Accurate and sensitive taxonomic profiling is essential for any metagenomic analysis to reveal microbial community structure and for potential functional prediction. Antimicrobial resistance (AMR) detection is also a critical task in the clinical diagnosis of infection and antimicrobial therapy. By incorporating Oxford Nanopore Technologies (ONT) sequencing, users benefit from the high-confidence alignment of long reads for taxonomic classification, even among bacteria with similar genomes. Portable ONT devices, such as VolTRAX with MinION, allow short turnaround time for detection and can be used in a lightweight laboratory setting. However, error-prone ONT sequencing reads are still challenging for existing software for accurate taxonomic classification of microbes and detection of AMR down to the drug level. In this paper, we present MegaPath-Nano, the successor to NGS-based MegaPath. It is a high-precision compositional analysis software with drug-level AMR detection for ONT metagenomic sequencing data. MegaPath-Nano performs 1) thorough multi-level filtering against decoy and human reads while removing noisy alignments, 2) alignment-based taxonomic classification with RefSeq down to strain-level, with an alignment-reassignment algorithm to tackle the challenge of non-unique alignments, based on global alignment distribution, and 3) comprehensive downstream drug-level AMR detection, integrating five AMR databases. In our benchmarks using the Zymo metagenomic dataset, MegaPath-Nano performed better than other existing software for taxonomic classification. We also sequenced five real patient isolates using MinION to benchmark its performance of AMR detection. MegaPath-Nano was the most accurate and provided the most comprehensive output at both the drug and class level of AMR prediction against other state-of-the-art software. MegaPath-Nano is open-source and available at https://github.com/HKU-BAL/MegaPath-Nano.

MOTIVATION: It remains challenging to unravel new susceptibility genes of complex diseases and the mechanisms in genome-wide association studies. There are at least two difficulties, isolation of the genuine susceptibility genes from many indirectly associated genes and functional validation of these genes. RESULTS: We first proposed a novel conditional gene-based association test which can use only summary statistics to isolate independently associated genes of a disease. Applying this method, we detected 185 genes of independent association with schizophrenia. We then designed an in-silico experiment based on expression/co-expression to systematically validate pathogenic potential of these genes. We found that genes of independent association with schizophrenia formed more co-expression pairs in normal post-natal but not pre-natal human brain regions than expected. Interestingly, no co-expression enrichment was found in the brain regions of schizophrenia patients. The genes with independent association also had more significant P-values for differential expression between schizophrenia patients and controls in the brain regions. In contrast, indirectly associated genes or associated genes by other widely-used gene-based tests had no such differential expression and co-expression patterns. In summary, this conditional gene-based association test is effective for isolating directly associated genes from indirectly associated genes, and the results insightfully suggest that common variants might contribute to schizophrenia largely by distorting expression and co-expression in post-natal brains. AVAILABILITY AND IMPLEMENTATION: The conditional gene-based association test has been implemented in a platform 'KGG' in Java and is publicly available at http://grass.cgs.hku.hk/limx/kgg/. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Summary: AC-DIAMOND (v1) is a DNA-protein alignment tool designed to tackle the efficiency challenge of aligning large amount of reads or contigs to protein databases. When compared with the previously most efficient method DIAMOND, AC-DIAMOND gains a 6- to 7-fold speed-up, while retaining a similar degree of sensitivity. The improvement is rooted at two aspects: first, using a compressed index of seeds with adaptive-length to speed-up the matching between query and reference sequences; second, adopting a compact form of dynamic programing to fully utilize the parallelism of the SIMD capability. Availability and implementation: Software source codes and binaries available at https://github.com/Maihj/AC-DIAMOND/. Supplementary information: Supplementary data are available at Bioinformatics online.

BACKGROUND: The recent release of the gene-targeted metagenomics assembler Xander has demonstrated that using the trained Hidden Markov Model (HMM) to guide the traversal of de Bruijn graph gives obvious advantage over other assembly methods. Xander, as a pilot study, indeed has a lot of room for improvement. Apart from its slow speed, Xander uses only 1 k-mer size for graph construction and whatever choice of k will compromise either sensitivity or accuracy. Xander uses a Bloom-filter representation of de Bruijn graph to achieve a lower memory footprint. Bloom filters bring in false positives, and it is not clear how this would impact the quality of assembly. Xander does not keep track of the multiplicity of k-mers, which would have been an effective way to differentiate between erroneous k-mers and correct k-mers. RESULTS: In this paper, we present a new gene-targeted assembler MegaGTA, which attempts to improve Xander in different aspects. Quality-wise, it utilizes iterative de Bruijn graphs to take full advantage of multiple k-mer sizes to make the best of both sensitivity and accuracy. Computation-wise, it employs succinct de Bruijn graphs (SdBG) to achieve low memory footprint and high speed (the latter is benefited from a highly efficient parallel algorithm for constructing SdBG). Unlike Bloom filters, an SdBG is an exact representation of a de Bruijn graph. It enables MegaGTA to avoid false-positive contigs and to easily incorporate the multiplicity of k-mers for building better HMM model. We have compared MegaGTA and Xander on an HMP-defined mock metagenomic dataset, and showed that MegaGTA excelled in both sensitivity and accuracy. On a large rhizosphere soil metagenomic sample (327Gbp), MegaGTA produced 9.7-19.3% more contigs than Xander, and these contigs were assigned to 10-25% more gene references. In our experiments, MegaGTA, depending on the number of k-mers used, is two to ten times faster than Xander. CONCLUSION: MegaGTA improves on the algorithm of Xander and achieves higher sensitivity, accuracy and speed. Moreover, it is capable of assembling gene sequences from ultra-large metagenomic datasets. Its source code is freely available at https://github.com/HKU-BAL/megagta .

BACKGROUND: Because of the short read length of high throughput sequencing data, assembly errors are introduced in genome assembly, which may have adverse impact to the downstream data analysis. Several tools have been developed to eliminate these errors by either 1) comparing the assembled sequences with some similar reference genome, or 2) analyzing paired-end reads aligned to the assembled sequences and determining inconsistent features alone mis-assembled sequences. However, the former approach cannot distinguish real structural variations between the target genome and the reference genome while the latter approach could have many false positive detections (correctly assembled sequence being considered as mis-assembled sequence). RESULTS: We present misFinder, a tool that aims to identify the assembly errors with high accuracy in an unbiased way and correct these errors at their mis-assembled positions to improve the assembly accuracy for downstream analysis. It combines the information of reference (or close related reference) genome and aligned paired-end reads to the assembled sequence. Assembly errors and correct assemblies corresponding to structural variations can be detected by comparing the genome reference and assembled sequence. Different types of assembly errors can then be distinguished from the mis-assembled sequence by analyzing the aligned paired-end reads using multiple features derived from coverage and consistence of insert distance to obtain high confident error calls. CONCLUSIONS: We tested the performance of misFinder on both simulated and real paired-end reads data, and misFinder gave accurate error calls with only very few miscalls. And, we further compared misFinder with QUAST and REAPR. misFinder outperformed QUAST and REAPR by 1) identified more true positive mis-assemblies with very few false positives and false negatives, and 2) distinguished the correct assemblies corresponding to structural variations from mis-assembled sequence. misFinder can be freely downloaded from https://github.com/hitbio/misFinder.

Predicting drug-target interaction using computational approaches is an important step in drug discovery and repositioning. To predict whether there will be an interaction between a drug and a target, most existing methods identify similar drugs and targets in the database. The prediction is then made based on the known interactions of these drugs and targets. This idea is promising. However, there are two shortcomings that have not yet been addressed appropriately. Firstly, most of the methods only use 2D chemical structures and protein sequences to measure the similarity of drugs and targets respectively. However, this information may not fully capture the characteristics determining whether a drug will interact with a target. Secondly, there are very few known interactions, i.e. many interactions are “missing” in the database. Existing approaches are biased towards known interactions and have no good solutions to handle possibly missing interactions which affect the accuracy of the prediction. In this paper, we enhance the similarity measures to include non-structural (and non-sequence-based) information and introduce the concept of a “super-target” to handle the problem of possibly missing interactions. Based on evaluations on real data, we show that our similarity measure is better than the existing measures and our approach is able to achieve higher accuracy than the two best existing algorithms, WNN-GIP and KBMF2K.

Bayesian networks (BNs) are popular for modeling conditional distributions of variables and causal relationships, especially in biological settings such as protein interactions, gene regulatory networks and microbial interactions. Previous BN structure learning algorithms treat variables with similar tendency separately. In this paper, we propose a grouped sparse Gaussian BN (GSGBN) structure learning algorithm which creates BN based on three assumptions: (i) variables follow a multivariate Gaussian distribution, (ii) the network only contains a few edges (sparse), (iii) similar variables have less-divergent sets of parents, while not-so-similar ones should have divergent sets of parents (variable grouping). We use L1regularization to make the learned network sparse, and another term to incorporate shared information among variables. For similar variables, GSGBN tends to penalize the differences of similar variables' parent sets more, compared to those not-so-similar variables' parent sets. The similarity of variables is learned from the data by alternating optimization, without prior domain knowledge. Based on this new definition of the optimal BN, a coordinate descent algorithm and a projected gradient descent algorithm are developed to obtain edges of the network and also similarity of variables. Experimental results on both simulated and real datasets show that GSGBN has substantially superior prediction performance for structure learning when compared to several existing algorithms.

Feature selection is important for many biological studies, especially when the number of available samples is limited (in order of hundreds) while the number of input features is large (in order of millions), such as eQTL (expression quantitative trait loci) mapping, GWAS (genome wide association study) and environmental microbial community study. We study the problem of multiple output regression which leverages the underlying common relationship shared by multiple output features and propose an efficient and accurate approach for feature selection. Our approach considers both intra- and inter-group sparsities. The intergroup sparsity assumes that only small set of input features are related to the output features. The intragroup sparsity assumes that each input features may relate to multiple output features which should have different kinds of sparsity. Most existing methods do not model the intragroup sparsity well by either assuming uniform regularization on each group, i.e. each input feature relates to similar number of output features, or requiring prior knowledge of the relationship of input and output features. By modelling the regression coefficients as a mixture distributions of Laplacian and Gaussian, we can shrink group regression coefficients to be small adaptively and learn the intergroup, intragroup sparsity and shrinkage estimation patterns. Empirical studies on the synthetic and real environmental microbial community datasets show that our model has better predictions on test dataset than existing methods such as Lasso, Elastic Net, dirty model and rMTFL (robust multi-task feature learning). Moreover, by using least angle regression or coordinate descent and projected gradient descent techniques for optimization, we can obtain the optimal regression efficiently.

MOTIVATION: RNA sequencing based on next-generation sequencing technology is effective for analyzing transcriptomes. Like de novo genome assembly, de novo transcriptome assembly does not rely on any reference genome or additional annotation information, but is more difficult. In particular, isoforms can have very uneven expression levels (e.g. 1:100), which make it very difficult to identify low-expressed isoforms. One challenge is to remove erroneous vertices/edges with high multiplicity (produced by high-expressed isoforms) in the de Bruijn graph without removing correct ones with not-so-high multiplicity from low-expressed isoforms. Failing to do so will result in the loss of low-expressed isoforms or having complicated subgraphs with transcripts of different genes mixed together due to erroneous vertices/edges. Contributions: Unlike existing tools, which remove erroneous vertices/edges with multiplicities lower than a global threshold, we use a probabilistic progressive approach to iteratively remove them with local thresholds. This enables us to decompose the graph into disconnected components, each containing a few genes, if not a single gene, while retaining many correct vertices/edges of low-expressed isoforms. Combined with existing techniques, IDBA-Tran is able to assemble both high-expressed and low-expressed transcripts and outperform existing assemblers in terms of sensitivity and specificity for both simulated and real data. AVAILABILITY: http://www.cs.hku.hk/~alse/idba_tran. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Next-generation sequencing allows us to sequence reads from a microbial environment using single-cell sequencing or metagenomic sequencing technologies. However, both technologies suffer from the problem that sequencing depth of different regions of a genome or genomes from different species are highly uneven. Most existing genome assemblers usually have an assumption that sequencing depths are even. These assemblers fail to construct correct long contigs. RESULTS: We introduce the IDBA-UD algorithm that is based on the de Bruijn graph approach for assembling reads from single-cell sequencing or metagenomic sequencing technologies with uneven sequencing depths. Several non-trivial techniques have been employed to tackle the problems. Instead of using a simple threshold, we use multiple depthrelative thresholds to remove erroneous k-mers in both low-depth and high-depth regions. The technique of local assembly with paired-end information is used to solve the branch problem of low-depth short repeat regions. To speed up the process, an error correction step is conducted to correct reads of high-depth regions that can be aligned to highconfident contigs. Comparison of the performances of IDBA-UD and existing assemblers (Velvet, Velvet-SC, SOAPdenovo and Meta-IDBA) for different datasets, shows that IDBA-UD can reconstruct longer contigs with higher accuracy. AVAILABILITY: The IDBA-UD toolkit is available at our website http://www.cs.hku.hk/~alse/idba_ud

MOTIVATION: Metagenomic binning remains an important topic in metagenomic analysis. Existing unsupervised binning methods for next-generation sequencing (NGS) reads do not perform well on (i) samples with low-abundance species or (ii) samples (even with high abundance) when there are many extremely low-abundance species. These two problems are common for real metagenomic datasets. Binning methods that can solve these problems are desirable. RESULTS: We proposed a two-round binning method (MetaCluster 5.0) that aims at identifying both low-abundance and high-abundance species in the presence of a large amount of noise due to many extremely low-abundance species. In summary, MetaCluster 5.0 uses a filtering strategy to remove noise from the extremely low-abundance species. It separate reads of high-abundance species from those of low-abundance species in two different rounds. To overcome the issue of low coverage for low-abundance species, multiple w values are used to group reads with overlapping w-mers, whereas reads from high-abundance species are grouped with high confidence based on a large w and then binning expands to low-abundance species using a relaxed (shorter) w. Compared to the recent tools, TOSS and MetaCluster 4.0, MetaCluster 5.0 can find more species (especially those with low abundance of say 6× to 10×) and can achieve better sensitivity and specificity using less memory and running time. AVAILABILITY: http://i.cs.hku.hk/~alse/MetaCluster/ CONTACT: [email protected].

MOTIVATION: With the rapid development of next-generation sequencing techniques, metagenomics, also known as environmental genomics, has emerged as an exciting research area that enables us to analyze the microbial environment in which we live. An important step for metagenomic data analysis is the identification and taxonomic characterization of DNA fragments (reads or contigs) resulting from sequencing a sample of mixed species. This step is referred to as 'binning'. Binning algorithms that are based on sequence similarity and sequence composition markers rely heavily on the reference genomes of known microorganisms or phylogenetic markers. Due to the limited availability of reference genomes and the bias and low availability of markers, these algorithms may not be applicable in all cases. Unsupervised binning algorithms which can handle fragments from unknown species provide an alternative approach. However, existing unsupervised binning algorithms only work on datasets either with balanced species abundance ratios or rather different abundance ratios, but not both. RESULTS: In this article, we present MetaCluster 3.0, an integrated binning method based on the unsupervised top--down separation and bottom--up merging strategy, which can bin metagenomic fragments of species with very balanced abundance ratios (say 1:1) to very different abundance ratios (e.g. 1:24) with consistently higher accuracy than existing methods. AVAILABILITY: MetaCluster 3.0 can be downloaded at http://i.cs.hku.hk/~alse/MetaCluster/.

MOTIVATION: Next-generation sequencing techniques allow us to generate reads from a microbial environment in order to analyze the microbial community. However, assembling of a set of mixed reads from different species to form contigs is a bottleneck of metagenomic research. Although there are many assemblers for assembling reads from a single genome, there are no assemblers for assembling reads in metagenomic data without reference genome sequences. Moreover, the performances of these assemblers on metagenomic data are far from satisfactory, because of the existence of common regions in the genomes of subspecies and species, which make the assembly problem much more complicated. RESULTS: We introduce the Meta-IDBA algorithm for assembling reads in metagenomic data, which contain multiple genomes from different species. There are two core steps in Meta-IDBA. It first tries to partition the de Bruijn graph into isolated components of different species based on an important observation. Then, for each component, it captures the slight variants of the genomes of subspecies from the same species by multiple alignments and represents the genome of one species, using a consensus sequence. Comparison of the performances of Meta-IDBA and existing assemblers, such as Velvet and Abyss for different metagenomic datasets shows that Meta-IDBA can reconstruct longer contigs with similar accuracy. AVAILABILITY: Meta-IDBA toolkit is available at our website http://www.cs.hku.hk/~alse/metaidba. CONTACT: [email protected].

BACKGROUND: With the rapid development of genome sequencing techniques, traditional research methods based on the isolation and cultivation of microorganisms are being gradually replaced by metagenomics, which is also known as environmental genomics. The first step, which is still a major bottleneck, of metagenomics is the taxonomic characterization of DNA fragments (reads) resulting from sequencing a sample of mixed species. This step is usually referred as "binning". Existing binning methods are based on supervised or semi-supervised approaches which rely heavily on reference genomes of known microorganisms and phylogenetic marker genes. Due to the limited availability of reference genomes and the bias and instability of marker genes, existing binning methods may not be applicable in many cases. RESULTS: In this paper, we present an unsupervised binning method based on the distribution of a carefully selected set of l-mers (substrings of length l in DNA fragments). From our experiments, we show that our method can accurately bin DNA fragments with various lengths and relative species abundance ratios without using any reference and training datasets. Another feature of our method is its error robustness. The binning accuracy decreases by less than 1% when the sequencing error rate increases from 0% to 5%. Note that the typical sequencing error rate of existing commercial sequencing platforms is less than 2%. CONCLUSIONS: We provide a new and effective tool to solve the metagenome binning problem without using any reference datasets or markers information of any known reference genomes (species). The source code of our software tool, the reference genomes of the species for generating the test datasets and the corresponding test datasets are available at http://i.cs.hku.hk/~alse/MetaCluster/.

BACKGROUND: DNA assembling is the problem of determining the nucleotide sequence of a genome from its substrings, called reads. In the experiments, there may be some errors on the reads which affect the performance of the DNA assembly algorithms. Existing algorithms, e.g. ECINDEL and SRCorr, correct the error reads by considering the number of times each length-k substring of the reads appear in the input. They treat those length-k substrings appear at least M times as correct substring and correct the error reads based on these substrings. However, since the threshold M is chosen without any solid theoretical analysis, these algorithms cannot guarantee their performances on error correction. RESULTS: In this paper, we propose a method to calculate the probabilities of false positive and false negative when determining whether a length-k substring is correct using threshold M. Based on this optimal threshold M that minimizes the total errors (false positives and false negatives). Experimental results on both real data and simulated data showed that our calculation is correct and we can reduce the total error substrings by 77.6% and 65.1% when compared to ECINDEL and SRCorr respectively. CONCLUSION: We introduced a method to calculate the probability of false positives and false negatives of the length-k substring using different thresholds. Based on this calculation, we found the optimal threshold to minimize the total error of false positive plus false negative.

The problem of discovering novel motifs of binding sites is important to the understanding of gene regulatory networks. Motifs are generally represented by matrices (position weight matrix (PWM) or position specific scoring matrix (PSSM) or strings. However, these representations cannot model biological binding sites well because they fail to capture nucleotide interdependence. It has been pointed out by many researchers that the nucleotides of the DNA binding site cannot be treated independently, e.g. the binding sites of zinc finger in proteins. In this paper, a new representation called Scored Position Specific Pattern (SPSP), which is a generalization of the matrix and string representations, is introduced which takes into consideration the dependent occurrences of neighboring nucleotides. Even though the problem of discovering the optimal motif in SPSP representation is proved to be NP-hard, we introduce a heuristic algorithm called SPSP-Finder, which can effectively find optimal motifs in most simulated cases and some real cases for which existing popular motif finding software, such as Weeder, MEME and AlignACE, fail.

The problem of finding motifs of binding sites is very important to the understanding of gene regulatory networks. Motifs are generally represented by matrices (PWM or PSSM) or strings. However, these representations cannot model biological binding sites well because they fail to capture nucleotide interdependence. It has been pointed out by many researchers that the nucleotides of the DNA binding site cannot be treated independently, e.g. the binding of zinc finger in proteins. In this paper, a new representation called Scored Position Specific Pattern (SPSP), which is a generalization of the matrix and string representations, is introduced which takes into consideration the dependent occurrences of neighboring nucleotides. Even though the problem of finding the optimal motif in SPSP representation is proved to be NP-hard, we introduce a heuristic algorithm called SPSP-Finder, which can effectively find optimal motifs in most simulated cases and some real cases for which existing popular motif-finding software, such as MEME and AlignACE fail

MOTIVATION: Finding common patterns, motifs, from a set of promoter regions of coregulated genes is an important problem in molecular biology. Most existing motif-finding algorithms consider a set of sequences bound by the transcription factor as the only input. However, we can get better results by considering sequences that are not bound by the transcription factor as an additional input. RESULTS: First, instead of using the simple hyper-geometric analysis, we propose to calculate the likelihood based on a more precise probabilistic analysis which considers motif length, sequence length and number of binding sites as input parameters for testing whether motif is found. Second, we adopt an heuristic algorithm bases on our analysis to find motifs. For the simulated and real datasets, our algorithm ALSE compares favorably against common motif-finding programs such as SeedSearch and MEME in all cases and performs very well, especially when each input sequence contains more than one binding site. AVAILABILITY: ALSE is available for download at the homepage http://alse.cs.hku.hk CONTACT: [email protected].

Finding common patterns, or motifs, from a set of DNA sequences is an important problem in molecular biology. Most motif-discovering algorithms/software require the length of the motif as input. Motivated by the fact that the motifs length is usually unknown in practice, Styczynski et al. introduced the extended (l,d)-motif problem (EMP), where the motifs length is not an input parameter. Unfortunately, the algorithm given by Styczynski et al. to solve EMP can take an unacceptably long time to run, e.g. over 3 months to discover a length-14 motif. This paper makes two main contributions. First, we eliminate another input parameter from EMP: the minimum number of binding sites in the DNA sequences. Fewer input parameters not only reduces the burden of the user, but also may give more realistic/robust results since restrictions on length or on the number of binding sites make little sense when the best motif may not be the longest nor have the largest number of binding sites. Second, we develop an efficient algorithm to solve our redefined problem. The algorithm is also a fast solution for EMP (without any sacrifice to accuracy) making EMP practical.

Finding motifs is an important problem in computational biology. Our paper makes two major contributions to this problem. Firstly, we better characterize the types of problem instances that cannot be solved by most existing methods of finding motifs. Secondly, we introduce a different method, which is shown to succeed for various problem instances for which popular existing methods fail.Most existing computational methods to finding motifs are based on the strong-signal model wherein only strong-signal sequences (i.e. those that are known to contain binding sites very similar to the motif) are considered as input and weak-signal sequences (i.e. those do not contain any sub-string similar to the motif) are disregarded.Buhler and Tompa have studied the limitations of methods based on the strong-signal model. They characterized the problem instances for which the motif is unlikely to be found in terms of the number of input (strong-signal) sequences needed under the assumption that each input sequence contains exactly one binding site. They further gave a method to calculate the minimum number of input sequences required.We re-characterize the limitations of the strong-signal model in terms of the minimum total number of binding sites, rather than the minimum number of strong-signal sequences, required to be in the input data set. We use a probability matrix to represent a motif instead of a string pattern to calculate the minimum total number of binding sites required. This new characterization is shown to be more general and realistic.Next, we introduce a more general and realistic energy-based model, which considers all available sequences (including weak-signal sequences) with varying degrees of binding strength to the transcription factors (as measured experimentally by observed color intensity). Given varying degrees of binding strength, our model can consider sequences ranging from those that contain more than one binding site to those that are weak sequences. By treating sequences with different degrees of binding strength differently, we develop a heuristic algorithm called EBMF (Energy-Based Motif Finding algorithm) using an EM-like approach to find motifs under our model. This EBMF algorithm can find motifs for data sets that do not even have the required minimum number of binding sites as previously derived for the strong-signal model. Our algorithm compares favorably with common motif-finding programs AlignACE and MEME, which are based on the strong-signal model. In particular, for some simulated and real data sets, our algorithm finds the motif when both AlignACE and MEME fail to do so.