Synthetic Genomes Benchmark
Year: 2026
Type: software
Abstract: Benchmarks for evaluating generated DNA against matched natural sequence, accompanying Limitations of genomic language models for realistic sequence generation. The suite works on matched pairs of windows, with each natural locus paired to the synthetic sequence generated from it, and measures: Composition and divergence: GC, CpG observed/expected, homopolymer content, low complexity, entropy, canonical nullomer fraction, k-mer Jensen-Shannon divergence and chaos game representation distance. Calibration: whether that divergence exceeds the variation real genomic windows already show among themselves, tested by permuting the natural/synthetic label within each matched pair. Detectability: how separable natural and synthetic are to shallow k-mer models and to a dilated 1D ResNet. Context decay: how detectability grows with distance from the conditioning prompt. Biological structure: transcription factor binding site abundance and non-B DNA motifs (Z-DNA, G-quadruplexes, inverted, direct and mirror repeats, short tandem repeats). Also included are reusable interventions for constraining generation at training time (structural auxiliary losses) and at sampling time (logit constraints), and the archival scripts behind the paper's figures and statistics.
Institution: The University of Texas at Austin
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7203868091