Skip to content

AlphaGenome — Avsec et al. (Google DeepMind)

One-line description — A unified DNA-sequence-to-function model that takes 1 Mb of DNA as input and emits 5,930 functional-genomics tracks at single-base-pair resolution across RNA expression, splicing, chromatin accessibility, histone marks, TF binding, and Hi-C contact maps — a U-Net with a transformer bottleneck, ~450M parameters, jointly trained on ENCODE + GTEx + FANTOM5 + 4D Nucleome for human and mouse, and matching or exceeding the strongest external models on 25 of 26 variant-effect-prediction evaluations.

What it does

AlphaGenome is a supervised sequence-to-track regressor — not a self-supervised foundation model like NT / DNABERT-2 / HyenaDNA / Evo. It maps a 1-Mb DNA window to thousands of predicted functional-genomics measurements at single-base resolution, jointly for human and mouse. The headline application is non-coding variant effect prediction: take a reference sequence and an alternate sequence carrying a candidate variant, run both through AlphaGenome, and compare the predicted track deltas at the variant locus. Outputs cover the major regulatory readouts that a wet-lab functional genomics core would assay — making the model a "Swiss-army knife for the non-coding 98% of the genome" (per the Trends in Genetics commentary by Spille et al. 2025) — and the first single model to deliver this range in a single inference call (per DeepMind blog).

How it works

Architecture. U-Net with transformer bottleneck (per DeepMind blog and architecture commentary in rewire.it 2025 and Hannigan 2025): 1. Convolutional encoder downsamples the 1-Mb input from 1-bp to 128-bp resolution, learning local sequence motifs (TF binding sites, splice signals, promoter elements). 2. Transformer bottleneck at the 128-bp resolution propagates information across the full 1-Mb window — modelling long-range enhancer–promoter interactions, isochore context, chromatin domain structure. 3. U-Net decoder upsamples back to 1-bp resolution with skip connections from the encoder, producing single-base predictions for the output tracks that need it (splice junctions, TSS coordinates). 4. Per-modality output heads emit the 5,930 human / 1,128 mouse functional tracks.

Implementation uses sequence parallelism across 8 interconnected TPU v3 devices — the 1-Mb input is partitioned into 131-kb chunks, with each TPU processing one interval, then aggregated for the transformer bottleneck (per Nature paper and the arxiviq deep-dive).

Parameter count. Approximately 450 million trainable parameters (per multiple secondary sources including Wasai Tech 2025 — the Nature paper does not report this in the abstract; the 450M figure traces to the DeepMind blog and is reported consistently across architecture explainer write-ups but is not a peer-reviewed primary number).

Tokenisation. No tokenisation in the LLM sense — input is one-hot encoded DNA at single-base resolution (4-channel: A, C, G, T, with N handled as the zero vector). This is the Enformer / Borzoi lineage rather than the NT / DNABERT / HyenaDNA tokenised-LM lineage.

Training data. - Reference genomes: human (GRCh38) and mouse (GRCm39) reference sequences. - Functional readouts (training targets): consortium-curated assay tracks grouped by assay type, cell ontology, and target factor: - ENCODE — RNA-seq, PRO-cap, DNase-seq, ATAC-seq, TF and histone ChIP-seq - GTEx — bulk-tissue RNA-seq (held out from public release for licensing; used in training) - FANTOM5 — CAGE (transcription start sites) - 4D Nucleome — Hi-C / Micro-C contact maps

Output modalities. 5,930 human tracks (1,128 mouse tracks), spanning: RNA expression, CAGE / PRO-cap (transcription initiation), splice sites + splice usage + splice junction strength, DNase / ATAC accessibility, histone-mark ChIP, TF binding ChIP, and Hi-C contact maps. Resolution: single base for RNA / splicing / chromatin tracks; per-bin (variable) for contact maps.

Training objective. Multi-task regression: per-track Poisson / Gaussian loss between predicted and observed coverage / signal, summed across modalities with modality-specific weights. Knowledge distillation is then applied — the published all_folds checkpoint is a student model distilled from an ensemble of four cross-validation-fold "teacher" models, balancing accuracy against inference cost.

Context length / input. 1,000,000 bp input window (extended internally by 4,096 bp padding — 2,048 bp each side — for boundary handling). Output predictions cover the central ~131 kb of the input at single-base resolution; the flanking sequence provides regulatory context but is not itself predicted. This is 5× Enformer's 196,608-bp context and ~2× Borzoi's 524-kb context.

Pretraining compute. Not reported precisely in the published paper. Training infrastructure is TPU v3 with sequence parallelism across 8 chips per replica; the multi-fold teacher-ensemble + distillation pipeline is significantly heavier than single-model training (per DeepMind blog — no peer-reviewed total TPU-hour figure available).

Downstream / variant scoring. Inference takes a reference 1-Mb window and an alt-allele 1-Mb window (typically centred on the variant), produces 5,930 tracks each, and reports per-track and aggregated effect scores. The DeepMind hosted API exposes this as a single REST call.

Headline benchmarks

  • Variant effect prediction: AlphaGenome matches or exceeds the strongest available external model on 25 of 26 evaluations spanning eQTL, sQTL, caQTL, MPRA, and clinical / GWAS variant scoring (per Nature paper abstract).
  • vs Enformer (DeepMind 2021): Direct successor; AlphaGenome ~5× longer context (1 Mb vs 196 kb), adds splicing and Hi-C output heads, unifies human and mouse, adds per-base resolution for splicing tracks. Outperforms Enformer across all comparable RNA-seq and chromatin tasks (per Nature paper).
  • vs Borzoi (Calico 2024): AlphaGenome's closest peer (Borzoi: 524-kb context, RNA-seq + epigenetic prediction, U-Net architecture). AlphaGenome wins on splicing-specific tasks and contact-map prediction (which Borzoi does not output); the two are comparable on bulk RNA-seq.
  • vs single-task specialists (Pangolin for splicing, SpliceAI for splice prediction, ChromBPNet for chromatin): AlphaGenome matches or beats the specialists in their own domain with a single unified model — the load-bearing methodological argument of the paper.
  • vs genomic FMs (NT, DNABERT-2, HyenaDNA, Evo): Not a fair comparison — those are self-supervised LMs evaluated by fine-tuning, AlphaGenome is supervised end-to-end. But on functional-track prediction as a downstream task, AlphaGenome dominates the genomic-FM fine-tunes because it was trained for it directly.

(Caveat: the "25 of 26" framing is from the Nature paper's own benchmarking suite, which DeepMind defines. Independent reproduction at this breadth is ongoing — the alphagenome_research code release in January 2026 [per STAT News, 28 Jan 2026] is what enables third-party replication. Per DeepMind blog — no peer-reviewed external validation of the full 26-task suite as of May 2026.)

Why it matters

AlphaGenome is the first model to unify the non-coding regulatory genomics stack (gene expression, splicing, chromatin, 3D contact) in a single end-to-end inference call at megabase scale — collapsing what was a dozen separate specialist models (Enformer, Pangolin, SpliceAI, ChromBPNet, Akita, Borzoi, …) into one. Three concrete consequences for the 2026 oncology and clinical-genomics translation effort:

  1. One model for non-coding variant interpretation. Pathogenic non-coding variants in cancer drivers (TERT promoter, MYC enhancers, PVT1) and Mendelian disease can now be scored uniformly across modalities — addressing the "98% of the genome is non-coding and we can't score it" problem that has dogged ACMG variant classification.
  2. Megabase context as the new default. Enformer's 196 kb was the previous ceiling; AlphaGenome at 1 Mb covers most TADs in a single window. The next generation of clinical variant pipelines is expected to default to 1 Mb context (Borzoi, AlphaGenome, Evo2).
  3. Free hosted API for non-commercial use. Unlike the prior AlphaFold release pattern (weights public, commercial use restricted), AlphaGenome adds a free hosted API alongside the weights — significantly lowering the integration cost for academic clinical-genomics groups. The non-commercial restriction is the binding constraint for industry uptake (see Limitations).

The model's relationship to the self-supervised FM lineage (NT, DNABERT-2, HyenaDNA, Evo, Evo2) is complementary, not competitive — AlphaGenome will not sample novel sequences or do in-context learning, and the FM lineage will not give per-base RNA-seq predictions out of the box. The expected 2026–2027 pattern is hybrid pipelines where Evo2-style FMs do variant discovery / sequence design and AlphaGenome scores the resulting variants for regulatory effect.

Known limitations

  • Non-commercial license is the binding industry constraint. The model weights are released under the AlphaGenome Model Terms — strictly non-commercial use, no distillation into derivative models, no commercial fine-tuning. This blocks pharma / biotech integration without a separate commercial agreement (per the HF model card).
  • No generative capability. AlphaGenome is purely a forward predictor; it cannot sample novel DNA sequences, do in-context learning, or score variants by sequence likelihood. The Evo2 / HyenaDNA lineage is needed for those use cases.
  • Trained on bulk-tissue / cell-line assays, not single-cell. Output tracks are bulk RNA-seq, bulk ChIP-seq, bulk Hi-C — not scRNA-seq, scATAC-seq, single-cell Hi-C. Cell-type-specific predictions are at the resolution of the input ENCODE/GTEx cell-type ontology (~hundreds of cell types and tissues), not single-cell resolution. This is the binding limitation for tumour-microenvironment applications where intra-tumoural heterogeneity is the question.
  • Human + mouse only. No fly, yeast, plant, or arbitrary-species support — unlike multispecies FMs (NT-multispecies-2.5B, Evo). Translational genomics outside the mouse-human axis requires either Borzoi-style retraining or a multispecies FM.
  • GTEx training data is held out from public release. Reproducibility of the GTEx-derived eQTL benchmarks requires GTEx access (which is dbGaP-controlled) — the released all_folds checkpoint includes GTEx-trained weights but the training tracks themselves are not redistributable.
  • The 450M-parameter figure is not in the peer-reviewed paper. It appears consistently in DeepMind blog text and secondary explainer write-ups but is not in the Nature abstract or the main-text architecture description; treat as DeepMind-asserted, not peer-validated.
  • Distillation to a single student model loses some teacher-ensemble accuracy. The published all_folds model is a distilled student; per-fold teacher models (fold-0 … fold-3) are available on HF Hub for users willing to ensemble at inference cost.
  • Context boundary effects. The 4,096-bp flanking pad helps but variants near the edge of the 1-Mb window have less surrounding regulatory context than variants in the centre.

Sources

Primary publications: - Nature paper (Jan 2026): Avsec et al., "Advancing regulatory variant effect prediction with AlphaGenome," Nature 649(8099), 2026. 10.1038/s41586-025-10014-0 - bioRxiv preprint (Jun 2025): 10.1101/2025.06.25.661532v1 — coincident with DeepMind blog release - DeepMind paper PDF mirror: storage.googleapis.com/deepmind-media/papers/alphagenome.pdf

DeepMind communications (vendor-only — flagged accordingly above): - Blog (Jun 2025): AlphaGenome: AI for better understanding the genome - Hosted API: deepmind.google.com/science/alphagenome

Code and weights: - API client (Apache 2.0): github.com/google-deepmind/alphagenome - Research code (Apache 2.0, released Jan 2026): github.com/google-deepmind/alphagenome_research - Weights (non-commercial): huggingface.co/google/alphagenome-all-folds — student model; per-fold teachers at google/alphagenome-fold-0 … fold-3

Press, commentary, and independent explainers: - STAT News (Jan 2026 source release): DeepMind releases AlphaGenome source code - STAT News (Jun 2025 launch): DeepMind launches AlphaGenome - Science / AAAS news: DeepMind's latest AI tool makes sense of changes in the human genome - Nature News: DeepMind's new AlphaGenome AI tackles the 'dark matter' in our DNA - C&EN (Jan 2026): Google's AlphaGenome predicts the function of a DNA sequence - Scientific American: DeepMind's AlphaGenome Uses AI to Decipher Noncoding DNA - Alzforum: Top Dog: AlphaGenome Predicts How Noncoding Variants Work - Trends in Genetics commentary: AlphaGenome, a Swiss-army knife for exploring non-coding DNA - Architecture deep-dives (independent): arxiviq.substack.com — rewire.it — livingbio.substack.com — binaryverseai.com