Advancing regulatory variant effect prediction with AlphaGenome

Avsec, Žiga; Latysheva, Natasha; Cheng, Jun; Novati, Guido; Taylor, Kyle R.; Ward, Tom; Bycroft, Clare; Nicolaisen, Lauren; Arvaniti, Eirini; Pan, Joshua; Thomas, Raina; Dutordoir, Vincent; Perino, Matteo; De, Soham; Karollus, Alexander; Gayoso, Adam; Sargeant, Toby; Mottram, Anne; Wong, Lai Hong; Drotár, Pavol; Kosiorek, Adam; Senior, Andrew; Tanburn, Richard; Applebaum, Taylor; Basu, Souradeep; Hassabis, Demis; Kohli, Pushmeet

doi:10.1038/s41586-025-10014-0

Download PDF

Article
Open access
Published: 28 January 2026

Advancing regulatory variant effect prediction with AlphaGenome

Žiga Avsec ORCID: orcid.org/0000-0003-3119-0567¹^na1,
Natasha Latysheva¹^na1,
Jun Cheng¹^na1,
Guido Novati ORCID: orcid.org/0000-0003-0681-4892¹^na1,
Kyle R. Taylor¹^na1,
Tom Ward¹^na1,
Clare Bycroft ORCID: orcid.org/0000-0002-1139-0732¹^na1,
Lauren Nicolaisen¹^na1,
Eirini Arvaniti¹^na1,
Joshua Pan¹^na1,
Raina Thomas ORCID: orcid.org/0000-0001-8027-3460¹,
Vincent Dutordoir¹,
Matteo Perino¹,
Soham De¹,
Alexander Karollus¹,
Adam Gayoso¹,
Toby Sargeant¹,
Anne Mottram¹,
Lai Hong Wong¹,
Pavol Drotár¹,
Adam Kosiorek¹,
Andrew Senior ORCID: orcid.org/0000-0002-2401-5691¹,
Richard Tanburn¹,
Taylor Applebaum¹,
Souradeep Basu¹,
Demis Hassabis¹ &
…
Pushmeet Kohli ORCID: orcid.org/0000-0002-7466-7997¹

Nature volume 649, pages 1206–1218 (2026)Cite this article

205 Altmetric
Metrics details

Subjects

Abstract

Deep learning models that predict functional genomic measurements from DNA sequences are powerful tools for deciphering the genetic regulatory code. Existing methods involve a trade-off between input sequence length and prediction resolution, thereby limiting their modality scope and performance^1,2,3,4,5. We present AlphaGenome, a unified DNA sequence model, which takes as input 1 Mb of DNA sequence and predicts thousands of functional genomic tracks up to single-base-pair resolution across diverse modalities. The modalities include gene expression, transcription initiation, chromatin accessibility, histone modifications, transcription factor binding, chromatin contact maps, splice site usage and splice junction coordinates and strength. Trained on human and mouse genomes, AlphaGenome matches or exceeds the strongest available external models in 25 of 26 evaluations of variant effect prediction. The ability of AlphaGenome to simultaneously score variant effects across all modalities accurately recapitulates the mechanisms of clinically relevant variants near the TAL1 oncogene⁶. To facilitate broader use, we provide tools for making genome track and variant effect predictions from sequence.

Main

Interpreting the impact of genome sequence variation remains a central biological challenge. Non-coding variants, which reside outside of protein-coding regions, are particularly challenging to interpret because of the diverse molecular consequences they can elicit. For example, non-coding variants can modulate genome properties such as chromatin accessibility, epigenetic modifications and three-dimensional chromatin conformation. Variants can further influence messenger RNA (mRNA) availability by altering expression levels or modifying sequence composition through splicing changes. Additionally, variants can exhibit cell-type-specific or tissue-specific effects. Given that more than 98% of observed genetic variation in humans is non-coding⁷, global characterization of the complex effects of this vast majority of variants remains intractable without computational predictions.

Computational methods can learn patterns from experimental data to predict and explain variant effects. One class of methods,sequence-to-function models^1,2,3,4,5, takes a DNA sequence as input and predicts genome tracks, a data format associating each DNA base pair with a value (representing read coverage, count or signal) derived from experimental assays performed in cell lines or tissues. Genome tracks span various data modalities measuring gene expression (with output types comprising RNA sequencing (RNA-seq), cap analysis of gene expression (CAGE) sequencing, and precision nuclear run-on analysis of capped RNA (PRO-cap)), splicing (splice sites, splice site usage and splice junctions), DNA accessibility (DNase I hypersensitive site sequencing (DNase-seq)) and assay for transposase-accessible chromatin sequencing (ATAC-seq)), histone modification (chromatin immunoprecipitation sequencing (ChIP-seq)), transcription factor binding or chromatin conformation (high-throughput chromosome (Hi-C) or micrococcal nuclease-based (Micro-C) conformation capture). Successfully trained sequence-to-function models accurately predict experimental measurements from input sequences. Furthermore, by comparing genome track predictions from an alternative sequence versus a reference sequence, these models can predict the molecular effects of variants.

Currently, deep learning-based sequence-to-function models face two fundamental trade-offs constraining their ability to predict how variants affect diverse modes of biological regulation. First, often owing to computational limitations, models must trade off between capturing long-range genomic interactions and achieving nucleotide-level predictive resolution. Although models such as SpliceAI⁴, BPNet⁸ and ProCapNet⁹ provide base-resolution predictions, they are restricted to short input sequences (for example, 10 kb or less), and thus may miss the influence of distal regulatory elements. Models such as Enformer¹ and Borzoi² can process longer sequences (approximately 200–500 kb) to capture broader context but at the cost of reducing output resolution (128-bp or 32-bp bins), which can blur fine-scale regulatory features such as splice sites, transcription factor footprints or polyadenylation sites.

A second trade-off exists between capturing diverse modalities versus specializing in one or a few. Several state-of-the-art (SOTA) models are highly specialized for single modalities, such as SpliceAI⁴ for splice site prediction, ChromBPNet¹⁰ for local chromatin accessibility and Orca³ for three-dimensional genome architecture. However, specialized models alone are insufficient for capturing the diverse molecular consequences of variants across modalities. Even within a single modality like splicing, specialized models such as SpliceAI⁴ or Pangolin¹¹ predict certain aspects (such as splice site prediction) while omitting others (such as splice junction prediction or competition between splice sites). Models like DeepSEA, Basenji, Enformer, Sei and Borzoi have demonstrated the utility and practicality of multimodal models. They allow users to use a single model for several modalities, instead of requiring several specialized models. Furthermore, their learned general sequence representation enables them to be readily fine-tuned for new tasks. However, these more generalist models can lag behind their specialized counterparts on certain tasks, such as splicing, or may lack particular modalities, such as contact maps.

Here we present AlphaGenome, a model that unifies multimodal prediction, long-sequence context and base-pair resolution into a single framework. The model takes 1 Mb of DNA sequence as input and predicts a diverse range of genome tracks across numerous cell types. The splicing predictions of AlphaGenome include a new splice junction prediction approach alongside splice site usage prediction. We evaluated the performance of AlphaGenome using a comprehensive set of benchmarks, covering both its ability to accurately predict genome tracks on previously unseen DNA sequences and its effectiveness in variant effect prediction tasks. AlphaGenome achieved SOTA performance on 22 of 24 genome track prediction tasks and 25 of 26 variant effect prediction tasks. We performed extensive ablations of target resolution, sequence length, distillation and modality combinations to explain the performance of AlphaGenome and inform design choices for future sequence-to-function models. We envisage that AlphaGenome will provide a powerful and extensible foundation for analysing the regulatory code within the genome.

We first present key technical details of the AlphaGenome data and training procedure, alongside a high-level summary of our evaluations (Fig. 1). We then demonstrate high-fidelity genome track prediction performance, a prerequisite for variant effect prediction (Fig. 2). Next, we focus on variant effect prediction with modality-specific deep dives into splicing (Fig. 3), gene expression (Fig. 4) and chromatin accessibility (Fig. 5). Finally, we highlight the model’s utility in cross-modality variant interpretation (Fig. 6) and dissect the impact of modelling choices on the performance of AlphaGenome (Fig. 7).

Fig. 1: **AlphaGenome model architecture, training regimes and comprehensive evaluation performance.**

Fig. 2: **Example of AlphaGenome track predictions and detailed performance evaluations.**

Fig. 3: **AlphaGenome is a SOTA splicing variant effect prediction model.**

Fig. 4: **AlphaGenome predicts the effect of variants on gene expression.**

Fig. 5: **AlphaGenome accurately predicts variant effects on chromatin accessibility and SPI1 transcription factor binding.**

Fig. 6: **Interpreting variant effects across modalities with AlphaGenome.**

Fig. 7: **Impact of resolution, sequence length, ensembling, distillation and multimodal training on AlphaGenome performance.**

Unifying DNA sequence-to-function model

AlphaGenome is a deep learning model designed to learn the sequence basis of diverse molecular phenotypes from human and mouse DNA (Fig. 1a). It simultaneously predicts 5,930 human or 1,128 mouse genome tracks across 11 modalities covering gene expression (RNA-seq, CAGE and PRO-cap), detailed splicing patterns (splice sites, splice site usage and splice junctions), chromatin state (DNase, ATAC-seq, histone modifications and transcription factor binding) and chromatin contact maps. These span a variety of biological contexts, such as different tissue types, cell types and cell lines (see Supplementary Table 1 for the summary and Supplementary Table 2 for the complete metadata). These predictions are made on the basis of 1-Mb of DNA sequence, a context length designed to encompass a substantial portion of the relevant distal regulatory landscape. For instance, 99% (465 of 471) of validated enhancer–gene pairs fall within 1 Mb (ref. ¹²).

AlphaGenome uses a U-Net-inspired^2,13 backbone architecture (Fig. 1a and Extended Data Fig. 1a) to efficiently process input sequences into two types of sequence representations: one-dimensional embeddings (at 1-bp and 128-bp resolutions), which correspond to representations of the linear genome, and two-dimensional embeddings (2,048-bp resolution), which correspond to representations of spatial interactions between genomic segments. The one-dimensional embeddings serve as the basis for genomic track predictions, whereas the two-dimensional embeddings are the basis for predicting pairwise interactions (contact maps). Within the architecture, convolutional layers model local sequence patterns necessary for fine-grained predictions, whereas transformer blocks model coarser but longer-range dependencies in the sequence, such as enhancer–promoter interactions. Base-pair-resolution training on the full 1-Mb sequence is enabled through sequence parallelism across eight interconnected tensor processing unit (v3) devices. Genomic track predictions are linear transformations of these sequence embeddings, aside from splice junction count prediction, which uses a separate mechanism that captures interactions between one-dimensional embeddings of donor–acceptor pairs (Extended Data Fig. 1).

We trained the model using a two-stage process: pretraining and distillation. The pretraining phase (Fig. 1b) used the observed experimental data to produce two types of models. Fold-specific models were trained using a 4-fold cross-validation scheme (Methods), with three fourths of the reference genome used for training and the remaining one fourth held out for validation and testing. These models were then used to evaluate the generalization of AlphaGenome by predicting genomic tracks on unseen (test) reference genome intervals (Fig. 1b). Additionally, all-fold models were trained on all available intervals of the reference genome and served as teachers in the second stage (distillation; Fig. 1c). In the distillation phase, a single student model, sharing the pretrained architecture, was trained to predict the output of an ensemble of all-fold teachers using randomly augmented input sequences (Methods). This distilled student model, as shown previously¹⁴, achieved improved robustness and variant effect prediction accuracy in a single model instance, making predictions across all modelled modalities and cell types with a single device call per variant. Taking less than 1 s on an NVIDIA H100 GPU, the student model is highly efficient for large-scale variant effect prediction relative to the alternative approach of ensembling several independently trained models.

Performance overview

To characterize the model performance of AlphaGenome, we first assessed its generalization to unseen genome intervals, a prerequisite for high-quality variant effect prediction. We conducted 24 genome track evaluations, encompassing all 11 predicted modalities (Methods and Supplementary Table 3). For out-of-fold evaluations, pretrained, fold-specific AlphaGenome models were used and compared with the strongest available external model for each respective task. AlphaGenome outperformed these external models on 22 of 24 evaluations (Fig. 1d, Extended Data Fig. 3 and Supplementary Table 3). Notably, AlphaGenome exhibited a +14.7% relative improvement in cell-type-specific gene-level expression log-fold change prediction compared with Borzoi², another multimodal sequence model (Fig. 1e and stratified metrics in Extended Data Fig. 3e). AlphaGenome also outperformed specialized single-modality models on their respective tasks, such as Orca³ on contact maps (contact map Pearson r +6.3%; cell-type-specific differences +42.3%; Fig. 1d and Extended Data Fig. 4), ProCapNet⁹ on transcription initiation tracks (+15% total counts Pearson r; Extended Data Fig. 3f) and ChromBPNet¹⁰ on accessibility (+1.6% for ATAC; +9.5% for DNase profile Jensen–Shannon divergence; Extended Data Fig. 3g).

We next evaluated the model’s performance on predicting variant effects. We assembled a second set of 26 variant effect prediction benchmarks across gene expression, splicing, polyadenylation, enhancer–gene linking, DNA accessibility and transcription factor binding. Again, we compared with the strongest externally available model on each task (Methods and Supplementary Table 4). For variant effect prediction, we used the distilled student model. AlphaGenome matched or outperformed the external models on 25 of 26 evaluations (Fig. 1e and Supplementary Table 4). This included strong performance in quantitative trait locus (QTL) evaluations, such as sign prediction for expression QTLs (eQTLs; +25.5% versus Borzoi²) and accessibility QTL (+8.0% versus ChromBPNet¹⁰, averaged across five datasets; Methods), demonstrating its strength against both multimodal and specialized single-modality baselines. Collectively, these results demonstrate that AlphaGenome more accurately models both genome tracks and variant effects.

Improved track prediction performance

Given the strong performance of AlphaGenome on genome track evaluations, we investigated its track predictions in more detail. Fold-specific, pretrained AlphaGenome models demonstrated high concordance between predicted and observed read coverage on unseen genome intervals (Fig. 2a). As an example, predicted HepG2 genome tracks over the LDLR gene showcased strand-specific, base-pair-resolution RNA-seq coverage over exons, along with predicted splice sites, splice site usage and splice junction read coverage (Fig. 2b). More examples illustrating splicing, gene expression and chromatin track predictions are provided in Supplementary Figs. 1–3, and finer delineation of genomic features such as exon boundaries is highlighted in Supplementary Fig. 4.

Quantitatively, we observed strong Pearson correlations (r) between predicted and observed signals for functional genomics tracks in both human and mouse genomes (Fig. 2c), both across all tracks and when subsetting by biosample types or data sources (Supplementary Fig. 5). Although overall expression levels are predicted well, accurately capturing cell-type-specific expression deviations remains a challenging task (Fig. 2d and Supplementary Fig. 2j).

On splicing (Extended Data Fig. 2a), AlphaGenome accurately predicts splice sites (Extended Data Fig. 2b) and splice site usage (Extended Data Fig. 2b,c). It also accurately predicts quantitative splice junction read coverage and PSI5 and PSI3 within various tissues, achieving high correlation with experimental measurements (Fig. 2e, Extended Data Fig. 2b,d,e and Methods). Although AlphaGenome demonstrates the ability to predict tissue-specific alternative splicing in some instances (Supplementary Fig. 1), further improvements are needed to precisely predict intermediate splicing efficiencies and to capture tissue-specific nuances (Extended Data Fig. 2c,e).

Improved splicing variant predictions

One of the main ways genetic variants cause disease is by disrupting splicing¹⁵, a process that produces mature RNA sequences by excising introns and ligating exons at splice junctions. Splicing outcomes can be modelled at three levels: the probability that any given nucleotide acts as a splice donor or acceptor (splice site prediction)^4,11,16, competitive selection among potential splice sites (splice site usage prediction)^11,16 and prediction of specific introns (splice junction prediction). AlphaGenome predicts all three of these quantities alongside direct RNA-seq coverage prediction, thereby providing a more comprehensive view of the splicing-related molecular consequences of variants (Fig. 3a).

To illustrate the capacity of AlphaGenome to simultaneously predict several relevant splicing variant effects, we first probed its ability to recapitulate known biological outcomes. We interrogated a 4-bp deletion (chr. 3: 197081044: TACTC>T), a variant empirically observed to cause exon skipping in tibial artery tissue in a sample from the Genotype–Tissue Expression (GTEx)¹⁷ project (Fig. 3b). AlphaGenome accurately predicted this established consequence across all levels: a substantial reduction in the predicted usage of the affected exon’s splice site, loss of predicted junctions linking the skipped exon edges, emergence of a putative junction bypassing the exon and strong decrease in predicted RNA-seq coverage of the exon. Similarly, the predictions of AlphaGenome accurately captured the new splice junction and extended exon induced by the chr. 21: 46126238: G>C variant, an effect observed in a heterozygous GTEx RNA-seq sample (Fig. 3c). Finally, in silico mutagenesis (ISM), which systematically predicts effects of all possible single nucleotide variations in a sequence region (Methods), revealed the sequence determinants of the splicing predictions. For example, ISM analysis of exon 9 of the U2SURP gene and its flanking introns highlighted recognizable splicing-related sequence motifs^18,19 (Fig. 3d). Further examples of experimentally validated splice-disrupting variants identified in individuals with autism spectrum disorder⁴ are shown in Supplementary Fig. 6.

Building on the multifaceted splicing predictions of AlphaGenome, we developed a unified splicing variant scorer to systematically detect splice-disrupting variants. Specifically, we designed a custom variant scoring strategy for each prediction modality (Fig. 3e and Methods) and summed the individual scores to provide a composite measure of a variant’s predicted effect. We benchmarked this composite scorer against existing methods on a wide range of splicing-related variant effect prediction tasks. AlphaGenome performed best on fine-mapped splicing QTL (sQTL) classification^2,20, including both single-nucleotide polymorphisms (SNPs) within 10 kb of the closest splice site and proximal variants within 200 bp of splice sites (Fig. 3f). Furthermore, it achieved the highest performance at predicting rare SNPs and indels associated with GTEx splicing outliers in both supervised and unsupervised settings (Fig. 3g and Methods). We also evaluated the performance of AlphaGenome in distinguishing pathogenic variants from benign variants in ClinVar²¹, in which its composite scores outperformed the existing best method in each category across all three variant categories: deep intronic and synonymous (area under the precision–recall curve (auPRC) 0.66 versus 0.64 by Pangolin), splice region (auPRC 0.57 versus 0.55 by Pangolin) and missense (auPRC 0.18 versus 0.16 by DeltaSplice and Pangolin; Fig. 3h and Supplementary Fig. 7). When assessing the ability of AlphaGenome to predict whether rare variants disrupt splicing using data from a massively parallel splicing minigene reporter assay (multiplexed functional assay of splicing using SORT-seq (MFASS))²², it was outperformed by Pangolin (auPRC 0.54 versus 0.51) but surpassed SpliceAI and DeltaSplice (both auPRC 0.49; Fig. 3i). Notably, the splice junction scorer alone outperformed previous approaches on all benchmarks, except ‘deep intronic and synonymous’ ClinVar and MFASS variants, underscoring the importance of modelling splicing at the junction level. In summary, AlphaGenome achieves SOTA splicing variant effect prediction on six of seven benchmarks, providing a more comprehensive view of altered splicing events and transcript structure.

Performance across gene expression tasks

Beyond impacting isoform composition through splicing modulation, non-coding variants can influence traits and cause diseases by altering gene expression^23,24. We evaluated the ability of AlphaGenome to predict the impact of variants on gene expression across a range of regulatory mechanisms, including eQTLs, enhancer–gene interactions and alternative polyadenylation (APA).

Improved prediction of eQTL effects

We first evaluated the ability of AlphaGenome to predict the impact of eQTLs, which are variants associated with gene expression variation. We developed a variant scoring strategy to quantify a variant’s predicted effect on a gene’s expression level (Fig. 4a and Methods). An illustrative example of a known eQTL/sQTL locus (rs9610445; chr. 22: 36201698: A>C) is shown in Fig. 4b. The alternative ‘C’ allele is associated with lower expression of APOL4 in GTEx data (sum of single effects (SuSiE) ‘β posterior’ = −0.709; posterior inclusion probability (PIP) > 0.9), and both the RNA-seq coverage and direction of effect are recapitulated by the predictions of AlphaGenome (variant score = −1.52; quantile score = −1.00; Methods). Furthermore, ISM around the variant indicates a possible nucleotide splice donor sequence motif, which is disrupted by the variant, thereby leading to an aberrant transcript and reduced expression (Fig. 4b; inset). More examples of eQTL variant prediction and mechanistic interpretation by ISM are shown in Supplementary Fig. 8.

Using fine-mapped GTEx eQTL as ground truth, we benchmarked AlphaGenome against the SOTA models Borzoi² and Enformer¹. AlphaGenome demonstrated improved prediction of both the magnitude (‘coefficient’; Spearman ρ with SuSiE²⁵ β posterior) and direction (‘sign’; area under the receiver operating characteristic (auROC) curve) of eQTL effects compared with the previous SOTA (Borzoi) (Fig. 4c–f). AlphaGenome improved the tissue-weighted mean Spearman ρ from 0.39 to 0.49 and the mean sign auROC from 0.75 to 0.80. These improvements in coefficient and sign prediction were observed broadly across most of GTEx tissues, variant-to-transcription start site (TSS) distance bins and variant functional annotation classes (Extended Data Fig. 5a–c).

As previously reported²⁶, performance decays with distance to the target gene across all eQTL tasks (Extended Data Fig. 5f). However, AlphaGenome exhibited mild improvement on sign/coefficient prediction for distal variants (greater than 35 kb; Fig. 4f and Extended Data Fig. 5b), and top-scoring predictions exhibited high sign accuracy across distance categories (Extended Data Fig. 5g). Additionally, AlphaGenome outperformed Borzoi on coefficient and sign prediction for indel (insertion or deletion) eQTLs (Fig. 4c,e and Extended Data Fig. 5d). Notably, the effect size predictions of AlphaGenome for high-confidence eQTLs (scores greater than 99th percentile of common variant effects) highly correlate with observed effects (Spearman ρ = 0.73; Extended Data Fig. 5h) and consistently surpassed Borzoi’s performance across various quantile score thresholds (such as Spearman ρ 0.73 versus 0.61 at approximately 99th percentile threshold; Extended Data Fig. 5i).

The overall improvement in prediction accuracy, particularly for the sign of a variant’s effect, translates to substantial gains in practical applications. At a score threshold yielding 90% sign prediction accuracy, AlphaGenome recovered over twice as many GTEx eQTLs (41%) as Borzoi (19%; Fig. 4g). Applying this improved sign prediction capability to genome-wide association study (GWAS) interpretation, we evaluated the ability of AlphaGenome to assign a direction of effect to candidate target genes for 18,537 GWAS credible sets (Methods). Using a threshold calibrated to 80% accuracy on eQTLs (Fig. 4g), AlphaGenome assigned a confident sign prediction for at least one variant in 49% of GWAS credible sets (11% using a conservative PIP-weighted scoring approach; Fig. 4h). AlphaGenome and a widely used co-localization method for sign prediction (COLOC H₄ > 0.95)²⁷ resolved the direction of effect for largely non-overlapping sets of loci (Fig. 4h). This indicates their complementary utility, collectively increasing the total yield of loci with determined effect directions. Furthermore, AlphaGenome resolved approximately 4-fold more credible sets in the lowest minor allele frequency quintile compared with COLOC, probably reflecting its reduced dependence on population genetics parameters that affect power to detect associations (Fig. 4h, stratified bars). Thus, AlphaGenome expands our ability to generate functional hypotheses about the direction of GWAS signals, particularly for low-frequency variants.

The performance of AlphaGenome on distinguishing fine-mapped eQTLs from distance-matched variants (‘causality’; auROC) was comparable with Borzoi (Fig. 4i). However, leveraging the predictions of AlphaGenome within a supervised framework boosted performance on the causality task; training a random forest model using the scores of AlphaGenome from several modalities increased the mean auROC from 0.68 to 0.75, also surpassing previous SOTA performance (mean auROC 0.71 for Borzoi; Fig. 4i). Notably, using features derived from variant scores across all predicted modalities provided a performance uplift in this supervised setting compared with using RNA-seq-derived scores alone (Extended Data Fig. 5e), highlighting the practical benefit of the multimodal predictions of AlphaGenome for identifying causal expression-modulating variants.

Competitive enhancer–gene linking

We then assessed whether AlphaGenome can link enhancer elements to their target genes, given that tissue-specific gene expression is modulated by enhancer–promoter interactions, often involving enhancers in distal genomic regions. For this task, we leveraged an independent CRISPRi perturbation dataset from the ENCODE–rE2G study¹². Evaluated zero-shot, AlphaGenome outperformed Borzoi in identifying validated enhancer–promoter links, particularly for enhancers located beyond 10 kb from the TSS of the target gene (Fig. 4j, Extended Data Fig. 7a and Methods), although both models still underestimate the impact of very distal enhancers (Extended Data Fig. 7d). Furthermore, the zero-shot performance of AlphaGenome was comparable (within 1% auPRC) with the ENCODE–rE2G (extended) model, which was explicitly trained on this task and cell line data (Fig. 4j). It also strongly outperformed the simpler DNase-based ENCODE–rE2G model and a distance-to-TSS baseline. Beyond their stand-alone predictive power, features derived from AlphaGenome improved the supervised enhancer–promoter linking models. Incorporating AlphaGenome predictions into the ENCODE–rE2G (extended) model yielded a new SOTA performance across all distance-to-TSS categories (Fig. 4j, Extended Data Fig. 7b,c and Methods).

Altogether, these results demonstrate the improved capacity of AlphaGenome to capture long-range functional regulatory connections directly from sequence, which is vital for interpreting distal genetic variants.

Improved prediction of 3′ polyadenylation QTL effects

We further evaluated AlphaGenome on variants affecting APA. APA is a process that generates transcript diversity by varying the 3′ untranslated region (3′ UTR) of mRNA molecules, often with concomitant impacts on RNA half-life and tissue specificity²⁸. By predicting RNA-seq coverage, AlphaGenome intrinsically models the competition between proximal and distal polyadenylation signals (PASs) and thus can be used to detect APA. We found that AlphaGenome achieves SOTA performance in predicting APA (Spearman ρ = 0.894 versus Borzoi’s 0.790; Extended Data Fig. 6a and Methods)². Given this, we next examined its ability to distinguish 3′ polyadenylation QTLs (paQTLs) from expression-matched negatives (Fig. 4k and Methods). AlphaGenome outperformed Borzoi on this task (all paQTLs within 10 kb of PAS, auPRC 0.629 versus Borzoi’s 0.621; proximal paQTLs within 50 bp of PAS, auPRC 0.762 versus Borzoi’s 0.727). AlphaGenome outperformed Borzoi at all considered distances, although the predictive accuracy of both models declined with distance to the polyadenylation site (Fig. 4k and Supplementary Table 5). ISM revealed that AlphaGenome learned the relevance of the canonical polyadenylation motif and could detect variants that disrupt or create this motif (Extended Data Fig. 6c–h), although it occasionally underpredicted the expression impact (Extended Data Fig. 6i,j). Overall, these results demonstrate the improved accuracy of AlphaGenome in predicting genetic variant effects on 3′ UTR processing, a distinct regulatory mechanism, using only RNA-seq predictions and without any explicit training on polyadenylation data or associated variants.

Improved prediction of chromatin accessibility, DNase sensitivity and binding QTLs

We next assessed the ability of AlphaGenome to predict the effects of variants on chromatin states, focusing on QTLs associated with chromatin accessibility (caQTLs), DNase sensitivity (dsQTLs) and transcription factor binding (bQTLs). Variant effects were scored by comparing model predictions for reference and alternative alleles in a local window around the variant (Fig. 5a and Methods). Benchmarking performance using fine-mapped QTLs from diverse ancestries (using benchmarks from ChromBPNet²⁹), AlphaGenome consistently achieved SOTA performance compared with both Borzoi and the specialized ChromBPNet model. This was observed across QTL types and ancestries, both in predicting QTL causality (average precision; Fig. 5b and Supplementary Fig. 9) and in correlating with experimentally determined effect sizes (‘coefficient’; Pearson r; Fig. 5c and Supplementary Fig. 9). This strong performance generalized across more datasets, including European ancestry caQTLs, Yoruba ancestry DNase sensitivity QTLs and predictions within specific cell types such as microglia and cardiac smooth muscle cells. To illustrate this performance and gain mechanistic insight, we examined specific examples, including fine-mapped caQTLs (African ancestry) and SPI1 bQTLs³⁰. In both cases, AlphaGenome outperformed alternative models in causality prediction (precision–recall curves; Supplementary Fig. 9e,h), and its predicted effect sizes correlated with observed values (Fig. 5d,g; Pearson r = 0.74 for caQTLs and r = 0.55 for SPI1 bQTLs). Furthermore, ISM applied to high-impact variants revealed that predicted changes in accessibility or transcription factor binding scores generally corresponded to altered motifs for transcription factors known to modulate chromatin accessibility, such as NF-κB (Fig. 5e,f). For SPI1-specific variants, ISM further highlighted altered SPI1 motifs within the local sequence context (Fig. 5h,i).

Beyond population-level QTLs, understanding the impact of local sequence context on gene regulation is crucial. We therefore assessed the performance of AlphaGenome on the 5th Critical Assessment of Genome Interpretation (CAGI5) saturation mutagenesis massively parallel reporter assay (MPRA) challenge^31,32. MPRAs measure the regulatory activity of short DNA sequences (typically through reporter gene expression), a process closely linked to local chromatin accessibility and transcription factor binding. We therefore evaluated this benchmark using DNase, RNA-seq and ChIP output types, comparing with Enformer, Borzoi and ChromBPNet (Methods and Extended Data Fig. 8). Notably, the cell-type-matched DNase predictions of AlphaGenome achieved performance comparable to both ChromBPNet and the Borzoi Ensemble (Fig. 5j, top; Pearson r = 0.57). Furthermore, aggregating features from DNase predictions across all modelled cell types using least absolute shrinkage and selection operator (LASSO) regression further improved performance over cell-type-matched DNase features and the analogous approach with the Borzoi Ensemble model (Fig. 5j, middle; Pearson r = 0.63). Finally, by integrating features across several modalities and all cell types with LASSO, AlphaGenome achieved SOTA performance on the CAGI5 benchmark (Fig. 5j, bottom; Pearson r = 0.65). Overall, these analyses highlight the strong performance of AlphaGenome in predicting variant effects on chromatin states in QTL benchmarks, as well as its ability to leverage accessibility predictions to model the regulatory activity of sequences as measured by MPRAs.

Multimodal view of variant effects

We next investigated how the unified modelling approach of AlphaGenome can be used to virtually screen a locus. We evaluated three groups of gain-of-function mutations affecting the TAL1 gene in T cell acute lymphoblastic leukaemia (T-ALL): a cluster of 5′ neo-enhancer mutations upstream of the TAL1 TSS^6,33, an intronic single-nucleotide variant (SNV)³⁴ and a 3′ neo-enhancer³⁵. All three variant groups were shown to converge on a common mechanism—upregulation of the TAL1 oncogene (Fig. 6a,b).

To assess AlphaGenome, we analysed predictions in CD34⁺ common myeloid progenitor (CMP) data, the closest available match to the T-ALL cell of origin. For the oncogenic mutation chr. 1: 47239296: C>ACG⁶, AlphaGenome predicted increases in the activating histone marks H3K27ac and H3K4me1 at the variant, consistent with experimentally observed neo-enhancer formation at that position^6,33 (Fig. 6b). It also predicted decreased levels of the repressive histone marks H3K9me3 and H3K27me3 near the TAL1 TSS and elevated active transcription mark H3K36me3 across the TAL1 gene body, both of which are concordant with its predicted increase in TAL1 mRNA expression levels (Fig. 6b).

We expanded our analysis to include all oncogenic variants across studies^6,33,34,35, comparing each variant to a background set of length-matched and sequence-shuffled control variants. For the sole intronic SNV, we compared it with other possible SNVs at that site. Oncogenic mutations were predicted to increase TAL1 expression in CD34⁺ CMPs more than shuffled controls (Fig. 6c). Across modalities, oncogenic variants exhibited a distinct predicted mechanism compared with shuffled controls, as shown through unsupervised clustering (Fig. 6d). Finally, these observations are tissue-specific. By reproducing this analysis across all RNA-seq tracks, we found that the difference in oncogenic versus shuffled mutations is most pronounced in T-ALL-relevant tracks such as ‘thymus’, ‘CMP’ and ‘haematopoietic multipotent progenitor cells’ (Supplementary Fig. 10).

To understand the sequence determinants of the predictions of AlphaGenome in CD34⁺ CMP, we performed ISM on the reference and alternative sequences of the oncogenic chr. 1: 47239296: C>ACG variant. No mutations within 40 bp of the variant had a predicted effect on TAL1 expression in the reference sequence (Fig. 6e). By contrast, the alternative sequence introduced a MYB motif at the variant position predicted to increase TAL1 expression, chromatin accessibility and H3K27ac chromatin marks (Fig. 6e), as discovered previously⁶. The model additionally identified a second ETS-like motif nearby, which affected TAL1 expression in the alternative sequence but not in the reference sequence. The role of this motif is currently unknown.

We then quantitatively evaluated the utility of AlphaGenome for analysing trait-altering non-coding variants. Although models of molecular effect do not directly predict phenotypic consequences (Supplementary Note (trait-altering variants) and Supplementary Fig. 11), the predictions of AlphaGenome could nevertheless prove useful for interpreting the gene-regulatory implications of such variants. Setting a high threshold on the quantitative scores of AlphaGenome (for example, predicted expression change) strongly enriched for causal variant candidates among matched controls (Fig. 6f), although at the cost of low recall, particularly for GWAS variants. Inspection of some of these predictions for Mendelian disease variants (Extended Data Fig. 9a–h) illustrates how the multimodal outputs of the model can be used to simultaneously elucidate the effects of a variant on transcription factor binding, accessibility and gene expression while also predicting the direction of variant-induced expression changes. This makes AlphaGenome a useful complement to conservation-based measures of deleteriousness, which are agnostic to the mechanism of action of a variant.

Model and data ablations

To understand the contributions of key design and training decisions to the performance of AlphaGenome, we conducted ablation studies evaluating the impact of target resolution, sequence length, ensembling, distillation and multimodal learning (Fig. 7). First, we found that base-pair resolution is important for achieving the highest performance. Training with target tracks at 1-bp resolution consistently yielded the best results, particularly for tasks requiring fine-scale detail, such as splicing (PSI5 and PSI3) and accessibility (ATAC), in which performance generally declined as target resolution became coarser (Fig. 7a). By contrast, performance on tracks with coarser assay-specific resolution, such as contact map correlation or histone ChIP-seq correlation, was relatively insensitive to the target resolution used during training. Similarly, variant effect prediction metrics, which often aggregate effects over larger regions (such as gene bodies or exons), were also robust to resolution changes.

We next evaluated the performance of AlphaGenome across various combinations of training sequence lengths and inference context lengths, finding that training with 1-Mb input sequences and subsequently performing inference using the full 1-Mb context yielded the best overall results (Fig. 7b). First, models trained on longer DNA sequences (up to the full 1 Mb) outperformed those trained on shorter sequences (32 kb or less), even when the latter were evaluated using 1-Mb contexts (Fig. 7b (purple crosses) and Methods), underscoring the benefits of providing more sequence context during training. Second, inference also benefited from longer context. The 1-Mb-trained AlphaGenome model performed optimally when evaluated on the full 1-Mb sequence, with its performance progressively declining as shorter segments were used for inference input (Fig. 7b (blue dots)). Finally, the 1-Mb-trained model, even when assessed on shorter inference contexts, often achieved comparable performance to models that were specifically trained and evaluated at those same matched shorter lengths (Fig. 7b (green triangles)). This implies that the same model trained on the full 1-Mb sequence can also be used with a more limited context at inference time, providing an option for further increasing prediction speed if needed.

We compared distillation (Fig. 1c) with standard ensembling as a strategy to efficiently achieve high performance (Fig. 7c). Distillation using many teacher models (for example, 64 (orange crosses)) produced single models with performance that was competitive with, and sometimes surpassed, mean ensembles of several independently pretrained models (for example, the four-model ensemble (blue dots)). Distillation of even just a single teacher model was beneficial for some variant effect prediction tasks, including caQTL, splicing outlier and eQTL sign prediction tasks. Finally, distillation without randomly mutating the input sequence resulted in a drop in student model performance (eQTL sign, −0.06; eQTL causality, −0.01; sQTL causality, −0.01; splicing outlier, −0.015), affirming the impact of our input perturbation strategy during distillation. Overall, distillation offers a way to achieve strong performance with reduced inference cost compared with evaluating large ensembles, facilitated by scalable training strategies such as assigning one unique teacher per training device (Methods).

Lastly, we assessed the importance of multimodal learning (Fig. 7d and Supplementary Fig. 12). The fully multimodal model generally outperformed models trained on single modality groups, confirming the overall benefit of integrating diverse data types for learning shared representations (Fig. 7d). This benefit was task-dependent; for example, predicting accessibility variants was effective using only accessibility data, whereas predicting eQTLs benefited from the full multimodal input (a more complete set of experiments is shown in Supplementary Fig. 12). Complementary experiments showed that excluding single modality groups during training typically resulted in only modest performance decreases, suggesting redundancy between modalities (Supplementary Fig. 12, left panel). Training on single modalities was often detrimental to variant effect prediction performance (Supplementary Fig. 12, middle panel). Cumulative addition experiments showed that although track predictions improved most when their relevant data types were added, variant effect tasks benefited most from the initial inclusion of expression and accessibility data, with diminishing returns from subsequent modality additions (Supplementary Fig. 12, right panel). Together, these results demonstrate the value of the unified multimodal approach for achieving broad and high performance, particularly for tasks integrating diverse regulatory signals.

Discussion

AlphaGenome advances efforts to decipher the regulatory code of the genome, offering a unified deep learning model that simultaneously predicts diverse functional genomic signals from megabase-scale DNA sequences. It matches or surpasses specialized SOTA models in regulatory variant effect prediction, underscoring the model’s relatively robust grasp of fundamental DNA regulatory principles and its value for mechanistically interpretating non-coding variation. A core strength is its efficient multimodal variant effect prediction, which simultaneously scores variant impacts across all predicted modalities in a single inference pass. This integrated capability is crucial for understanding variants with complex mechanisms, as illustrated by the recapitulation of oncogenic TAL1 variant effects, and could power large-scale analyses that dissect regulatory sequence elements genome-wide. Furthermore, the new capability of AlphaGenome to directly model splice junctions enables a more holistic view of splice-disrupting variants.

Going forward, AlphaGenome holds promise across diverse biological disciplines. For molecular biology, AlphaGenome can serve as an engine for in silico experimentation, enabling rapid hypothesis generation and prioritization of resource-intensive wet-lab experiments. For rare disease diagnostic research, the improved variant effect predictions of AlphaGenome could provide further functional evidence to current annotation pipelines, offering new diagnostic avenues for non-coding variants of uncertain significance. The improved splicing, expression and accessibility predictions of AlphaGenome could be used ‘in the loop’ to accelerate sequence design applications, such as therapeutic antisense oligonucleotides³⁶ and tissue-specific enhancers^37,38. Finally, AlphaGenome can complement the capabilities of generative models trained on DNA sequences by predicting functional properties of newly generated sequences³⁹. To widely enable these broad applications, we provide accessible tooling to access AlphaGenome through a hosted model and API.

Despite its advances, AlphaGenome shares challenges common to current sequence-based models and has specific scope limitations. Accurately capturing the influence of distal regulatory elements (more than 100 kb away) remains a continuing objective. Moreover, although the model predicts tissue-specific and cell-type-specific genome tracks with some success, accurately recapitulating tissue-specific patterns across cellular contexts and predicting condition-specific variant effects remain challenging. Both our training data and evaluations are heavily focused on protein-coding genes, and future work can be done to improve coverage of non-coding genes such as microRNAs. Our species coverage remains limited to human and mouse, and the evaluations in this study are primarily human-focused. We have not yet benchmarked the model on personal genome prediction, which is a known weakness of models in this space^40,41. Finally, application to complex trait analysis is limited because AlphaGenome predicts molecular consequences of variants, whereas these phenotypes involve broader biological processes (including gene function, development and environmental factors) and gene-to-disease effects⁴² beyond the direct sequence-to-function scope of the model.

Addressing these limitations motivates several future research directions. Data generation efforts could increase input genome diversity by assaying more species or by perturbing non-coding regulatory elements at scale to help build the next generation of variant effect prediction models. Key computational avenues include refining variant prediction accuracy and utility (such as through task-specific calibration, fine-tuning on perturbational datasets or integration of single-cell data^43,44), incorporating a broader range of data modalities (such as DNA methylation and RNA structural features) and pursuing foundational model improvements (such as leveraging DNA language models^45,46,47, expanding multi-species capabilities and developing robust methods for assay bias correction^29,48). Moreover, integrating AlphaGenome predictions with other measures of variant effect, such as conservation-based scores, as well as existing data on gene function and biological pathways could prove useful in advancing common and rare-variant analysis. Finally, estimates of model certainty would aid in better interpreting predictions.

In summary, AlphaGenome provides a powerful and unified model for analysing the regulatory genome. It advances our ability to predict molecular functions and variant effects from DNA, offering valuable tools for biological discovery and enabling applications in biotechnology. Ultimately, AlphaGenome serves as a foundational step towards the broader scientific goal of deciphering the complex cellular processes encoded in DNA sequences.

Methods

For detailed descriptions of data acquisition and processing, model architecture, training procedures (pretraining and distillation) and benchmarking protocols against existing methods, refer to the Supplementary Information.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Data availability

All primary experimental datasets used for the training and evaluation of AlphaGenome in this study were obtained from publicly accessible sources. A comprehensive manifest detailing these data sources, including specific repositories (such as ENCODE portal, GTEx portal, 4D Nucleome portal, ClinVar and gnomAD), individual accession numbers, relevant version information and direct URLs where applicable, is provided in Supplementary Table 2. This study did not generate new primary experimental data requiring deposition.

Code availability

AlphaGenome is available for non-commercial use through an online API at http://deepmind.google.com/science/alphagenome, with an accompanying Python software development kit provided to interact with the model. We also provide a genome interpretation suite to facilitate the exploration and interpretation of AlphaGenome. This offers a range of functionalities, such as streamlined variant scoring with quantile calibration and identification of critical sequence regions through contribution scores from ISM-based experiments. The model source code, weights, variant scoring implementations and a selection of variant evaluation datasets and predictions are available at https://github.com/google-deepmind/alphagenome_research.

References

Avsec, Ž. et al. Effective gene expression prediction from sequence by integrating long-range interactions. Nat. Methods 18, 1196–1203 (2021).
Article CAS PubMed PubMed Central Google Scholar
Linder, J., Srivastava, D. & Yuan, H. Predicting RNA-seq coverage from DNA sequence as a unifying model of gene regulation. Nat. Genet. 57, 949–961 (2025).
Article CAS PubMed PubMed Central Google Scholar
Zhou, J. Sequence-based modeling of three-dimensional genome architecture from kilobase to chromosome scale. Nat. Genet. 54, 725–734 (2022).
Article CAS PubMed PubMed Central Google Scholar
Jaganathan, K. et al. Predicting splicing from primary sequence with deep learning. Cell 176, 535–548 (2019).
Article ADS CAS PubMed Google Scholar
Kelley, D. R., Snoek, J. & Rinn, J. L. Basset: learning the regulatory code of the accessible genome with deep convolutional neural networks. Genome Res. 26, 990–999 (2016).
Article CAS PubMed PubMed Central Google Scholar
Mansour, M. R. et al. Oncogene regulation. An oncogenic super-enhancer formed through somatic mutation of a noncoding intergenic element. Science 346, 1373–1377 (2014).
Article ADS CAS PubMed PubMed Central Google Scholar
Halldorsson, B. V. et al. The sequences of 150,119 genomes in the UK Biobank. Nature 607, 732–740 (2022).
Article ADS CAS PubMed PubMed Central Google Scholar
Avsec, Ž. et al. Base-resolution models of transcription-factor binding reveal soft motif syntax. Nat. Genet. 53, 354–366 (2021).
Article CAS PubMed PubMed Central Google Scholar
Cochran, K. et al. Dissecting the cis-regulatory syntax of transcription initiation with deep learning. Preprint at bioRxiv https://doi.org/10.1101/2024.05.28.596138 (2024).
Trevino, A. E. et al. Chromatin and gene-regulatory dynamics of the developing human cerebral cortex at single-cell resolution. Cell 184, 5053–5069 (2021).
Article CAS PubMed Google Scholar
Zeng, T. & Li, Y. I. Predicting RNA splicing from DNA sequence using Pangolin. Genome Biol. 23, 103 (2022).
Article CAS PubMed PubMed Central Google Scholar
Gschwind, A. R. et al. An encyclopedia of enhancer-gene regulatory interactions in the human genome. Preprint at bioRxiv https://doi.org/10.1101/2023.11.09.563812 (2023).
Ronneberger, O., Fischer, P. & Brox, T. U-Net: Convolutional networks for biomedical image segmentation. In Proc. International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI 2015) 234–241 (Springer, 2015).
Zhou, J., Rizzo, K., Tang, Z. & Koo, P. K. Uncertainty-aware genomic deep learning with knowledge distillation. Preprint at bioRxiv https://doi.org/10.1101/2024.11.13.623485 (2024).
López-Bigas, N., Audit, B., Ouzounis, C., Parra, G. & Guigó, R. Are splicing mutations the most frequent cause of hereditary disease? FEBS Lett. 579, 1900–1903 (2005).
Article PubMed Google Scholar
Xu, C. et al. Reference-informed prediction of alternative splicing and splicing-altering mutations from sequences. Genome Res. 34, 1052–1065 (2024).
Article CAS PubMed PubMed Central Google Scholar
GTEx Consortium The GTEx Consortium atlas of genetic regulatory effects across human tissues. Science 369, 1318–1330 (2020).
Article Google Scholar
Wang, Z. & Burge, C. B. Splicing regulation: from a parts list of regulatory elements to an integrated splicing code. RNA 14, 802–813 (2008).
Article CAS PubMed PubMed Central Google Scholar
Fairbrother, W. G., Yeh, R.-F., Sharp, P. A. & Burge, C. B. Predictive identification of exonic splicing enhancers in human genes. Science 297, 1007–1013 (2002).
Article ADS CAS PubMed Google Scholar
Alasoo, K. et al. Genetic effects on promoter usage are highly context-specific and contribute to complex traits. eLife 8, e41673 (2019).
Article PubMed PubMed Central Google Scholar
Landrum, M. J. et al. ClinVar: improving access to variant interpretations and supporting evidence. Nucleic Acids Res. 46, D1062–D1067 (2018).
Article CAS PubMed Google Scholar
Chong, R. et al. A multiplexed assay for exon recognition reveals that an unappreciated fraction of rare genetic variants cause large-effect splicing disruptions. Mol. Cell 73, 183–194 (2019).
Article PubMed Google Scholar
Albert, F. W. & Kruglyak, L. The role of regulatory variation in complex traits and disease. Nat. Rev. Genet. 16, 197–212 (2015).
Article CAS PubMed Google Scholar
Maurano, M. T. et al. Systematic localization of common disease-associated variation in regulatory DNA. Science 337, 1190–1195 (2012).
Article ADS CAS PubMed PubMed Central Google Scholar
Wang, G., Sarkar, A., Carbonetto, P. & Stephens, M. A simple new approach to variable selection in regression, with application to genetic fine mapping. J. R. Stat. Soc. B 82, 1273–1300 (2020).
Article MathSciNet Google Scholar
Karollus, A., Mauermeier, T. & Gagneur, J. Current sequence-based models capture gene expression determinants in promoters but mostly ignore distal enhancers. Genome Biol. 24, 56 (2023).
Article PubMed PubMed Central Google Scholar
Giambartolomei, C. et al. Bayesian test for colocalisation between pairs of genetic association studies using summary statistics. PLoS Genet. 10, e1004383 (2014).
Article PubMed PubMed Central Google Scholar
Tian, B. & Manley, J. L. Alternative polyadenylation of mRNA precursors. Nat. Rev. Mol. Cell Biol. 18, 18–30 (2017).
Article CAS PubMed Google Scholar
Pampari, A. et al. ChromBPNet: bias factorized, base-resolution deep learning models of chromatin accessibility reveal cis-regulatory sequence syntax, transcription factor footprints and regulatory variants. Preprint at bioRxiv https://doi.org/10.1101/2024.12.25.630221 (2025).
Tehranchi, A. K. et al. Pooled ChIP-seq links variation in transcription factor binding to complex disease risk. Cell 165, 730–741 (2016).
Article CAS PubMed PubMed Central Google Scholar
Kircher, M. et al. Saturation mutagenesis of twenty disease-associated regulatory elements at single base-pair resolution. Nat. Commun. 10, 3583 (2019).
Article ADS PubMed PubMed Central Google Scholar
Shigaki, D. et al. Integration of multiple epigenomic marks improves prediction of variant impact in saturation mutagenesis reporter assay. Hum. Mutat. 40, 1280–1291 (2019).
Article CAS PubMed Google Scholar
Liu, Y. et al. The genomic landscape of pediatric and young adult T-lineage acute lymphoblastic leukemia. Nat. Genet. 49, 1211–1218 (2017).
Article CAS PubMed PubMed Central Google Scholar
Liu, Y. et al. Discovery of regulatory noncoding variants in individual cancer genomes by using cis-X. Nat. Genet. 52, 811–818 (2020).
Article CAS PubMed PubMed Central Google Scholar
Smith, C. et al. TAL1 activation in T-cell acute lymphoblastic leukemia: a novel oncogenic 3′ neo-enhancer. Haematologica 108, 1259–1271 (2023).
Article CAS PubMed PubMed Central Google Scholar
Celaj, A. et al. An RNA foundation model enables discovery of disease mechanisms and candidate therapeutics. Preprint at bioRxiv https://doi.org/10.1101/2023.09.20.558508 (2023).
de Almeida, B. P. et al. Targeted design of synthetic enhancers for selected tissues in the Drosophila embryo. Nature 626, 207–211 (2023).
Article PubMed PubMed Central Google Scholar
Taskiran, I. I. et al. Cell-type-directed design of synthetic enhancers. Nature 626, 212–220 (2023).
Article ADS PubMed PubMed Central Google Scholar
Brixi, G. et al. Genome modeling and design across all domains of life with Evo 2. Preprint at bioRxiv https://doi.org/10.1101/2025.02.18.638918 (2025).
Sasse, A. et al. Benchmarking of deep neural networks for predicting personal gene expression from DNA sequence highlights shortcomings. Nat. Genet. 55, 2060–2064 (2023).
Article CAS PubMed Google Scholar
Huang, C. et al. Personal transcriptome variation is poorly explained by current genomic deep learning models. Nat. Genet. 55, 2056–2059 (2023).
Article CAS PubMed PubMed Central Google Scholar
Finucane, H. K. et al. Variant scoring performance across selection regimes depends on variant-to-gene and gene-to-disease components. Preprint at bioRxiv https://doi.org/10.1101/2024.09.17.613327 (2024).
Hingerl, J. C. et al. Scooby: modeling multimodal genomic profiles from DNA sequence at single-cell resolution. Nat. Methods 22, 2275–2285 (2025).
Lal, A. et al. Decoding sequence determinants of gene expression in diverse cellular and disease states. Preprint at bioRxiv https://doi.org/10.1101/2024.10.09.617507 (2025).
Dalla-Torre, H. et al. Nucleotide transformer: building and evaluating robust foundation models for human genomics. Nat. Methods 22, 287–297 (2025).
Article CAS PubMed Google Scholar
Karollus, A. et al. Species-aware DNA language models capture regulatory elements and their evolution. Genome Biol. 25, 83 (2024).
Article CAS PubMed PubMed Central Google Scholar
Benegas, G., Batra, S. S. & Song, Y. S. DNA language models are powerful predictors of genome-wide variant effects. Proc. Natl Acad. Sci. USA 120, e2311219120 (2023).
Article CAS PubMed PubMed Central Google Scholar
Hu, Y. et al. Multiscale footprints reveal the organization of cis-regulatory elements. Nature 638, 779–786 (2025).
Article ADS CAS PubMed PubMed Central Google Scholar
Ling, J. P. et al. ASCOT identifies key regulators of neuronal subtype-specific splicing. Nat. Commun. 11, 137 (2020).
Article ADS CAS PubMed PubMed Central Google Scholar
Wagner, N. et al. Aberrant splicing prediction across human tissues. Nat. Genet. 55, 861–870 (2023).
Article CAS PubMed Google Scholar
Cheng, J. et al. Accurate proteome-wide missense variant effect prediction with AlphaMissense. Science 381, eadg7492 (2023).
Article CAS PubMed Google Scholar
Mountjoy, E. et al. An open approach to systematically prioritize causal variants and genes at all published human GWAS trait-associated loci. Nat. Genet. 53, 1527–1533 (2021).
Article CAS PubMed PubMed Central Google Scholar
Rauluseviciute, I. et al. JASPAR 2024: 20th anniversary of the open-access database of transcription factor binding profiles. Nucleic Acids Res. 52, D174–D182 (2024).
Article CAS PubMed Google Scholar
Newburger, D. E. & Bulyk, M. L. UniPROBE: an online database of protein binding microarray data on protein–DNA interactions. Nucleic Acids Res. 37, D77–D82 (2009).
Article CAS PubMed Google Scholar
Benegas, G., Eraslan, G. & Song, Y. S. Benchmarking DNA sequence models for causal regulatory variant prediction in human genetics. Preprint at bioRxiv https://doi.org/10.1101/2025.02.11.637758 (2025).
Mohammadi, P., Castel, S. E., Brown, A. A. & Lappalainen, T. Quantifying the regulatory effect size of cis-acting genetic variation using allelic fold change. Genome Res. 27, 1872–1884 (2017).
Article CAS PubMed PubMed Central Google Scholar
Matsunaga, A. et al. Compound heterozygosity for an apolipoprotein A1 gene promoter mutation and a structural nonsense mutation with apolipoprotein A1 deficiency. Arterioscler. Thromb. Vasc. Biol. 19, 348–355 (1999).
Article CAS PubMed Google Scholar
Yüregir, G. T. et al. Hb H disease in a Turkish family resulting from the interaction of a deletional α-thalassaemia-1 and a newly discovered poly A mutation. Br. J. Haematol. 80, 527–532 (1992).
Article PubMed Google Scholar
Heibel, S. K. et al. N-carbamylglutamate enhancement of ureagenesis leads to discovery of a novel deleterious mutation in a newly defined enhancer of the NAGS gene and to effective therapy. Hum. Mutat. 32, 1153–1160 (2011).
Article CAS PubMed Google Scholar
Kaneko, K. et al. Identification of a novel erythroid-specific enhancer for the ALAS2 gene and its loss-of-function mutation which is associated with congenital sideroblastic anemia. Haematologica 99, 252–261 (2014).
Article CAS PubMed PubMed Central Google Scholar
Horn, S. et al. TERT promoter mutations in familial and sporadic melanoma. Science 339, 959–961 (2013).
Article ADS CAS PubMed Google Scholar
Pollard, K. S., Hubisz, M. J., Rosenbloom, K. R. & Siepel, A. Detection of nonneutral substitution rates on mammalian phylogenies. Genome Res. 20, 110–121 (2010).
Article CAS PubMed PubMed Central Google Scholar
Collins, F. S. et al. Concordance of a point mutation 5′ to the ^Gγ globin gene with Gγβ⁺ hereditary persistence of fetal hemoglobin in the black population. Blood 64, 1292–1296 (1984).
Article CAS PubMed Google Scholar
Graham, S. E. et al. The power of genetic diversity in genome-wide association studies of lipids. Nature 600, 675–679 (2021).
Article ADS CAS PubMed PubMed Central Google Scholar
Sun, B. B. et al. Plasma proteomic associations with genetics and health in the UK Biobank. Nature 622, 329–338 (2023).
Article CAS PubMed PubMed Central Google Scholar
Mbatchou, J. et al. Computationally efficient whole-genome regression for quantitative and binary traits. Nat. Genet. 53, 1097–1103 (2021).
Article CAS PubMed Google Scholar
Chen, J. et al. WNT7B promotes bone formation in part through mTORC1. PLoS Genet. 10, e1004145 (2014).
Article PubMed PubMed Central Google Scholar

Download references

Acknowledgements

We thank D. Hariharan, C. Taylor and O. Bertolli for product and program management support, and M. Beck and U. Okereke for legal counsel. We are grateful to Y. Assael and A. Botev for their contributions to modelling and engineering and to A. Trostanetski, L. Tenório, V. Johnston, R. Green and S. Chakera for their work on the API release. We also thank K. Tunyasuvunakool for feedback on the paper; R. Tremlett for assistance with graphic design and figure preparation; and former interns I. I. Taskiran, A.-A. Muşat, R. Khan and R. Yi for their contributions to the earlier stages of this research. Gemini (Google) was used for assistance with language editing and improving the clarity of the paper. This study extensively used publicly available datasets. We extend our sincere gratitude to the numerous research consortia (including, but not limited to, ENCODE, GTEx, FANTOM5, ACMG, OMIM and the 4D Nucleome Project), their members, data contributors and all study participants whose efforts in generating and openly sharing these foundational genomic resources were essential for this study. This study was funded by Google DeepMind.

Author information

These authors contributed equally: Žiga Avsec, Natasha Latysheva, Jun Cheng, Guido Novati, Kyle R. Taylor, Tom Ward, Clare Bycroft, Lauren Nicolaisen, Eirini Arvaniti, Joshua Pan

Authors and Affiliations

Google DeepMind, London, UK
Žiga Avsec, Natasha Latysheva, Jun Cheng, Guido Novati, Kyle R. Taylor, Tom Ward, Clare Bycroft, Lauren Nicolaisen, Eirini Arvaniti, Joshua Pan, Raina Thomas, Vincent Dutordoir, Matteo Perino, Soham De, Alexander Karollus, Adam Gayoso, Toby Sargeant, Anne Mottram, Lai Hong Wong, Pavol Drotár, Adam Kosiorek, Andrew Senior, Richard Tanburn, Taylor Applebaum, Souradeep Basu, Demis Hassabis & Pushmeet Kohli

Authors

Žiga Avsec
View author publications
Search author on:PubMed Google Scholar
Natasha Latysheva
View author publications
Search author on:PubMed Google Scholar
Jun Cheng
View author publications
Search author on:PubMed Google Scholar
Guido Novati
View author publications
Search author on:PubMed Google Scholar
Kyle R. Taylor
View author publications
Search author on:PubMed Google Scholar
Tom Ward
View author publications
Search author on:PubMed Google Scholar
Clare Bycroft
View author publications
Search author on:PubMed Google Scholar
Lauren Nicolaisen
View author publications
Search author on:PubMed Google Scholar
Eirini Arvaniti
View author publications
Search author on:PubMed Google Scholar
Joshua Pan
View author publications
Search author on:PubMed Google Scholar
Raina Thomas
View author publications
Search author on:PubMed Google Scholar
Vincent Dutordoir
View author publications
Search author on:PubMed Google Scholar
Matteo Perino
View author publications
Search author on:PubMed Google Scholar
Soham De
View author publications
Search author on:PubMed Google Scholar
Alexander Karollus
View author publications
Search author on:PubMed Google Scholar
Adam Gayoso
View author publications
Search author on:PubMed Google Scholar
Toby Sargeant
View author publications
Search author on:PubMed Google Scholar
Anne Mottram
View author publications
Search author on:PubMed Google Scholar
Lai Hong Wong
View author publications
Search author on:PubMed Google Scholar
Pavol Drotár
View author publications
Search author on:PubMed Google Scholar
Adam Kosiorek
View author publications
Search author on:PubMed Google Scholar
Andrew Senior
View author publications
Search author on:PubMed Google Scholar
Richard Tanburn
View author publications
Search author on:PubMed Google Scholar
Taylor Applebaum
View author publications
Search author on:PubMed Google Scholar
Souradeep Basu
View author publications
Search author on:PubMed Google Scholar
Demis Hassabis
View author publications
Search author on:PubMed Google Scholar
Pushmeet Kohli
View author publications
Search author on:PubMed Google Scholar

Contributions

Z.A. conceptualized the study with input from N.L., J.C., C.B., T.W., G.N., K.R.T., P.K. and D.H. Z.A. led the project, with N.L. leading manuscript preparation, G.N. leading overall model development, J.C. conceiving and leading the splicing-related research and K.R.T. and T.W. leading engineering. Z.A. and G.N. conceived the device-distributed model architecture with input from R.T. G.N. conceived the contact-map pair stack. J.C. and G.N. conceived and developed the splice junction prediction head, with contributions from Z.A. Z.A. and G.N. conceived model distillation with input from J.C., V.D. and K.R.T. G.N. developed the model with contributions from Z.A., J.C., K.R.T., T.W., E.A., V.D., S.D., A.G., A.M., A. Kosiorek and A.S. Z.A., J.C. and G.N. curated, processed and prepared the training data with contributions from L.N., K.R.T., T.W., J.P., M.P., T.S. and A.M. T.W. and T.S. developed the data pipeline with contributions from Z.A., J.C., G.N., K.R.T. and L.N. Z.A., J.C., K.R.T., T.W., L.N. and E.A. designed the variant scoring infrastructure with contributions from N.L., C.B., R.T., T.A. and G.N. G.N., K.R.T., T.W., L.N., E.A. and V.D. optimized the model and variant scoring runtime. C.B. conceived and developed the variant score normalization with contributions from Z.A. and T.W. T.W. conceived and developed the model serialization and inference infrastructure with contributions from Z.A., K.R.T. and L.N. L.N. developed the evaluation infrastructure with contributions from N.L., K.R.T., T.W., R.T., V.D. and A.M. Z.A., N.L., J.C., G.N., C.B., L.N., E.A., J.P., R.T., V.D., M.P., A. Kosiorek, A.G., A.M. and P.D. evaluated the model. Specific evaluation areas were covered as follows: Z.A. and V.D. covered track evaluation; J.C. and E.A. covered splicing track and variant evaluation; N.L. and C.B. covered eQTL; E.A. and J.P. covered enhancer-to-gene linking; A. Karollus covered eQTL indel effects; N.L. covered accessibility QTL; J.P. covered multimodal examples and ACMG case studies; N.L. and G.N. covered contact maps; C.B. covered GWAS and complex trait analysis with help from A. Karollus, J.P. and T.A.; R.T. and M.P. covered paQTL and CAGI5; A. Karollus covered TraitGym. Z.A., N.L., J.C., G.N., K.R.T., T.W., C.B., L.N., E.A., J.P., R.T., V.D., M.P. and A. Karollus analysed data and prepared figures. Z.A., N.L., J.C., G.N., K.R.T., T.W., C.B., L.N., E.A., J.P., R.T., V.D., M.P., A. Karollus, A.G. and S.B. wrote the paper. L.H.W. managed and coordinated the project planning and execution. Z.A. and P.K. managed the project. All authors reviewed the paper.

Corresponding authors

Correspondence to Žiga Avsec or Pushmeet Kohli.

Ethics declarations

Competing interests

This study was conducted in the course of employment at Google DeepMind, with no competing financial interests.

Peer review

Peer review information

Nature thanks Jeremy Guez, Konrad Karczewski and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Extended data figures and tables

Extended Data Fig. 1 AlphaGenome model architecture.

(a) The architecture follows a U-Net-like structure with an Encoder, a central Transformer Tower, and a Decoder processing a 1-Mb DNA input sequence. The Encoder uses convolutional blocks and max pooling to progressively downsample the sequence resolution (from 1 bp to 128 bp) while increasing feature channels. The Transformer Tower operates at 128 bp resolution, iteratively refining sequence representations and generating pairwise (2D) representations. The Decoder uses convolutional blocks and upsampling, incorporating skip connections (dashed lines) from corresponding Encoder stages, to restore sequence resolution up to 1 bp. An Output Embedder performs final processing before feeding representations to task-specific output heads. (b) Internal structure of key component blocks used repeatedly within the architecture overview shown in (a). Diagrams detail the layers within the convolutional blocks (Conv block, Upres block), the Transformer blocks, and the blocks responsible for generating and updating pairwise representations (Pair update block, Sequence to pair block). Tensor shapes are shown excluding the batch dimension. Abbreviations: r = log-resolution, c = channels.

Extended Data Fig. 2 Splicing track performance.

(a) Schematic overview of splice site (SS) classification, splice site usage (SSU) prediction, and splice junction (SJ) read count prediction tasks. (b) (left) Performance comparison (AUPRC) of SS classification and SJ classification against reference methods. ‘Baseline’ means the fraction of positive splice junctions in the evaluated data. Splice site classification is evaluated with both GTF (GENCODE v46) annotated splice sites only and also splice sites derived from GTEx RNA-seq data (Methods). Splice junction classification discriminates between true splice junctions observed from RNA-seq data versus false junctions not observed from RNA-seq (but where the splice sites are observed). Splice junction classification was evaluated per tissue and then the mean AUPRC across tissues were reported. (right) Performance comparison (Pearson r) of predicted vs. measured SSU and SJ counts (log(1+x) transformed). (c) Scatter plot between predicted and measured donor SSU across seven example human tissues (from GTEx). Pearson r in each tissue is displayed as text. (d) Scatter plot between predicted and measured splice junction counts across seven human tissues (from GTEx). Pearson r in each tissue is displayed as text. (e) Distribution of Pearson correlation coefficients between predicted and measured PSI3 per tissue (left), PSI5 per tissue (middle), and junction counts across tissues (measuring tissue specificity of the splice junction predictions).

Extended Data Fig. 3 Track-level performance benchmarking.

Performance comparison of AlphaGenome with Enformer and Borzoi on held-out genomic track prediction. (a, b) Comparison of AlphaGenome test set performance on Enformer human tracks (each dot is one track) against Enformer models either (a) not fine-tuned or (b) fine-tuned on human data (the main released Enformer version). AlphaGenome model was re-trained for direct comparability using matched training intervals and an additional Enformer prediction head (Methods). (c) Evaluation of RNA-seq prediction performance at base and gene resolution using the same source of RNA-seq data as Borzoi, but processed at base-resolution and not scaled (Methods). Borzoi’s 32 bp RNA-seq predictions were upsampled and unscaled to the original scale for comparison. The larger performance difference observed on the normal scale (first column) likely reflects resolution differences at exon-intron boundaries. This difference decreased when using log(1+x) transformed values (second column), suggesting better agreement on overall gene expression levels. A similar trend was observed when aggregating expression per gene (average exon coverage, third column). Cell-type specificity was evaluated by correlating quantile-normalized, mean-subtracted expression profiles across genes (fourth column) and across tracks (fifth column). (d) Test set performance comparison of AlphaGenome against Borzoi (fold 1) on Borzoi track data at 32 bp resolution (each dot is one track). AlphaGenome was fine-tuned with an additional Borzoi head at matched resolution (Methods). (e) Stratification of cell-type specific prediction accuracy. The per-gene log-fold change correlation performance (from panel c, fourth column) was stratified by gene characteristics: median expression level across tissues (Median TPM; quintile breakpoints: 5.5×10⁻⁹, 4.1×10⁻⁴, 8.1×10⁻⁴, 0.17, 4.1, 3.6×10⁴ TPM), number of tissues with the gene expressed (TPM ≥ 0.001; quintile breakpoints: 9.4×10⁻⁸, 9.4×10⁻⁴, 8.0, 52, 54, 54 tissues), and housekeeping gene status. Sample sizes in brackets are the number of genes in each category. Box plots display the median (center line), the 25th and 75th percentiles (box bounds), and the whiskers extend to 1.5 times the interquartile range from the box bounds; points beyond whiskers indicate outliers. (f, g) Performance comparison of AlphaGenome against (f) ProCapNet (on PRO-Cap data) and (g) ChromBPNet (on ATAC and DNase). Evaluation was performed on ProCapNet fold 5 and ChromBPNet fold 0 test peak regions, respectively, where regions overlapping with AlphaGenome fold 0 training intervals were excluded. Performance is quantified by track Pearson r, Pearson r on the log total count, and Jensen-Shannon distance (JSD; lower indicates better performance). AlphaGenome outperforms the baselines across all metrics, modalities and cell-lines. For (g), only tracks with matching experiment accessions between AlphaGenome and ChromBPNet training sets were considered.

Extended Data Fig. 4 Contact map prediction and comparison with Orca.

(a) Comparison of predicted chromatin contact maps for five randomly selected 1-Mb intervals from the validation set. Rows show ground truth experimental data (H1-hESC Micro-C), predictions from the baseline model Orca, and predictions from AlphaGenome. Pearson correlation coefficients (r) between prediction and ground truth are displayed in the upper right corner of each corresponding predicted map. (b) Scatterplot comparing overall contact map prediction performance of AlphaGenome and Orca across all 1-Mb test intervals, for the two cell-types predicted by Orca. Each point represents an interval, plotting the Pearson R achieved by AlphaGenome (y-axis) versus Orca (x-axis) when comparing predictions to the ground truth. Points above the diagonal indicate intervals where AlphaGenome achieved higher correlation than Orca.

Extended Data Fig. 5 Additional eQTL analysis.

Further characterization and stratification of AlphaGenome’s eQTL prediction performance. (a) Performance comparison on eQTL tasks (Coefficient Spearman R, Sign auROC, Causality auROC zero-shot, Causality RF auROC) across individual GTEx tissues (see Methods for specific scorer configurations). For zero-shot tasks, AlphaGenome’s default variant scorer is used (Methods; Fig. 4a). For the supervised causality task (‘eQTL Causality (RF)’), results use the best-performing feature set (‘All scorers combined’; e). For Borzoi comparison, the published L2_DIFF scorer across all output tracks is used. (b) Performance on eQTL tasks (Coefficient, Sign, Causality) stratified by distance from variant to target gene TSS, compared to Borzoi and a 1/distance baseline for the causality task. (c) Performance on eQTL tasks stratified by broad functional classes based on variant location. TG = Target Gene; VTG = Variant target gene. (d) Performance on indel eQTLs for the Sign and Coefficient tasks stratified by indel size, comparing AlphaGenome and Borzoi. (e) Random forest performance on the eQTL causality task using different feature sets. Horizontal bar plot compares RF performance (auROC from Fig. 4g) when trained using features derived from all scorers combined or individual modality scorers (also comparing AlphaGenome to Borzoi outputs). (f) Observed versus predicted eQTL log2 fold change plotted against variant-TSS distance (within 262 kb, ensuring target TSS is within receptive field context). Observed values are allelic log-fold changes from Mohammadi et al.⁵⁶, which were used to allow direct comparison with the scale of AlphaGenome’s predicted log-fold changes (Methods). (g) Sign accuracy as well as the fraction of eQTLs that pass a certain quantile-score cutoff visualized for all possible thresholds and four eQTL distance categories. The quantile score cutoff which leads to an overall (across distance bins) sign accuracy of 90% is highlighted. At this cutoff, sign accuracy is similar across distance categories, although recall is notably lower for distal variants. (h) Scatterplot comparing observed versus predicted effect sizes for 8,678 fine-mapped eQTLs filtered for high AlphaGenome prediction scores (>99th percentile of common variants). (i) Relationship between AlphaGenome’s quantile score threshold (x-axis; Methods) and the resulting performance (y-axis) for eQTL Sign accuracy and Coefficient Spearman R tasks.

Extended Data Fig. 6 Improved prediction of 3’ UTR polyadenylation and paQTLs with AlphaGenome.

(a) Polyadenylation predicted vs observed coverage ratio (COVR) for each gene’s most distal and proximal PAS for Borzoi and AlphaGenome. (b) Polyadenylation variants scoring scheme. (c,d) Example of variant disrupting a polyadenylation motif in the TP53 gene. (c) Distribution of scores for the indicated variant across GTEx tissues (gray swarmplot) with the highest and lowest scoring tissues and average score highlighted (red dots). (d) Observed and predicted RNA-seq for the tissues highlighted in (c), and in silico saturation mutagenesis scores (ISM) of the 72 bp flanking the variant. ISM clearly highlights the relevance of the polyadenylation motif exclusively in the reference background, where the variant does not compromise the motif. (e,f) As (c) and (d) but for a variant generating a new polyadenylation motif in the RIGI gene. (g,h) As (c) and (d) but for a paQTL with a low variant score. The variant disrupts the PAS but a cryptic one emerges nearby with limited effect on gene expression. (i,j) As (c) and (e) but for a failure case. AlphaGenome correctly identifies the emergence of a novel PAS but fails to correctly predict the effect on gene expression.

Extended Data Fig. 7 AlphaGenome improves enhancer-gene linking using input gradients and shows enhanced sensitivity to distal enhancers.

(a) Zero-shot performance of AlphaGenome and Borzoi on the ENCODE-rE2G benchmark. Bars indicate the area under the precision-recall curve (auPRC) for predicting enhancer-gene links. Two scoring methods derived from each model were evaluated: input gradient scores and RNA-seq variant effect scores. (b) Impact of incorporating AlphaGenome’s input gradient score as a feature in the ENCODE-rE2G extended logistic regression model, evaluated on the ENCODE-rE2G benchmark. ENCODE-rE2G is a logistic regression model trained to predict enhancer-gene interactions from features¹². Precision-recall curves are shown, colored by the feature sets used for training the regression model (auPRC values indicated in the legend). Feature sets are: ‘rE2G extended with AlphaGenome features’: All ENCODE-rE2G extended model features plus a single AlphaGenome’s input x gradient score; ‘AlphaGenome features only’: The AlphaGenome input x gradient score alone; ‘TSS distance with AlphaGenome features’: AlphaGenome input x gradient score plus the distance to TSS feature; ‘rE2G extended’: All features from the ENCODE-rE2G extended model¹²; ‘TSS distance’: Distance to TSS feature from¹²; ‘ABC features only’: Subset of ‘rE2g extended’, with only features related to the Activity-By-Contact (ABC) model¹². (c) Precision-recall curves for the ENCODE-rE2G benchmark, similar to panel (b), evaluating the ENCODE-rE2G extended regression model with different feature sets. Area under the precision-recall curve (auPRC) values for the different feature sets are indicated in the legend. In this configuration, ‘AlphaGenome features’ consist of a more comprehensive set of K562 cell line-specific variant effect scores. These include Allele-Specific Activity Scores (AAS) and variant effect scores calculated as the difference between alternate (ALT) and reference (REF) allele predictions (ALT-REF Diff scores). These scores were derived from AlphaGenome for the following genomic assays: RNA-seq of the target gene, ChIP-TF EP300, ChIP-Histone H3K27ac, CAGE, PRO-cap, H1-ESC contact maps. (d) Relationship between enhancer perturbation effects (ENCODE-rE2G dataset¹²) and enhancer-promoter distance. The scatter plot shows experimentally observed percentage changes in gene expression upon enhancer knockout (grey points and trend line) versus the genomic distance between the enhancer and the target gene’s Transcription Start Site (TSS). Overlaid are trend lines for AlphaGenome’s (AG, dark blue) and Borzoi’s (green) predictions of these expression changes, derived from their respective model input gradient scores. Each point corresponds to a validated enhancer-gene pair. Error bars (grey) represent 95% confidence intervals from the linear regression.

Extended Data Fig. 8 AlphaGenome performance on the CAGI5 MPRA challenge.

Comparison of models predicting the effects of regulatory variants on gene expression using CAGI5 MPRA data. (a) Pairwise comparisons of zero-shot, cell-type matched performance using raw DNase model outputs. Scatter plots show per-gene or per-context Pearson correlations. AlphaGenome shows comparable or slightly improved performance compared to Enformer, ChromBPNet, and Borzoi Ensemble based on overall correlations across compared loci (details in plot annotations; cell-type matching strategy adapted from Enformer/Borzoi papers, see Methods). Note that ChromBPNet only reports performance for TERT-HEK293T, thus TERT-GBM is excluded in the AlphaGenome vs ChromBPNet comparison. (b) Pairwise comparisons of cell-type matched (see a) and cell type agnostic performance using LASSO regression on spatially summed DNase features (Methods). Comparison of the original Borzoi scorer (upper row) and AlphaGenome recommended scoring strategy (bottom row). AlphaGenome outperforms Borzoi in all settings and shows that cell type matching is not required when applying LASSO regression. (c) Pairwise comparisons of cell-type agnostic performance using LASSO regression with different input features: left column shows the composite scorer described in Borzoi (DNase + Histone ChIP-seq + RNA), right column shows the composite scorer described in Enformer (DNase + CAGE). AlphaGenome generally outperforms Enformer and Borzoi for all input feature settings.

Extended Data Fig. 9 Interpreting trait-affecting variants using AlphaGenome.

(a) AlphaGenome predictions for the variant chr11:116837649:T>G, previously reported to result in underexpression of APOA1 and Hypoalphalipoproteinemia⁵⁷. The variant strongly reduces the predicted expression of the APOA1 transcript in heart tissue. ISM of the reference sequence suggests that the variant disrupts a TATA-like motif. (b) AlphaGenome predictions for the variant chr16:173692:A>G, previously reported to cause Hemoglobin H disease⁵⁸. The model predicts that the variant strongly reduces the predicted expression of the HBA2 transcript. ISM of the reference sequence suggests that the variant disrupts the polyadenylation hexamer. (c) AlphaGenome predictions for the variant chr17:44001600:C>A, previously reported to result in reduced binding of HNF1, underexpression of NAGS and ultimately N-acetylglutamate Synthase Deficiency⁵⁹. AlphaGenome predicts reduced binding of HNF1A, local DNA accessibility, as well as a strong reduction in expression of the NAGS transcript in liver and related cell types. ISM of the reference sequence suggests that the variant disrupts a HNF1 motif. (d) AlphaGenome predictions for the variant chrX:55028202:A>G, previously reported to result in reduced GATA binding and reduced expression of ALAS2, thereby possibly causing Sideroblastic Anemia⁶⁰. The model predicts reduced binding of GATA factors (GATA2 shown), as well as TAL1, and drastically reduced expression of the ALAS2 transcript in K562. ISM of the reference sequence suggests that the variant disrupts a GATA-TAL composite motif. (e) AlphaGenome predictions for the variant chr5:1295046:T>G, previously reported to cause TERT overexpression associated with malignant melanoma⁶¹. AlphaGenome predicts increased binding of ETS factors (ELF1 shown) and an increase in TERT expression in melanocytes. Comparing the ISM around the reference and alternative alleles suggests that the variant creates an ETS factor binding motif. The phyloP⁶² score for this position is −3.29, consistent with the fact that gain-of-function variants can occur in fast-evolving parts of the genome. (f) AlphaGenome predictions for the variant chr11:5254983:G>C, previously implicated in hereditary persistence of fetal hemoglobin⁶³. The model predicts increased CTCF binding, as well as increased HBG2 expression (shown for CD14+ monocytes). Comparing the ISM around the reference and alternative alleles suggests that the variant creates a CTCF binding motif. The phyloP score for this position is −0.67, indicating that the variant position is under relatively neutral selection (given the low magnitude of the score). (g) AlphaGenome predictions for the variant rs12740374, previously found to be associated with low-density lipoprotein cholesterol levels⁶⁴. Additionally, the variant is a plasma pQTL for CELSR2, as well as a GTEx eQTL for CELSR2 and PSRC1^17,65. Consistent with the observed molecular QTL effects, AlphaGenome predicts increased expression of both of these genes in liver tissue. Moreover the model predicts increased CEBP binding on the alternative allele. Comparing the ISM around the reference and alternative alleles suggests that the variant creates a CEBP binding motif. The phyloP score for this position is −0.20, indicating that the variant position is under relatively neutral selection (given the low magnitude of the score). (h) AlphaGenome predictions for the variant rs570639864, previously found to be associated with reduced bone mineral density⁶⁶. AlphaGenome predicts decreased binding of JUN/FOS factors (JUNB shown), as well as a strong reduction in expression of WNT7B, a gene implicated in bone formation⁶⁷. ISM of the reference sequence suggests that the variant disrupts a JUN/FOS motif. (i) AlphaGenome predictions for the variant chrX:101386224:T>C, previously reported to result in a lack of functional BTK and causing X-linked Agammaglobulinemia (OMIM #300755). The model predicts reduced binding of SPI1 at the BTK promoter, diminished local DNA accessibility, as well as a reduction in expression of the BTK transcript in B & lymphoblastoid cells. ISM of the reference sequence suggests that the variant disrupts a SPI-like motif (SPI1.H13.0.PB shown). Also shown is the 100-vertebrate PhyloP conservation track, which shows high conservation around the variant.

Supplementary information

Supplementary Information

This file contains Supplementary Methods (detailed description of data acquisition and processing, model architecture, training procedures (pretraining and distillation) and benchmarking protocols against existing methods), figures, notes and references.

Reporting Summary

Supplementary Tables

This file contains Supplementary Tables 1–11.

Peer Review File

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and permissions

About this article

Cite this article

Avsec, Ž., Latysheva, N., Cheng, J. et al. Advancing regulatory variant effect prediction with AlphaGenome. Nature 649, 1206–1218 (2026). https://doi.org/10.1038/s41586-025-10014-0

Download citation

Received: 16 May 2025
Accepted: 04 December 2025
Published: 28 January 2026
Version of record: 28 January 2026
Issue date: 29 January 2026
DOI: https://doi.org/10.1038/s41586-025-10014-0