2026-09-28
MIT applies Enformer across modern human, archaic, and great-ape personalized genomes; supplement tables list 7,366 hominin-gain and 3,699 hominin-loss linCREs.
Millions of derived alleles separate humans from great apes, but only a thin slice is likely to move traits or disease risk, and most of that slice sits in noncoding DNA. Prior scans hunted regions of elevated sequence divergence and treated acceleration as a proxy for functional change. High-impact single mutations in otherwise quiet sequence fall through that net.
The bioRxiv PDF could not be retrieved because the host rate-limited this session. Method details below come from the authors' public code repository; counts come from the accompanying supplementary tables. Experimental thresholds, full statistics, and the discussion remain in the unread paper and are not invented here.
The public repo describes a genome-wide chromatin-accessibility scan over modern human, archaic hominin, and great-ape personalized genomes. The sequence-to-function backbone is DeepMind's Enformer; variant-effect scoring uses AlphaGenome. Predicted signal is peak-called with MACS3, then compared across a 35-way alignment to mark linCREs (lineage-specific cis-regulatory elements). Downstream steps include GREAT enrichment, EpiMap target linking, luciferase reporter tests, and model-interpretability tools that nominate motif-creating derived alleles.
The design choice is to compare predicted function on personalized genomes first, rather than gating on sequence acceleration. Sites that barely move at the sequence level can still enter the candidate set.
The abstract's qualitative claims: linCREs are shorter, less pleiotropic, and less conserved than conserved elements, and they enrich for neurodevelopmental pathways; many sit in low-divergence sequence that acceleration scans would miss; luciferase assays confirm lineage-specific enhancer activity, including one hominin-specific neurodevelopmental enhancer driven by a single motif-generating derived allele nominated by interpretability tools.
Supplementary tables from the official repo give these element counts (row counts of coordinate lists):
| Class | Table | N |
| Hominin gain | TableS02 | 7,366 |
| Hominin loss | TableS03 | 3,699 |
| Archaic gain | TableS04 | 76 |
| Archaic loss | TableS05 | 61 |
| Modern gain | TableS06 | 88 |
| Modern loss | TableS07 | 46 |
| Luciferase candidates | TableS13 | 15 |
Hominin gain and loss sets are about two orders of magnitude larger than archaic- or modern-specific sets. The luciferase candidate table names predicted targets such as CACNA1C, GRID1, and GRIN3B. Many GREAT rows have high adjusted p-values, so the neurodevelopmental-enrichment sentence should stay at abstract strength until the paper body is in hand.
For sequence-to-function and comparative-genomics work, this treats predicted functional difference as the evolutionary scan statistic, instead of using sequence speed as a stand-in. Personalized genomes plus Enformer can surface candidates outside accelerated regions. The 35-way alignment, peak sets, and scripts are public enough for others to rerun the scan.
Without full-text numbers, this should not be read as a finished mechanistic account of human-specific brain enhancers.
The PDF was not retrieved. Training-data makeup, differential-accessibility cutoffs, overlap with HAR/HAQER catalogs, luciferase effect sizes, and cell lines remain in the unread text. Enformer miscalibration on cross-species personalized genomes would manufacture false linCREs. Archaic coverage, contamination, and great-ape reference quality all bias the scan. Table row counts are not a substitute for the paper's own significance claims or experimental replicates.