henhaobio
Sign in

Genome biologyBioinformaticsPeer-reviewed journal article

Rapid and sensitive detection of genome contamination at scale with FCS-GX

Alexander Astashyn, Eric S. TvedteiD, Deacon Sweeney, Victor Sapojnikov, Nathan Bouk, Victor Joukov, Eyal Mozes, Pooja K Strope, Pape M. Sylla, Lukas Wagner, Shelby L. Bidwell, Larissa C. Brown, Karen ClarkiD, Emily W. DavisiD, Brian J. Smith-White, Wratko HlavinaiD, Kim D. PruittiD, Valérie SchneideriD, Terence D. MurphyiD

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

Abstract

Assembled genome sequences are being generated at an exponential rate. Here we present FCS-GX, part of NCBIs Foreign Contamination Screen (FCS) tool suite, optimized to identify and remove contaminant sequences in new genomes. FCS-GX screens most genomes in 0.1-10 minutes. Testing FCS-GX on artificially fragmented genomes demonstrates sensitivity >95% for diverse contaminant species and specificity >99.93%. We used FCS-GX to screen 1.6 million GenBank assemblies and identified 36.8 Gbp of contamination (0.16% of total bases), with half from 161 assemblies. We updated assemblies in NCBI RefSeq to reduce detected contamination to 0.01% of bases. FCS-GX is available at https://github.com/ncbi/fcs/.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (6 Jun 2023) · CC0 1.0

FCS-GX, a hashed k-mer cross-species aligner with a 709 Gbp curated database, screens most genomes in 0.1-10 minutes and flagged 36.8 Gbp of contamination in GenBank.

Foreign DNA in public genome assemblies spreads annotation errors and false evolutionary or lateral-transfer conclusions, yet NCBI's legacy VecScreen/BLAST screen misses much of it and most existing tools are not built for automated removal. FCS-GX offers a fast, automated sequence-level screen that found fourfold more contamination than the legacy pipeline and was used to clean RefSeq.

Key results

  1. ShownOn artificially fragmented, highly contiguous genomes, FCS-GX flags contaminants from species in its database with high sensitivity and rare false positives. Sensitivity drops for simulated novel eukaryotic contaminants.

    1 kbp fragments: 76% of prokaryote (663 species) and 91% of eukaryote (370 species) datasets exceeded Sn=95%; dropping same-species tax-ids cut median Sn to 89% (prokaryotes) and 17-63% (eukaryotes); 95% of prokaryote datasets had Sp=100% (88% without same species); sequence-level Sp >99.93% in all scenarios. · FCS-GX accurately detects contaminants with few…; Computing sensitivity and specificity

  2. ShownScreening is fast enough for archive-scale and per-submission use, and it costs far less CPU time than NCBI's legacy megaBLAST screen.

    After the database was loaded on one 64 vCPU server, 28,774 eukaryote genomes (15.7 Tbp) were screened in 18 days; prokaryote batches ran at 1.94 sec/genome; megaBLAST used 135x more CPU time while aligning against 80% less sequence. · Results; FCS-GX enables high-throughput contamination…; GenBank and RefSeq genome datasets

  3. ShownGenBank contamination is concentrated: a few highly fragmented genomes carry half of it, most contaminants are short prokaryotic sequences, and the contaminated fraction has not fallen over time.

    1,545,312 prokaryote and 30,053 eukaryote assemblies (22.4 Tbp): 36.8 Gbp from 23,451,091 sequences (0.16% of bases, 1.30% of sequences); half from 161 genomes; 82% of contaminants ≤1 kbp; prokaryote sources 26.2 Gbp (71%); contig N50 >1 Mbp genomes 0.028% vs 0.209% for others. · FCS-GX detects extensive contamination in NCBI…; Extreme genome contamination

  4. ShownFCS-GX finds more contamination than NCBI's legacy screen and the all-by-all Conterminator scan; the authors attribute the gain to its larger reference database, cross-species h-mer alignment and ability to detect same-kingdom contaminants.

    On 14,344 eukaryote and 194,995 prokaryote genomes from the past 2.5 years: 0.163% contamination by length vs 0.038% for the legacy screen; confirmed 88% of Conterminator suspect sequences in 16,232 GenBank and 7,023 RefSeq assemblies and found an additional 138-146%. · Comparison of contamination detected by FCS-GX…; GenBank and RefSeq genome datasets

  5. ShownFCS-GX, combined with manual review and outreach to submitters, sharply reduced contamination in RefSeq and now screens new eukaryote genomes before they enter RefSeq.

    124 eukaryote genomes cleaned (79,593 sequences, 548 Mbp, including 34,337 genes and 30,356 proteins); 1,284 more prokaryote genomes removed; residual contamination 0.018% of prokaryote and 0.003% of eukaryote sequence; contaminant bases down 90% (eukaryotes) and 53% (prokaryotes) from 2020 peaks. · Cleanup of RefSeq genomes; Cleaning RefSeq genomes

Methods, data and reproducibility

Query scaffolds are split at N-runs and into 100 kbp chunks, masked for low-complexity sequence (and for transposon-like repeats in eukaryotes only), then aligned in two passes using codon-wobble-tolerant, purine/pyrimidine-reduced hashed k-mers (h-mers) against a 709 Gbp curated database; per-division coverage and scores are compared with the user tax-id to assign cleaning actions.

  • FCS-GX reference database r2023-01-24: curated genomes plus common contaminants (709 Gbp, 47,754 taxa), with contaminant sequences iteratively removed · multiple (prokaryotes, fungi, protists, animals, plants) · n = 28,213 RefSeq + 19,541 GenBank assemblies
  • Benchmark genomes: non-reference assemblies (contig N50 ≥100 kbp), one per genus; sequences ≥1 Mbp split into 1, 10 and 100 kbp fragments · multiple (prokaryotes and eukaryotes in five kingdom groups) · n = 663 prokaryote + 370 eukaryote assemblies
  • GenBank/INSDC genome assemblies listed in NCBI assembly_summary.txt as of 4/15/2023 (22.4 Tbp) · multiple (prokaryotes and eukaryotes) · n = 1,545,312 prokaryote + 30,053 eukaryote assemblies
  • Genomes released in the last 2.5 years that have legacy megaBLAST submission screening results · multiple (prokaryotes and eukaryotes) · n = 14,344 eukaryote + 194,995 prokaryote genomes
  • Assemblies with contaminant sequences flagged in the published 2019 Conterminator results · multiple (prokaryotes and eukaryotes) · n = 16,232 GenBank + 7,023 RefSeq assemblies · figshare project Conterminator/77346
  • Current NCBI RefSeq genome collection after FCS-GX cleanup · multiple (prokaryotes and eukaryotes) · n = 283,221 prokaryote + 1,616 eukaryote genomes
  • proGenomes2 genomes with family-level or higher lateral gene transfer events (HGT_data.txt from the promge.embl.de mobile genetic elements resource, Khedkar et al.) · prokaryotes · n = 3,986 genomes
Methods & tools
FCS-GX v0.4.0, FCS-GX reference database r2023-01-24, megaBLAST (legacy NCBI submission screen), VecScreen, BLAST, Conterminator, BlobToolKit, seqkit v.0.11.0, NCBI eutils v.19.0, ANI-based RefSeq screening, proGenomes2 lateral gene transfer annotations (promge.embl.de HGT_data.txt)
Software & environment
FCS-GX v0.4.0, written in C++ and Python and shipped as Docker and Singularity containers; database build r2023-01-24; seqkit v.0.11.0 and eutils v.19.0 used to prepare the benchmark.
Assay conditions & compute
Database sized for a 512 GiB server; loading takes 4-30+ minutes. Eukaryote screens ran on a Dell PowerEdge c6525 with 64 vCPUs and 1 TiB RAM (16 cores per job, up to 4 parallel jobs); prokaryotes ran on shared servers using 1-4 cores each.
Reproducibility gaps
No analysis scripts for the benchmark or the GenBank/RefSeq surveys are linked, and survey results were kept in a custom SQL database; screening needs the roughly 470 GiB database loaded in memory (sized for a 512 GiB server).
Limitations (6)
  • Sensitivity falls for contaminants from eukaryotic groups that are sparse in the database (median 17-63% for novel-species 1 kbp eukaryote fragments). This is the cost of keeping the database small enough for institutional clusters, and lab contaminants may need a much larger database. · stated by the authors
  • The tool is not built to resolve contaminants at lower taxonomic ranks, such as prokaryotic sequences below family level in multiisolates, MAGs or uncultured samples. It also ignores secondary evidence such as GC content or read depth. · stated by the authors
  • Intra-Metazoan calls may include false positives that are still under investigation (90% of such sequences are below 1 kb, median 279 bp). To limit false positives, prokaryote-in-prokaryote calls under 1% of genome length are not flagged for removal. · stated by the authors
  • Archived INSDC contamination can only be removed with the submitter's cooperation. For ENA-submitted genomes or unreachable submitters, only RefSeq sequences were suppressed, and chimeric sequences were revised only in significant cases. · stated by the authors
  • The benchmark's ground truth assumes long contiguous spans are contaminant-free and uses swapped tax-ids rather than real mixed assemblies. Outlier review found true contaminants among the apparent false positives, and no other tool was run on the same benchmark. · noted in this summary
  • The reported lateral gene transfer chimera rates (0.002% for 8/3,920 genomes; 0.004% for 2,227/534,869) look inconsistent with their own counts. Also, 3,986 genomes were collected but 3,920 were tested, so these rates should be rechecked against Table S15. · noted in this summary

Read this if you assemble, curate or reuse public genome assemblies and need a fast, automated screen for foreign sequences, or want quantitative estimates of contamination across GenBank and RefSeq.

Citations
481
FWCI
64.74
Field percentile
top 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
THISSteinegger '20Cornet '18Delafoy '21Alvarez '22Orakov '21Bálint '24
59% semantic matchBioinformatics
Martin Steinegger, Steven L. Salzberg·Genome biology·May 12, 2020· 247 citations