henhao
Sign in
GenomicsbioRxiv
Preprint

AlphaVaR: an R framework for the statistical interpretation of AlphaGenome variant-effect predictions

Karim MarhabaiD, Carlo MajiD, Johannes SchumacheriD, Pouria DasmehiD

Abstract

SummaryAlphaGenome (Google DeepMind) scores a DNA variant across thousands of functional tracks at single-base resolution, reporting both the magnitude of each predicted effect and its rarity against a genome-wide background. That volume is itself the obstacle to biological interpretation. Here we present AlphaVaR, an R package that gives AlphaGenomes output a typed structure together with the statistical methods and visualizations needed to interpret it. The output schema is identical for every variant, so the same tests apply throughout it. AlphaVaR provides localization tests with multiple-testing correction and effect sizes, a specificity index measuring how far an effect concentrates on a few elements of a chosen variable, and a transparent prioritization that ranks candidates across interpretable criteria and maps each to a target gene. Results feed a plot library, reproducible reports and a code-free Shiny application. Applied to rs1427407, the lead common variant for fetal-haemoglobin level, AlphaVaR recovers the established biology of the BCL11A erythroid enhancer. Availability and implementationhttps://github.com/KarimMarhaba/AlphaVaR, released under the MIT licence, R [≥] 4.2, with documentation at https://karimmarhaba.github.io/AlphaVaR/. The released version is archived at Zenodo (doi:10.5281/zenodo.21939265); the AlphaGenome scores analysed here are archived as a separate dataset (doi:10.5281/zenodo.21920988), and the scripts that regenerate every figure and reported number are in the repository (Supplementary Section S5). Contactpouria.dasmeh@uni-marburg.de Supplementary informationSupplementary data are available online.

PDFDOICode
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
0
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (21 Aug 2026) · CC BY 4.0

AlphaVaR structures AlphaGenome's per-track variant scores in R and adds multiple-testing-corrected enrichment and shift tests, a tissue-specificity index and transparent ranking; on rs1427407 it recovers BCL11A erythroid-enhancer biology.

AlphaGenome returns thousands of per-track scores for one variant, annotated by modality, tissue and assay, with no built-in way to reduce them to a variant-level readout. The paper describes existing tools as running AlphaGenome and plotting its tracks without statistical testing, and single-score predictors such as CADD, REVEL and AlphaMissense as collapsing variants to one scalar; AlphaVaR adds corrected tests, effect sizes and an interpretable ranking over that structure.

Key results

  1. ShownFor rs1427407, extreme scores concentrated in chromatin and active-regulatory modalities and were depleted in splice junctions, with a predicted gain in regulatory activity.

    One-vs-rest Fisher's exact test at the top-5% rarity default: ATAC odds ratio (OR) 18 (log2 OR 4.2), CAGE log2 OR 3.2, histone marks log2 OR 2.0; splice junctions OR 0.009. Raw signed scores indicated gains in all 11 modalities, with the same sign in 96% of ATAC tracks; 71,420 track predictions over two alleles. · 3 Application: rs1427407 in the BCL11A erythroid…; Which modalities carry the effect?; Figure legends

  2. ShownTissue specificity depended on which modality was read. Among the enriched modalities DNase gave the highest tissue specificity, and its ten top-scoring biosamples were all blood-forming.

    Specificity index tau across biosamples ranged from 0.06 (ATAC) to 0.30 (DNase) among enriched modalities. On quantile scores a hematopoietic multipotent progenitor led the DNase biosamples; on the raw score DNase tau rose to 0.84, with a CD34-positive common myeloid progenitor on top. · In which tissue/cell type does it act?; Figure legends

  3. SuggestedRanking transcription factors (TFs) by their effect on TF-binding tracks placed the erythroid TAL1, CBFA2T3 (ETO2), GATA1 and CBFA2T2 complex first, consistent with the variant lying in the enhancer's GATA1/TAL1 occupancy peak.

    Descriptive ranking over 751 profiled factors across all tissues, with no minimum-coverage filter. GATA1 and TAL1 are each profiled on only four K562 tracks, so applying such a filter would remove them and put a broadly profiled factor first. · Which regulators drive it?; Figure legends

  4. ShownBCL11A was identified as the target gene: the Tier-1 call depended on allele and aggregation method, and BCL11A's lead as top gene per hematopoietic tissue disappeared when tissues were pooled.

    Of 23 annotated genes in the 1 Mb interval (three protein-coding: BCL11A, PAPOLG, REL), av_map_targets gave BCL11A as Tier-1 target for T>C under mean, median and max aggregation, but for T>G only under max. Per hematopoietic tissue BCL11A ranked first in 14 of 34; with all tissues pooled it ranked sixth, behind neighbouring pseudogenes. · Which gene is the target?; Figure legends

  5. ShownImporting the case study takes under a second, and every later analysis finishes within 2 s, also on a 3.6-million-track cohort.

    Importing the 71,420 tracks of the case study took 0.33 s, and no subsequent analysis exceeded 2 s, including on a 3.6 million-track cohort (Supplementary Section S3); hardware not stated. · A typed object for the output

Methods, data and reproducibility

An R package loads AlphaGenome's long-format variant scoring CSV into an S3 object (AlphaVarSet) and runs one-vs-rest Fisher's exact and rank-based shift tests (Wilcoxon-Mann-Whitney, Kruskal-Wallis) with Benjamini-Hochberg correction and effect sizes, Yanai's tau specificity index, two-tier target-gene mapping and mean-percentile-rank prioritization with a Pareto front, applied to both alleles of rs1427407.

  • AlphaGenome variant scores for rs1427407 (chr2:60,490,908, hg38), alleles T>C and T>G, over 11 modalities and up to 714 biosamples · human (hg38 coordinates) · n = 71,420 track predictions · doi:10.5281/zenodo.21920988
  • Transcription-factor binding tracks within these scores, used for regulator ranking · human (hg38 coordinates; GATA1/TAL1 tracks in K562 cells) · n = 751 profiled transcription factors
  • Cohort used for runtime testing (Supplementary Section S3) · n = 3.6 million tracks
Methods & tools
AlphaVaR, AlphaGenome, R, Shiny, Fisher's exact test, Wilcoxon-Mann-Whitney test, Kruskal-Wallis test, Benjamini-Hochberg correction, tau specificity index (Yanai et al. 2005)
Code
https://github.com/KarimMarhaba/AlphaVaRhttps://karimmarhaba.github.io/AlphaVaR/
Software & environment
R >= 4.2; MIT licence; released version archived on Zenodo (doi:10.5281/zenodo.21939265); the GitHub repository includes scripts to rebuild all figures and numbers in the paper (see Supplementary Section S5).
Experimental conditions
Import of the 71,420-track case study took 0.33 s and no later analysis exceeded 2 s, including on a 3.6 million-track cohort (Supplementary Section S3); hardware not stated.
Reproducibility gaps
The analysed AlphaGenome scores for rs1427407 are deposited as a separate Zenodo dataset, doi:10.5281/zenodo.21920988, given as a DOI rather than a URL; hardware behind the timing figures is not stated; Bioconductor submission is future work.
Limitations (6)
  • Because one variant is scored on thousands of tracks, cross-track p-values suffer pseudo-replication and mostly reflect assay coverage, so interpretation rests on effect sizes; the minimum-coverage filter can also remove real regulators measured on few tracks. · stated by the authors
  • Specificity is diluted when pooled over modalities, so lineage emerges only after conditioning on the modality that defines the signal; quantile scores of many tracks hit their ceiling, and the sequence-only model predicts a broader footprint than the verified erythroid restriction. · stated by the authors
  • Tier-2 target assignment by distance to the transcription start site is a deliberate simplification; the priority rank only orders variants inside one locus and has no significance interpretation; phenotype-level causality is not inferred. · stated by the authors
  • No quantitative benchmark against single-score predictors (CADD, REVEL, AlphaMissense) is reported; the authors argue these collapse each variant to one value and that no established baseline exists for this kind of prioritization. · stated by the authors
  • Evaluation rests on a single positive-control locus (two alleles of rs1427407) with known biology; ranking several candidate variants with av_prioritize and the Pareto front is described but not exercised in the main-text case study, and no negative-control loci are tested. · noted in this summary
  • Key readouts shifted with analysis choices (T>G gave BCL11A only under max aggregation; pooling tissues dropped BCL11A to sixth; the TF ranking required dropping the coverage filter), and these choices were judged against an answer known in advance. · noted in this summary

Who should read:Read this if you run AlphaGenome on candidate non-coding variants and want an R workflow with corrected enrichment tests, effect sizes, specificity indices and target-gene calls instead of scanning thousands of tracks by hand.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISQu '26authors: '26Latysheva '26Kumar '26Cavus '26Kopanos '18
59% semantic matchGenomics
Hui‐Qi Qu, Hakon H. Hakonarson·bioRxiv·Oct 1, 2026