Copy-Number Pitch: Grouping 16S Variants into Genomes by Integer Abundance Ratios

Most bacteria carry several copies of the 16S ribosomal RNA gene, the standard barcode used to identify microbes. These copies are often not identical. When a community is sequenced, one organism therefore shows up as several distinct "exact sequence variants," and its abundance is counted several times over. The usual response is to throw away resolution, either by clustering similar sequences or by trimming to a short region where copies look alike. Yet Johnson et al. (2019) show that full-length 16S sequencing resolves these intragenomic copies, and that the exact combination of variants inside one genome is a strain fingerprint. It separates, for example, harmless E. coli K-12 from pathogenic O157:H7. The open problem is to recover those fingerprints from a community without reference genomes: which variants belong to the same genome, and in what copy numbers?

We propose to treat this as a pitch-estimation problem. A voiced sound puts energy at integer multiples of one fundamental frequency. In the same way, a genome puts reads at its variants in integer multiples of one "per-copy abundance": a strain with five copies of variant A and two of variant B yields A and B counts near a 5:2 ratio in every sample it appears in. Speech processing already has robust tools for finding a shared fundamental under noise: spectral autocorrelation, conditioning of the spectrum, accumulation over time frames, and tolerance of a missing fundamental. We adapt them to find, across samples, groups of variants whose counts share a common integer-multiple structure. Each such group is read as one genome, with its copy numbers.

Core Claim

Current practice uses sequence similarity to decide which 16S variants to merge. That fails in two directions. Copies within one genome can differ by several substitutions and so stay separate, while variants from distinct but closely related strains can differ by just as much and get wrongly merged. Our claim is that abundance structure carries information that sequence similarity lacks. Variants from the same genome keep a fixed small-integer count ratio across every sample, because they are physically linked on one chromosome. Variants from ecologically co-occurring but distinct organisms can be strongly correlated, but their ratios drift and are not integer-locked. A harmonic estimator that looks for shared fundamentals across samples should therefore beat both similarity clustering and plain co-abundance clustering at recovering genome-level units and copy-corrected abundances.

The generative picture is simple. Let $y_{s,v}$ be the read count of variant $v$ in sample $s$, $a_{s,g}$ the per-copy abundance of genome $g$ in sample $s$, and $c_{g,v}\in\{0,1,\dots,15\}$ the number of copies of variant $v$ in genome $g$. Then

$$ \mathbb{E}[y_{s,v}] = b_v \sum_g a_{s,g}\, c_{g,v}, $$

where $b_v$ is a per-variant amplification bias that we assume is constant across samples. The integer constraint on $c_{g,v}$ is the "harmonic" structure, and $a_{s,g}$ plays the role of the fundamental. Two consequences follow from Poisson counting noise. First, telling copy number $c$ from $c+1$ requires relative error well below $1/(2c)$, so the per-copy count $a_{s,g}$ must be well above roughly $4c$. This sets a sequencing-depth threshold we can test directly. Second, the constant bias $b_v$ distorts integer ratios but does not make them vary across samples. It can therefore be calibrated on a mock community rather than silently breaking the method.

Research Program

The method has three stages. Each borrows a specific mechanism from pitch tracking.

The first stage builds a per-sample "abundance spectrum." For a candidate set of variants (neighbors within a sequence-distance cutoff generous enough to contain intragenomic divergence), we place impulses along the count axis at each variant's bias-corrected count. We then compute an autocorrelation along that axis, using a harmonic-sieve variant that scores how well all counts fall on multiples of a candidate fundamental. This is the direct analogue of spectral autocorrelation in Rahman and Shimamura (2010), who shift a conditioned amplitude spectrum against itself and read pitch from the lag of the first major peak. We also reuse their conditioning chain. Clipping at a low fraction of the maximum removes low-count error variants, much as their 10–15% clipping removes spurious spectral peaks. Their flattening step becomes a variance-stabilizing transform, so that high-copy variants do not dominate the score. Following Zahorian et al. (2006), who recover pitch from the spacing of higher harmonics when the fundamental itself is absent, we do not require any variant to be at copy number one. A genome with only 2- and 3-copy variants must still be detectable, and a naive greatest-common-divisor rule would fail on it.

The second stage accumulates across samples. Lee et al. (2010) sum spectral peaks over consecutive frames so that random noise peaks cancel while stable harmonic combs reinforce. Our "frames" are samples. After normalizing each sample's candidate by its estimated fundamental, the integer copy pattern $c_{g,\cdot}$ should stack across samples. Ecologically correlated but unlinked variants, whose ratios wander, should blur out. The accumulated salience of a candidate group is

$$ S(G) = \sum_s w_s \max_{a>0} \sum_{v\in G} \exp\!\left(-\frac{(\tilde y_{s,v}/a - \mathrm{round}(\tilde y_{s,v}/a))^2}{2\sigma_{s,v}^2}\right), $$

where $\tilde y_{s,v}$ is the bias-corrected count, $\sigma_{s,v}$ its Poisson-derived tolerance on the multiple scale, and $w_s$ down-weights samples where the group is near the depth threshold. Only groups whose salience exceeds a null built from shuffled cross-sample pairings are kept. The search is greedy, as in multi-pitch estimation: we extract the most salient group, subtract its explained counts, and repeat.

The third stage handles overlapping sources. Two strains of one species often share some variants, so a shared variant's count is a sum of two combs. This is the situation Kurth (2016) addresses with the spectral shift-autocorrelation, which separates overlapping repetitive animal calls where the classical autocorrelation fails. We use the extracted fundamentals and copy patterns to initialize a small integer-constrained non-negative factorization of the equation above, which jointly refines $a$, $c$, and $b$ and splits shared variants between strains. The output is a set of inferred genomes, each with an integer copy fingerprint, and per-sample abundances in cells rather than in gene copies.

Baselines

The naive reference point treats every exact sequence variant as its own taxon, the default in modern amplicon pipelines. It inflates richness by the number of distinct copies per genome and weights abundance by copy number. Beating it means lower richness error and lower abundance error against truth. That result is almost guaranteed, so this baseline mainly calibrates the scale of the problem.

The similarity baseline clusters variants into operational taxonomic units at 97% identity and at stricter full-length thresholds, sweeping the threshold on the same data and reporting the best one. This is how the field currently collapses intragenomic copies. It is the right competitor for the question "does abundance add anything beyond sequence?" Beating it means higher agreement between inferred groups and true genomes at every threshold, especially where two strains of one species are present.

The strongest competitor for the specific claim is co-abundance clustering. Here we group variants by the proportionality of their counts across samples, using log-ratio variance within the same sequence-distance neighborhoods and the same depth filtering, and tune the clustering threshold with a budget equal to ours. It uses the same cross-sample signal we do, minus the integer constraint. Beating it is the core evidence for the proposal.

The key ablation runs our full pipeline with the integer rounding replaced by a continuous ratio, keeping conditioning and accumulation. It also runs the integer factorization with random initialization instead of harmonic initialization. These isolate whether the gain comes from the integer prior, from the harmonic search, or both.

We also report reference fingerprint matching in the style of Johnson et al. (2019) as an upper bound when genomes are in the reference set. For novel strains it is inapplicable, which is the regime we target. For the grouping task itself, we found no reported result on Lacuna for grouping 16S variants into genomes with copy numbers, so all baseline numbers will be produced here under matched conditions.

Why not just…?

Why not just use correlation-based co-abundance clustering? This is the real threat. With many samples and deep sequencing, linked variants will be near-perfectly proportional, and proportionality alone may suffice. The integer prior should matter in three situations: few samples, closely related strains with correlated dynamics, and splitting shared variants between strains, since a real-valued ratio cannot be uniquely decomposed into two combs. The continuous-ratio ablation settles this directly. If it matches the full method across sample counts, we will report that integer structure adds nothing, and the contribution reduces to copy-number calibration.

Why not just map variants to reference genomes with known operon sequences? Where references exist, this is better, and we include it as an upper bound. But Johnson et al. (2019) themselves note that curated 16S databases underrepresent intragenomic variation, and uncultivated strains have no reference at all. Our held-out-genome experiment measures exactly the case where matching has nothing to match.

Why not just train a deep sequence model to classify reads? Read2Pheno (Zhao et al., 2021) reaches 95.9% genus accuracy on SILVA with a convolutional, recurrent, and attention network. Woloszynek et al. (2018) get 0.962 balanced body-site accuracy from word2vec embeddings of k-mers. Both assign each read to a label independently. Neither can tell that two different sequences come from one chromosome, because that fact lives in cross-sample count ratios, not in any single read. They are complementary: our copy-corrected genome units could serve as their inputs.

Why not just Fourier-transform the nucleotide sequence itself? This is the obvious "harmonics meets DNA" move, and we deliberately avoid it. As the objections compiled in our search note, there is no evidence that 16S sequences carry evenly spaced periodic combs. A position-free spectrum also discards the site-specific substitutions that distinguish strains, which k-mer profiles already capture. The harmonic structure we exploit is physical, since copies are integers, rather than imposed by a numeric encoding.

Proposed Experiments

The first experiment uses the synthetic 36-species mock community and the cyclist stool samples of Johnson et al. (2019), sequenced with full-length PacBio circular consensus reads. In the mock community every genome and copy pattern is known, including the K-12 versus O157:H7 contrast. We use part of the mock community to estimate the bias $b_v$ and evaluate on the remainder. Metrics are the adjusted Rand index between inferred and true variant-to-genome assignments, the exact-match rate of copy fingerprints, and Bray–Curtis distance between inferred and true cell abundances. The decisive check is whether K-12 and O157:H7 come out as two groups with correct fingerprints without reference matching. On stool, where truth is partial, we check whether inferred copy-corrected abundances agree better with the shotgun metagenomic profiles in the same study than raw variant abundances do. We also check whether Bacteroides intestinalis, which that study found only with full-length reads, is recovered as a single genome unit.

The second experiment is a simulation built from complete bacterial genomes with annotated 16S operons. We create communities over controlled numbers of samples (5 to 200), sequencing depths, ecological correlation between strains, and fractions of genomes withheld from any reference. We add per-variant amplification bias and homopolymer indel errors, since Johnson et al. report these as the characteristic long-read failure. This is where the decisive ablation lives. We plot grouping accuracy against the number of samples and against strain relatedness for the full method, the continuous-ratio ablation, and co-abundance clustering. We predict the integer prior wins most at few samples and among co-varying congeneric strains. We will also test the predicted depth threshold: copy-number accuracy should collapse when the per-copy count falls below a small multiple of the copy number, as the Poisson argument above implies.

The third experiment asks whether the output matters downstream. On a longitudinal or multi-sample full-length dataset, we compare diversity estimates and differential-abundance calls computed from raw variants, clustered units, and our genome units. The question is whether copy correction changes which taxa appear to shift. The proposal fails if, in the simulation, the continuous-ratio ablation matches the full method across sample counts and strain relatedness, or if the mock-community bias is not stable enough across runs for integer ratios to survive calibration. Both outcomes would be clear negative results about how rigid copy-number structure is in real amplicon data.