Advancing regulatory variant effect prediction with AlphaGenome

1Interpreting 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-coding, global characterization of the complex effects of this vast majority of variants remains intractable without computational predictions.

2Computational methods can learn patterns from experimental data to predict and explain variant effects. One class of methods,sequence-to-function models–, 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.

3Currently, 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 SpliceAI, BPNet and ProCapNet 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 Enformer and Borzoi 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.

4A 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 SpliceAI for splice site prediction, ChromBPNet for local chromatin accessibility and Orca 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 SpliceAI or Pangolin 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.

5Here 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.

6We 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).

7AlphaGenome 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. ).

8AlphaGenome uses a U-Net-inspired, 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).

9We 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 previously, 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.

10To 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 Borzoi, 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 Orca on contact maps (contact map Pearson r +6.3%; cell-type-specific differences +42.3%; Fig. 1d and Extended Data Fig. 4), ProCapNet on transcription initiation tracks (+15% total counts Pearson r; Extended Data Fig. 3f) and ChromBPNet on accessibility (+1.6% for ATAC; +9.5% for DNase profile Jensen–Shannon divergence; Extended Data Fig. 3g).

11We 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 Borzoi) and accessibility QTL (+8.0% versus ChromBPNet, 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.

12Given 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.

13Quantitatively, 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).

14On 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).

15One of the main ways genetic variants cause disease is by disrupting splicing, 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),,, competitive selection among potential splice sites (splice site usage prediction), 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).

16To 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) 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, (Fig. 3d). Further examples of experimentally validated splice-disrupting variants identified in individuals with autism spectrum disorder are shown in Supplementary Fig. 6.

17Building 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,, 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 ClinVar, 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)), 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.

18Beyond impacting isoform composition through splicing modulation, non-coding variants can influence traits and cause diseases by altering gene expression,. 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).

19We 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.

20Using fine-mapped GTEx eQTL as ground truth, we benchmarked AlphaGenome against the SOTA models Borzoi and Enformer. AlphaGenome demonstrated improved prediction of both the magnitude (‘coefficient’; Spearman ρ with SuSiE 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).

21As previously reported, 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).

22The 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) 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.

23The 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.

24We 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 study. 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).

25Altogether, 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.

26We 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 specificity. 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). 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.

27We 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 ChromBPNet), 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 bQTLs. 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).

28Beyond 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,. 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.

29We 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,, an intronic single-nucleotide variant (SNV) and a 3′ neo-enhancer. All three variant groups were shown to converge on a common mechanism—upregulation of the TAL1 oncogene (Fig. 6a,b).

30To 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>ACG, AlphaGenome predicted increases in the activating histone marks H3K27ac and H3K4me1 at the variant, consistent with experimentally observed neo-enhancer formation at that position, (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).

31We expanded our analysis to include all oncogenic variants across studies,–, 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).

32To 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 previously. 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.

33We 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.

34To 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.

35We 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.

36We 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).

37Lastly, 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.

38AlphaGenome 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.

39Going 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 oligonucleotides and tissue-specific enhancers,. Finally, AlphaGenome can complement the capabilities of generative models trained on DNA sequences by predicting functional properties of newly generated sequences. To widely enable these broad applications, we provide accessible tooling to access AlphaGenome through a hosted model and API.

40Despite 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,. 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 effects beyond the direct sequence-to-function scope of the model.

41Addressing 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,), 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–, expanding multi-species capabilities and developing robust methods for assay bias correction,). 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.

42In 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.

43For 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.

44Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

45Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41586-025-10014-0.