K-mer Conservation Patterns Reveal Evolutionary Selection in Genomes

Published in iScience, the study presents a new metric called the kGain score, which captures the enrichment of specific k-mers near single-nucleotide variants (SNVs). By analyzing long-term evolution experiments in Escherichia coli and Saccharomyces cerevisiae, as well as natural evolutionary data from SARS-CoV-2, the team provides evidence that variant selection correlates with systematic shifts in genome-wide k-mer frequencies. This work suggests that intra-genomic conservation patterns represent a previously underappreciated layer of evolutionary selection that complements traditional cross-species conservation models.

The central goal of evolutionary genomics is to understand how genetic variation translates into phenotypic consequences and adaptive fitness. While early variant effect prediction methods relied on evolutionary conservation signals from cross-species alignments, recent deep learning frameworks have advanced the field by capturing complex patterns of selection in protein-coding regions. However, these techniques often focus on interspecies conservation and may overlook subtle intra-genomic patterns of selection that govern the organization and evolution of the entire genome within a species. The study posits that analyzing k-mer frequencies within a single genome offers a complementary window into how selection shapes genomic architecture, reflecting local selection pressures acting on various genomic elements.

To test this hypothesis, the researchers developed the kGain score, which quantifies the relative abundance of k-mers associated with alternative alleles compared to reference alleles. For each variant, two sequences are generated: one with the reference allele and one with the alternate allele at the central position. The kGain score is calculated by summing the natural log of the fold change between the genomic frequencies of the k-mer containing the alternate allele and the reference allele across all k-mers. This metric allows for the assessment of whether local sequence composition influences the retention of genetic variants across diverse evolutionary landscapes, including laboratory evolution, natural viral adaptation, and controlled antibiotic stress.

They categorized mutations into three classes based on allele-frequency trajectories: positively selected, negatively selected, and neutral. The analysis revealed that mutations under directional selection were associated with higher kGain scores compared to both neutral and purifying selection mutations. Furthermore, a gradual increase in the abundance of alternate k-mers was observed over time in both mutator and non-mutator backgrounds, reflecting the progressive accumulation of mutations associated with high-frequency alternate k-mers under differing mutation rates.

The study also examined the relationship between kGain scores and the functional impact of mutations. Essential genes, which are generally resilient to mutations due to their critical role in organismal survival, preferentially accumulated variants with higher kGain scores. Fixed mutations, defined as those persisting over multiple generations, consistently exhibited higher kGain values than non-fixed mutations. This pattern suggests that mutations which reach fixation are typically under stronger selection, consistent with the expected outcomes of adaptive evolution. Additionally, parallel genes, which are recurrently mutated across independent populations, showed elevated median kGain scores, indicating that recurrent adaptive evolution may be facilitated by favorable motif architectures.

To validate the framework in a eukaryotic system, the researchers performed a meta-analysis of a genomic sequencing dataset from a Saccharomyces cerevisiae long-term evolution experiment. They computed kGain scores for 133,538 substitutions across 90 populations under study. The results showed a gradual increase in both median fitness and median kGain scores over successive generations. Essential genes consistently exhibited higher kGain scores than non-essential genes across all generations, underscoring the greater evolutionary constraint these genes face in maintaining functional integrity. Fixed mutations were also assigned higher kGain scores compared to those not fixed, reinforcing the link between k-mer enrichment and evolutionary success.

The applicability of the kGain score to natural viral evolution was assessed using global genomic surveillance data from SARS-CoV-2. The team analyzed approximately 0.5 million high-quality genome sequences from the GISAID database, yielding around 54,000 unique single-nucleotide polymorphisms. They observed a significant contrast between control variants and Variants of Concern (VOCs) and Variants of Interest (VOIs), with the latter exhibiting substantially higher kGain scores. This difference was particularly pronounced within accessory and structural genes. Variants with kGain scores greater than or equal to zero were found to be present in considerably higher numbers of COVID-19 genome sequences, suggesting that the score could indicate the type of selection pressure experienced by a mutation.

Direct functional validation was provided through a single-colony bottleneck passage experiment in E. coli under sublethal antibiotic pressure. Wild-type E. coli K-12 MG1655 lineages were subjected to five successive serial passages in incrementally increased kanamycin concentrations. In one adaptive lineage, a Y515N substitution in the fusA gene, which encodes elongation factor G, emerged and fixed by the fifth passage. This mutation, characterized by a high kGain score, conferred kanamycin resistance without compromising growth fitness. In contrast, fusA variants with lower or negative kGain scores in other populations provided only partial resistance and were associated with diminished growth. These results demonstrate that k-mer-based selection biases can have tangible fitness consequences under controlled selection regimes.

To further explore the computational predictability of kGain, the team trained an attention-based neural network to predict kGain scores from sequence windows centered on the mutation site. The model achieved a Pearson’s correlation coefficient of 0.81 in-domain and 0.82 out-of-domain, indicating that kGain reflects a consistent compositional property of mutation context rather than a strain-specific frequency artifact. The attention maps revealed that the model focuses most strongly on the mutation site while also considering neighboring bases, consistent with the rolling window approach used for kGain calculation. This suggests that kGain can be approximated without exhaustive full-genome frequency computation, opening the door for applying kGain inference across species using learned representations.

The study also applied the kGain metric to deep mutational scanning data for three essential E. coli genes: murA, fabZ, and lpxC. The analysis showed that tolerant mutations, which are less likely to impair protein function, were associated with higher kGain scores compared to intolerant mutations. Residues with high relative solvent accessibility, indicating surface-exposed positions, frequently aligned with higher kGain values. This suggests that surface-exposed residues may be more permissive to beneficial substitutions in enriched k-mer contexts. Although kGain alone is not a substitute for functional assays, it captures a meaningful dimension of mutational tolerance that may complement structural and biochemical predictors.

The authors acknowledge several limitations of the study. The kGain score is computed relative to a single reference genome per organism, and results may differ for organisms with high intra-species sequence diversity. The analyses are restricted to single-nucleotide substitutions, and the interaction of insertions, deletions, or horizontally transferred elements with k-mer-based compositional selection remains an open question. Furthermore, while the parallel between kGain and thermodynamic effective-energy landscapes is conceptually appealing, this connection remains one of analogy rather than an established mechanism. The kGain score is an empirical measure derived from k-mer frequencies and should not be conflated with a direct measurement of physical free energy or duplex stability.

Collectively, these insights propose an expanded view of molecular evolution that considers the potential role of k-mer frequencies within the same genome as a critical layer of selection. By tracking k-mer dynamics through the kGain metric, the study establishes a new avenue for identifying variants that may be key to adaptation. As genome sequencing efforts broaden across diverse organisms and ecological contexts, this framework holds promise for enhancing our understanding of how genomic architecture evolves under selective pressures, complementing the well-established narratives of cross-species conservation and protein-focused evolutionary dynamics.

Subject of Research: Intra-DNA k-mer conservation patterns encode evolutionary selection of variants

Article Title: Intra-DNA k-mer conservation patterns encode evolutionary selection of variants

Article References: Mathew, B., Halder, A., Jaiswal, N., Panda, S., Pramanik, D., Peela, S. C. M., Garg, A., Tripathi, S., Samanta, S., Gupta, P., Malhotra, V., Ahuja, G., & Sengupta, D. (2026). Intra-DNA k-mer conservation patterns encode evolutionary selection of variants. iScience, 29(11), Article 118142. https://doi.org/10.1016/j.isci.2026.118142

Image Credits: AI Generated

DOI: 10.1016/j.isci.2026.118142

Keywords: Intra-DNA, k-mer, conservation, patterns, encode, evolutionary, selection, variants, scientific research