Signals of microbial growth learned from single amplicon samples

Philippe Chlenski, Deirdre Ricaurte, Itsik Pe’erbioRxiv 2023 · DOI

Abstract

Irregularities in metagenomic whole-genome shotgun (WGS) read coverage can arise in quickly replicating microbial populations. These irregularities, summarized as peak-to-trough ratios (PTRs), are correlated with growth rates. This study seeks to explore the presence of similar coverage irregularities in 16S amplicon datasets, where multicopy diverged 16S genes provide an opportunity to explore coverage at different positions on the bacterial chromosome. To this end, we propose a model of Operational Taxonomic Unit (OTU) observations under replication and sequence similarity, from which we derive a method for simultaneous copy number correction and dynamics estimation by gradient descent. We conduct a series of benchmarks on synthetic data, determining a set of heuristics for when such methods may be applied, and compare our method with WGS-based methods on a real dataset. We find no correlation between coPTR estimates and our method, suggesting further modifications to our method may be required.

Introduction

In microbial ecology, distinguishing between the presence of a species and its active growth is a critical challenge. Traditionally, researchers have relied on Whole-Genome Shotgun (WGS) sequencing to estimate growth rates by calculating Peak-to-Trough Ratios (PTRs). This method exploits the fact that rapidly replicating bacteria have more DNA copies near the origin of replication (OOR) than at the terminus (RT). However, WGS is often prohibitively expensive and computationally demanding.

Most microbiome studies instead utilize 16S amplicon sequencing, which targets a single marker gene. While cost-effective, 16S data is typically viewed as a static snapshot of community composition, lacking information on replication dynamics. This research investigates whether the growth signal—the "pulse" of the microbiome—is detectable within 16S amplicon samples by leveraging the existence of multicopy, diverged 16S genes.

Methodology: Leveraging Genomic Architecture

The core innovation of this work lies in the observation that many bacteria possess multiple copies of the 16S gene, often with slight sequence variations, distributed at different distances from the OOR. If these copies are sufficiently diverged, their relative counts in an amplicon sample should reflect the replication "slope" of the chromosome.

The researchers developed a generative model that predicts Operational Taxonomic Unit (OTU) counts based on a bacterium's abundance, its PTR, and the known genomic locations of its 16S genes. This relationship is captured in a sequence-sharing matrix ($S$). To infer growth rates from observed data, they implemented an inference engine using PyTorch that employs gradient descent to simultaneously estimate true abundance and PTR.

Theoretical read probability across chromosome locations for different Peak-to-Trough Ratios (PTR).

A critical component of the model is the inclusion of a learnable "amplicon bias" term. This accounts for PCR-induced variations where certain sequences are preferentially amplified, a factor that can otherwise lead to significant errors in growth estimation.

Benchmarking and the "1,000-Read Rule"

To validate the model, the authors conducted extensive synthetic benchmarks, simulating microbial communities with varying growth rates and sequencing depths. The results demonstrated that the model could effectively recover true PTRs under controlled conditions.

However, these simulations revealed a significant threshold for accuracy: the "1,000-read rule." The method requires at least 1,000 reads per OTU per sample to achieve high precision. Below this depth, the signal of replication is often drowned out by stochastic noise.

PTR error as a function of the number of reads, illustrating the convergence of accuracy at higher sequencing depths.

Real-World Application and Challenges

The method was applied to the DIABIMMUNE dataset, which contains paired 16S and WGS samples from human infant gut microbiomes. This allowed for a direct comparison between 16S-based estimates and gold-standard WGS-based coPTR estimates.

The results were sobering: the 16S-based method showed no correlation with the WGS estimates. In several instances, the model produced extreme, non-biological PTR values. The primary culprit appears to be inadequate coverage; most genomes in the real-world dataset did not meet the 1,000-read threshold required for reliable signal detection. Furthermore, the use of short V4 region reads limited the detectable divergence between 16S copies.

Conclusion and Future Directions

While the current implementation faces hurdles in real-world metagenomic datasets, this study establishes a rigorous mathematical framework for extracting growth signals from amplicon data. Future improvements may include moving to full-length 16S sequencing (e.g., PacBio) to capture more divergence, improving reference databases for gene locations, and testing on lower-complexity in vitro communities where coverage is higher.

This work represents a foundational step toward making microbial growth estimation as accessible as composition profiling.

Summary updated

Abstract

Abstract Irregularities in metagenomic whole-genome shotgun (WGS) read coverage can arise in quickly replicating microbial populations. These irregularities, summarized as peak-to-trough ratios (PTRs), are correlated with growth rates. This study seeks to explore the presence of similar coverage irregularities in 16S amplicon datasets, where multicopy diverged 16S genes provide an opportunity to explore coverage at different positions on the bacterial chromosome. To this end, we propose a model of Operational Taxonomic Unit (OTU) observations under replication and sequence similarity, from which we derive a method for simultaneous copy number correction and dynamics estimation by gradient descent. We conduct a series of benchmarks on synthetic data, determining a set of heuristics for when such methods may be applied, and compare our method with WGS-based methods on a real dataset. We find no correlation between coPTR estimates and our method, suggesting further modifications to our method may be required.
BibTeX
@misc{chlenski2023signals,
  author = {Philippe Chlenski and Deirdre Ricaurte and Itsik Pe’er},
  title = {{Signals of microbial growth learned from single amplicon samples}},
  year = {2023},
  doi = {10.1101/2023.11.07.565791},
  howpublished = {bioRxiv},
}
Figure