KrakenHLL: Confident and fast metagenomics classification using unique k-mer counts

Abstract

Motivation False positive identifications are a significant problem in metagenomics. Spurious identifications can attract many reads that often aggregate in the genomes. Genome coverage may be used to filter false positives, but fast k-mer based metagenomic classifiers only provide read counts as metrics, and re-alignment is expensive. We propose using k-mer coverage, which can be computed during classification, as proxy for genome base coverage.

Results We present KrakenHLL, a metagenomics classifier that records the number of unique k-mers as well as coverage for each taxon. KrakenHLL is based on the ultra-fast classification engine Kraken and combines it with HyperLogLog cardinality estimators. We demonstrate that more false-positive identifications can be filtered using the unique k-mer count, especially when looking at species of low abundance. Further enhancements include mapping against multiple databases, plasmid and strain identification using an extended taxonomy, and inclusion of over 100,000 additional viral strain sequences. KrakenHLL runs as fast as Kraken, and sometimes faster.

Availability and Implementation KrakenHLL is implemented in C++ and Perl, and available under the GPL v3 license at https://github.com/fbreitwieser/krakenhll.

Contact .

Introduction

Metagenomic classifiers attempt to assign taxon identifiers to each read in a sample. Typically, this is done using mapping rather than alignment, which returns the read classifications but not the aligned positions in the genomes (as reviewed by Breitwieser, et al., 2017). However, read counts can be deceiving. Sequence contamination of the samples - introduced from laboratory kits or the environment during sample extraction, handling or sequencing - can yield high numbers of spurious identifications (Salter, et al., 2014; Thoendel, et al., 2017). Having only small amounts of input material can further compound the problem of contamination. In clinical diagnosis of infectious diseases, for example, often less than 0.1% of the DNA sequenced is from microbes of interest (Brown, et al., 2018; Salzberg, et al., 2016). Furthermore, spurious matches can result from low-complexity regions of genomes, and contamination in the database genomes themselves (Mukherjee, et al., 2015).

Such false positive reads typically match only small portions of the genome. Reads from microbes that are truly present should distribute relatively uniformly across the genome rather than be concentrated in one or a few locations. Genome alignment can reveal this information. However, it is resource intensive, requires the selection of specific genomes, and it is difficult to extrapolate from the alignment of one genome to higher levels in the taxonomic tree. Some metagenomics methods use coverage information for better mapping or quantification, but usually require results from much slower alignment methods as input (Dadi, et al., 2017). Notably, assembly-based methods also work, but only for highly abundant species (Quince, et al., 2017).

Here, we present KrakenHLL, a novel method that combines fast k-mer based classification with fast k-mer cardinality estimation. KrakenHLL is based on the Kraken metagenomics classifier (Wood and Salzberg, 2014) and implements fast counting of the number of unique k-mers identified for each taxon using the efficient probabilistic cardinality estimation algorithm HyperLogLog (Ertl, 2017; Flajolet, et al., 2007; Heule, et al., 2013). The count and percentage of the taxon’s unique k-mers in the database that are covered by read k-mers can be used to discern false positive from true-positive sequences. Furthermore, KrakenHLL implements other new features for better metagenomics classifications: (a) searches can be done against multiple databases hierarchically, (b) the taxonomy can be extended to include nodes for strains and plasmids, thus enabling their detection, and (c) database build script enables adding over 100 thousand viral strains from the NCBI Viral Genome Resource (Brister, et al., 2015). Notably, KrakenHLL, which provides a superset of the information of Kraken, is as fast or faster than Kraken while using very little additional memory during classification.

Results

KrakenHLL was developed to provide efficient k-mer coverage information for all taxa identified in a metagenomics experiment. The main workflow is as follows: As reads are processed, each k-mer is assigned a taxa from the database (Figure 1 (A)). KrakenHLL instantiates a HyperLogLog data sketch for each taxon, and adds the k-mers to it (Figure 1 (B)). After classification, KrakenHLL traverses up the taxonomic tree and merges the estimators of the child taxa to the parent. KrakenHLL reports the number of unique k-mers, and the breadth and depth of k-mer coverage for each taxon in the taxonomic tree in the classification report (Figure 1 (C)).

Figure 1:

KrakenHLL algorithm and report. (A) The taxon mappings for each k-mer of a read are queried from the database. For each taxon, a unique k-mer counter is instantiated, and the observed k-mers are added to it. (B) Unique k-mer counting is implemented with the probabilistic estimation method HyperLogLog (HLL) with below 1% error in 16KB of memory per counter. (C) The number of unique k-mers, duplicity (average time each k-mer has been seen) and coverage are reported for each taxon in the taxonomic tree, enabling assessment of the classification.

Efficient k-mer cardinality estimation with HyperLogLog algorithm

Exact counting of the number of unique values (cardinality) in the presence of duplicates requires memory proportional to the cardinality. Very accurate estimation of the cardinality, however, can be achieved using only a small amount of fixed space. The HyperLogLog algorithm (HLL), originally described by (Flajolet, et al., 2007), is currently one of the most efficient cardinality estimators, and lends itself to k-mer counting (Irber Junior and Brown, 2016). The main idea behind the method is that long runs of leading zeros are unlikely in random hashes. E. g., it’s expected to see every fourth hash start with one 0-bit before the first 1-bit (012), and every 32^nd hash starts with 00001₂. The algorithm saves a sketch of observed data based on hashes of the k-mers in 2^p one byte registers (in our implementation), where p is the precision parameter. The relative error of the estimate is 1/sqrt(2^p). With p=14, the sketch uses 2¹⁴ one-byte registers, i.e. 16KB of space and has a relative error less than 1% (Figure 2).

Figure 2:

Cardinality estimation on randomly sampled microbial k-mers using HyperLogLog. (Left) Standard deviations on the relative errors of the estimate with precision p ranging from 10 to 18. As expected, higher values of p give lower relative error, and no systematic bias is apparent. Up to cardinalities of about 2^p/4 the relative error is near zero, and at higher cardinalities the error boundaries stay constant. (Right) The size of the registers, space requirement, and expected relative error for HyperLogLog cardinality estimates with different values of p. For example, with a precision p=14, the expected relative error is less than 1%. The counter only requires 16 kB of space, which is three orders of magnitude less than that of an exact counter (at a cardinality of a million). Up to cardinalities of 2^p/4, a sparse representation of the counter is used with a higher precision of 25 and an effective relative error rate of about 0.02%.

Generating the sketch: Each k-mer is first hashed into a 64-bit string H. The sketch starts out in sparse representation which has an effective p of 25, using 4 bytes per element. See (Heule, et al., 2013) for more details on the encoding. Once m/4 distinct elements have been observed, we switch to the standard representation of (Flajolet, et al., 2007): The first p bits of H are used as index i into the registers M. The later 64-p=q bits are used to define the rank based on the position of the first 1-bit (or, equivalently, the count of leading zeros plus one). If all q bits are zero, the rank is q+1. The register M[i] is updated if the rank is higher than the current value of M[i].

When the read classification is finished, KrakenHLL aggregates the taxon sketches up the taxonomy tree. Each taxon’s sketch is merged with its children’s sketches. The cardinality estimate is computed using a recently reported improved method (Ertl, 2017) that does not require empirically determined thresholds to account for biases and switching between linear counting and HLL estimator (Supplementary Figures 1 and 2). Figure 2 shows the performance and memory usage of KrakenHLL’s cardinality estimator for up to one million k-mers. Suppl. Methods Section 1 contains a more in-depth description of the algorithm and implementation.

Results on simulated and biological data

Simulated test datasets are invaluable in assessing the performance of bioinformatics algorithms. Read simulators can create arbitrarily complex artificial communities and we know the source of every read. However, simulated datasets do not necessarily represent biological data. Specifically, laboratory and environmental contamination, a main reason behind false identifications in metagenomics samples (Salter, et al., 2014), are hard to model. Biological test datasets that are generated by mixing bacterial isolates at known quantities, on the other hand, usually have very few species and limited complexity.

(McIntyre, et al., 2017) recently reviewed eleven metagenomics classifiers and compiled a list of simulated and biological test datasets from 16 distinct sources (McIntyre-Mason, Suppl. Table 2). Eleven of these datasets were from biological mock communities. The largest biological datasets consist of 23 species that were mixed at even proportions (Human Microbiome Project mock communities, sequenced with Illumina and 454 machines). We tested KrakenHLL on ten biological and 21 synthetic datasets to see if better separation of false positives and true positives can be achieved using unique k-mer counts instead of read counts (see Suppl. Table 3). Our main measure for comparison is the maximum F1 score, defined as 2*precision*recall/(precision + recall).

Unique k-mer count thresholds worked very well in biological datasets, performing better than the read count threshold in nine out of ten datasets, with a tie in one (Figure 3 and Suppl. Table 3). On average, the maximum F1 was 0.05 higher when using k-mer instead of read thresholds, improving from 0.87 to 0.92. As expected, the difference was not as clear in simulated datasets, even though the k-mer count still performed better than the read count. In eight out of the 21 datasets, both metrics performed equally well, as the datasets were easily separated into true and false identifications. In eight datasets k-mer count achieved better F1 scores, and in five read count achieved better F1 scores. The average F1 with k-mer count was slightly higher with 0.945 against 0.940. This difference in difference in performance is likely due to simulated datasets lacking some features of biological data.

Figure 3:

Using unique k-mers as thresholds instead of reads can give higher F1 scores. Each dot is a dataset described in McIntyre, et al. Excludes (for visual purposes) dataset LC5 of Segata et al. with a F1 read score of 0.73 and F1 Kmer score of 0.75.

Figure 4 shows the results on two simulated and one biological datasets. In simple simulated and biological datasets, the true species often separate nearly perfectly using either a read count or am unique k-mer count threshold (Figure 3 (A)). In more complex datasets, however, read count thresholds often contain more false species than k-mer thresholds (Figure 3 (B) and (C)).

Figure 3:

Unique k-mer counts separate true identifications better from false ones. The plot shows the number of reads vs the number of unique k-mers in two simulated and one biological test datasets. Each point is a species identification. Blue dots are the ‘true’ species, red dots are false positive or background identifications. (A) Relatively easy case of a simulated test dataset with 100 genomes. The true and false positives separate nearly perfectly with either read or k-mer count. (B) Largest simulated test dataset shows better separation by unique k-mer counts. (C) Largest biological dataset also separates well between true and false positives.

Results on biological samples for infectious disease diagnosis

Metagenomics is increasingly used to find species of low abundance. A special case is the emerging use of metagenomics for the diagnosis of infectious diseases (Simner, et al., 2017; Zhang, et al., 2015). Host tissue or body fluids are used to find the likely culprit of a disease. Usually, most (often 95% and more) of the reads match to the host, and maybe 10 to 100 out of the millions of reads are matched to the target species. Skin bacteria from the patient, physician or lab personal and other contamination from sample collection or preparation can easily accumulate a similar number of reads, and thus cloud the detection of the pathogen.

To assess if the unique k-mer count metric can be used to rank and identify pathogen identification, we reanalyzed ten patient samples (Salzberg, et al., 2016). (See Supplementary Methods for details on the database, which also contains over 100 thousand viral strain sequences.). (Salzberg, et al., 2016) sequenced spinal cord mass and brain biopsies from ten patients in the intensive care unit, for whom routine tests for pathogens returned inconclusive. In three out of the ten cases, a likely diagnosis could be made with the help of metagenomics, and in a fourth case, a diagnosis could be made with an updated database. For confirmation of metagenomics class, the authors re-aligned pathogen reads to individual genomes.

Table 1 shows the results of our reanalysis for the confirmed identifications in the four patients, including the number of reads and unique k-mers of the pathogen, as well as the number of covered bases of a re-alignment. Even though the read numbers are low in some cases, the number of unique k-mers suggests that they are distributed across the genome. For example, in PT8, 15 reads are matching 1570 k-mers, and re-alignment shows 2201 covered base pairs. In contrast, Table 2 shows examples of identifications in the same dataset that are not well supported by a high unique k-mer count.

Table 1:

Pathogen identifications in patients with suspected neurological infections. The pathogens were identified with as little as 15 reads, but those mapped to a high number of unique k-mers, indicating random distribution of the reads on the genome. “Bases” are the number of covered bases in the re-alignment of a selected genome. Interestingly, the k-mer count in PT5 reveals that there seems to be more than one viral strain present, as the k-mers cover more than one genome.

Table 2:

Dubious identifications have few k-mers. Note that the viral identifications in PT4 and PT10 stem from non-RefSeq viral genomes form the NCBI Viral Genome Resource. Since the KrakenHLL reports sequence-level matches, the source genomes are easy to find.

Storing strain genomes with assembly project and sequence accessions

Kraken stores a NCBI taxonomic identifier for each k-mer in its database. This strategy worked well when new taxonomy IDs were assigned to each new microbial strain in GenBank. However, in 2014 the NCBI Taxonomy project stopped giving new IDs to microbial strains – only novel species get new taxonomy IDs (Federhen, et al., 2014). New strains, therefore, have the taxonomy ID of the species, or the taxonomy ID of a strain that was added before 2014.

Microbes that have been intensively surveyed, such as Escherichia coli or Salmonella spp., have up to hundreds of genomes indexed with the same taxonomy ID, and are thus indistinguishable by Kraken. The new way of identifying microbial strains is to use the Bioproject, Biosample and Assembly accession codes (Breitwieser, et al., 2017). KrakenHLL thus adds new nodes to the taxonomy tree as children of the assigned taxon. A taxonomic node may also be added for each sequence – e.g. specific bacterial chromosomes or plasmids. Those new nodes in the taxonomy tree are given taxonomy IDs starting at 1,000,000,000. Having these extended nodes can help identify specific strains as well as bad database sequences (see Table 2 and Suppl. Table 3).

Hierarchical read classification with multiple databases

KrakenHLL allows using multiple databases hierarchically in order of confidence. In the following example each k-mer is matched first against the HOST, then the PROK, then the EUK_DRAFT database.

krakenhll —db HOST —db PROK —db EUK_DRAFT

Note that all database need to share the same taxonomy database. If taxIDs are added for genomes or sequences, then it is necessary that the databases are consecutively constructed with the same taxonomy database.

Timing and memory requirements

The additional features of KrakenHLL come without a runtime penalty. In fact, due to code improvements, KrakenHLL can run faster than Kraken especially when most of the reads are from one species (See Suppl. Table 2 for timings on patient data, Suppl. Table 3 for timings on the test datasets). On the patient data, the processing speed (base-pairs per minute) was on average 57% higher with KrakenHLL compared to Kraken, while it was 8% higher. Overall wall clock time was slower, too, when comparing the runtime of both kraken and kraken-report with krakenhll (which generates the report with the classification binary). The average additional memory requirements were less than 1GB. On the patient datasets, the average maximum memory usage went from 118 to 118.35GB, and for the test datasets, the usage went up from 46.28 to 46.99GB.

Conclusions

We present a novel method that combines fast k-mer based classification with efficient cardinality estimation. We demonstrated that unique k-mer counts can help discard false identifications in real samples. When the reads from a species yield many unique k-mers, we are more confident that the taxon is truly present, while a low number of unique k-mers suggests a possible false positive identification. It is important to note that choice of the appropriate threshold will depend on the application. For example, in infectious disease diagnosis, unique k-mers can be used for ranking of the identifications. Conversely, in microbial ecology, a global threshold on the number of unique k-mers can be applied at any desired taxonomic rank. We believe that the ability to summarize to higher levels of the tree is a great advantage of the k-mer count over using covered bases in a genome alignment. In summary, KrakenHLL gives more confident identifications by reporting the unique k-mer count and coverage, without any runtime penalty.

Funding

This work was supported in by the National Institutes of Health [grant number R01-HG007196]; and the U. S. Army Research Office [grant number W911NF-14-1-0490].

Acknowledgements

We’d like to thank Jen Lu, as well as David Karig, Susan Bewick, Peter Thielen, Thomas Mehoke for valuable discussions. Furthermore, we’d like to thank Jessica E. Atwell for proofreading of the manuscript.