dna-origin-benchmark

Labeled DNA sequence fragments for source classification, each 300 bp. The label is the source of the sequence: human, other eukaryote, bacterial, viral, engineered/synthetic, or undatabased (environmental sequence absent from reference databases). 24,950 fragments across 382 organisms.

Schema

field	description
id	unique fragment id
sequence	300 bp DNA fragment (ACGT)
class	natural, nonhost, engineered, or undatabased
subclass	human, eukaryote, bacteria, virus, real, recoding scheme, or metagenome
organism	source organism (or construct type / recoding scheme / metagenome)
split	train, test, test_novel, or undatabased

test_novel holds entire organisms and clades absent from train, for measuring generalization to unseen taxa. undatabased holds environmental fragments that Kraken2 against RefSeq cannot classify, for measuring reference-free behavior where database methods have no match.

Composition

class	subclass	fragments
natural	human	1,500
natural	eukaryote	3,760
nonhost	bacteria	5,587
nonhost	virus	3,087
engineered	real (NCBI synthetic/vectors)	7,302
engineered	recodings (ecoli/yeast/gc/motif/sampled)	1,560
undatabased	metagenome	2,154

split	fragments
train	16,323
test	4,181
test_novel	2,292
undatabased	2,154

Construction

Natural and non-host sequences are RefSeq records sampled across 94 eukaryote taxa (vertebrates through protists), 140 bacterial genomes, and 140 viral genomes, plus human. Engineered positives are NCBI synthetic-construct, codon-optimized, expression- and cloning-vector, and synthetic-plasmid records, together with six synonymous recoding schemes of the human set (E. coli- and yeast-optimized, GC-max, GC-min, human-frequency-sampled, and restriction-site-avoiding). Undatabased fragments are drawn from environmental metagenomes and kept only where Kraken2 against the RefSeq Standard-8 database returns unclassified.

Fragments are clustered with CD-HIT-est at 90% identity; splits are assigned at the cluster level so no cluster spans train and test, and 20% of the taxa in each non-host group are held out entirely as test_novel. The build is reproducible from a fixed seed.

Tasks

• Origin classification across the five labeled sources.
• Engineered vs natural detection.
• Human vs non-host detection.
• Reference-free behavior on undatabased sequence.

Baselines and head-to-head

AUROC for the reference-free closed-form classifier (dna-origin-classifier), beside the two database tools run on the same splits.

task	closed-form (reference-free)	Kraken2 (RefSeq DB)
human vs non-host (test)	0.993 AUROC	sens 0.972 / spec 1.000
human vs non-host (novel taxa)	0.990 AUROC	spec 1.000 (databased)
engineered vs natural (test)	0.919 AUROC	n/a
five-class origin accuracy (test)	0.708	n/a

Coverage on the undatabased split, where the sequence is absent from RefSeq:

method	coverage
Kraken2 (RefSeq)	0.0%
BLAST (vs train)	6.6%
closed-form classifier	100% (99.1% called non-human)

Database methods classify almost nothing here because there is no reference to match; the composition classifier needs none. On databased taxa Kraken2 is at least as strong; the closed-form classifier matches it while being reference-free and about 50,000x smaller.

Source

NCBI RefSeq, NCBI synthetic-construct records, and environmental metagenome records (source in the organism field).

License

CC-BY-4.0 for the compilation. Sequence data is from NCBI.