Applying PCR cycle autonormalization to PacBio full-length 16S rRNA library preparations: impacts on error rates and sequence distributions

ABSTRACT

The bacterial 16S rRNA gene is widely used to characterize host-associated and environmental microbiomes, most commonly through sequencing short hypervariable regions. Recent improvements in PacBio sequencing chemistry and concatenation approaches can now enable high-throughput, full-length 16S rRNA gene sequencing with high accuracy and depth. However, errors introduced during library preparation remain a major limitation, particularly during PCR amplification of full-length amplicons, where error accumulation may be elevated due to longer sequence lengths. These challenges are amplified when samples vary widely in microbial biomass, making it difficult to select a single optimal number of PCR cycles. Here, we evaluated PCR cycle autonormalization for PacBio Kinnex full-length 16S rRNA gene sequencing across seven agriculturally relevant specimen types. We compared conventional fixed-cycle PCR protocols (20, 24, and 30 cycles) with an autonormalization approach in which individual reactions were terminated during exponential amplification based on real-time fluorescence thresholds. Under the workflow tested here, autonormalized libraries generally retained a high proportion of sequences following denoising and chimera removal, exhibited low residual error rates (<0.005%), and yielded relatively even read distributions across heterogeneous sample inputs. Overamplified reactions (30 cycles) showed elevated residual error rates and greater sequence loss, particularly in samples with higher microbial biodiversity, whereas low-cycle libraries produced more variable read output among specimens. Importantly, the PCR protocol had relatively minor effects on overall community composition compared with specimen type. These results support PCR cycle autonormalization as a useful workflow strategy for heterogeneous full-length 16S library preparation, while also highlighting the importance of library design, pooling strategy, and downstream processing in shaping technical outcomes.

IMPORTANCE

Amplicon-based sequencing of the 16S rRNA gene is a foundational tool in microbiome research, yet PCR amplification remains a major source of library-preparation error. This challenge is magnified for full-length 16S rRNA sequencing and for workflows that process specimen types with widely varying microbial biomass. Selecting a single PCR cycle number can underamplify low-biomass samples or overamplify high-titer samples, increasing artifacts and sequence loss during downstream processing. Here, we show that PCR cycle autonormalization can be integrated into a PacBio full-length 16S rRNA workflow and, under the conditions tested, provides low residual error rates and relatively even sample representation across heterogeneous inputs. Autonormalization also enables blind pooling of amplicons without post-PCR quantification or equimolar normalization, reducing hands-on time and sample loss. These benefits make cycle autonormalization particularly valuable for high-throughput and production-scale library preparation applications handling diverse specimen types.

INTRODUCTION

Amplicon-based sequencing of the bacterial 16S rRNA SSU (16S) gene has greatly advanced the study of microbial communities across host-associated and environmental systems (1, 2). Most microbiome studies rely on short-read sequencing of partial 16S regions generated using sequence-by-synthesis (SBS) chemistries, which provide high per-base accuracy and scalability by taking advantage of the gene’s hypervariable regions for taxonomic classification (3–5). For some taxa, short fragments are insufficient to resolve closely related lineages, leading to ambiguous classification and reduced taxonomic precision (3, 6, 7). Sequencing near-full length 16S offers substantially improved taxonomic inferences and the potential to better resolve bacterial communities (8–16). Historically, adoption of long-read sequencing approaches has been constrained by lower read quality and throughput relative to short-read platforms. However, recent advances in PacBio circular consensus sequencing (CCS) have enabled the generation of highly accurate long reads (Q > 30, HiFi reads) (17, 18). In parallel, MAS-Seq concatenation strategies implemented through PacBio Kinnex protocols allow high-throughput full-length 16S sequencing at rates and costs comparable to traditional SBS-based short-read amplicon sequencing (19). These developments enable hundreds to thousands of samples to be pooled into a single sequencing run, increasing the importance of consistent library preparation.

Upstream from sequencing, PCR is a major source of bias and error in amplicon-based analyses and can create artifacts that influence 16S sequence data (20–22). Polymerase choice, thermocycle number, and template concentration all influence substitution errors, chimera formation, and amplification bias (23–25). Excessive thermocycling can damage templates and propagate errors into high-abundance artifacts that survive downstream quality filtering (23, 26–28). For full-length 16S sequencing, PCR-induced errors may be especially consequential because each read spans ~5–7× more bases than short-read amplicons, increasing the probability that errors lead to sequence loss during QA/QC and processing into amplicon sequence variants (ASVs) or operational taxonomic units (OTUs).

Mitigating excessive PCR cycles is particularly challenging when processing heterogeneous specimens. Microbial biomass and 16S gene copy number can vary by orders of magnitude between specimen types, making a fixed-cycle number suboptimal for mixed sample sets (27, 29). Underamplification reduces yield in low-biomass samples, while overamplification potentially increases error rates and chimera formation in high-biomass samples. Having knowledge of sample composition can aid in mitigating cycling-related issues, but this can be challenging in large-scale production practices. Most workflows involve blind receipt and sample processing such that specimen-specific protocols are potentially unrealistic.

In this study, we evaluated an implementation of PCR cycle autonormalization using iconPCR (where individual reactions are automatically terminated during exponential amplification based on real-time fluorescence thresholds) (28). By integrating iconPCR with full-length 16S microbiome analysis using PacBio Kinnex concatenation, we evaluated the robustness of this workflow on a range of samples. Using agriculturally relevant arthropod, plant, and soil specimens with varying microbial diversity and titers, we compared autonormalized PCR with defined-cycle protocols across multiple metrics of sequence quality, retention, and community inference (Fig. 1). We hypothesized that excessive PCR cycling would increase residual error and sequence loss during denoising, particularly in more diverse or high-titer samples, whereas insufficient cycling could adversely impact amplicon generation in lower-biomass samples. We further anticipated that sample-specific autonormalization would help mitigate these opposing constraints while preserving biological signal in long-read 16S sequencing.

Fig 1.

Schematic of experimental workflow. Cycling protocols varied between autonormalization and a priori designated protocols of 30, 24, and 20 cycles. Samples were barcoded using dual indices and pooled prior to concatenation and sequencing. The specimens tested were included to provide a range in putative bacterial biodiversity and were agriculturally relevant. We performed no upstream QA/QC of sample types to simulate blind sample receipt and processing.

MATERIALS AND METHODS

Samples

Seven environmental specimen types were selected to represent a wide range of microbial richness, diversity, composition, and dominant taxonomic groups (Table S1). Samples included five arthropod species from distinct orders (pickleworm, Diaphania nitidalis[Lepidoptera]; Lema beetle,Lemasp.[Coleoptera]; termite, Cryptotermes brevis[Blattodea]; honey bee, Apis mellifera[Hymenoptera]; and medfly, Ceratitis capitata[Diptera]), coffee (Coffea arabicavar. typica) leaf phyllosphere swabs, and soil collected at the base of coffee trees. Environmental sample sizes ranged from 6 to 8 replicates per group. Termites, Lema beetles, and pickleworm caterpillars were surface sterilized and whole-body homogenized prior to DNA extraction, whereas honey bees and Mediterranean fruit flies were surface sterilized and dissected prior to DNA extraction. Because samples were drawn from independent projects, DNA extraction methods differed among specimen types (Table S1). We retained this heterogeneity intentionally to evaluate workflow performance under realistic mixed-input conditions rather than under a fully standardized extraction design. We used commercial positive controls alongside experimental samples, including two DNA standards (2 ng per reaction, Zymobiomics Microbial DNA Standard Lot ZRC193007 Zymo Research Corp, Irvine, CA, USA; henceforth,“DNA control 1and2”) and one community standard processed through DNA extraction (Zymobiomics Microbial Community Standard Lot ZCR213683; henceforth,“extraction control”).

PCR amplification procedures

PCRs were performed using the iconPCR instrument (n6 Tec, Pleasanton, CA, USA) (28). The iconPCR instrument possesses individual well control such that amplification cycles may be ceased on a well-by-well basis following a designated change in fluorescence via intercalating dyes. Reactions (15 µL) included 0.25 µM dual-indexed 27F and 1492R primers contained in-line 10-mer barcodes and indices for concatenation PCR, as described in the PacBio Kinnex 16S Kit (Pacific Biosciences, Menlo Park, CA, USA). Q5 polymerase was chosen over other options because, in preliminary optimization of amplification protocols, there was minimal incidence of primer-dimer formation and false-positive signal on the iconPCR instrument compared to another high-fidelity polymerase (Fig. S1).

The number of PCR cycles was varied to assess the impact of autonormalization vs predetermined cycle number (Fig. 1). PCR cycling conditions were 98℃ for 30 s,followed by a variable number of cycles consisting of 98℃ for 10 s, 57℃ for 15 s, and 72℃ for 2 min, followed by a final extension at 72℃ for 10 min. For autonormalization procedures, cycling was terminated according to changes in per-well fluorescence, with a maximum limit of 30 cycles. The iconPCR software “Slope mode” monitors fluorescence during amplification and terminates reactions during the exponential phase before signal plateau (see Results, Fig. 3and Table S1). Briefly, the fluorescence signal recorded for each well, and a baseline is calculated as the average signal over the first cycles. After baseline estimation, amplification onset is detected when the rate of signal increase exceeds a threshold relative to the baseline or when the absolute signal exceeds a baseline-adjusted threshold. Once this condition is met, the reaction is classified as entering the amplification phase. Amplification is then monitored until the rate of signal increase declines, indicating that the reaction has passed its optimal amplification phase and amplification is terminated for that well. In theory, this enables adaptive, per-sample control of amplification, allowing each reaction to reach an appropriate endpoint based on its input characteristics rather than on a predefined cycle number. For these analyses, we used a minimum threshold of 1,000 RFUs as a baseline-adjusted threshold. Fixed-cycle protocols of 20, 24, and 30 cycles were performed for comparison. All protocols were performed on the iconPCR instrument.

Pooling, library preparation, and sequencing

Following amplification, 5 µL of each reaction was pooled blindly for the autonormalized and 20-cycle protocols and purified using AMPure beads before fluorometric quantification of the resulting pools (DeNovix Inc., Wilmington, DE, USA). For the 24- and 30-cycle protocols, reactions were first AMPure-purified, individually quantified, and then pooled at equimolar concentrations to reflect a commonly used manual normalization workflow. Following pool quantification, the four amplicon libraries were combined at equivalent DNA concentrations for Kinnex PCR and concatenation. Manufacturer kit protocols were followed, except that incubation times were doubled. The ~19 kb concatenated amplicon library was sequenced on a single Revio 24M SMRT cell on a Pacific Biosciences Revio system. HiFi reads resulting from the run were desegmented and demultiplexed using Lima v.2.13.0, implemented in PacBio SMRTLink v.13, before further processing. Lima default parameters were modified to ensure complete match between anticipated barcodes and individual samples.

Sequence processing

HiFi reads were pre-filtered to retain reads with mean read quality score ≥ Q30 and per-base quality score ≥ Q20 using fastplong v.0.4.1 (30). For per-base quality score, we designated a 2% cutoff rate;if the proportion of bases with a quality score below this level exceeded that cutoff, the entire read would be culled. We then discarded sequences by length (min = 1,300; max = 1,550) before using cutadapt v.5.1 (31) to trim primers and reorient reads into the same direction. Reads lacking both forward and reverse primers were removed from the data set. Amplicon sequence variants (ASVs) were generated with DADA2 in R v.4.4 (32, 33). Reads were filtered with DADA2 with expected error = 2 (34) and dereplicated. PacBio Revio default sequencing uses binned quality scores, so we used the PacBio DADA2 error-learning algorithm adjusted for binned quality scores (PacBio Revio quality scores: 3, 10, 17, 22, 27, 35, and 40). After estimating error rates, DADA2 denoising algorithms were used to infer sequences, and chimeric sequences were removed using dada2::isBimeraDenovo in pooled samples. ASVs were taxonomically classified using the Naïve Bayesian Classifier implemented in DADA2 against the SILVA database v.138.2 (35–37).

Error rate determination

As noted previously for commercial mock communities (8, 38), there can be sequence-level discrepancies between data resulting from sequencing the Zymo commercial standards and the corresponding vendor-provided reference sequence file. Specifically, not all produced sequences matched the commercial standards but were clearly consistent variants of those taxa. We therefore followed the general strategy outlined by Callahan et al. (2019) to refine the control reference fasta (8). ASVs recovered from positive controls were searched against the NCBI nucleotide core collection (nr/nt) database (accessed 3/18/2026), excluding non-isolate, uncultured, and environmental entries. Sequences with complete query coverage and exact taxonomic matches were treated as bona fide references rather than PCR-derived errors. Residual non-reference ASVs were then interpreted as putative errors or contaminants. Because Kinnex library preparation and dual-index barcoding may permit occasional barcode carryover, switching, or misassignment, all non-reference sequences were retained in the residual error assessment rather than excluded a priori.

Per-base residual error rates were computed in mothur v.1.48.3 with the seq.error command using the revised reference set (39). After DADA2 processing, ASV count and fasta file were converted into mothur-friendly formats for end-of-pipelineprocessing to determine control error rates under different conditions. We used manufacturer-provided reference sequences updated with bona fide sequence matches, as described above, to compute these errors (File S1).

Data analysis

Statistical analyses were performed using R v. 4.4 (39). Relationships between fluorescence and read yields were analyzed using Spearman correlation coefficients with the R package rstatix (40). Alpha-biodiversity metrics were computed using the R package vegan (41). To accommodate differences in sequence depths, we computed metrics using independent rarefaction events (999 iterations, 10,000 reads subsampled) that were then averaged across all separate iterations. Samples with fewer than 10,000 reads were removed from subsequent statistical analysis. Relationships between sample Shannon H′ and sequence retention were performed using Spearman correlation coefficients. PCR cycling protocol and alpha diversity metrics were compared using pairwise Wilcox rank-sum tests with FDR correction. Beta-diversity indices were computed after repeated rarefaction using vegan::avgdist, with 10,000 reads subsampled per iteration. Bray-Curtis dissimilarities were calculated using the raw counts, and Jaccard dissimilarities were calculated after transforming to presence/absence. Non-metric multidimensional scaling was performed on the resulting matrices to determine separation of samples. PERMANOVA was performed with vegan::adonis2 using specimen type and PCR protocols as fixed effects. Because the same biological samples were subjected to multiple PCR treatments, sample identity was included as a strata term to constrain permutations within sample. Givenstrong overall differences among specimen types, we also conducted pairwise PERMANOVA within specimen groups using the package pairwiseAdonis (42), again constraining permutations by sample identity. Pairwise P-values were then adjusted using a Holm adjustment.

RESULTS

Sequencing output following initial processing

The sequenced library produced 3.6M HiFi reads and 63.6 Gb of raw sequence data, with 94.3% of reads having an average base quality ≥Q30. The mean and median lengths of concatenated reads were 17.9 kb and 18.9 kb, respectively. Desegmentation yielded 37,699,404 16S reads with an average length of 1,554 bp. Autonormalization of libraries generated the largest number of processed reads after quality control compared with fixed protocols (10.5M for autonormalization, 8.3M for 30 cycles of PCR, 6.7M for 24 cycles of PCR, and 9.2M for 20 cycles of PCR) (Fig. 2).

Fig 2.

Library-level effects of PCR protocols on sequences. (A) Library preparation protocol exhibited different levels of sequence loss at the different filter steps. All protocols exhibited approximately the same sequence loss (~30%) due to low-quality bases. Protocols with higher (30) PCR cycles had more sequences culled with DADA2 denoising and chimera checking protocols compared to the other library protocols. (B) Distribution of the number of starting fully processed reads between different PCR protocols. Data are plotted on a log₁₀(y) scale, and samples are distributed by rank abundance of the final reads along the x-axis.

Effects of library preparation protocol on sequence retention

Because pooling strategy differed among protocols, metrics such as read evenness and downstream sequence retention should be interpreted as workflow-level outcomes rather than as effects attributable solely to PCR cycle control. Library preparation protocol strongly influenced read retention during denoising and chimera removal. While all PCR protocols lost ~30% of reads during initial base-quality filtering, samples amplified for 30 cycles experienced substantially greater losses during denoising than autonormalized reactions. The 20- and 24-cycle procedures were intermediate between autonormalization and 30 cycles in overall retention (Fig. 2A). After full processing, protocols also differed in the distribution of reads per environmental sample (Fig. 2B). The 20- and 24-cycle workflows showed the greatest spread in final read counts, whereas the 30-cycle workflow was more compressed. The autonormalized workflow yielded a comparatively tight distribution of final read counts across samples. Because pooling strategy differed among protocols, these differences in read evenness should be interpreted as workflow-level outcomes rather than as effects of thermocycle control alone.

Library protocol and specimen-specific effects on read yield and retention

In the generation of the amplicon libraries, we evaluated different library preparation strategies for implementation, including quantification-based guidance (24 cycles and 30 cycles) and blind pooling (20 cycles following Kinnex protocols, autonormalization). Fluorescence was monitored for all reactions, which indicated some samples were under- and overamplified (Fig. 3A). In the 20-cycle library preparation, this materialized in a wildly different number of reads following demultiplexing (Fig. 3B). For the 24-cycle protocol, even though they were quantified, there was still variation between samples. One caveat in the 24-cycle protocol was that termite samples amplified successfully (Fig. 3A) but were mistakenly omitted during pooling for that library.Accordingly, their low read counts in this protocol reflect a laboratory pooling error rather than failed amplification or poor barcode performance. The 30-cycle protocol was more even than the second pool. For the individual autonormalized specimens, there was a very tight and even distribution of sample yields, despite differences in sample type and DNA extraction.

Fig 3.

Library preparation protocol on the number of demultiplexed reads for different specimens. (A) Amplification curves of samples across different cycling protocols. (B) Number of sequences yielded for each specimen type by thermocycling protocol. (C) Impacts of sample autonormalization fluorescence and stop cycle on full-length 16S yield. The left panels show the relationship between fluorescence and read output, and the right panels show therelationshipbetween stop cycle and read output.

Because autonormalization uses fluorescence to determine reaction termination, we examined whether fluorescence metrics predicted read output (Fig. 4). Maximum fluorescence and stop cycle were not significantly correlated (σ = −0.21, P = 0.216), consistent with the dynamic behavior of slope-mode termination. After excluding reactions with RFU values < 2,000, maximum fluorescence at termination was positively correlated with both demultiplexed read count (σ = 0.66, P < 0.001) and final processed read count (σ = 0.55, P < 0.001) (Fig. 4A). In contrast, among reactions that reached the termination threshold, stop cycle itself was not significantly associated with read output (demultiplex σ = −0.23, P = 0.102; processed (σ = −0.18, P = 0.208) (Fig. 4B).

Fig 4.

PCR protocol effects on retained reads (%) after filtering steps for different specimen types. Only samples with ≥3,000 sequenced reads are displayed. AN, autonormalization of PCR thermocycles.

Percent sequence retention varied by both specimen type and library workflow (Fig. 4 andTable S2). Samples with greater microbial biodiversity, particularly soil samples, appeared to be disproportionately affected by higher cycle protocols, with fewer than 10% of starting reads retained after 30 cycles. Across protocols, sequence retention was negatively correlated with Shannon diversity (20: σ = −0.33, P = 0.021; 24: σ = −0.61, P < 0.001; 30: σ = −0.78, P < 0.001; autonormalized: σ = −0.66, P < 0.001) (Fig. 5), indicating that more diverse samples were generally more susceptible to downstream sequence loss during denoising and chimera removal (Table S3). This trend remained broadly similar even when soil samples were excluded, although the relationship was weaker in the 20-cycle workflow (P = 0.078). Some arthropod samples (honeybee, termite, and pickleworm) had higher percentages of sequences filtered at amplicon length and primer detection step compared to others.

Fig 5.

Relationships between sample ASV alpha-diversity and sequence loss. Generally, samples with higher Shannon diversity had fewer sequences retained, with higher thermocycle procedures exacerbating the retention of samples. Only samples with sequences ≥3,000 at start of process are displayed. Shannon H′ was computed using 10,000 sequences rarefied and averaged for 100 iterations using autonormalization PCR methods.

Positive control performance

Library workflow influenced positive-control read retention, residual error rate, and the fraction of non-reference ASVs (Fig. 6). Autonormalized and 20-cycle workflows retained the greatest fraction of reads after processing, and 20-cycle controls exhibited the lowest residual per-base error rates (Fig. 5A). Residual error rates for autonormalized controls were well below 0.005%, whereas those for 20-cycle controls were below 0.002%. In contrast, the 30-cycle controls retained the fewest sequences and showed the highest residual error rates (all above 0.004%). Autonormalized controls terminated between 19 and 22 cycles (Table S1). The fraction of non-reference ASVs remained low across workflows, and the taxonomic composition of the control communities was broadly consistent among protocols (Fig. 6B). Importantly, the revised reference framework indicated that some previously flagged non-refence ASVs likely represent bona fide sequence variants rather than PCR-derived artifacts (Table S4). Despite some differences in error and sequence loss, the taxonomic distribution and ASV reference matches were largely consistent between the different library preparations (Fig. 6C). The “extraction” control clearly had different genus-level abundances than the others, but the ratios of processed ASV sequences were comparable.

Fig 6.

Impact of PCR library protocol on positive controls. (A) Bars showing non-chimeric sequences (%) retained following different protocols (left y-axis) and shapes showing the error rate (%) compared to modified reference sequences (right y-axis). (B) Total bars showing the percentage of non-target ASVs in each sample. Main bars show the percent of non-target sequences in the data set including all potential contaminants. The filled fraction is the percent of sequences that were taxonomically classified as one of the controls, but had NCBI nr/nt matches with discrepancies from the references. (C) Relative abundances of genera in the Zymo controls. Line-level breakdowns represent the bona fide ASVs in the reference and distribution across the control samples.

Negative control performance

After full sequence processing, negative control read counts ranged from 19 to 1,132 reads (Table S5). Putative contaminants were sparse and inconsistently distributed among controls, and no single contaminant taxon was detected across all negative control reactions. These patterns do not support a dominant reagent-wide contaminant background in this data set. Instead, some taxa appeared associated with particular barcode combinations or plate positions, consistent with low-level barcode carryover or sporadic cross-assignment during library preparation or demultiplexing (Table S5 and Fig. S2).

Community compositions and diversity metrics

Ordination analyses clustered samples primarily by specimen type rather than by PCR protocol for both Bray-Curtis and Jaccard dissimilarities (Fig. 7). PERMANOVA detected statistically significant protocol effects, but these explained only a small fraction of total variation (R²< 0.01), whereas specimen type accounted for the majority of variance (R²> 0.24; Table S6). Residual variation remained high, indicating that most compositional differences reflected specimen-level heterogeneity rather than the protocol alone. Pairwise analyses within specimen groups showed that protocol effects were inconsistent and generally modest, although some more diverse specimens, including termites and soil, appeared somewhat more sensitive to workflow differences (Table 1). Overall, these results suggest that PCR workflow influenced technical performance more strongly than broad ecological interpretation in this data set. Some specific comparisons could not be conducted due to loss or limited samples following library preparation. Alpha-diversity metrics were largely consistent across protocols (Fig. 7C and D;Table S3), with limited specimen-specific differences. Termite samples had greater Shannon metrics (P = 0.03) and ASV richness (P = 0.016) in the autonormalized reactions compared with 30 cycles.

Fig 7.

Ordination and biodiversity metrics of environmental samples. (A) NMDS using Bray-Curtis dissimilarities indicate strong clustering by sample type, with little overt impact of PCR protocol. (B) NMDS of specimens using Jaccard (presence/absence) indicates a first grouping of samples. (C) Shannon diversity and (D) ASV richness of environmental samples at 30 cycles vsautonormalized protocols. Large circles represent the mean within each group of specimens. For both alpha- and beta-biodiversity metrics, soil samples did not obtain enough processed read counts (less than 10,000) to adequate comparisons at 30 PCR cycles with the other sample types.

TABLE 1.

Pairwise PERMANOVA from Bray-Curtis and Jaccard dissimilarities^a

Dissimilarity and sample	Overall	20 vs 24	20 vs 30	24 vs 30	AN vs 20	AN vs 24	AN vs 30
Bray-Curtis
Medfly	P = 0.002	0.156	0.156	0.156	0.156	0.156	0.047
	R2 = 0.007	0.007	0.009	0.002	0.007	0.001	0.002
Beetle	P = 0.571	1.000	1.000	1.000	1.000	1.000	0.563
	R2 = 0.019	0.013	0.011	0.007	0.021	0.008	0.017
Pickleworm	P = 0.259	1.000	1.000	1.000	1.000	1.000	0.375
	R2 = 0.030	0.058	0.008	0.008	0.041	0.040	0.004
Termite	P < 0.001		0.023		0.023		0.023
	R2 = 0.047		0.043		0.044		0.022
Coffee leaf	P = 0.026	1.000	1.000	1.000	1.000	0.750	1.000
	R2 = 0.109	0.135	0.208	0.071	0.131	0.035	0.025
Soil	P < 0.001	0.063			0.047	0.063
	R2 = 0.041	0.035			0.026	0.032
Positive controls	P = 0.222	1.000	1.000	1.000	1.000	1.000	1.000
	R2 = 0.007	0.001	0.033	0.022	0.006	0.032	0.004
Jaccard
Medfly	P = 0.002	0.156	0.047	0.188	0.047	0.188	0.188
	R2 = 0.077	0.053	0.068	0.049	0.057	0.044	0.046
Beetle	P = 0.186	1.000	0.750	0.781	1.000	1.000	1.000
	R2 = 0.112	0.086	0.102	0.067	0.092	0.063	0.067
Pickleworm	P = 0.075	1.000	1.000	1.000	0.625	0.875	0.094
	R2 = 0.070	0.065	0.062	0.043	0.056	0.031	0.041
Termite	P < 0.001		0.031		0.023		0.023
	R2 = 0.072		0.057		0.045		0.062
Coffee leaf	P = 0.360	1.000	1.000	1.000	1.000	1.000	0.750
	R2 = 0.144	0.199	0.167	0.098	0.119	0.066	0.066
Soil	P < 0.001	0.063			0.047	0.063
	R2 = 0.069	0.062			0.045	0.054
Positive controls	P = 0.008	1.000	1.000	1.000	1.000	1.000	1.000
	R2 = 0.310	0.218	0.252	0.256	0.205	0.264	0.176

^a Top values are pairwise P-values, and bottom values are R² values. Bold represents statistically significant Holm-adjusted P-values. Honeybee samples and blank slots are indicative of data with too few observations >10,000 reads. AN,autonormalized PCR libraries.

DISCUSSION

Long-read sequencing techniques and their application to microbiome or 16S rRNA amplicon sequencing have been hampered in part by poor read quality and sequencing yields. Advances in PacBio sequencing chemistry coupled with MAS-Seq concatenation strategies have made high-throughput, full-length 16S sequencing increasingly accessible and cost-effective. Our study produced read output and base-pair quality commonly observed with SBS techniques for partial hypervariable regions. However, PCR-based library preparation remains a critical point in processing, particularly for workflows handling heterogeneous specimens to be combined into a single sequencing pool. In this study, we found that higher PCR cycle numbers were associated with greater sequence loss during denoising and chimera removal, particularly in more diverse samples. Under the workflow tested, autonormalized libraries generally maintained low residual error rates and relatively even read representation across heterogeneous inputs. Because denoising algorithms infer true sequences based on abundance and error models (8, 32), substitution errors introduced early in PCR can propagate into abundant but erroneous variants that are subsequently removed. This effect is magnified for full-length amplicons, where even low per-base error rates increase the probability of read rejection during processing.

Following sequence processing, residual error rates in positive controls were low and were comparable to, or lower than, values commonly reported for other full-length 16S workflows (25, 32, 43, 44) and were in line, or improved relative, to past-generation non-Kinnex PacBio methods (8, 12, 45). Protocols using fewer effective PCR cycles generally retained more reads and produced lower residual error estimates. We also observed sparse, inconsistent non-reference taxa in negative controls and some clustering by barcode combination or well position, suggesting that low-level carryover, barcode cross-assignment, or related library-preparation artifacts may contribute to a fraction of residual non-refence reads. This possibility is especially relevant for low-abundance features and should be considered when interpreting rare taxa in Kinnex-style concatenated libraries, but could be ameliorated with UMI-based methods (38, 46). At the same time, the broad ASV/genus structure of the positive controls remained consistent across workflows, suggesting that the dominant biological composition of these standards was stable despite these technical differences. The ratios of the PCR-based controls differ from commercial manual's theoretical makeup, which may be due to differences in amplification efficiencies between individual taxa.

PCR workflow had relatively minor effects on community-level inference compared with specimen identity. Although some PERMANOVA tests were statistically significant, protocol effect sizes were small relative to specimen-type and sample-level variation. This distinction is important because large sample numbers can yield significant P-values even when the magnitude of the effect is limited. Within individual specimen groups, protocol-associated compositional shifts were inconsistent and generally modest, although more biodiverse samples such as termites and soils appeared somewhat more sensitive. These findings align with previous short-read studies showing limited compositional effects of PCR cycle number under controlled conditions (25, 27, 47, 48). However, others have indicated that overamplification can impact rare amplicons in diverse samples (28, 49), such that the impacts of thermocycling protocols may be case or sample-specific. Taken together, these results suggest that workflow differences can influence technical metrics without necessarily producing large distortions in higher-level ecological interpretation. It should be noted, though, that PCR-induced bias can vary between metrics.

The fact that more diverse samples were more susceptible to sequence loss likely results from a combination of factors. First, the most diverse sample we used also possessed the greatest titers (soil); therefore, these samples received up of 11 extra cycles that may have introduced additional errors compared to the autonormalization protocol. Second, samples possessing naturally high gut microbiome biodiversity may increase the number of erroneous sequences or ASVs inferred after amplification and denoising, thereby increasing the proportion of reads removed during downstream processing. Regardless of the root cause for this pattern of sequence loss, these observations are especially important when considering PCR protocols for diverse samples in full-length 16S sequencing.

Currently, there are few specific protocols or best practices for handling full-length 16S data, and the technical aspects of data quality control (i.e., error modeling) were outside of the specific goals of this study. Upstream from amplicon QA/QC, default demultiplexing parameters using the PacBio-provided Lima demultiplexing tool (80% match) have the potential to produce more errant mismatches and impact perceived contamination rates;therefore, stricter demultiplexing, either with Lima or by utilizing alternative software tools, may be desirable. Additionally, while default PacBio Revio sequence parameters produce FASTQ files with binned quality scores, flags at the sequencing step can allow for the production of continuous per-base quality scores. Full-length 16S inference remains sensitive to decisions about prefiltering, error modeling, denoising, chimera removal, and ASV-versus OTU-based summarization, so the core framework we used here may not be the best for all instances. In particular, full-length data sets can split individual genomes into multiple ASVs (8, 50, 51), as also illustrated by the positive controls here, and clustering or OTU-based approaches may be preferable for some applications (52, 53). However, such approaches can become computationally intensive for the size of these full-length datasets. While our workflow yielded control samples with low levels of residual error, future work should evaluate how laboratory workflow effects interact with alternative downstream pipelines, especially for highly diverse or low-biomass sample types.

Several limitations should be noted. First, pooling strategy differed among workflows, which limits strict attribution of read evenness and retention outcomes solely to PCR cycle control. Second, specimen types were drawn from independent projects and used different extraction procedures, so this study evaluates workflow behavior under realistic heterogeneous inputs rather than under a tightly standardized extraction design. Third, all data derive from a single Kinnex library sequenced on one Revio SMRT cell, so additional run-level replication would be needed to assess broader reproducibility and scalability.

Auto-termination of PCR reactions not only has impacts on sequence retention but also results in measurable improvements in starting library evenness. Our results demonstrate this across a swath of sample types extracted using distinct protocols, but inhibitor-laden samples or those with highly concentrated DNA may require some initial QA/QC to ensure efficient autonormalization. Like all fluorescence-based quantification and PCR, these methods are sensitive to potential inhibitors and primer dimerization. High starting DNA concentrations, for instance, can adversely affect accurate quantification and termination of libraries. Furthermore, while the PCR protocols employed here were effective with autonormalization techniques, different primer designs and/or polymerases employed can produce different results and require troubleshooting for implementation.

Conclusions

Full-length 16S gene sequencing offers clear advantages for taxonomic resolution and phylogenetic interpretation, but broader adoption depends on reliable and scalable upstream workflows. When specimens vary widely in microbial biomass or amplification behavior, selecting a single fixed PCR cycle number can be suboptimal: low-biomass samples may be underamplified, whereas high-titer samples may be overamplified and incur greater downstream sequence loss. Individual well control, like that of the iconPCR instrument, helps to mitigate both issues by controlling the number of thermocycles for specific samples and allowing for real-time adjustment of parameters. Our results demonstrate that PCR cycle autonormalization aids PacBio full-length 16S sequencing by producing more even sample representation across heterogeneous sample types, improving read retention, mitigating residual errors, and eliminating the need for post-PCR quantification and normalization. Importantly, this approach enhances both data quality and laboratory efficiency and has the potential to expedite library preparations across several applications.

DATA AVAILABILITY

Demultiplexed FASTQ files have been deposited to the NCBI SRA under the accession number PRJNA1330381. Data processing scripts, R scripts for figure generation and statistical analysis, and processed count, fasta, and taxonomy files are available through the USDA NAL at https://doi.org/10.15482/USDA.ADC/30964789.

SUPPLEMENTAL MATERIAL

The following material is available online at https://doi.org/10.1128/msphere.00295-26.

File S1. msphere.00295-26-s0001.fasta.

nr/nt supplemented Zymo fasta references used for assessing residual error.

Supplemental figures. msphere.00295-26-s0002.pdf.

Fig. S1 and S2.

Supplemental tables. msphere.00295-26-s0003.xlsx.

Tables S1 to S6.

ASM does not own the copyrights to Supplemental Material that may be linked to, or accessed through, an article. The authors have granted ASM a non-exclusive, world-wide license to publish the Supplemental Material files. Please contact the corresponding author directly for reuse.