During 2015–2021, we collected clinical samples from infected children and nasal swab samples from healthy children in southern China (Guangxi and Guangdong Provinces). From hospitalized infected children, we collected 349 nonrepetitive pneumococcal isolates (e.g., blood, bronchoalveolar lavage fluid, sputum, middle ear fluid), of which 342 were noninvasive and 7 invasive. The eligibility criteria for infected children were having clinical infectious manifestations such as cough, respiratory secretions, abnormal lung sounds, dyspnea, or fever >38°C, with or without infiltrates seen on chest radiographs; having S. pneumoniae infection diagnosed by clinical doctors on the basis of signs and symptoms; and having S. pneumoniae isolated from clinical infection sites. In terms of asymptomatic carriage isolates, we sampled 434 isolates from healthy children in kindergarten.

We performed high-throughput genome sequencing on a Hiseq 2000 machine (Illumina, https://www.illumina.com) to obtain paired-end 150-bp reads. We assessed the quality of the raw sequenced reads by using FastQC version 0.11.5 (https://github.com/s-andrews/FastQC) and trimmed for low quality reads and adaptor regions by using Trimmomatic version 0.36 (https://github.com/usadellab/Trimmomatic). We then assembled trimmed reads by using SPAdes version 3.6.1 (https://github.com/ablab/spades). We used PathogenWatch (https://pathogen.watch) to predict global pneumococcal sequencing cluster (GPSC), multilocus sequence typing (MLST), and serotyping for all genomes.

To generate the variant sites with SNPs, we mapped assembled contigs to a standard reference genome S. pneumoniae R6 by using Snippy version 4.4.5 (https://github.com/tseemann/snippy). We used the generated core SNP alignment to construct a maximum-likelihood phylogenetic tree by using the generalized time reversible plus gamma model and 100 bootstrap replicates with FastTree version 2.1.10 (http://www.microbesonline.org/fasttree). We visualized and annotated the phylogenetic tree by using ChiPlot (https://www.chiplot.online).

We scanned all k-mers that were 9- to 100-bp long from all assembled reads by using fsm-lite (https://github.com/nvalimak/fsm-lite) and filtered them to obtain 10,591,337 k-mers seen on 1%–99% of the total samples. To identify the relevant genes by using BWA-MEM (the Burrows-Wheeler Aligner with maximal exact matches alignment tool, https://github.com/lh3/bwa), we mapped all k-mers to 10 S. pneumoniae reference genomes (CGSP14, D39, Hungary^19A-6, R6, Taiwan^19F-14, TIGR4, Spain^23F-ST81, ATCC 49619, EF3030, and MDRSPN001) obtained from the Virulence Factor Database (http://www.mgc.ac.cn/VFs) and previous studies. We determined gene ontology annotations by using the UniProt (https://beta.uniprot.org).

To explore the genomewide associations between genetic elements (k-mers) and S. pneumoniae disease status (infection or carriage), and thus to identify infection-associated k-mers, we used GWAS methods. Because of the high-dimensional genomic data structures, we used multiple GWAS methods: the linear mixed model (LMM; (https://github.com/mgalardini/pyseer), phylogenetic-based approach (Scoary; https://github.com/AdmiralenOla/Scoary), variable selection using random forests (VSURF; https://github.com/robingenuer/VSURF), and least absolute shrinkage and selection operator (LASSO; https://scikit-learn.org/stable/modules/generated/sklearn.linear_model.Lasso.html) regression.

In brief, we used a 2-stage analysis process to detect the infection-associated k-mers by comprehensive GWAS analyses (Figure 1). First, we fitted a univariate LMM to initially screen infection-associated k-mers by using the Pyseer tool (version 1.3.10) (15). To correct for the population structure, we used the similarity pyseer command of Pyseer, which computes a similarity kinship matrix on the basis of the core genome SNPs. For covariates, the GWAS analysis used host age (years) and sex. Second, we used multiple methods (Scoary, LASSO, and VSURF) to minimize false-positive associations and identify consensus infection-associated k-mers by Venn diagram. In the GWAS analyses, we used the Bonferroni correction (α/N) to control for false-positive rates resulting from multiple comparisons of 1,418,815 k-mers (adjusted p value threshold 3.52 × 10⁻⁸). Scoary is an ultrafast software tool for GWAS analyses that uses a phylogenetic-based method to adjust population structure. The LASSO regression is suitable for high-dimensional data structures, and the coefficients of nonrelevant variables can be compressed to zero to solve the problem of model overfitting (16). We used VSURF, based on random forest (RF), to perform a 2-step feature selection on the variables (17). Initially, VSURF ranks the variables according to the importance measure by using the RF permutation-based score of importance to obtain a subset of important variables, and then it uses a stepwise forward strategy for variable introduction based on the smallest out-of-bag error. More precisely, a variable is added only if the error decrease is larger than a threshold. We ranked the importance of k-mers by the mean decrease in impurity (mean decrease Gini), which is a measure of the predictor’s contribution to the correct sample classification. We compiled associated phenotype data for all 783 isolates (Appendix Table 1) and deposited sequences in the National Center for Biotechnology Information Sequence Read Archive database (https://www.ncbi.nlm.nih.gov/sra; projection no. PRJNA976286). The k-mer sequences and output results files from several GWAS analyses are publicly available (https://doi.org/10.6084/m9.figshare.24466606.v3).

Of the 349 children with S. pneumoniae infection, 342 (98.0%) had noninvasive disease (264 pneumonia, 49 bronchitis, 13 otitis media, 9 upper respiratory infection, 6 nasosinusitis, and 1 corneal ulcer), and 7 (2.0%) had invasive disease (6 bacteremia and 1 endocarditis). χ² test results indicated no differences between infection and carriage isolates with regard to host sex (p = 0.359) but significant differences with regard to age (p<0.001) (Appendix Table 2).

The most prevalent GPSCs for infection isolates were GPSC1 (45.9%), GPSC321 (9.2%), and GPSC852 (5.4%); the predominant GPSCs for carriage isolates were GPSC321 (16.1%), GPSC1 (15.4%), and GPSC23 (15.0%). In terms of sequence types (STs), the most common genotypes for infection isolates were ST271 (29.2%), ST320 (9.5%), and ST902 (7.2%); the predominant genotypes for carriage isolates were ST902 (15.9%), ST90 (13.8%), and ST271 (8.5%). The most prevalent serotypes for infection isolates were 19F (43.0%), 6B (15.2%), and 23F (8.3%); and the predominant serotypes for carriage isolates were 6B (32.7%), 19F (13.1%), and 15A (11.1%). The results indicated potential genotype differences between infection and carriage isolates. In addition, the phylogenetic tree based on core SNPs revealed that several genotypes (GPSCs/STs/serotypes) from infection and carriage isolates clustered in the same branches (Figure 2). Moreover, we found statistically significant differences in the proportion of specific GPSCs/STs/serotypes between infection and carriage isolates (Table 1), indicating that these isolates are associated with infection.

We identified 10,591,337 k-mers from the assemblies of 783 S. pneumoniae isolates and then filtered out low-frequency k-mers for a reduced matrix with 1,418,815 k-mers. Using those k-mers for GWAS, we performed a univariate LMM analysis to initially identify 22,790 infection-associated k-mers; 10,713 k-mers were successfully mapped to 1,215 unique genes (Figure 3, panel A; Appendix Figure 1). In the initial model with 10,713 k-mers, we used the RF model to assess the prediction effect. The classification balanced accuracy based on cross-validation was 93.60% (95% CI 91.48%–95.72%) (Table 2); the area under the curve (AUC), based on the out-of-bag risk scores of the classifier, was 0.98. In the LMM analysis, the QQ-plot indicated that population structure was well controlled at low p values (p<0.01) (Appendix Figure 2). Because of the considerable redundancy among the genetic elements in risk prediction, studying all k-mer combinations had little benefit; therefore, we used a simpler model with 886 k-mers successfully mapped to 52 antibiotic resistance or virulence genes (Appendix Table 3). The classification balanced accuracy was 91.28% (95% CI 89.34%–93.22%) (Table 2); the AUC was 0.96, suggesting that the power of these 886 k-mers for predicting disease status was close to that of the model with 10,713 k-mers.

In addition, we sorted the 886 disease-associated k-mers according to estimated importance (Figure 3, panel B). The k-mers were mainly associated with antimicrobial resistance (34%), adherence (20%), immune modulation (17%), and exoenzyme (1%). Moreover, the k-mers were divided into 3 functional gene ontology categories. Among those categories, proteolysis and cell wall organization were the largest subcategories in the biological process, membrane was the most enriched term in the cellular component, and serine-type D-Ala-D-Ala carboxypeptidase activity was the top term in the molecular function (Figure 3, panel C).

To reduce the complexity of the model, we used 3 methods to identify consensus infection-associated k-mers (Figure 4). On the basis of the 886 k-mers screened above, we observed consensus on genomewide statistically significant associations for pathogenicity k-mers; 8 k-mers were identified by all 3 methods. When we used the simplest model with the 8 k-mers, the classification balanced accuracy was 90.89% (95% CI 89.48%–92.31%) (Table 2), and the AUC value was 0.93 (Figure 5, panel A), suggesting that the power of the 8 k-mers to predict disease status was comparable to that of the model with 886 k-mers. Of note, the k-mer predictors still exhibited high classification balanced accuracy in the predominant GPSCs (95.34% for GPSC1 and 92.79% for GPSC321). The importance of the selected k-mers in the final model indicated that these predictors were mainly associated with adherence function (Figure 5, panel B). The highest ranked predictor (Kmer_9823 in sortase [srtG1]) achieved a classification accuracy of 79.57% on its own and also showed high classification accuracy in the predominant GPSCs (70.04% for GPSC1 and 85.29% for GPSC321). In addition, the best predictor (in srtG1) was associated with GPSC1 and GPSC321 (all p<0.05). For the additional validation analysis that used the best RF classifier k-mer (in srtG1), 2 independent datasets of S. pneumoniae genomes with genotype distribution similar to that of our study were available on the National Center for Biotechnology Information Assembly database (https://www.ncbi.nlm.nih.gov/assembly (data1: 60 noninvasive vs. 60 carriage isolates; data2: 60 invasive versus 60 carriage isolates; the prevalence of the predominant GPSCs [GPSC1 and GPSC321] was 58.3% for noninvasive, 55.0% for invasive and 30.0% for carriage isolates) (Appendix Table 4). Classification accuracy was 75.83% for data1 and 74.17% for data2, similar to that in the larger primary dataset in our study.

The proportion of k-mers differed significantly between infection and carriage isolates (all p<0.05) (Figure 5, panel C), indicating that the proportion of k-mers was substantially higher in infection isolates than in carriage isolates. The effect of each k-mer on the estimated risk score (Figure 5, panel D), indicated by a point above the diagonal, indicates that the risk score is increased when the k-mer profile is present. The presence of k-mers associated with adherence genes markedly increased the risk for S. pneumoniae infection (odds ratio [OR] 1.88 for Kmer_9823, OR 1.65 for Kmer_10039, and OR 1.69 for Kmer_10431) (Table 3).