Protein-level assembly increases protein sequence recovery from metagenomic samples manyfold


The open-source de novo protein-level assembler, Plass (, assembles six-frame-translated sequencing reads into protein sequences. It recovers 2–10 times more protein sequences from complex metagenomes and can assemble huge datasets. We assembled two redundancy-filtered reference protein catalogs, 2 billion sequences from 640 soil samples (soil reference protein catalog) and 292 million sequences from 775 marine eukaryotic metatranscriptomes (marine eukaryotic reference catalog), the largest free collections of protein sequences.

Access options

Rent or Buy article

Get time limited or full article access on ReadCube.


All prices are NET prices.

Fig. 1: Plass workflow.
Fig. 2: Sensitivity and precision of protein sequences assembled from synthetic reads and associated false discovery rate (FDR) of functional annotations.
Fig. 3: Plass assembles more protein sequences from various environments than the state of the art.

Data availability

The assembled protein sequence sets are available in FASTA format under a Creative Commons Attribution CC-BY 4.0 License at All scripts and benchmark data including command-line parameters necessary to reproduce the benchmark and analysis results presented are available at

Code availability

Plass is GPLv3-licensed open-source software. The source code and binaries for Plass can be downloaded at


  1. 1.

    Howe, A. C. et al. Proc. Natl. Acad. Sci. USA 111, 4904–4909 (2014).

  2. 2.

    Li, J. et al. Nat. Biotechnol. 32, 834–841 (2014).

    CAS  Article  Google Scholar 

  3. 3.

    Sunagawa, S. et al. Science 348, 1261359 (2015).

    Article  Google Scholar 

  4. 4.

    Nielsen, H. B. et al. Nat. Biotechnol. 32, 822–828 (2014).

    CAS  Article  Google Scholar 

  5. 5.

    Zerbino, D. & Birney, E. Genome Res. 18, 821–829 (2008).

    CAS  Article  Google Scholar 

  6. 6.

    Li, D. et al. Bioinformatics 31, 1674–1676 (2015).

    CAS  Article  Google Scholar 

  7. 7.

    Nurk, S. et al. Genome Res. 27, 824–834 (2017).

    CAS  Article  Google Scholar 

  8. 8.

    Ye, Y. & Tang, H. J. Bioinform. Comput. Biol. 7, 455–471 (2009).

    CAS  Article  Google Scholar 

  9. 9.

    Yang, Y. et al. Bioinformatics 31, 1833–1835 (2015).

    CAS  Article  Google Scholar 

  10. 10.

    Steinegger, M. & Söding, J. Nat. Commun. 9, 2542 (2018).

    Article  Google Scholar 

  11. 11.

    Sczyrba, A. et al. Nat. Methods 14, 1063–1071 (2017).

    CAS  Article  Google Scholar 

  12. 12.

    Kashtan, N. et al. Science 344, 416–420 (2014).

    CAS  Article  Google Scholar 

  13. 13.

    Berube, P. M. et al. Sci. Data 5, 180154 (2018).

    CAS  Article  Google Scholar 

  14. 14.

    Hyatt, D. et al. BMC Bioinforma. 11, 119 (2010).

    Article  Google Scholar 

  15. 15.

    van der Walt, A. J. et al. BMC Genom. 18, 521 (2017).

    Article  Google Scholar 

  16. 16.

    Huerta-Cepas, J. et al. Mol. Biol. Evol. 34, 2115–2122 (2017).

    CAS  Article  Google Scholar 

  17. 17.

    Tian, W. & Skolnick, J. J. Mol. Biol. 333, 863–882 (2003).

    CAS  Article  Google Scholar 

  18. 18.

    Lee, S. T. M. et al. Microbiome 5, 50 (2017).

    Article  Google Scholar 

  19. 19.

    Carradec, Q. et al. Nat. Commun. 9, 373 (2018).

    Article  Google Scholar 

  20. 20.

    Ovchinnikov, S. et al. Science 355, 294–298 (2017).

    CAS  Article  Google Scholar 

  21. 21.

    Magoc, T. & Salzberg, S. L. Bioinformatics 27, 2957–2963 (2011).

    CAS  Article  Google Scholar 

  22. 22.

    Sheetlin, S. et al. Bioinformatics 32, 304–305 (2016).

    CAS  PubMed  Google Scholar 

  23. 23.

    Mirdita, M. et al. Nucleic Acids Res. 45, D170–D176 (2017).

    CAS  Article  Google Scholar 

  24. 24.

    Kanehisa, M. et al. Nucleic Acids Res. 45, D353–D361 (2016).

    Article  Google Scholar 

  25. 25.

    Steinegger, M. & Söding, J. Nat. Biotechnol. 35, 1026–1028 (2017).

    CAS  Article  Google Scholar 

  26. 26.

    Huerta-Cepas, J. et al. Nucleic Acids Res. 44, D286–D293 (2016).

    CAS  Article  Google Scholar 

  27. 27.

    Frith, M. Nucleic Acids Res. 39, E23 (2011).

    Article  Google Scholar 

  28. 28.

    Hingamp, P. ISME J. 7, 1678–1695 (2013).

    CAS  Article  Google Scholar 

Download references


We are grateful to C. Notredame and C. Seok for hosting M.S. at the Centre for Genomic Regulation and Seoul National University for 12 and 30 months, respectively. We thank S. Sunagawa, F. Meyer and A. Sczyrba for helpful discussions, and T. Brown for his early analysis and detailed feedback on Plass results. We thank all who contributed metagenomic datasets used to build SRC and MERC, in particular contributors to the TARA ocean project and the US Department of Energy Joint Genome Institute ( This work was supported by the EU’s Horizon 2020 Framework Programme (Virus-X, grant no. 685778).

Author information




M.S. and J.S. designed the research study. M.S. and M.M. developed code and performed the analyses. M.S. and J.S. wrote the manuscript.

Corresponding authors

Correspondence to Martin Steinegger or Johannes Söding.

Ethics declarations

Competing interests

The authors declare no competing interests.

Additional information

Peer review information: Lei Tang was the primary editor on this article and managed its editorial process and peer review in collaboration with the rest of the editorial team.

Publisher’s note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Integrated supplementary information

Supplementary Figure 1 Schematic comparison of a nucleotide- and a protein assembly.

On top is the final protein assembly followed by the stacked overlapping protein reads. The small gray section highlights the multiple protein sequence alignment of the overlapping reads and below the respective nucleotide alignment. Less ambiguity is visible on the protein level due to conservative mutations (mutations with similar biochemical properties) compared to the nucleotide level, resulting in an assembly that is more robust to microdiversity in the population.

Supplementary Figure 2 Overlap of assemblies of Plass with Megahit and metaSPAdes.

(a) left: A fraction 38.3% of amino acids in the Plass-assembled proteins of set 1 is covered by alignments to proteins in the Megahit assembly at a minimum sequence identity cut-off of 99%. Conversely, 83.2% of proteins in the Megahit-assembled set 1 is covered by alignments with Plass-assembled proteins. Right: Same as on the left but comparing the Plass assembly with the metaSPAdes assembly. (b) Same as (a) but for protein set 2.

Supplementary Figure 3 Effect of neural network filter to remove wrong translation frames.

Sensitivity and precision in set 1 (a) and set 2 (b). Top: assembly sensitivity is the fraction of reference sequence amino acids that matches to an assembled protein sequence. Bottom: assembly precision is the fraction of assembled amino acids that matches to a reference protein at the minimum sequence identity on the x-axis. Plass uses a minimum sequence identity for merging fragments of 90%, Plass-97 uses a threshold of 97%.

Supplementary Figure 4 Comparison of Megahit assignment using the 2bLCA protocol.

The MMseqs2 taxonomy assignment workflow uses three steps to assign a taxonomic label to a query sequence. (1) We search with the query sequence against a reference database and extract the aligned subsequence of the best hit. (2) This sequence is matched again against the reference database. Each hit with an E-value smaller than the best hit E-value from the previous search is accepted. (3) We compute the lowest common ancestor based on the taxonomic labels of all accepted hits.

Supplementary Figure 5 Plass ORF extraction and start codon prediction (ORF calling).

Plass extracts two sets of ORFs. ORF set 1 contains all translated ORFs with at least 45 codons. ORF set 2 contains all translated ORFs with at least 20 codons starting with a putative ATG start codon that is the first ATG codon after a stop codon in the same frame. (Start codon prediction) Plass predicts start codons with a consensus method using a multiple sequence alignment of ORF set 1 and 2. Wherever at least 20% of all methionines in one column are marked by a prepended asterisk, it removes the preceding residues from all other sequences and prepends an asterisk to all sequences to mark the start.

Supplementary Figure 6 Taxonomy evaluation of the soil metagenome assembly.

(a) We investigate the taxonomic composition of the 8 most abundant taxa (all other taxa are pooled in ‘Others’) in the soil assemblies from Fig. 2d (blue: Megahit, red: Plass) and the assemblies of the 12 soil samples from Fig. 2e (light blue: Megahit, light red: Plass). On top we show the read count ratios between Plass and Megahit, for both the single and 12 soil assemblies. The inset gives the fraction of reads in the single and the 12 soil samples that could be mapped to an assembled protein sequence. (b) We show the count of assembled amino acids within various coverage ranges for Megahit (blue) and Plass (red) in the single soil sample.

Supplementary information

Source data

Rights and permissions

Reprints and Permissions

About this article

Verify currency and authenticity via CrossMark

Cite this article

Steinegger, M., Mirdita, M. & Söding, J. Protein-level assembly increases protein sequence recovery from metagenomic samples manyfold. Nat Methods 16, 603–606 (2019).

Download citation

Further reading


Nature Briefing

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

Get the most important science stories of the day, free in your inbox. Sign up for Nature Briefing