA sequencing for phenotype prediction. In addition, k-mer features predicted representative 16S rRNA gene sequences of 18 ecological environments, and 5 organismal environments with high macro-F1 scores of 0.88 and 0.87. For large datasets, deep learning outperformed classical methods such as Random Forest and Support Vector Machine.Availability and implementationThe software and datasets are available at https://llp.berkeley.edu/micropheno.Supplementary informationSupplementary data are available at Bioinformatics online.

Deep learning models for bacteria taxonomic classification of metagenomic data

BackgroundAn open challenge in translational bioinformatics is the analysis of sequenced metagenomes from various environmental samples. Of course, several studies demonstrated the 16S ribosomal RNA could be considered as a barcode for bacteria classification at the genus level, but till now it is hard to identify the correct composition of metagenomic data from RNA-seq short-read data. 16S short-read data are generated using two next generation sequencing technologies, i.e. whole genome shotgun (WGS) and amplicon (AMP); typically, the former is filtered to obtain short-reads belonging to a 16S shotgun (SG), whereas the latter take into account only some specific 16S hypervariable regions. The above mentioned two sequencing technologies, SG and AMP, are used alternatively, for this reason in this work we propose a deep learning approach for taxonomic classification of metagenomic data, that can be employed for both of them.ResultsTo test the proposed pipeline, we simulated both SG and AMP short-reads, from 1000 16S full-length sequences. Then, we adopted a k-mer representation to map sequences as vectors into a numerical space. Finally, we trained two different deep learning architecture, i.e., convolutional neural network (CNN) and deep belief network (DBN), obtaining a trained model for each taxon. We tested our proposed methodology to find the best parameters configuration, and we compared our results against the classification performances provided by a reference classifier for bacteria identification, known as RDP classifier. We outperformed the RDP classifier at each taxonomic level with both architectures. For instance, at the genus level, both CNN and DBN reached 91.3% of accuracy with AMP short-reads, whereas RDP classifier obtained 83.8% with the same data.ConclusionsIn this work, we proposed a 16S short-read sequences classification technique based on k-mer representation and deep learning architecture, in which each taxon (from phylum to genus) generates a classification model. Experimental results confirm the proposed pipeline as a valid approach for classifying bacteria sequences; for this reason, our approach could be integrated into the most common tools for metagenomic analysis. According to obtained results, it can be successfully used for classifying both SG and AMP data.