Using k-mer frequencies to map reads to homologous protein clusters

The main motivation for kMermaid is that short metagenomic reads are rarely mapped to a single protein and instead often map to multiple functionally similar proteins. Obtaining a unique read-to-protein mapping requires grouping these homologous proteins into some broader functional unit. Such functional units become especially critical for downstream quantitative analyses to prevent issues such as double counting multi-mapped reads. We find that most reads map to at least five protein hits using BLASTX, which is the minimum recommended value for the number of hits reported, while only 7% of the reads can be uniquely mapped by alignment to a single protein by either BLASTX or DIAMOND in BLASTX mode (Fig 2a). In contrast, by aggregating multi-mapped hits from single proteins into homologous protein clusters (see Methods), we find that more than 93% of the reads can be uniquely mapped to a single cluster or functional unit. In other words, for more than 93% of the reads, all BLASTX hits for the read belong to a single cluster. DIAMOND follows a similar trend to BLASTX. Together, this demonstrates that our clusters resolve the majority of ambiguous alignments without loss of information from multimapping.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Protein clusters underlying kMermaid mitigate multi-mapping and allow cluster-specific prevalences of k-mers.

(a) The reduction in the number of reads mapped to >1 protein using default configurations of BLASTX and DIAMOND compared to cluster assignments, when employed to the BLASTX and DIAMOND outputs, respectively, across 29 human fecal samples. (b) The percent of co-clustered proteins using our clustering process also co-clustered in the NCBI PCLA prokaryotic protein clusters for all overlapping proteins, plotted as bars. The x-axis contains individual proteins ordered by cluster, i.e., proteins in the same cluster are plotted next to each other, and the color of the bar corresponds to the percent similarity. (c) The distribution (Kernel density estimation, KDE) of keyword percentage, i.e., percent of cluster members with the most common word from all names of proteins in the cluster, across all clusters (yellow), and for clusters with specific common keywords of interest. (d) Visual representation of the kMermaid cluster frequency model sorted by the number of clusters a k-mer is present in. Colors represent the number of clusters in which a unique k-mer is found, and the y-axis corresponds to the k-mer frequency in the cluster. The panel shows a representative random 50K subset of all k-mers.

https://doi.org/10.1371/journal.pcbi.1013470.g002

We developed kMermaid to uniquely and efficiently map microbial short read sequencing into homologous protein clusters using this underlying clustering framework. A user can provide a file with nucleotide sequences from whole-genome shotgun sequencing to kMermaid, for querying against our precomputed model of 1,793,361 proteins aggregated into 32,308 clusters. kMermaid will uniquely assign each nucleotide read to a cluster and provide a readable functional annotation, i.e., protein label, based on its cluster representative. Because approximately 25% of our cluster representatives had non-descriptive names, e.g., “hypothetical protein,” in RefSeq, we employed HH-suite3 [27] remote homology detection to produce descriptive cluster names. In total, we reannotated 8,617 proteins, of which 6,488 were with high confidence (S1 Table). These include 601 phage proteins, 436 membrane proteins, 230 transcriptional proteins, and 205 lipoproteins (S1 Table). The composition of kMermaid’s clusters is also highly consistent with preexisting, smaller-scale cluster annotations. We verified that 96% of proteins share 100% similarity with existing NCBI protein clusters [28], i.e., all proteins in the kMermaid cluster also co-occur in the broader NCBI-derived clusters (Fig 2b). In addition, using keywords of NCBI-assigned protein names, we show that the functional annotations of the proteins are highly homogenous within clusters (Fig 2c). Therefore, we concluded that kMermaid’s cluster model and corresponding annotations are sufficiently biologically accurate and highly reflective of the cluster content.

kMermaid’s internal pipeline assigns sequencing reads to clusters according to a frequency-based assignment score calculated from amino acid (AA) k-mer frequencies. A higher assignment score indicates that k-mers within a query sequence are more frequently observed in the assigned cluster than other clusters in the underlying model. For some query sequences, a k-mer may uniquely determine cluster assignment, while other query sequences may contain multiple k-mers that have a higher combined frequency in the assigned (maximal) cluster compared with any other cluster. With this in mind, we reasoned that a k-mer found in many clusters may be less informative and its contribution to the assignment score for that cluster is noisy. In the improbable case that all k-mers were present in all clusters at a similar frequency, our cluster assignment would be close to random chance. On the other hand, k-mers that are only found in one or two clusters allow for deterministic classification and enhance our confidence in use for read assignments. Out of approximately 2.5 million AA 5-mers in our model, presence in a single cluster was the most common scenario (22%) and 81% of all AA 5-mers were found in <10 clusters (Fig 2d). We refer to these 5-mers as “deterministic,” such that the presence of a deterministic k-mer in a query sequence is highly informative for cluster assignment. We also found that the AA 5-mers that are present in many clusters have relatively similar frequencies across the clusters when examining the top 10 clusters where they are most frequently found, implying that common 5-mers should not bias the assignment score. As anticipated, given the relative percentage of 5-mers that tend to be deterministic and that the assignment score considers multiple k-mers for each query sequence, kMermaid clusters fully resolve read alignment ambiguity or multi-mapping in more than 95% of the cases (Fig 2a).

Performance evaluation on simulated data with known labels

To provide validation for our model, we comprehensively benchmarked the accuracy and sensitivity of kMermaid against the most widely used metagenomic protein mapping methods: BLASTX, DIAMOND, and MMseqs2. We first benchmarked kMermaid using simulated data sets with varying rates of point mutations selected to mimic biological mutation rates and sequencing error rates. To compare against ground truth, the reads were simulated from the RefSeq data used in the second step of the clustering procedure (see Methods) so that their true cluster labels would be known. kMermaid, DIAMOND, and MMseqs2 were almost always able to map a read to the correct protein at the level of their reporting, i.e., either the single protein (BLASTX, DIAMOND, MMseqs2) or the protein cluster (kMermaid) (Fig 3a). As expected, when BLASTX and DIAMOND were restricted to reporting only one match per read, their performance dropped considerably as highly homologous proteins will have the same alignment scores. The performance of methods that align to a single protein increases substantially when viewing the singly-aligned results at the cluster level in concordance with the notion that clustering proteins by homology resolves ambiguous alignments in most cases (Fig 2a). Encouragingly, kMermaid was able to assign reads to the correct cluster in nearly all cases, albeit with a slight decrease in coverage (percentage of reads classified) when query sequences are > 500 nucleotides and the mutation rates are high (Fig 3b). Because kMermaid uses a cumulative scoring model, it is expected that the assignments for long, highly mutated sequences are noisier, especially at the default scoring threshold which is tuned to short reads. Similar trends were observed when we simulated reads guaranteed to have a certain number of mutations rather than using a probabilistic rate (S2a and S2b Fig). Thus, by resolving multi-mapped reads at the cluster level, kMermaid shows improved sensitivity compared to methods that assign reads only to individual proteins.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. kMermaid sensitivity and resource benchmarking on simulated microbial protein data.

(a) The percentage of reads classified correctly by kMermaid compared with leading read-to-protein mapping tools averaged across 10 simulated datasets per each combination of read length and mutation rate. (b) The number of reads classified by each tool normalized by the number of input reads averaged across 10 simulated datasets per each combination of read length and mutation rate. (c) kMermaid (green) provides up to a 25-fold decrease in runtimes (in seconds, log-transformed) compared to BLASTX and has comparable runtimes to DIAMOND (blue). The y-axis has been truncated and tools that exceeded a 24-hour run time for larger input sizes are denoted with an asterisk. (d) kMermaid (green) requires a fixed, low memory allocation in comparison to other read-to-protein mapping tools. BLASTX was excluded from comparisons with more than 1 million sequences due to the infeasible running times. Methods exceeding 16GB of RAM are denoted with an asterisk.

https://doi.org/10.1371/journal.pcbi.1013470.g003

Computational efficiency of kMermaid

Shotgun metagenomics experiments commonly yield tens of millions of sequences per sample, and each must be queried against a large reference database for classification purposes. Computational efficiency remains a challenge. BLASTX is perhaps the most established nucleotide-to-protein aligner, but its computational time is infeasible for typical large files, necessitating the development of alternatives that can process these reads in a reasonable timeframe. We benchmarked kMermaid’s single-CPU runtime and RAM usage again against BLASTX, DIAMOND, and MMseqs2 (Fig 3c, d). DIAMOND is regarded as one the fastest accurate approaches for protein mapping of metagenomic reads and was in part developed to address the runtime limitations of BLASTX. kMermaid ran 1,600 times faster than BLASTX on files with 100,000 sequences while DIAMOND ran 1,000 times faster. Both methods also provided substantial reductions in running time over MMseqs2. Both kMermaid and DIAMOND were able to classify 500K sequences in a minimum of 2.7 minutes and 3.7 minutes, respectively. The same input when given to BLASTX took over six days to complete. Further, classifying 40M sequences took kMermaid 3.3 hours (compared to 1.9-2.2 hours for DIAMOND), which highlights its usability for experimental shotgun metagenomic sequencing files. Like some other methods, since kMermaid classifies reads independently of other reads in the same input file, it easily lends itself to parallelization by means of splitting input files into smaller chunks, allowing for further speed improvement when resources are available.

Along with speed, RAM usage is another potentially limiting factor when input files are large. We have developed kMermaid to be highly memory efficient. We therefore compared kMermaid to leading tools including DIAMOND, which has excellent running times, but achieves this performance in speed at the expense of higher memory and multiple CPU utilization. With these limitations in mind, kMermaid performs read assignments in a way that only requires the precomputed k-mer frequency model, and not the input data, to be loaded into memory. As such, kMermaid requires a fixed amount (2GB) of memory per run regardless of file size, whereas DIAMOND and MMseqs2 generally require memory to scale with the increasing input file size (Fig 3d). BLASTX was excluded from comparisons with more than 1 million sequences due to long running times.

Using kMermaid to map new or unseen proteins

A key challenge in metagenomics is the classification of unknown microbial sequences that are not present in existing reference databases. To understand how kMermaid performs on unknown microbial proteins, we classified segments of 22,435 new RefSeq protein sequences deposited between January and May 2025, after our frequency model was developed. We compared the resulting mappings to BLASTX alignments with the same RefSeq database used to construct kMermaid’s underlying database. Importantly, we observed that the kMermaid assignment score is correlated with BLASTX percent identity (Spearman r = 0.83, 0.82, 0.8 for reads of length 125, 150, and 200, respectively; S3a, b Fig), highlighting the importance of choosing a more stringent score threshold when classifying reads which are likely to originate from unseen microbes. We further used the area under the receiver operating curve (AUROC) to assess the kMermaid’s ability to correctly classify reads, where a correct classification was based on BLASTX alignment. The kMermaid was able to achieve AUROCs of 0.93 and ≥0.96 for reads matching BLASTX results and higher confidence BLASTX results filtered at a more stringent percent-identity threshold of 66.6%, respectively (Fig 4a). We observed that kMermaid scores ranging from 6.3-8, depending on the read length with longer reads requiring higher scores, were able to achieve false positive rates ≤0.05 while still maintaining true positive rates around 80% (S2 Table). BLASTX was able to classify a slightly higher percentage of reads (13.7-14.9%) compared to kMermaid (11.7%-13.6%), but kMermaid was found to be highly concordant with BLASTX results, with around 95% of assignments correct assuming BLASTX as the gold standard (S2 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Biological applications, function-specific performance, and evidence of remote homology detection of kMermaid.

(a) Receiver operating curves demonstrating the ability of kMermaid’s assignment score to correctly classify reverse-translated nucleotide segments of varying lengths from unseen protein sequences that were added to RefSeq in early 2025. (b) Agreement of BLASTX alignments and kMermaid protein assignments on 29 fecal samples from ulcerative colitis patients. (c) Boxplot showing kMermaid agreement with BLASTX for clusters with specific functional annotations. (d) Violin plots showing the k-mer frequency scores for six clusters of reads unclassified with BLASTX that were correctly functionally classified by kMermaid.

https://doi.org/10.1371/journal.pcbi.1013470.g004

kMermaid is highly sensitive for protein cluster mapping of human fecal samples

Even though kMermaid performed well on simulated reads, it is difficult to account for the additional challenges and noise associated with real, experimental data by simulations alone. Therefore, we performed additional testing using real sequencing data from 29 publicly available human fecal samples of ulcerative colitis patients [29], comparing kMermaid assignments against BLASTX. On average, kMermaid results agreed with BLASTX alignments 83.3% of the time and the agreement rate was highly consistent across the 29 samples (Fig 4b). Assuming BLASTX hits to be the ground truth, kMermaid was able to maintain a balance between retaining a high percentage of the assignments that agree with BLASTX hits as well as a high ratio of assignments that agree with BLASTX to assignments that disagree with BLASTX (Figs 4b and S3c) at the default kMermaid assignment score ≥ 3. Given that the overwhelming majority of BLASTX hits belong to a single cluster (Fig 2a), the consistent agreement between BLASTX and kMermaid provides strong evidence for kMermaid’s ability to accurately and sensitively classify short reads from human fecal metagenomic sequencing.

Cluster specific results

A primary objective of kMermaid is to achieve high performance for classification at the read level. To comprehensively assess kMermaid’s performance, we also evaluated its cluster specific performance. We compared the kMermaid read assignments of experimental metagenomic sequencing input samples used for benchmarking to the assignment by BLASTX, within each cluster. Interestingly, we find that clusters related to restriction, toxins, transposons, and those in the GCN5-related N-acetyltransferases family (GNAT) had high agreement with BLASTX, whereas clusters related to ABC transporters tended to have relatively low agreement with BLASTX, and therefore likely lower accuracy (Fig 4c). We also confirmed that the proportion of reads concordant with BLASTX was correlated with the mean kMermaid assignment score for all reads mapping to the cluster, a trend that was not confounded by the number of reads mapped to the cluster (S3d Fig). Importantly, by investigating reads that were assigned with a high kMermaid assignment score but were not classified by BLASTX, we identified reads with remote homology to proteins within kMermaid clusters. We verified a correct functional classification of reads assigned to six such kMermaid clusters using both PSI-BLAST [30] and HHblits3 from HH-suite3 [27] and validated these kMermaid functional annotations (Fig 4d and S1 File).

Forming functionally similar microbial protein clusters using a two-step clustering procedure

kMermaid is foremost designed to address ambiguity in read alignment, where most shotgun metagenomic reads align against multiple microbial proteins with similar alignment scores and are therefore classified as multiple functionally related proteins by the alignment (Fig 2a). To circumvent this issue, kMermaid uses a well-defined concept and groups functionally related proteins into functional units prior to the assignment such that a read can be uniquely classified into a single functional unit. We therefore constructed a set of comprehensive and high confidence clusters of microbial proteins by employing a two-step clustering procedure using CD-HIT [43,44]. CD-HIT uses an incremental greedy algorithm to identify representative sequences and cluster remaining sequences by sequence similarity using short word filtering. We first clustered 43,176 proteins from the NCBI RefSeq non-redundant microbial protein database [45] using CD-HIT with a similarity threshold of 65% and a word size of 5. This first clustering step resulted in 32,308 clusters allowing non-redundant clusters. In the second step, to further expand and diversify protein members within these clusters, we applied CD-HIT-2D [44] with a 70% sequence similarity threshold to cluster all RefSeq microbial proteins against the previously selected cluster representatives (N = 1,797,426 proteins from RefSeq, dataset downloaded in May 2023). A set of expanded clusters was created from this process such that the final dataset clusters 1,793,361 proteins into 32,308 functional groups.

Assigning functions to clusters of hypothetical or uncharacterized microbial proteins

kMermaid annotates its underlying protein clusters based on the name of the representative sequence as determined in the initial CD-HIT phase. Even though these proteins are in the representative microbial protein database, 8,617 (approximately 25%) of the cluster representatives are best annotated by NCBI or RefSeq as hypothetical proteins, i.e., proteins of unknown or unverified function. To annotate these protein clusters and assign them with protein names, we used HHblits from HH-suite3 [27] for remote homology detection against two databases from The Protein Databank and UniProtKB (PD70 [46] and Uniclust30 [47] v2023, respectively) and selected the match with lowest e-value across the databases that did not map to hypothetical, unknown, or uncharacterized proteins. We were able to confidently assign a protein name or function to 6,488 of the 8,617 hypothetical clusters (e-value < 0.01) and the rest are assigned with lower confidence.

Creating kMermaid’s k-mer frequency cluster model using nested hashing

The goal of kMermaid is to functionally classify short reads using the previously defined clusters of functionally similar microbial proteins. To this end, we built a k-mer frequency model by obtaining all amino acid (AA) k-mers of all protein sequences in each of 32,308 clusters and computing the cluster-level frequency of each AA k-mer (S1 Fig). The value of k was chosen through hyperparameter search, by simulating truncations of each protein in the database into 50 overlapping AA segments. We evaluated k values of 3, 4, 5 and 6 for classifying truncated protein sequences, measuring accuracy as the fraction of truncated sequences which were correctly assigned to their cluster. Both k = 5 and k = 6 achieved similarly high accuracy (>0.99), but the number of k-mers increased sharply from 2,574,615 for k = 5 to 12,043,100 for k = 6. Therefore, k = 5 was selected, achieving high accuracy with substantially fewer parameters.

We then obtained overlapping 5-mers of each protein amino acid sequence. The k-mer frequencies for each cluster were defined by the count of the k-mer in the cluster divided by the total number of proteins in the cluster (note that frequency can be > 1 if k-mers appear multiple times on average in the proteins of a cluster). The underlying model is then stored in a two-level hash map where the top-level map stores for each k-mer , a map of clusters containing , and the second level maps the clusters containing , to the frequency of in , (S1 Methods, Algorithm 1). The resulting nested hash map can be written as .This precomputed model, consisting of 2,574,615 unique 5-mers (all the naturally occurring AA 5-mers in the underlying protein cluster database), is then used to determine the cluster to which a query sequence is assigned. The map is distributed along with the kMermaid package and is implemented as the default frequency model.

Assigning protein maps to reads using the pre-computed k-mer model

To assign a read into a protein cluster, a six-frame translation is applied to a query sequence and translations containing a stop codon (truncated frames) are discarded. Next, all overlapping AA 5-mers are extracted from the non-truncated frames. A score representing the strength of a match between and each cluster is then calculated by the summation of the precomputed model k-mer frequencies for each k-mer in the query sequence. The score for query sequence and each cluster C is computed as:

where is the model frequency of each k-mer w from in cluster C, i.e., the average occurrence of w in proteins of C. Finally, the sequence is then assigned to the cluster with the global maximum score across all clusters (Fig 1; S1 Methods, Algorithm 2). kMermaid annotates the query sequence by the cluster representative for the cluster corresponding to this maximum assignment score. This scoring approach effectively assigns higher confidence to scores when k-mers within a query sequence are more frequently observed in a cluster. kMermaid uses a k of 5 AA chosen via hyperparameter search, as described previously, for both the base model construction and the assignment procedure described previously and reports assignments with an assignment score >3.

Evaluating the clusters of functionally similar microbial proteins

As the protein clusters lie at the base of kMermaid’s approach, we validated their correctness using two orthogonal analyses aimed at verifying that the clusters produced contain homologous proteins with shared biological function.

1. Compatibility with NCBI protein clusters. To demonstrate kMermaid’s ability to construct biologically relevant clusters, we compared the results of the two-step CD-HIT clustering to the datasets from the NCBI Protein Clusters [28], which groups together proteins by sequence similarity. A subset of 102,380 of the proteins contained in our expanded cluster model was also clustered through the prokaryotic PCLA protein clusters dataset within this database. Proteins that were in this overlapping subset and were also in a non-singleton kMermaid cluster were used for comparison (N = 102,367, mapped to 9,984 kMermaid clusters). For each of these proteins, we evaluated their tendency to be co-clustered with the same proteins in both PCLA and kMermaid clusters by computing the percent of co-clustered proteins by kMermaid clustering that were also co-clustered in PCLA. The number of kMermaid clusters was chosen as the denominator for this evaluation metric to verify the correctness of kMermaid clusters rather than to assess its ability to maximize clusters, which is not an objective of this approach.
2. Within-cluster keyword similarity. High-throughput text analysis was performed on the protein name annotations to further investigate the similarity and homogeneity of the clusters. Trends and frequencies of word presence in clusters were used to evaluate cluster functional homogeny. After removing ubiquitous and generic words (e.g., “bacteria” or “protein”), we computed the frequency of the most common keyword found in each cluster, i.e., the fraction of proteins in the cluster containing the most common keyword in that cluster. Most clusters demonstrated a common keyword frequency of 1, indicating that our clusters are highly homogenous in key functions.

Benchmarking against established protein mapping tools

We benchmarked kMermaid against popular methods that can be used for protein mapping—BLASTX, DIAMOND, and MMseqs2. Each method was run against an underlying database containing the same 1,797,426 RefSeq protein sequences described previously. For consistency, we used the default or recommended configurations of each method. BLASTX and DIAMOND in BLASTX mode were run with e-values of 1e-4 and 1e-3 (default), respectively. MMseqs2 was run with a min-length set to 16 and e-value set to 1e-4. Each method except for BLASTX was set to report the default number of matches based on the e-value. BLASTX was set to report a maximum of 1 match for simulated data and 5 matches for experimental data due to its excessive running time when reporting all matches. DIAMOND was additionally run with a maximum of 1 match to demonstrate the consistency of mapping when using higher level groups rather than exact matches as well as to provide a fairer comparison resource benchmarking.

Analysis of unseen RefSeq microbial protein data

Since kMermaid was first implemented in 2023, we were able to test the ability of our tool to classify unknown microbial proteins by using 22,435 annotated protein sequences deposited in RefSeq between January 2025 and May 2025 (obtained May 2025). Because our method uses nucleotide reads as input, we had to perform a reverse translation of the amino acid sequences. If an amino acid reverse mapped to multiple tri-nucleotide sequences without a stop codon, a tri-nucleotide sequence was selected at random, which created an even more challenging classification task. We also removed sequences with missing or non-standard amino acid sequences at this stage. We then created datasets containing randomly selected 125, 150, and 200 base pair substrings of the reverse translated sequences with 10 different seeds each. We then ran BLASTX with max_target_seqs = 5 and evalue = 0.0001 and kMermaid with default configurations. Spearman correlation was computed on all BLASTX and kMermaid results; datapoints were down sampled to 150,000 across all 30 datasets for visualization (S2a Fig). We additionally performed post-hoc filtering of BLASTX results using percent identity ≥66.6 to obtain higher confidence hits for comparison with kMermaid. Because the reference databases for each method did not contain the truth assignment, we considered a correct hit, or true positive, to be a sequence for which the BLASTX and kMermaid protein map matched. We computed AUROC using the kMermaid scores for each read assigned by both BLAST and kMermaid. We then the kMermaid scoring threshold at which the false positive rate fell below 0.05 for each read length. These thresholds were used for subsequent comparisons with BLASTX (S2 Table). We note that while kMermaid assignment scores generally correlate with higher confidence for data of consistent read length, the specific scores thresholds reported are likely specific to this analysis given the cluster-specific trends and biases in recently deposited protein sequences, i.e., over-representation of specific proteins.

Protein mapping of reads from fecal samples from ulcerative colitis patients

The goal of kMermaid is to efficiently map reads to proteins, while maintaining the accuracy and sensitivity of BLASTX. We benchmarked kMermaid’s functional read classification against BLASTX using paired end reads from ulcerative colitis patients enrolled in the LOTUS fecal matter transplant clinical trial [29] available on the NCBI’s Sequence Read Archive (SRA). Two samples were excluded from analysis based on data incompletion (low read depth). Due to the infeasibly long running times incurred by BLASTX, we randomly subsampled each fasta file to 100,000 reads. We used a small, representative subset (n = 3) of the available samples for in-depth follow-up analyses of all reads in the samples. BLASTX was set to max_target_seqs = 5, evalue = 1e-4 and to max_target_seqs = 3, evalue = 0.01 for these analyses, respectively. We additionally filtered BLASTX results at percent identity >66.6% where specified. We compared the overall coverage, defined as the percentage of reads that were able to be classified by both BLASTX and kMermaid, as well as the percentage of BLASTX hits that kMermaid was able to classify. We further investigated the correctness of kMermaid’s assignments using BLASTX results as a gold standard for correct read assignment, by examining reads which were assigned by both methods. Correct assignment by kMermaid was defined for reads as a non-empty overlap between proteins in the BLASTX hits and the assigned kMermaid clusters.

Cluster-specific results and remote homology detection

To evaluate kMermaid’s performance for specific biological functions, we calculated the accuracy across clusters with similar functional annotations. We used the LOTUS clinical trial data to evaluate function specific performance in a real metagenomic sequencing cohort. To this end, for every cluster we calculated the ratio of reads correctly assigned to that cluster (i.e., assigned to that cluster by BLASTX and kMermaid), out of all the reads assigned to that cluster by kMermaid. Then, we evaluated the distribution of cluster performances for distinct functions, i.e., clusters named with common keywords (S3 Table).

To evaluate kMermaid’s ability to identify sequences with remote homology to proteins in the database, we examined reads that were not classified by BLASTX from the LOTUS clinical trial data and were assigned with high kMermaid assignment scores (>20). We explored reads that were confidently assigned to six clusters that failed to be classified with BLASTX, and carefully verified that these reads have remote homology to their kMermaid assigned clusters using PSI-BLAST and HHblits3 from HH-suite3 [27] (S1 File).