Every base of the human genome swapped in silico: about 3.12 billion positions form a directed dependency network with seven recurring architectures
Synopsis
Using the DNA language model Nucleotide Transformer v3 on the JUPITER supercomputer, the team substituted each of roughly 3.12 billion positions of the telomere-to-telomere human reference genome with each of the three alternative bases and recorded how model predictions changed elsewhere, yielding a genome-wide directed map of nucleotide dependencies. These dependencies preferentially link evolutionarily conserved positions and repeatedly assemble into seven geometric structures called Relons, some outside current annotation. Combined with CADD, dependency features raised intronic-variant pathogenicity AUPRC from 0.299 to 0.356.
Interpretation
The human genome now has a base-resolution directed dependency map: each substitution is recorded as a directed edge to positions whose model predictions change, so the genome can be read as one weighted network. The genome had been described mainly through linear sequence, epigenetic state and three-dimensional contacts, with individual positions largely interpreted on their own; local nucleotide dependencies were established only in small contexts such as regulatory motifs and splice signals, and whether a genome-wide dependency organization exists was unclear. Dependency edges preferentially fall between evolutionarily conserved positions: against distance-matched controls, conserved-to-conserved edges are enriched 1.91-fold (95% CI 1.88–1.94) at 1–99 bp, rising to 23.3-fold (22.4–24.1) at 4–5 kb and remaining 10.1-fold (9.2–11.1) at 8.0–8.2 kb.
The dependencies are not spread uniformly; they follow different rules by genomic distance, chromosomal environment and sequence identity, and repeatedly assemble into seven nameable geometric structures. Retention of dependencies with distance differs by region: gene-rich and gene-poor regions keep only 1.8% and 1.0% of possible position pairs at 4–8 kb, whereas heterochromatic and centromeric regions keep 41% and 60%; after excluding repeat-associated pairs, retention at 4–8 kb falls to 2.0%, close to the 1.3% of repeat-free pairs in gene-rich regions. Under the strongest 0.1% edge filter, 1,286,832 Relons were identified genome-wide, with all seven classes present on every chromosome; class proportions are stable across chromosomes within two percentage points, and anti-diagonals are most frequent at 33.5%.
The dependency map is associated with human genetic variation and adds information to an established pathogenicity score, especially for intronic variants. Positions carrying pathogenic ClinVar variants have more outgoing (2.15-fold, 95% CI 2.09–2.22) and incoming (2.73-fold, 2.65–2.81) dependencies than benign positions; among phyloP-adjusted missense positions the ratios remain 1.52-fold (1.40–1.66) and 1.48-fold (1.36–1.61). Across 1,035,405 ClinVar variants, CADD reaches AUPRC 0.955, NTv3/T2T 0.620 and their combination 0.957; for intronic variants (282,095), NTv3/T2T reaches 0.412 versus CADD's 0.299, and the combination 0.356.
Perspective
The demonstrated scope is the T2T-CHM13v2.0 reference genome, 8 kb windows, simulated substitution responses from Nucleotide Transformer v3, and variant associations based on ClinVar and gnomAD; the authors propose comparisons across models, species and human haplotypes plus experimental perturbations to establish which patterns correspond to real biological interactions.
Dependencies reflect model responses to in silico substitutions rather than experimentally established molecular interactions, and edge direction does not establish biological causation; sequence homology, repetitive elements and model-specific properties may influence the observed patterns, and Relon detection depends on the applied thresholds and background models. The authors note that whether the enrichment of non-B DNA motifs in particular Relon classes reflects DNA structural organization remains to be explored.