AetherXeno
AetherXeno is a receptor-aware regulatory-variant interpretation platform for human drug-response circuitry. It exhaustively maps every single-nucleotide substitution across experimentally defined xenobiotic-receptor regulatory elements in primary human hepatocytes. Receptor-finetuned sequence models quantify allele-specific binding effects, which are then connected to population frequency, ClinVar, GTEx liver eQTLs, GWAS, motif evidence and a pharmacogene-to-drug mechanism layer called CAMIP.
Try a variant id like chr16_31093557_G_A or chr2_233757136_G_A, or a gene like CYP1A2 or CYP2C19. Coordinates are GRCh38, the hg38 assembly.
How it works
From condition-specific receptor occupancy, to saturation variant effects, to an evidence-linked mechanism layer. Select a stage to see what it does.
Primary human hepatocyte ChIP-seq maps receptor occupancy under the source study conditions: pooled DMSO and rifampicin data for PXR, GW4064 for FXR, and TCDD for AhR. These measurements define the regulatory search space.
What a score means
Each score is ALT minus REF. The model predicts receptor binding for the reference sequence and for the sequence carrying the variant, then sums the difference across the central bins. One score per receptor.
The CAMIP causal chain
For curated pharmacogenes, each chain explains how a regulatory variant can reach a drug, in mechanism only.
What is inside
17,267,831 scored variant effects, one row for every possible single base substitution inside a receptor regulatory element, each with a score for PXR, FXR and AhR.
8,691 regulatory elements: the top ChIP-seq peak summits by signal, used as the scored regions, up to 3,000 per receptor.
28,227 atlas variants also carry a ClinVar clinical annotation joined onto their receptor scores.
14,324 variants fall inside the high priority receptor peak tiers used for the curated analyses.
Regulatory elements per receptor
scored regulatory elements
scored regulatory elements
scored regulatory elements
Each element is a top ChIP-seq peak summit by signal, capped at 3,000 per receptor.
How the predictions are checked
The platform is evaluated with chromosome-held-out receptor occupancy and orthogonal functional evidence. What each check tests is below; full quantitative benchmarks are reported in the paper and supplement.
Gold-standard recovery
Known functional PXR luciferase response SNPs rank among the most strongly scored variants in their regions, so the model recovers established response elements.
Liver eQTL agreement
Variants the model scores highly are enriched for being GTEx liver eQTLs, an expression readout never seen in training. Browse them on the Evidence page.
Motif grammar
The bases the model is most sensitive to concentrate on known nuclear-receptor sequence motifs such as RXRA and NR1I2, rather than on flanking DNA.
Full benchmark statistics are in the paper. See the Documentation for methods.