Home / Documentation
Documentation
How the atlas is built, what each field means, and how to cite it. For the live REST API and an interactive console, see the API page. For the external resources we link to, see Databases.
Data dictionary
| Field | Definition |
|---|---|
| ft_head_delta_{R} | Raw center bin sum of ALT minus REF on receptor R's CHIP_TF head |
| ft_head_delta_{R}_norm | Raw Δ divided by receptor-specific p99(|Δ|), clipped to [−1, 1]; not a percentile rank |
| effect_direction_{R} | LoF (< −0.1), GoF (> 0.1), or Neutral |
| ahr_exploratory | TRUE when the AhR head was trained on a single replicate at Gate 6B |
| in_{R}_peak | Variant overlaps receptor R ChIP-seq peak |
| multi_receptor_overlap | Comma list of receptors whose peaks overlap the variant |
| region_tier | receptor_confirmed (peak union) vs ATAC_only (extended) |
| gnomad_af | gnomAD v4.1 allele frequency |
| is_singleton | Legacy ingest flag for an observed row with AF=0; normally FALSE in the gnomAD-derived layer |
Methods
ChIP-seq on PXR, FXR and AhR in primary human hepatocytes is aligned with BWA MEM, peaks called with MACS2 and converted to BigWig. The AlphaGenome trunk is frozen while a CHIP_TF head is finetuned per receptor with Adam on an L0 plus center and peak weighted loss. Variant effect is the center bin sum of ALT minus REF from predict_sequence. Validation covers chromosome-held-out prediction of measured receptor occupancy against native AlphaGenome and Enformer, functional-variant recovery, GTEx liver eQTL enrichment and motif-core enrichment. The luciferase variants used during PXR input selection are reported as supporting recovery rather than as an independent test set.
Data availability and sources
AetherXeno is built entirely on public data. The receptor binding maps are reanalyses of published primary human hepatocyte ChIP-seq, all retrievable from their original archives. We add the finetuned variant effect layer and the cross-annotations; we credit and link every source.
| Layer | Source study | Accession | Condition |
|---|---|---|---|
| PXR ChIP-seq | Smith et al. 2014, PHH | SRR1642055/56/57 | input / DMSO / rifampicin |
| FXR ChIP-seq | Zhan et al. 2014, PHH | GSE57227 | GW4064 |
| AhR ChIP-seq (exploratory) | Bhattacharya et al. 2023, PHH | GSE205502 | TCDD, single replicate |
| Backbone model | AlphaGenome | alphagenome_research | non-commercial terms, finetuned per receptor |
Motif annotation uses the non-redundant JASPAR 2024 vertebrate collection: NR1I2 MA1533.2, Nr1H4 MA1110.3, Ahr::Arnt MA0006.2, RXRA MA0512.2, HNF4A MA0114.5 and FOXA1 MA0148.5. The latter collection profiles are sequence models and are used for FIMO overlap annotation; receptor-effect scores do not depend on motif calls.
The variant effect predictions are AlphaGenome Output, used under the AlphaGenome Output Terms of Use (non-commercial). Cross-annotation sources and their licenses are listed on the Databases page. The full derived atlas and annotation layers are deposited on Zenodo (see Download).
Citation
AetherXeno: receptor-aware saturation mapping and interpretation of regulatory variants in human drug-response circuitry. Variant effect predictions are AlphaGenome Output, subject to the AlphaGenome Output Terms of Use (non-commercial); AetherXeno annotations CC-BY-NC; code MIT. Permanent data archive: Zenodo DOI 10.5281/zenodo.20753827.