Advancing regulatory variant effect prediction with AlphaGenome

More than 98% of observed genetic variation in humans is non-coding.

“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.”

• cites [7] Halldorsson, B. V. et al. The sequences of 150,119 genomes in the UK Biobank. Nature 607, 732–740 (2022). sourceSupported

AlphaGenome's ability to simultaneously score variant effects across all modalities accurately recapitulates the mechanisms of clinically relevant variants near the TAL1 oncogene.

“The ability of AlphaGenome to simultaneously score variant effects across all modalities accurately recapitulates the mechanisms of clinically relevant variants near the TAL1 oncogene6.”

• cites [6] 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). sourcePartly supported

  The supplied passages (labeled Abstract in our copy) describe TAL1-region mutations in T-ALL samples and their mechanism: newly created MYB binding sites nucleate a super-enhancer that increases TAL1 expression. They do not evaluate AlphaGenome, its scoring across modalities, or whether its predictions accurately recapitulate those mechanisms. (LLM, conf=0.88)

  Experimental — accuracy not yet measured

Existing sequence-to-function methods involve a trade-off between input sequence length and prediction resolution, limiting their modality scope and performance.

“Existing methods involve a trade-off between input sequence length and prediction resolution, thereby limiting their modality scope and performance1–5.”

• cites [1] Avsec, Ž. et al. Effective gene expression prediction from sequence by integrating long-range interactions. Nat. Methods 18, 1196–1203 (2021). sourcePartly supported

  The Methods (Model architecture; page unavailable) describe a trade-off for Enformer: finer prediction resolution would increase the sequence length entering its attention layers and make computation intractable on available hardware. The Results (Enformer improves gene expression prediction; page unavailable) also say previous models reached less-distant elements. The paper does not establish that existing methods generally trade input length against resolution or that this trade-off limits their modality scope. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [2] 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). sourcePartly supported

  The Results (RNA-seq model design; page unavailable) explicitly describe a computational trade-off between sequence length and prediction resolution. The Methods (Model; page unavailable) explain how insufficient length or coarse resolution can obstruct learning, and the Main section describes earlier models’ narrower assay focus. The paper does not explicitly establish that this trade-off is what limits existing methods’ modality scope and performance overall. (LLM, conf=0.82)

  Experimental — accuracy not yet measured
• cites [3] Zhou, J. Sequence-based modeling of three-dimensional genome architecture from kilobase to chromosome scale. Nat. Genet. 54, 725–734 (2022). sourcePartly supported

  The Introduction (page unavailable) says earlier sequence models did not predict large-scale genome organization beyond 1 Mb, limiting predictions of large structures and structural-variant effects. The Results describe using coarser resolution at longer distances and report a small improvement in local predictions from larger sequence context. The paper does not establish that a length–resolution trade-off in existing methods limits their modality scope or overall performance. (LLM, conf=0.83)

  Experimental — accuracy not yet measured
• cites [4] Jaganathan, K. et al. Predicting splicing from primary sequence with deep learning. Cell 176, 535–548 (2019). sourceNot checked — no full text of the source
• cites [5] 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). sourcePartly supported

  The Discussion says prior sequence-kernel methods lose positional information, while position-specific kernels can suffer in predictive performance; it also says Basset offers finer nucleotide-level resolution than annotation-overlap approaches. It does not describe a trade-off between input sequence length and prediction resolution or say that such a trade-off limits modality scope. (LLM, conf=0.78)

  Experimental — accuracy not yet measured

99% (465 of 471) of validated enhancer-gene pairs fall within 1 Mb.

“For instance, 99% (465 of 471) of validated enhancer–gene pairs fall within 1 Mb (ref. 12).”

• cites [12] 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). sourcePartly supported

  The Genome-wide maps of regulatory interactions section (page unavailable) says most predicted enhancer–gene interactions span short distances and that this is consistent with CRISPR perturbation results. The Supervised classification section reports 472 CRISPR-positive pairs, but our copy does not give the claimed count of 465 of 471 or establish that 99% of validated pairs fall within 1 Mb. (LLM, conf=0.83)

  Experimental — accuracy not yet measured

For an oncogenic insertion (chr. 1: 47239296: C>ACG) characterized in ref. 6, the ALT sequence increases expression of the TAL1 gene 7.5 kb away.

“Detailed ALT–REF predictions for an oncogenic insertion (chr. 1: 47239296: C>ACG) characterized in ref. 6. Shown are differences between AlphaGenome predictions between the ALT and REF sequences of the variant in CD34+ CMP tracks. The ALT sequence increases expression of the TAL1 gene 7.5 kb away.”

• cites [6] 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). sourcePartly supported

  The paper’s main text (page unavailable in our copy) describes oncogenic insertions at a hotspot 7.5 kb upstream of TAL1 and concludes that they drive TAL1 overexpression. It does not identify the insertion as chr. 1:47239296:C>ACG, so our copy does not establish that this exact ALT allele is one it characterized. The AlphaGenome predictions are the citing authors’ analysis, not a finding attributed to this source. (LLM, conf=0.78)

  Experimental — accuracy not yet measured

Three groups of gain-of-function mutations affecting TAL1 in T-ALL-a cluster of 5′ neo-enhancer mutations upstream of TAL1 TSS, an intronic SNV and a 3′ neo-enhancer-converge on a common mechanism of upregulation of the TAL1 oncogene.

“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 TSS6,33, an intronic single-nucleotide variant (SNV)34 and a 3′ neo-enhancer35.”

• Also stated in: § Multimodal view of variant effects · boosted

  “All three variant groups were shown to converge on a common mechanism—upregulation of the TAL1 oncogene (Fig. 6a,b).”

• cites [6] 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). sourceSupported
• cites [33] Liu, Y. et al. The genomic landscape of pediatric and young adult T-lineage acute lymphoblastic leukemia. Nat. Genet. 49, 1211–1218 (2017). sourceSupported
• cites [34] Liu, Y. et al. Discovery of regulatory noncoding variants in individual cancer genomes by using cis-X. Nat. Genet. 52, 811–818 (2020). sourceSupported
• cites [35] Smith, C. et al. TAL1 activation in T-cell acute lymphoblastic leukemia: a novel oncogenic 3′ neo-enhancer. Haematologica 108, 1259–1271 (2023). sourceSupported

AlphaGenome has not yet been benchmarked on personal genome prediction, a known weakness of models in this space.

“We have not yet benchmarked the model on personal genome prediction, which is a known weakness of models in this space40,41.”

• cites [40] 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). sourceNot checked — no full text of the source
• cites [41] Huang, C. et al. Personal transcriptome variation is poorly explained by current genomic deep learning models. Nat. Genet. 55, 2056–2059 (2023). sourceSupported

AlphaGenome unifies multimodal prediction, long-sequence context and base-pair resolution into a single framework.

“Here we present AlphaGenome, a model that unifies multimodal prediction, long-sequence context and base-pair resolution into a single framework.”

AlphaGenome's 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.

“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.”

AlphaGenome's splicing predictions include a new splice junction prediction approach alongside splice site usage prediction.

“The splicing predictions of AlphaGenome include a new splice junction prediction approach alongside splice site usage prediction.”

AlphaGenome processes 1 Mb of DNA sequences and species identity (human/mouse) to predict 5,930 human or 1,128 mouse genome tracks across diverse cell types and 11 output types at specific resolutions.

“AlphaGenome processes 1 Mb of DNA sequences and species identity (human/mouse) to predict 5,930 human or 1,128 mouse genome tracks across diverse cell types and 11 output types at specific resolutions (far right).”

Trained on human and mouse genomes, AlphaGenome matches or exceeds the strongest available external models in 25 of 26 evaluations of variant effect prediction.

“Trained on human and mouse genomes, AlphaGenome matches or exceeds the strongest available external models in 25 of 26 evaluations of variant effect prediction.”

AlphaGenome is a unified DNA sequence model that takes 1 Mb of DNA sequence as input and predicts thousands of functional genomic tracks up to single-base-pair resolution across diverse modalities.

“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.”

AlphaGenome showed +25.5% versus Borzoi for sign prediction for eQTLs.

“such as sign prediction for expression QTLs (eQTLs; +25.5% versus Borzoi2)”

AlphaGenome showed +8.0% versus ChromBPNet for accessibility QTL, averaged across five datasets.

“and accessibility QTL (+8.0% versus ChromBPNet10, averaged across five datasets; Methods)”

AlphaGenome matched or outperformed external models on 25 of 26 variant effect prediction evaluations.

“AlphaGenome matched or outperformed the external models on 25 of 26 evaluations (Fig. 1e and Supplementary Table 4).”

AlphaGenome exhibited mild improvement on sign/coefficient prediction for distal variants (greater than 35 kb).

“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).”

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.

“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).”

AlphaGenome predictions accurately captured the new splice junction and extended exon induced by the chr. 21: 46126238: G>C variant.

“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).”

AlphaGenome simultaneously predicts 5,930 human or 1,128 mouse genome tracks across 11 modalities covering gene expression, splicing, chromatin state and chromatin contact maps.

“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.”

At a score threshold yielding 90% sign prediction accuracy, AlphaGenome recovered over twice as many GTEx eQTLs (41%) as Borzoi (19%).

“At a score threshold yielding 90% sign prediction accuracy, AlphaGenome recovered over twice as many GTEx eQTLs (41%) as Borzoi (19%; Fig. 4g).”

Fold-specific, pretrained AlphaGenome models demonstrated high concordance between predicted and observed read coverage on unseen genome intervals.

“Fold-specific, pretrained AlphaGenome models demonstrated high concordance between predicted and observed read coverage on unseen genome intervals (Fig. 2a).”

There were strong Pearson correlations between predicted and observed signals for functional genomics tracks in both human and mouse genomes.

“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).”

AlphaGenome outperformed Orca3 on contact maps (contact map Pearson r +6.3%; cell-type-specific differences +42.3%).

“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)”

AlphaGenome classifies pathogenic versus benign ClinVar variants on the basis of splicing effects for deep intronic, synonymous, splice site region and missense variants predicted as likely_benign by AlphaMissense.

“Classifying pathogenic versus benign ClinVar variants on the basis of splicing effects for deep intronic (more than 6 bp from splice sites) and synonymous (more than 3 bp from splice sites) variants, variants in the splice site region (within 6 bp intronic or 3 bp exonic) and missense variants predicted as ‘likely_benign’ by AlphaMissense51.”

• cites AlphaMissense51Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

AlphaGenome outperformed ChromBPNet on accessibility (+1.6% for ATAC; +9.5% for DNase profile Jensen-Shannon divergence).

“ChromBPNet10 on accessibility (+1.6% for ATAC; +9.5% for DNase profile Jensen–Shannon divergence; Extended Data Fig. 3g).”

AlphaGenome outperformed ProCapNet on transcription initiation tracks (+15% total counts Pearson r).

“ProCapNet9 on transcription initiation tracks (+15% total counts Pearson r; Extended Data Fig. 3f)”

AlphaGenome accurately predicts quantitative splice junction read coverage and PSI5 and PSI3 within various tissues, achieving high correlation with experimental measurements.

“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).”

ISM analysis of exon 9 of U2SURP and flanking introns highlighted recognizable splicing-related sequence motifs.

“For example, ISM analysis of exon 9 of the U2SURP gene and its flanking introns highlighted recognizable splicing-related sequence motifs18,19 (Fig. 3d).”

For causal SPI1 bQTLs using the GM12878 SPI1 ChIP-seq track, predicted versus observed effect sizes gave signed Pearson r = 0.55; unsigned Pearson r = 0.12.

“Predicted versus observed effect sizes for causal SPI1 bQTLs using the GM12878 SPI1 ChIP-seq track. Signed Pearson r = 0.55; unsigned Pearson r = 0.12.”

AlphaGenome predicts the effect of variants on gene expression.

“AlphaGenome predicts the effect of variants on gene expression.”

The splice junction scorer alone outperformed previous approaches on all benchmarks, except deep intronic and synonymous ClinVar and MFASS variants.

“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.”

Top-scoring AlphaGenome predictions exhibited high sign accuracy across distance categories.

“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).”

AlphaGenome outperformed external models on 22 of 24 genome track evaluations.

“AlphaGenome outperformed these external models on 22 of 24 evaluations (Fig. 1d, Extended Data Fig. 3 and Supplementary Table 3).”

For causal caQTLs (African ancestry), predicted versus observed effect sizes using GM12878 cell line DNase predictions gave signed Pearson r = 0.74; unsigned Pearson r = 0.45.

“Predicted versus observed effect sizes for causal caQTLs (African ancestry). The scatterplot displays GM12878 cell line DNase predictions. Signed Pearson r = 0.74; unsigned Pearson r = 0.45.”

AlphaGenome predictions span a variety of biological contexts, such as different tissue types, cell types and cell lines.

“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).”

AlphaGenome outperformed Borzoi on coefficient and sign prediction for indel eQTLs.

“Additionally, AlphaGenome outperformed Borzoi on coefficient and sign prediction for indel (insertion or deletion) eQTLs (Fig. 4c,e and Extended Data Fig. 5d).”

Improvements in coefficient and sign prediction were observed broadly across most GTEx tissues, variant-to-TSS distance bins and variant functional annotation classes.

“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).”

AlphaGenome performed best on fine-mapped sQTL classification, including SNPs within 10 kb of the closest splice site and proximal variants within 200 bp of splice sites.

“AlphaGenome performed best on fine-mapped splicing QTL (sQTL) classification2,20, 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).”

AlphaGenome achieved highest performance at predicting rare SNPs and indels associated with GTEx splicing outliers in both supervised and unsupervised settings.

“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).”

AlphaGenome 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.

“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.”

AlphaGenome is a deep learning model designed to learn the sequence basis of diverse molecular phenotypes from human and mouse DNA.

“AlphaGenome is a deep learning model designed to learn the sequence basis of diverse molecular phenotypes from human and mouse DNA (Fig. 1a).”

AlphaGenome is a SOTA splicing variant effect prediction model.

“AlphaGenome is a SOTA splicing variant effect prediction model.”

For oncogenic mutation chr. 1: 47239296: C>ACG6, AlphaGenome predicted increases in activating histone marks H3K27ac and H3K4me1 at the variant, consistent with experimentally observed neo-enhancer formation at that position.

“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 position6,33 (Fig. 6b).”

• AssumptionRests on an assumption: CD34+ common myeloid progenitor (CMP) data is the closest available match to the T-ALL cell of origin.

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.

“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.”

Oncogenic mutations were predicted to increase TAL1 expression in CD34+ CMPs more than length-matched, sequence-shuffled control variants.

“Oncogenic mutations were predicted to increase TAL1 expression in CD34+ CMPs more than shuffled controls (Fig. 6c).”

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), for which performance generally declined as target resolution became coarser.

“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).”

Aggregating features from DNase predictions across all modelled cell types using LASSO regression further improved performance over cell-type-matched DNase features and the analogous approach with the Borzoi Ensemble model (Pearson r = 0.63).

“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).”

• AssumptionRests on an assumption: MPRAs measure the regulatory activity of short DNA sequences, a process closely linked to local chromatin accessibility and transcription factor binding.

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.

“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)).”

AlphaGenome resolved approximately 4-fold more credible sets in the lowest minor allele frequency quintile compared with COLOC.

“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).”

No mutations within 40 bp of the oncogenic chr. 1: 47239296: C>ACG variant had a predicted effect on TAL1 expression in the reference sequence.

“No mutations within 40 bp of the variant had a predicted effect on TAL1 expression in the reference sequence (Fig. 6e).”

Incorporating AlphaGenome predictions into the ENCODE-rE2G (extended) model yielded new SOTA performance across all distance-to-TSS categories.

“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).”

Variant effect prediction metrics, which often aggregate effects over larger regions such as gene bodies or exons, were robust to resolution changes.

“Similarly, variant effect prediction metrics, which often aggregate effects over larger regions (such as gene bodies or exons), were also robust to resolution changes.”

Both the training data and evaluations of AlphaGenome are heavily focused on protein-coding genes.

“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.”

AlphaGenome zero-shot performance was comparable (within 1% auPRC) with the ENCODE-rE2G (extended) model, which was explicitly trained on this task and cell line data.

“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).”

AlphaGenome identified a second ETS-like motif nearby that affected TAL1 expression in the alternative sequence but not in the reference sequence.

“The model additionally identified a second ETS-like motif nearby, which affected TAL1 expression in the alternative sequence but not in the reference sequence.”

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.

“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.”

AlphaGenome achieves SOTA performance in predicting APA (Spearman ρ = 0.894 versus Borzoi's 0.790).

“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.”

Training on single modalities was often detrimental to variant effect prediction performance.

“Training on single modalities was often detrimental to variant effect prediction performance (Supplementary Fig. 12, middle panel).”

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.

“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.”

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.

“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).”

Using features derived from variant scores across all predicted modalities provided a performance uplift in the supervised causality setting compared with using RNA-seq-derived scores alone.

“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.”

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.

“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)).”

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).

“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.”

By integrating features across several modalities and all cell types with LASSO, AlphaGenome achieved SOTA performance on the CAGI5 benchmark (Pearson r = 0.65).

“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).”

• AssumptionRests on an assumption: MPRAs measure the regulatory activity of short DNA sequences, a process closely linked to local chromatin accessibility and transcription factor binding.

The fully multimodal model generally outperformed models trained on single modality groups.

“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).”

Cell-type-matched DNase predictions of AlphaGenome achieved performance comparable to both ChromBPNet and the Borzoi Ensemble (Pearson r = 0.57).

“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).”

• AssumptionRests on an assumption: MPRAs measure the regulatory activity of short DNA sequences, a process closely linked to local chromatin accessibility and transcription factor binding.

The difference between oncogenic and shuffled mutations is most pronounced in T-ALL-relevant RNA-seq tracks such as thymus, CMP and haematopoietic multipotent progenitor cells.

“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).”

• AssumptionRests on an assumption: The RNA-seq tracks labelled thymus, CMP and haematopoietic multipotent progenitor cells are relevant to T-ALL biology.

Both AlphaGenome and Borzoi still underestimate the impact of very distal enhancers.

“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).”

Training with 1-Mb input sequences and subsequently performing inference using the full 1-Mb context yielded the best overall results.

“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).”

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.

“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).”

• AssumptionRests on an assumption: Tissue-specific gene expression is modulated by enhancer–promoter interactions, often involving enhancers in distal genomic regions.

For SPI1-specific variants, ISM further highlighted altered SPI1 motifs within the local sequence context.

“For SPI1-specific variants, ISM further highlighted altered SPI1 motifs within the local sequence context (Fig. 5h,i).”

Across modalities, oncogenic variants exhibited a distinct predicted mechanism compared with shuffled controls, as shown by unsupervised clustering.

“Across modalities, oncogenic variants exhibited a distinct predicted mechanism compared with shuffled controls, as shown through unsupervised clustering (Fig. 6d).”

AlphaGenome performance on distinguishing fine-mapped eQTLs from distance-matched variants (causality; auROC) was comparable with Borzoi.

“The performance of AlphaGenome on distinguishing fine-mapped eQTLs from distance-matched variants (‘causality’; auROC) was comparable with Borzoi (Fig. 4i).”

AlphaGenome and COLOC resolved the direction of effect for largely non-overlapping sets of loci.

“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).”

AlphaGenome outperformed Borzoi on distinguishing paQTLs from expression-matched negatives (all paQTLs within 10 kb of PAS, auPRC 0.629 vs 0.621; proximal paQTLs within 50 bp of PAS, auPRC 0.762 vs 0.727).

“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).”

Models must trade off between capturing long-range genomic interactions and achieving nucleotide-level predictive resolution.

“First, often owing to computational limitations, models must trade off between capturing long-range genomic interactions and achieving nucleotide-level predictive resolution.”

Several state-of-the-art 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.

“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.”

• cites SpliceAI4Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites ChromBPNet10Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Orca3Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

Global characterization of the complex effects of the vast majority of variants remains intractable without computational predictions.

“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.”

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).

“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).”

• cites SpliceAI4Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• cites Pangolin11Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

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.

“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.”

Specialized models alone are insufficient for capturing the diverse molecular consequences of variants across modalities.

“However, specialized models alone are insufficient for capturing the diverse molecular consequences of variants across modalities.”

ISM-derived sequence logos for REF/ALT alleles suggest variant disruption or modulation of transcription factor binding motifs.

“ISM-derived sequence logos for REF/ALT alleles from e, suggesting variant disruption or modulation of transcription factor binding motifs. Putative binding factors and JASPAR53 matrix IDs (MA0105.1 and MA0105.3) are indicated on the right.”

• cites JASPAR53Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

ISM-derived sequence logos for REF and ALT alleles of example SPI1 bQTLs suggest potential impacts such as creation or disruption of SPI1 or related motifs.

“ISM-derived sequence logos for REF and ALT alleles of example SPI1 bQTLs from h, suggesting potential impacts such as creation or disruption of SPI1 or related motifs. The putative binding factors and JASPAR matrix IDs (MA0081.2 and MA0080.5) are indicated on the right.”

• cites JASPAR53Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

AlphaGenome achieves SOTA splicing variant effect prediction on six of seven benchmarks.

“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.”

Application of AlphaGenome to complex trait analysis is limited because it 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 model's direct sequence-to-function scope.

“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.”

• cites [42] 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). sourceNot checked — no full text of the source

A core strength of AlphaGenome is its efficient multimodal variant effect prediction, which simultaneously scores variant impacts across all predicted modalities in a single inference pass.

“A core strength is its efficient multimodal variant effect prediction, which simultaneously scores variant impacts across all predicted modalities in a single inference pass.”

AlphaGenome's reduced dependence on population genetics parameters probably explains why it resolved more credible sets in the lowest minor allele frequency quintile compared with COLOC.

“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).”

Distillation offers a way to achieve strong performance with reduced inference cost compared with evaluating large ensembles.

“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).”

The unified multimodal approach achieves broad and high performance, particularly for tasks integrating diverse regulatory signals.

“Together, these results demonstrate the value of the unified multimodal approach for achieving broad and high performance, particularly for tasks integrating diverse regulatory signals.”

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.

“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.”

AlphaGenome has improved capacity to capture long-range functional regulatory connections directly from sequence.

“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.”

AlphaGenome has improved accuracy in predicting genetic variant effects on 3′ UTR processing using only RNA-seq predictions and without explicit training on polyadenylation data or associated variants.

“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.”

AlphaGenome matches or surpasses specialized state-of-the-art models in regulatory variant effect prediction.

“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.”

AlphaGenome is a useful complement to conservation-based measures of deleteriousness, which are agnostic to the mechanism of action of a variant.

“This makes AlphaGenome a useful complement to conservation-based measures of deleteriousness, which are agnostic to the mechanism of action of a variant.”

AlphaGenome expands ability to generate functional hypotheses about direction of GWAS signals, particularly for low-frequency variants.

“Thus, AlphaGenome expands our ability to generate functional hypotheses about the direction of GWAS signals, particularly for low-frequency variants.”