Skip to content

AlphaGenome

Summary

AlphaGenome takes up to 1 Mb of DNA sequence and predicts 5,930 human (or 1,128 mouse) functional genomic tracks across 11 output types, at resolutions down to a single base pair. It is the first model in the sequence-to-function family to break both of the field's standing trade-offs simultaneously — long context versus fine resolution, and multimodal breadth versus single-modality specialism — matching or exceeding the strongest available external model on 22 of 24 genome-track benchmarks and 25 of 26 variant-effect benchmarks.[1] Its practical value is that a single inference pass scores a variant across expression, splicing, chromatin and 3D contact simultaneously, which is what makes mechanistic (rather than merely deleteriousness-ranked) interpretation of non-coding variants possible.

Architecture

AlphaGenome uses a U-Net-inspired backbone that converts 1 Mb of input sequence, plus a species identity token (human or mouse), into two families of representations:[1]

Representation Resolution Used for
1D sequence embeddings 1 bp and 128 bp linear-genome track heads (expression, splicing, accessibility, ChIP)
2D pairwise embeddings 2,048 bp chromatin contact map prediction
  • Convolutional layers model local sequence grammar (motifs, splice sites, footprints); transformer blocks model coarser long-range dependencies such as enhancer–promoter interactions.[1]
  • Base-pair-resolution training over the full 1-Mb window is made tractable by sequence parallelism: the input is split into 131-kb chunks distributed across eight interconnected TPU v3 devices.[1]
  • Most track heads are linear transformations of the sequence embeddings. The exception is splice junction count prediction, which uses a dedicated mechanism modelling interactions between the 1D embeddings of candidate donor–acceptor pairs — the architectural novelty that lets the model reason about introns as objects rather than about splice sites in isolation.[1]

The 1-Mb context length was chosen because 99% (465 of 471) of experimentally validated enhancer–gene pairs fall within 1 Mb, per the ENCODE–rE2G CRISPRi encyclopedia.[1] See ENCODE-rE2G.

Predicted output types

Gene expression (RNA-seq, CAGE, PRO-cap); splicing (splice sites, splice site usage, splice junctions); chromatin state (DNase-seq, ATAC-seq, histone ChIP-seq, transcription-factor ChIP-seq); and DNA contact maps — 11 output types spanning many tissues, cell types and cell lines.[1] RNA-seq, PRO-cap, DNase, ATAC and all three splicing outputs are predicted at 1 bp; histone and TF ChIP-seq at 128 bp; contact maps at 2,048 bp.[1]

Training: pretraining then distillation

Training proceeds in two stages, and the distinction matters when reading the benchmark tables — different model variants are used for different evaluations.[1]

  1. Pretraining. 1-Mb intervals are sampled from the reference genome, augmented by random shift and 50% reverse-complementing, and fitted against observed experimental tracks. This yields two products: fold-specific models trained under 4-fold cross-validation (three quarters of the genome for training, one quarter held out) used for genome-track generalisation evaluations, and all-fold models trained on all intervals that serve as teachers.
  2. Distillation. A single student model with the same architecture is trained to reproduce the predictions of an ensemble of frozen all-fold teachers, on augmented and randomly mutated input sequences. The distilled student is the model used for variant effect prediction.

The distilled student scores a variant across every modality and cell type in a single device call in under 1 s on an NVIDIA H100, making genome-scale variant screening affordable relative to ensembling several independently trained models.[1]

Benchmark performance

Genome track prediction (fold-specific models, held-out intervals)

AlphaGenome outperformed the strongest external model on 22 of 24 track evaluations.[1] Selected relative improvements:

Task Comparator Relative improvement
Cell-type-specific gene-level expression log-fold change Borzoi +14.7%
Contact map, cell-type differences Orca +42.3%
Contact map, Pearson r Orca +6.3%
Transcription initiation (PRO-cap) total counts ProCapNet +15%
DNase profile Jensen–Shannon divergence ChromBPNet +9.5%
ATAC profile Jensen–Shannon divergence ChromBPNet +1.6%

Overall expression levels are predicted well, but capturing cell-type-specific expression deviations remains the harder task and the authors flag it as unsolved.[1]

Variant effect prediction (distilled student)

AlphaGenome matched or outperformed the best external model on 25 of 26 variant-effect benchmarks.[1]

Splicing — AlphaGenome predicts splicing at three levels (site probability, competitive site usage, and junction identity/strength) and sums per-modality scores into a composite variant scorer. It is SOTA on 6 of 7 splicing benchmarks:[1]

Benchmark AlphaGenome Best comparator
ClinVar deep intronic and synonymous (auPRC) 0.66 0.64 (Pangolin)
ClinVar splice site region (auPRC) 0.57 0.55 (Pangolin)
ClinVar missense, AlphaMissense-benign (auPRC) 0.18 0.16 (DeltaSplice, Pangolin)
MFASS minigene MPRA (auPRC) 0.51 0.54 (Pangolin)

MFASS is the single splicing benchmark AlphaGenome loses — to Pangolin, a model with a hundredth of its context window — though it still beats SpliceAI and DeltaSplice (both 0.49).[1] Notably, the splice junction scorer alone outperformed all previous approaches on every benchmark except deep-intronic/synonymous ClinVar and MFASS — direct evidence that junction-level modelling, not merely better site prediction, is where the gain originates.[1]

Gene expression — benchmarked against fine-mapped GTEx eQTLs:[1]

  • Effect-size magnitude: tissue-weighted mean Spearman ρ improved from 0.39 (Borzoi) to 0.49.
  • Direction of effect ("sign"): mean auROC improved from 0.75 (Borzoi) to 0.80.
  • On 17,675 fine-mapped GTEx eQTL SNVs, signed Spearman ρ = 0.50 against the SuSiE β posterior (unsigned ρ = 0.10; signed Pearson r = 0.39).
  • For high-confidence predictions (score above the 99th percentile of common-variant effects), Spearman ρ = 0.73 versus 0.61 for Borzoi.
  • Improvements held across most GTEx tissues, variant-to-TSS distance bins and functional annotation classes, and extended to indels.

For causality (distinguishing fine-mapped eQTLs from distance-matched variants), zero-shot performance was merely comparable to Borzoi; a supervised random forest over AlphaGenome scores from several modalities raised mean auROC from 0.68 to 0.75, exceeding Borzoi's 0.71 — and multimodal features beat RNA-seq-derived features alone, quantifying the benefit of the unified model for this task.[1]

Enhancer–gene linking — on the ENCODE–rE2G CRISPRi dataset, zero-shot AlphaGenome outperformed Borzoi (particularly beyond 10 kb from the TSS) and came within 1% auPRC of the ENCODE-rE2G (extended) model that was explicitly trained on this task and cell line. Adding AlphaGenome-derived features to ENCODE–rE2G (extended) set a new SOTA across all distance-to-TSS categories.[1]

Alternative polyadenylation — despite no explicit training on polyadenylation data, AlphaGenome achieves APA Spearman ρ = 0.894 (Borzoi 0.790) and beats Borzoi on paQTL discrimination at all distances (auPRC 0.629 vs 0.621 within 10 kb of the polyadenylation site; 0.762 vs 0.727 within 50 bp). In-silico mutagenesis shows it learned the canonical polyadenylation motif.[1]

Chromatin and TF binding — SOTA against both Borzoi and the specialist ChromBPNet across caQTL, dsQTL and bQTL benchmarks, across ancestries (African, European, Yoruba) and cell types (microglia, cardiac smooth muscle).[1] For African-ancestry causal caQTLs, predicted versus observed effect sizes gave signed Pearson r = 0.74 (unsigned 0.45); for SPI1 bQTLs, signed r = 0.55 (unsigned 0.12). In-silico mutagenesis attributed these to disrupted NF-κB and SPI1 motifs respectively.[1]

MPRA — on the CAGI5 saturation-mutagenesis MPRA challenge, cell-type-matched DNase predictions gave Pearson r = 0.57 (comparable to ChromBPNet and the Borzoi ensemble); LASSO over DNase features from all cell types raised this to 0.63; and LASSO over multimodal features across all cell types reached 0.65, a new SOTA.[1]

Variant scoring strategies

The paper defines a distinct scoring recipe per modality; the scores are not interchangeable:[1]

  • Expression: mask to the target gene's exons, then take the ALT-minus-REF difference in predicted coverage — an explicitly gene-directed score.
  • Splicing: take the maximum absolute REF-versus-ALT difference across splice sites or splice junctions; the composite scorer sums per-modality scores.
  • Accessibility and ChIP: a centre-mask strategy summing log-transformed signal in a local window around the variant.
  • Quantile calibration: raw scores are converted to quantiles against the distribution of common-variant effects, so a "quantile score of 1.0" means an effect exceeding 99% of common variants. The thresholds used in the paper's GWAS analyses are calibrated this way.

In-silico mutagenesis (ISM) over a window is used throughout to recover the sequence determinant of a prediction, producing motif logos matchable to JASPAR/UniPROBE matrices.

Multimodal locus interpretation: the TAL1 case

AlphaGenome was used to virtually screen three groups of non-coding gain-of-function mutations upstream of, within and downstream of TAL1 in T-cell acute lymphoblastic leukaemia, analysed in CD34⁺ common myeloid progenitor tracks (the closest available proxy for the T-ALL cell of origin).[1]

For the oncogenic insertion chr1:47239296 C>ACG — 7.5 kb from TAL1 — the model simultaneously predicted increased activating marks (H3K27ac, H3K4me1) at the variant, decreased repressive marks (H3K9me3, H3K27me3) near the TAL1 TSS, elevated H3K36me3 across the gene body, and increased TAL1 mRNA.[1] ISM on the alternative sequence recovered the MYB motif created by the insertion (matching the experimentally established mechanism) plus a second, nearby ETS-like motif of unknown role that was active only in the alternative sequence.[1] Across all reported oncogenic variants at the locus, predicted effects separated cleanly from length-matched shuffled controls, and the separation was strongest in T-ALL-relevant tracks (thymus, CMP, haematopoietic multipotent progenitors).[1]

This is the paper's clearest demonstration of why the multimodal design matters: no single-modality model reconstructs a chain of motif creation → neo-enhancer → chromatin state change → oncogene upregulation from one inference pass.

Application to trait-altering variants

Applying quantile thresholds to candidate causal trait-altering variants (338 Mendelian, 1,140 complex) versus matched controls (3,042 and 10,260), enrichment rose monotonically with threshold stringency — to 10.7-fold for Mendelian and 5.1-fold for complex variants at the 0.005 percentile — but at severe cost to recall, particularly for GWAS variants.[1]

Note on interpretation (stated by the authors): AlphaGenome predicts molecular consequences, not phenotypic ones. The authors explicitly caution that application to complex-trait analysis is limited because those phenotypes involve gene function, development, environment and gene-to-disease effects outside the sequence-to-function scope.[1] See Sequence-to-Function Genomic Models for how this boundary constrains downstream use.

Ablations: what actually drives performance

The ablation study is unusually informative for anyone designing or choosing such a model:[1]

  • Base-pair target resolution matters, selectively. 1-bp targets gave the best results, with the gain concentrated in tasks needing fine detail (splicing PSI5/PSI3, ATAC). Contact-map and histone ChIP correlation were largely insensitive to target resolution, as were variant-effect metrics that aggregate over gene bodies or exons.
  • Long context helps at both training and inference time. Models trained at 1 Mb beat models trained at ≤32 kb even when the latter were evaluated with 1-Mb context, and the 1-Mb-trained model degraded as inference context was shortened. Usefully, the 1-Mb-trained model evaluated at shorter contexts was often competitive with models trained at those matched shorter lengths — so context can be traded for speed at inference.
  • Distillation substitutes for ensembling. Distilling from 64 teachers produced single models competitive with, and sometimes better than, mean ensembles of several independently pretrained models. Distilling from even one teacher helped some variant-effect tasks.
  • Input mutation during distillation is essential. Removing random input mutation dropped eQTL sign by 0.06, eQTL causality by 0.01, sQTL causality by 0.01 and splicing outlier by 0.015 — the student must see perturbed sequences to learn variant effects rather than reference-genome memorisation.
  • Multimodal training helps, unevenly. The full multimodal model generally beat single-modality-group models, but the benefit was task-dependent: accessibility variant prediction was fine with accessibility data alone, whereas eQTL prediction benefited from the full multimodal signal. Excluding any single modality group caused only modest degradation, indicating substantial redundancy between modalities.

Limitations

Stated by the authors:[1]

  • Distal regulatory elements beyond ~100 kb remain poorly captured.
  • Tissue-specific and condition-specific variant effects are still difficult; cell-type-specific expression deviation is the weakest track-prediction axis.
  • Training data and evaluations are heavily protein-coding-centric; non-coding genes such as microRNAs are under-covered.
  • Species coverage is human and mouse only, and evaluation is primarily human.
  • Not benchmarked on personal-genome prediction, a known weakness of this model class.
  • No calibrated uncertainty estimates are provided.

Lab application

Internal analysis: AlphaGenome has been used in this group to interrogate loci where proximity-based variant-to-gene mapping fails. For rs603424, AlphaGenome predicts a large loss of DNase accessibility across stromal, fibroblast and adipocyte contexts and adipose-restricted RNA downregulation (quantile scores > 0.99) of both SCD and PKD2L1, supporting an adipocyte enhancer assignment that standard pipelines (and Open Targets) misattribute to PKD2L1 alone. The variant sits at a non-conserved base (phyloP100way = −0.48) — precisely the regime where a sequence model adds information over conservation-based deleteriousness scores. This analysis is unpublished; see the variant page for the supporting GTEx eQTL evidence.

Availability

See Also

  • Sequence-to-Function Genomic Models — the model family AlphaGenome belongs to, and the trade-offs it resolves.
  • Direction-of-Effect Assignment at GWAS Loci — how AlphaGenome's sign predictions complement colocalization.
  • Pangolin — the splicing specialist that beats AlphaGenome on MFASS despite a ±5 kb receptive field.
  • Enformer — the architectural ancestor of this lineage.
  • Borzoi — the predecessor RNA-seq-coverage model and AlphaGenome's principal multimodal comparator throughout the benchmarks above.
  • SpliceAI — the specialist splicing predictor AlphaGenome benchmarks against and surpasses on most tasks.
  • RNA-seq-Derived Molecular Phenotypes — what the sQTL/eQTL benchmarks above actually measure, and the u-sQTL gap AlphaGenome does not yet cover.
  • ENCODE-rE2G — the CRISPRi enhancer–gene benchmark AlphaGenome is evaluated on, and which AlphaGenome features improve.
  • AlphaMissense — the coding-variant counterpart from the same lab; AlphaGenome covers the non-coding complement and additionally detects splicing effects among variants AlphaMissense calls likely benign.
  • Google DeepMind — developer.

Citations

[1] Avsec, Latysheva, Cheng, Novati, Taylor et al. (2026), "Advancing regulatory variant effect prediction with AlphaGenome", Nature 649:1206–1217. Supports: architecture, training regime, all benchmark figures, variant scoring strategies, TAL1 case study, trait-variant enrichment, ablations, limitations and availability above. Location: Main text ("Unifying DNA sequence-to-function model", "Performance overview", "Improved splicing variant predictions", "Performance across gene expression tasks", "Improved prediction of chromatin accessibility, DNase sensitivity and binding QTLs", "Multimodal view of variant effects", "Model and data ablations", "Discussion"); Figs. 1–7; Data and Code availability statements. Source paper: s41586-025-10014-0.pdf