henhaobio
Sign in

bioRxivGenomicsPreprint

MAT-classifier: A memory-efficient pipeline for accurate genus level profiling from ancient metagenomic data

Aranya DhibariD, Mikhail V. MatziD

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Background: Advances in sequencing technology have expanded opportunities to recover microbial DNA from ancient samples and reconstruct past environments and host-microbe interactions. However, the field remains constrained by computational challenges and accuracy problems, as rare ancient microbial DNA must be distinguished from abundant modern contaminants. Moreover, existing pipelines demand substantial computational resources, particularly memory, limiting their accessibility. Results: Here, we present MAT-classifier, a genus-level profiling workflow for detecting ancient microbial taxa from metagenomic projects, designed to reduce computational requirements while increasing accuracy. Unlike its counterparts, MAT-classifier first consolidates candidate references at the genus level and then performs independent alignments using conventional short-read aligners instead of metagenomic aligners. Using simulated datasets, we showed that this approach achieves more accurate classification of ancient taxa while requiring substantially less memory and shorter runtime than a modern counterpart, the aMeta pipeline. Benchmarking on multiple empirical ancient datasets further confirmed its low memory footprint and practical utility. Conclusions: MAT-classifier provides a reliable, computationally efficient, and accessible framework for ancient microbiome profiling. It lowers computational barriers while maintaining robust classification performance, facilitating broader application of ancient microbial DNA analysis.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (20 Sept 2026) · CC BY 4.0

Per-genus Bowtie2 or BWA-aln alignments in place of MALT cut alignment-stage memory from about 140 GB (aMeta) to about 7 GB and gave higher simulated-data AUC (0.87 vs 0.79).

Alignment-based ancient metagenome classifiers such as HOPS and aMeta depend on MALT competitive alignment, which holds the whole reference database in memory; aMeta recommends 256 GB of RAM even for small projects. MAT-classifier keeps the Kraken2 candidate screen of hybrid workflows but replaces MALT with parallel, independent per-genus alignments using Bowtie2 or BWA-aln, then scores each genus on damage, fragment length, Kraken2 support and edit distance.

Key results

  1. ShownAlignment-stage memory falls from about 140 GB (aMeta's MALT step) to about 7 GB. Whole-pipeline peak memory was also lower on the 1-TB test system, because Kraken2 replaces aMeta's KrakenUniq step, though the authors note this k-mer-step gap may shrink on systems with limited memory.

    On 10 simulated samples (20 cores, up to 1 TB RAM), Kraken2 peaked at ~97 GB, while aMeta's KrakenUniq step came close to the 1-TB limit despite a 200-GB preload. Downstream peak was ~7 GB for MAT-classifier vs ~140 GB used repeatedly by aMeta during MALT. Peak RAM was lower on all four empirical datasets, with and without the k-mer classification step. · Computational Performance of MAT-classifier…; Computational performance on real-world datasets; Figure legends

  2. ShownWall-clock runtime is far shorter than aMeta's. The gain comes from parallel per-genus jobs, not from less total compute: CPU time on empirical data was similar to or higher than aMeta's.

    On simulated data the full workflow (trimming and Kraken2 included) took ~30 min vs ~15 h for aMeta on the same allocation; Bowtie2 and BWA-aln had similar memory profiles and runtimes. On four empirical datasets: lower wall-clock time and memory than aMeta, with similar or higher total CPU time. · Computational Performance of MAT-classifier…; Computational performance on real-world datasets; Trade-offs of independent mapping

  3. SuggestedOn simulated data, MAT-classifier separates ancient genera from modern or absent ones better than aMeta, both in threshold-independent ROC AUC and at each tool's recommended cutoff. The choice of aligner made little difference.

    ROC on raw genus scores across the simulated samples: AUC 0.87 vs 0.79 for aMeta, similar for Bowtie2 and BWA-aln. F1 at the recommended cutoffs (>3 for MAT-classifier, >7 for aMeta) was higher across samples (one-sided Wilcoxon signed-rank p = 0.0010). Aligners gave similar F1 except samples 1 and 9 (Bowtie2 better) and 6 (BWA-aln better). · Classification accuracy on simulated data; Accuracy at recommended cutoffs; Figure legends

  4. SuggestedWhen ancient and modern species of the same genus co-occur, MAT-classifier usually still recovers the ancient genus. aMeta never recovered ancient Pseudomonas. MAT-classifier's sensitivity drops for low-abundance ancient taxa.

    Recovery of the ancient genus by MAT-classifier, by co-occurrence: Mycobacterium 6 of 7, Pseudomonas 7 of 9, Streptococcus 5 of 8. aMeta missed ancient Pseudomonas every time but found two more Streptococcus occurrences than MAT-classifier. · Simulation of ancient metagenomic data; Accuracy at recommended cutoffs

  5. SuggestedOn empirical ancient data, MAT-classifier calls many more candidate ancient genera than aMeta and recovers most genera that the source studies reported. Some of the extra calls may be false positives.

    Genera not detected by aMeta: 652 (sedaDNA), 108 (enriched 16S), 173 (dental calculus shotgun), 97 (coral). MAT-classifier detected Ruegeria, the main genus reported for the coral data, which aMeta missed. Agreement with the source study was lowest for enriched 16S. The extra calls are not independently verified. · Computational performance on real-world datasets; Accuracy and limitations of MAT-classifier

Methods, data and reproducibility

MAT-classifier takes trimmed reads plus external Kraken2 reports and builds per-genus NCBI reference sets for candidate taxa. It aligns reads independently to each set with Bowtie2 or BWA-aln and scores each genus 0 to 5 using pyDamage, PMDtools, fragment length, Kraken2 support and edit distance. It was benchmarked against aMeta on 10 gargammel-simulated samples and four empirical ancient datasets.

  • Simulated ancient metagenomes (gargammel): 36 bacterial species (18 ancient, 18 modern) plus modern and ancient human DNA; ~500,000 ancient and ~500,000 modern microbial fragments per sample · bacteria and human (simulated) · n = 10 samples · 10.5281/zenodo.22822502
  • Deeply sequenced shotgun libraries from an ancient coral project (~60 million reads each) · coral (host) · n = 2 samples · PRJNA757238 (SRR16608038, SRR16608039)
  • Enriched 16S rRNA libraries from ancient dental calculus · n = 8 libraries · PRJNA1031168
  • Shotgun libraries from ancient human dental calculus · human (host) · n = 10 libraries · PRJNA445215
  • Ancient DNA shotgun libraries from Lake Lama sediments (sedaDNA) · n = 6 libraries · PRJEB80877
Methods & tools
MAT-classifier, Kraken2, KrakenUniq, NCBI Datasets CLI, Bowtie2, BWA-aln, PMDtools, pyDamage, gargammel, aMeta, MALT, psrecord
Assay conditions & compute
Simulated benchmark: 20 CPU cores, up to 1 TB RAM. Empirical runs: a 1-TB-RAM cluster (20 cores per Methods, 40 per Results). 4 threads per alignment job. Kraken2 peak ~97 GB. On simulated data, MAT-classifier took ~30 min vs ~15 h for aMeta.
Reproducibility gaps
No software versions, containers or environment files are stated. Read trimming and Kraken2 (the most memory-hungry step) run outside the automated workflow and must be set up by the user. The core count for the empirical runs is reported inconsistently (20 vs 40).
Limitations (6)
  • Classification stops at genus level by design. Short ancient reads, similarity among related species and shifting microbial taxonomy make species-level calls unreliable. · stated by the authors
  • The many additional genera called on empirical data may partly be false positives. The authors advise treating scores as a screen and checking important taxa against pyDamage damage plots. · stated by the authors
  • Independent mapping raises the risk of cross-mapping. Genus binning and depth/breadth scoring reduce this but cannot remove it between closely related genera. · stated by the authors
  • The design trades memory for CPU: total CPU time was similar to or higher than aMeta's, and runtimes may grow on machines with fewer cores. · stated by the authors
  • Sensitivity falls for low-abundance ancient taxa, and there is no dedicated pathogen-screening module. Agreement with the original study was lowest for the enriched 16S data. · stated by the authors
  • Accuracy rests on 10 simulated samples generated by the authors, with one comparator (aMeta). The Kraken2 thresholds were also tuned on simulations, and the text does not say whether those were held out. No independent verification of the extra genera called on empirical data is reported. · noted in this summary

Read this if you profile microbial taxa in ancient shotgun or sedimentary DNA data and lack the 256 GB of RAM that aMeta recommends for small projects.

Citations
1

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISPochon '23Velsko '18Marques '25Oskolkov '25Cárdenas '22Li '22
79% semantic matchBioinformatics
Zoé Pochon, Nora Bergfeldt, Emrah Kırdök+10 more·Genome biology·Oct 23, 2023· 65 citations