Overview
When we built Evo we created the field of generative genomics. Now we are creating the field of generative multi-omics. We built Omnii-1 to jointly model DNA and multimodal experimental assay data. It can reconstruct assay tracks, predict the effects of non-coding variants, and link regulatory elements to target genes. Omnii-1 outperforms leading sequence-to-function models on nearly every benchmark we tested. It is the most accurate of the models we compared at identifying causal QTL variants and at linking enhancers to their target genes. It also predicts DNA methylation levels at single-CpG resolution.
Less than 2% of the human genome codes for protein. Much of the rest encodes the logic for how and where these proteins are expressed. This includes diverse regulatory elements embedded and chemical modifications. This regulatory genome gives rise to the myriad cell types, tissues and structures of the human body. A central challenge in biology, and a major bottleneck in therapeutic development, is understanding exactly how one's genome impacts their phenotype via these complex hierarchies of DNA regulation.
Omnii-1, our multimodal genome language model, is designed to help address these challenges. Omnii-1 is generative, with unprecedented abilities to read, interpret and write the regulatory genome. In this post, we highlight Omnii-1's ability to flexibly reconstruct held-out assay data (which we refer to as track reconstruction), identify causal variants associated with differential gene expression, splicing, and chromatin accessibility, and link enhancers to the genes that they regulate (Figure 1). Together, these tasks lay the groundwork for future applications that require us to write new information into the genome, rewiring the genotype to phenotype relationship in diverse contexts.
Our team designed the first genome language models (gLMs)123 and earlier this year we subsequently released a new state-of-the-art gLM Omnii preview4. This model excelled at specific types of variant effect prediction, especially as it relates to variant pathogenicity. However, gLMs learn from DNA sequence alone; they can score how unusual a variant is but cannot predict what it does, such as altering gene expression or splicing in a specific cell type. By contrast, sequence-to-function models such as AlphaGenome5 and Borzoi6 learn to predict regulatory activity from DNA sequence. Experimental assay data are provided only as training targets; at prediction time, their input is DNA sequence alone. This fundamentally limits their ability to respond to new data modalities and generalize to out-of-distribution cellular contexts. Omnii-1 overcomes both of these limitations. We trained Omnii-1 on trillions of tokens of DNA sequence and experimental assay data, and the model can both accept assay data as an input and predict it. Omnii-1 can be post-trained on new tasks and new modalities, including assays it has never seen as input. Table 1 compares Omnii-1's capabilities with those of the sequence-to-omics models we benchmark against and gLMs.
In this post, we highlight specific aspects of Omnii-1's capabilities on previously published benchmarks for variant effect prediction and track reconstruction. We show that Omnii-1 has state-of-the-art performance compared to previously published models. Through these enhanced modeling capabilities, we hope to greatly improve upon our ability to demystify variation in the regulatory genome, which makes up most of the genetic signal behind both rare and common diseases.
Multimodal reconstruction
Nearly every cell in the human body carries the same genome, yet diverse cell types and structures arise from the many ways cells regulate their gene expression. These regulatory mechanisms leave measurable traces: which stretches of DNA are open or tightly packed in heterochromatin, where transcription factors bind, and what types of chemical tags are deposited on DNA or nearby histones. Whereas the DNA sequence is the same in every cell, these -omics modalities reveal how the genome is packaged and expressed. Over the past several decades, scientists have devised assays that convert these regulatory marks into DNA that can be sequenced, for example by capturing open stretches of chromatin, pulling down DNA bound to a particular protein, or chemically marking which cytosines are methylated. Aligning each assay's sequenced fragments to the genome yields a track, or a signal measured at every position along the same coordinates as the DNA sequence.
Omnii-1 models DNA sequence together with these sequence-aligned multi-omic modalities and with evolutionary conservation at each base (Figure 2). Because every track shares the genome's coordinate system, we are able to jointly model interactions between DNA sequence and other modalities. In this section, we evaluate how well Omnii-1 predicts these tracks in a cell-type- and tissue-specific fashion.
Here, we compare Omnii-1 against ChromBPNet7, Borzoi and AlphaGenomeaa Given our status as a commercial entity we are not able to run AlphaGenome. As such, all reported AlphaGenome metrics in this post are derived directly from Avsec, et al. (2026)5. on head-to-head track benchmarks using held-out genomic loci defined by Borzoi. For each track, we compute the Pearson correlation between predicted and observed signal over the held-out intervals and then average across biosamples†† (Table 2).
Methods
Held-out data
a Given our status as a commercial entity we are not able to run AlphaGenome. As such, all reported AlphaGenome metrics in this post are derived directly from Avsec, et al. (2026)5.
b A biosample is the biological source of a measurement: a cell line, primary cell type, tissue or organoid, sometimes under a particular treatment. Each track is measured in one biosample.
Held-out data
RNA-seq, CAGE, ChIP-seq and splicing are scored on Borzoi's held-out genome sections (folds 3 and 4 of its sequences_human.bed.gz), which AlphaGenome's fold-1 model also holds out. ATAC-seq and DNase-seq are scored on ChromBPNet's fold-0 test peaks, less any region that overlaps AlphaGenome's training intervals. The metric is Pearson r computed per track over the held-out intervals and averaged across tracks.
Tracks and biosamples
Tracks and biosamples
Tracks are Borzoi's human targets: 1,543 RNA-seq tracks from 955 stranded samples, kept where the 1 bp reprocessed data match Borzoi's at r ≥ 0.99; 1,276 CAGE tracks from 638 stranded FANTOM5 samples; and 1,885 TF and 2,001 histone ChIP-seq tracks. DNase-seq covers four ENCODE cell lines (K562, HepG2, GM12878 and IMR-90) and ATAC-seq one (IMR-90). Splicing is per-tissue splice site usage derived from RNA-seq. The TF and histone ChIP-seq rows of Table 2 use a subset of 8 TF and 4 histone tracks: [LIST THE 12 TRACKS AND THEIR BIOSAMPLES].
Resolution
Resolution
CAGE and ChIP-seq are scored at 32 bp, Borzoi's output resolution; for this comparison AlphaGenome is fine-tuned with a 32 bp head on Borzoi's tracks. RNA-seq is scored at 1 bp on log1p coverage, with Borzoi's 32 bp predictions upsampled to match. ATAC-seq and DNase-seq are scored at 1 bp, and DNA methylation at each CpG.
Baselines
Baselines
All AlphaGenome numbers are taken directly from the AlphaGenome paper in Nature5, as are the Borzoi and ChromBPNet numbers it reports for RNA-seq, CAGE, splicing, ATAC-seq and DNase-seq (its Supplementary Table 3: AlphaGenome's fold-1 model against Borzoi replicate 1, and against ChromBPNet on the same peaks). Borzoi's TF and histone ChIP-seq numbers on the 12-track subset are our own: [DESCRIBE THE BORZOI RUN: REPLICATE, INTERVALS, SCRIPT].
Scoring Omnii-1
Scoring Omnii-1
[DESCRIBE HOW OMNII-1 PREDICTS AND IS SCORED ON EACH MODALITY.]
Predicting DNA methylation patterns
DNA methylation is a chemical modification written directly onto DNA. In human cells, it adds a methyl group to a cytosine, almost always at a CpG site (a cytosine immediately followed by a guanine on the same strand). Methylating the CpG sites in a gene's promoter tends to switch that gene off, so methylation patterns across promoters strongly influence which genes a cell expresses. Because methylation is copied to the new strand during DNA replication, it acts as a durable memory of cell identity across cell division: it differs between cell types, shifts with age, and is disrupted early in cancer,8 which makes it one of the most useful signals in cell-free DNA for cancer detection.9 It is primarily measured by whole-genome bisulfite sequencing (WGBS),cc Bisulfite converts unmethylated cytosines to uracil, which is read as thymine, while methylated cytosines are protected. At each CpG, the fraction of reads that still carry a cytosine gives β. A sample pools many cells, each carrying two copies of the genome, so β usually falls between 0 and 1: it is the share of copies methylated at that CpG. which gives the methylation level, or β, at each CpG from 0 (never methylated) to 1 (always methylated).
WGBS is costly and must be sequenced deeply to measure methylation reliably at each CpG, so most biosamples have far more data from cheaper assays such as RNA-seq and ATAC-seq than from WGBS. A model that predicts biosample-specific methylation from DNA and these other tracks can fill in unmeasured methylation and is a step toward asking how sequence variation alters methylation in a given cell type. None of the sequence-to-omics models we compare against (ChromBPNet, Borzoi and AlphaGenome) predict DNA methylation.
Omnii-1 predicts β at single-CpG resolution. Across our evaluation set, predicted and measured β agree with a Pearson r of 0.879. A harder question is whether the model knows which biosamples are methylated at a given region, rather than which regions tend to be methylated everywhere. We averaged β over each promoter (defined as regions within 1kb of a transcription start site, or TSS), removed each promoter's mean across biosamples, and correlated predicted values with measured values across the ten biosamples. This normalized cross-track r is 0.793 over 124 promoters. Figure 3 shows one promoter in detail for 10 biosamples.
Omnii leverages local context for accurate track prediction
Omnii-1 predicts a track from DNA together with the measured tracks of the same biosample, so a prediction can draw on local context from both DNA sequence and chromatin context. To see which inputs a prediction uses, we attribute each predicted base of RNA-seq back to the 12 kb locus around it (Figure 4). In this vignette, we show the RNA-seq coverage prediction for the gene ORM2 (an acute-phase plasma protein made almost entirely in the liver). In the liver, where ORM2 is highly expressed, the strongest regions are the gene's own promoter, exon 2 and the last exon; conservation, DNA and Pol II (POLR2A) occupancy carry most of their influence. In skeletal muscle, H3K27ac, H3K4me3 and Pol II signals are at least an order of magnitude lower at the ORM2 promoter, and the model correctly predicts almost no coverage, as measured.
On these track reconstruction benchmarks, Omnii-1 matches or outperforms Borzoi and AlphaGenome, and is the only model that predicts DNA methylation. Because it conditions on whatever was measured in a biosample, it can fill in the assays that weren't run, and its predictions follow the biosample's state: the same DNA sequence gives high ORM2 expression in liver and almost none in skeletal muscle. In the next section, we ask how well it predicts the effects of changes to the DNA itself.
Variant effect prediction for quantitative trait loci
Most genetic variants associated with human traits and disease fall outside protein-coding regions, where they act through gene regulation rather than protein structure. Predicting what such a variant does means predicting how it changes the molecular readouts above: chromatin accessibility, transcription factor binding, splicing or gene expression, and in which tissues and cell types. One way to study these effects is through quantitative trait loci (QTLs), which are mapped by associating genetic variation with variation in a molecular trait, such as gene expression (eQTLs). Here, we benchmark Omnii-1 against other models on how accurately they predict whether specific variants are QTLs for a given molecular trait. These benchmarks encompass eQTLs, splicing QTLs (sQTLs), chromatin accessibility QTLs (caQTLs), DNase-sensitivity QTLs (dsQTLs) and binding QTLs (bQTLs).
Expression QTLs
We used the AlphaGenome zero-shot eQTL causality panel to test how accurately each model separates variants that cause a change in gene expression from variants that do not. The panel is a set of test cases, each pairing a potentially causal variant with a gene and a GTEx tissue. In a positive case, statistical fine-mappingdd Statistical fine-mapping takes the variants associated with a gene's expression and estimates, for each one, the probability that it causes the association rather than riding along with a causal variant inherited alongside it. This probability is the posterior inclusion probability (PIP). The panel's labels come from SuSiE fine-mapping of GTEx eQTLs: a variant with PIP ≥ 0.9 is a positive, and one with PIP ≤ 0.01 is a negative. These labels are statistical inferences from population genetic data, not variants whose effects have been validated by experiment. confidently identifies the variant as causal for that gene's expression in that tissue. In a negative case, fine-mapping rules the variant out. Negatives are chosen at the same distances from the gene as positives, so distance alone cannot separate them. AlphaGenome reports only overall scores on this panel (0.713 AUROC zero-shot and 0.799 with a supervised random forest, Table 3), so it is absent from the slices below. We scored Borzoi with its published eQTL recipe and sliced its scores and Omnii-1's by distance from the variant to the target gene's nearest transcription start site (TSS, where transcription of the gene begins), by variant class (Figure 6), and by gene type (Figure 7). Borzoi reaches 0.664 tissue-weighted AUROCee AUROC is the probability that a model scores a randomly chosen positive above a randomly chosen negative: 0.5 is chance and 1 is perfect. Tissue-weighted AUROC computes it within each tissue and averages, weighting each tissue by its number of test cases. overall. It holds between 0.66 and 0.70 out to 50 kb from the TSS and falls to near chance beyond it (0.492). Omnii-1 reaches 0.810 on the same panel and gains accuracy with distance, from 0.746 under 500 bp to 0.854 beyond 50 kb. By variant class,ff Variant classes come from the GENCODE gene annotation: exonic (inside an exon), splice-proximal (within 200 bp of an exon edge), promoter (within 2 kb of the TSS), intronic (elsewhere inside the gene) and flank (outside the gene). its lead over Borzoi is smallest on exonic and splice-proximal variants (0.07 and 0.08) and largest on promoter, intronic and flank variants (0.14 to 0.26).
The panel mixes two questions: whether a gene carries a causal variant at all, and which of its variants is the causal one. To test each question on its own, we grouped genes by the labels of their variants (Figure 7). A gene's variants are all the variants the panel pairs with it, each labeled causal or not. In single-variant and homogeneous genes, all of a gene's variants share one label, so a model can score well just by recognizing which genes carry a causal variant. These categories test the model's ability to recover gene-level signal. In contrast, mixed genes include both causal and non-causal variants, so the model must recover variant-level signal to tell them apart. Mixed genes with nine or more variants make up a fifth of the panel, and three in four of their rows are non-causal; there Borzoi falls below chance (0.447) and Omnii-1 scores 0.738. Omnii-1's largest margins are on homogeneous genes, where it scores 0.824 to 0.892 against 0.544 to 0.705 for Borzoi.
Methods
Panel
The AlphaGenome zero-shot eQTL causality set: 21,948 variant–gene–tissue triples over 12,532 variants, 3,759 genes and 48 GTEx11 tissues, 50% positives. Labels come from SuSiE fine-mapping of GTEx eQTLs in the eQTL Catalogue12: variant–gene pairs with posterior inclusion probability (PIP) ≥ 0.9 are causal and those with PIP ≤ 0.01 are non-causal. Non-causal triples are downsampled within ten log-distance-to-TSS bins (distance to the gene's median TSS) so that distance alone cannot separate the classes. The metric is AUROC computed per tissue and averaged with weights equal to the rows per tissue, as in the AlphaGenome paper.
Slices
Distance is measured from the variant to the nearest transcript TSS of the target gene (GENCODE v46 exons, v47 strand). Variant class is derived from the same annotation: exonic, splice-proximal (within 200 bp of an exon edge), promoter (within 2 kb of the TSS), intronic (inside the gene span) and flank. Gene cells count the distinct variants the panel pairs with a gene, across all tissues and both labels; a gene is mixed when its rows carry both labels. The AlphaGenome paper stratifies its eQTL results by TSS distance, variant location and tissue (its Extended Data Fig. 5); the bin edges here are ours. A 1/distance baseline scores 0.579 overall while sitting at chance inside every distance bin and variant class, so the panel's debiasing removes distance as a within-bin cue but not as a gene-level one.
Scoring Borzoi
We ran Borzoi through its Flashzoi port13, which shares Borzoi's 7,611 output tracks and 524 kb input, and scored it exactly as Borzoi's published GTEx eQTL benchmark scores Borzoi (borzoi_sed.py -u --rc followed by borzoi_gtex_coef_sed.py): predicted coverage is un-squashed per track, forward and reverse-complement predictions are averaged on the squashed scale with strand pairs swapped, exon bins are taken by majority overlap, and the score is log₂(Σ alt + 1) − log₂(Σ ref + 1) over the target gene's exon bins, averaged over the tissue's GTEx RNA-seq tracks and passed through arcsinh. Causality uses the magnitude. Four replicates (johahi/flashzoi-replicate-0 to 3) are averaged at the score level, as the paper averages its fold models. This recipe and a simpler one (mean over exon bins, natural log, squashed coverage) agree on the variant-level cell to within 0.013 and differ on the gene-ranking cell by 0.038, where un-squashing weights high-coverage exons more.
Scoring Omnii-1
Its scores cover 21,908 of the 21,948 rows; the remaining 40 are not scored, and every Omnii-1 number in this section is computed on the 21,908. Slices use the same annotation as the other models: on these rows the distance bins hold exactly the panel's rows less the 40, and each variant class differs from the panel's count by 24 rows or fewer. Omnii-1 reaches 0.810 tissue-weighted AUROC and 0.811 tissue-weighted AUPRC overall.
Comparisons
On the full panel, 699 rows have no exon within Borzoi's 512 kb; these score zero and are kept, which accounts for most of its loss beyond 50 kb. On the rows it can reach, Borzoi scores 0.679. All numbers are point estimates of tissue-weighted AUROC; like the AlphaGenome paper, we report no intervals. For scale, the four Borzoi replicates alone span 0.04 on the mixed cell, so differences of a few hundredths should be read with care.
Limitations
Short inputs are N-padded, which Borzoi never saw in training; that 16 kb matches 512 kb argues the padding is harmless, and a shuffled-flank control is the remaining check. The slices use one annotation and one TSS convention; the AlphaGenome paper's own bin edges are in its supplementary material and were not reproduced.
Splicing QTLs
A splicing QTL (sQTL) is a variant that changes how a gene's transcripts are spliced, for instance, how often an exon is kept. The spliceosome finds each intron using short sequence signals: a GT at its 5′ end (the donor site), an AG at its 3′ end (the acceptor site), and a branch point and pyrimidine-rich tract a few dozen bases upstream of the acceptor. Variants in these regions most clearly affect splicing, but sQTLs can also sit deeper in exons and introns, where they alter splicing enhancers and silencers or create new donor or acceptor sites.
We used the AlphaGenome sQTL causality panel on its test chromosomes, which is drawn from Borzoi's GTEx sQTL set: 14,458 test cases, each a variant in one tissue, over 6,815 variants, 2,836 genes and 47 tissues. Each causal variant is paired with a non-causal one in the same tissue, with the same substitution and the same distance to the splice site. We compare against Pangolin10, a splicing model trained on splice-site usage in four tissues, and Borzoi, which predicts RNA-seq coverage and so sees splicing through exon coverage. They score 0.721 and 0.693 tissue-weighted AUROC on this panel; Omnii-1 scores 0.814.
Figure 8 groups variants by their position relative to the nearest annotated splice site. Both models change little between classes: outside the canonical dinucleotides, Pangolin ranges from 0.722 to 0.785 and Borzoi from 0.693 to 0.789. Both are highest in the branch-point window (0.785 and 0.789), the one class where Borzoi edges ahead; Pangolin leads Borzoi in every other class, by 0.016 to 0.038. Near-exonic, near-intronic and deep-intronic variants, seven in ten test cases, sit at 0.722 to 0.727 for Pangolin and 0.693 to 0.710 for Borzoi. Variants in the GT or AG itself are the smallest class, 119 test cases over 28 variants and 78% causal; Pangolin scores 0.673 there, and Borzoi is at chance (0.487). Omnii-1 leads both models in every class, from 0.759 on deep exonic variants, where its margin over Pangolin is smallest (0.009), to 0.864 in the branch-point window. Its largest gain is on the canonical dinucleotides, 0.820 against Pangolin's 0.673, and on near-intronic variants it scores 0.832 against 0.722.
By distance (Figure 9), Pangolin is most accurate 11 to 200 bp from the splice site (0.757 and 0.755) and less accurate both at the site itself (0.719 within 2 bp) and beyond 1 kb (0.713). Borzoi is the stronger of the two within 2 bp (0.763), trails Pangolin by 0.014 to 0.021 from 3 bp to 1 kb, and falls furthest beyond 1 kb (0.658). The 1/distance baseline stays between 0.43 and 0.54 in every bin, so the panel's matching removes distance as a cue within each bin as well as overall. Omnii-1 leads in every bin and follows the same shape as Pangolin, highest 11 to 50 bp from the site (0.844) and lowest within 2 bp (0.791) and beyond 1 kb (0.785). Beyond 1 kb, it keeps a margin of 0.073 over Pangolin and 0.128 over Borzoi.
Methods
Panel
The AlphaGenome sQTL causality set restricted to its test chromosomes (3, 6, 9, 12, 16, 18, 19 and 21): 14,458 variant–tissue rows over 6,815 variants, 2,836 genes and 47 GTEx tissues, 50.9% causal. The rows come from Borzoi's sQTL benchmark: GTEx v8 sQTLs from the eQTL Catalogue12, quantified as txrevise14 internal-exon events and fine-mapped per tissue. A causal row is a variant with posterior inclusion probability above 0.9 for an event of its gene. Each causal row has one non-causal partner in the same tissue with the same REF→ALT substitution and, to within 2%, the same distance to the event's nearest splice site. The metric is AUROC computed per tissue and averaged with weights equal to the rows per tissue, as for eQTLs. On these rows Pangolin, Borzoi and Omnii-1 score 0.737, 0.707 and 0.818 tissue-weighted AUPRC. Table 3's sQTL row uses the same 14,458 rows for every model except AlphaGenome, whose published value is on its full release of 26,795 variant–gene–tissue rows over 15 chromosomes.
Slices
Position class and distance are measured to the nearest splice site of the row's gene over all its GENCODE v41 basic transcripts, on the gene's strand. Classes are mutually exclusive and assigned in this order: canonical dinucleotide (intron +1 and +2, or −2 and −1); donor region (the last three exonic bases to intron +6); acceptor region (intron −20 to exon +3); branch-point window (intron −44 to −21); then exonic or intronic by whether the variant falls in an annotated exon, near when within 50 bp (exonic) or 250 bp (intronic) of a splice site and deep beyond. The panel matches distance to the txrevise event's splice site rather than to the nearest annotated one: the 1/distance baseline scores 0.485 overall and 0.43 to 0.54 in every distance bin, but it is not at chance inside every position class, so it is not drawn in Figure 8. Tissues that carry only one label inside a slice are left out of that slice's average. This matters only for the canonical dinucleotides, where 73 of 119 rows enter the metric; with 28 variants, that class should be read with care. All numbers are point estimates.
Scoring Pangolin
We ran the released Pangolin ensemble of 40 checkpoints, five replicates of each of its eight outputs (splice-site probability and usage in heart, liver, brain and testis). For each tissue the score is the largest gain or loss in predicted splice-site probability within 2 kb of the variant, with strand from GENCODE v41. A row's score is the maximum over the four tissues, so it does not depend on the row's GTEx tissue. Probability rather than usage was chosen on chromosome 22, outside the panel.
Scoring Borzoi
Borzoi is scored through Flashzoi with the Borzoi paper's sQTL score: the largest absolute change between ALT and REF in gene-normalised predicted RNA-seq coverage over the gene body, on the GTEx RNA-seq tracks matched to the row's tissue, with forward and reverse-complement predictions averaged. Four replicates (johahi/flashzoi-replicate-0 to 3) are averaged at the score level.
Scoring Omnii-1
[DESCRIBE HOW OMNII-1 SCORES A VARIANT–TISSUE ROW.]
Chromatin Accessibility QTLs
A chromatin accessibility QTL (caQTL) is a variant whose alleles change how open the chromatin is at a nearby regulatory element, measured by ATAC-seq across individuals. We used the AlphaGenome caQTL panels from lymphoblastoid cell lines (LCLs) of African15 and European ancestry: 26,042 and 31,621 variants, of which 8.5% and 8.3% are caQTLs. A variant is a caQTL when its genotype is significantly associated with accessibility of its peak. We compare against Borzoi and ChromBPNet trained on GM12878 ATAC-seq. On the African panel they score 0.794 and 0.792 AUROC and Omnii-1 0.898; on the European panel 0.710, 0.680 and 0.802. On AUPRC, Omnii-1 reaches 0.587 and 0.394 against AlphaGenome's 0.564 and 0.364 (Table 3).
A variant changes accessibility mostly by changing how well a transcription factor binds there, and the factors that open chromatin bind near the centre of an accessible region, where they push nucleosomes aside. Figure 10 sorts variants by where they sit relative to the nearest GM12878 accessible peak. Omnii-1 leads both public models in every class on both panels. Where these factors bind, it leads by 0.084 AUROC at the summit on the African panel (0.896 against Borzoi's 0.812) and 0.064 on the European (0.828 against 0.764), and by 0.112 and 0.111 in the peak body. On the African panel caQTLs are most frequent outside GM12878 peaks (17%), plausibly because some regions open in African LCLs are closed in the European GM12878 line; there the public models fall to 0.735 and 0.755 and Omnii-1 holds 0.866. On the European flanks Borzoi and ChromBPNet are close to chance (0.577 and 0.545) and Omnii-1 scores 0.696.
caQTLs are specific to a cell type: a variant can only change accessibility in cells where the factor it affects is present and the element is in use. Figure 11 groups variants by the cell types their accessible element is most active in, across 733 ENCODE cell and tissue samples. caQTLs concentrate in elements of blood cells, the cell type of these panels: 15% of variants in lymphoid elements and 18% in elements of other blood cells are caQTLs on the African panel, against 3% in elements open in most cell types (13%, 14% and 6% on the European panel). Omnii-1 leads both public models in every group on both panels, by 0.067 to 0.094 AUROC on the African panel and 0.055 to 0.099 on the European; in lymphoid elements it scores 0.856 against ChromBPNet's 0.769 and 0.805 against Borzoi's 0.706.
Methods
Panels
The AlphaGenome caQTL panels on its test chromosomes (3, 6, 9, 12, 16, 18, 19 and 21). caQTL African: 26,042 variants, 2,219 caQTLs, from the ChromBPNet paper's benchmark of caQTLs mapped in 100 LCLs of the Esan, Gambian, Luhya, Mende and Yoruba 1000 Genomes populations and HapMap Maasai. caQTL European: 31,621 variants, 2,618 caQTLs. The metric is AUROC over all variants of a slice. Intervals in the hover notes are bootstraps over 1 Mb genomic blocks, because nearby variants share peaks and linkage.
Slices
Position is measured to the nearest ENCODE GM12878 ATAC-seq summit and its peak: summit (inside the peak, within 50 bp of the summit), peak body (inside the peak, further from the summit), flank (outside any peak, within 1 kb of a summit) and outside (more than 1 kb from any summit). Cell-type groups come from the ENCODE DHS index16, which assigns each accessible element, across 733 DNase-seq samples, to the group of cell types it is most active in. We merge its groups into lymphoid, other blood (myeloid and erythroid), other tissues, and open everywhere (its tissue-invariant group). Variants outside any indexed element are left out of Figure 11.
Scoring Borzoi
Borzoi is scored through Flashzoi (johahi/flashzoi-replicate-0): the absolute log2 fold change between ALT and REF of predicted GM12878 DNase signal summed over 501 bp, with forward and reverse-complement predictions averaged.
Scoring ChromBPNet
We ran the ENCODE GM12878 ATAC-seq ChromBPNet model: the absolute change in predicted log counts between ALT and REF, averaged over its five folds and both strands. On the caQTL African panel we use our own run of this model (0.792 AUROC); the leaderboard run in Table 3 scores 0.784.
Scoring Omnii-1
[DESCRIBE HOW OMNII-1 SCORES A caQTL VARIANT.]
Modeling long-range regulatory interactions
Enhancer-to-gene interactions
Enhancers are stretches of DNA that bind transcription factors and increase a gene's expression, often from tens or hundreds of kilobases away, by looping into contact with the gene's promoter. An enhancer does not always act on the nearest gene. It can skip over neighbors to reach a gene farther away, and one enhancer can regulate several genes (Figure 12). Enhancer-to-gene (E2G) prediction takes a candidate enhancer and a gene and asks whether the enhancer regulates that gene. Solving such a task has broad implications for genetics research, especially for genome-wide association studies (GWAS). Most variants that GWAS link to disease lie outside genes, often in enhancers, so the association alone does not say which gene a variant affects. Identifying that gene is a major bottleneck in turning genetic associations into drug targets.
We evaluate Omnii-1 on this task using a previously published benchmark of CRISPR interference (CRISPRi) experiments in K562 cells, curated for the ENCODE-rE2G project.17 CRISPRi guides a catalytically inactive Cas9 protein to a chosen stretch of DNA, where it recruits a repressor that silences that stretch without cutting it. Silencing each candidate enhancer in turn and measuring the expression of nearby genes reveals which enhancers regulate which genes. A pair is positive when silencing the enhancer significantly lowers the gene's expression. Only 471 of the 10,353 tested pairs are positive, so we report AUPRC, where chance performance is 0.045.
Table 4 shows performance of various models on this benchmark. The distance baseline ranks pairs by how close the enhancer is to the TSS. Borzoi scores a pair by the input × gradient of the gene's predicted expression over the enhancer. ENCODE-rE2G and its extended version ("rE2G extended") are logistic regression models trained on a curated set of features for each pair, including chromatin accessibility, 3D contact and activity-by-contact18. Omnii-1 + rE2G extended is a trained model: it adds Omnii-1's K562 predictions to the ENCODE-rE2G extended features and is trained on the same pairs, with each chromosome predicted by a model fit on the others. We then froze it and scored it once on a separate cohort of 1,709 K562 CRISPRi pairs that none of the models were fit on, as an external check that its gain does not come from fitting the main benchmark.
Long-range pairs are the hard part of this task. Most tested pairs sit 100 kb or more from the TSS (8,550 of 10,353), but only 91 of them are positive, about 1%. Over these distal pairs, Omnii-1 reaches an AUPRC of 0.491 vs. 0.439 for rE2G extended (Figure 13a), and 0.567 vs. 0.464 for ENCODE-rE2G on the external cohort. Omnii-1 leads rE2G extended at every distance (Figure 13a), but the gain is not evenly distributed across distal pairs. When no other gene sits between the enhancer and its target, the two models are tied (0.80 each). When the enhancer has to skip over other genes to reach its target, Omnii-1 pulls ahead, and with six or more genes in between it has three times the AUPRC (0.110 vs. 0.037).
MYC shows what this looks like at one locus (Figure 13b). Of the 105 candidate enhancers tested against MYC, 9 regulate it, from 163 kb to nearly 2 Mb away. We let each model call a link wherever its score clears the cutoff at which it reaches 50% precision over the whole benchmark. Omnii-1 links 7 of the 9, including four of the five in a cluster about 1.9 Mb downstream; rE2G extended links 3 and none of that cluster. Omnii-1 also makes more false links, 10 against 3, most of them closer to the gene.
Future directions
The results in this post are about reading, reconstructing and interpreting the regulatory genome. Our next goal is to write it: to design changes to DNA sequence that produce a chosen regulatory outcome in a chosen cell type. For example, the model could propose an edit to an enhancer that lowers its target gene's expression in one tissue while leaving the gene unchanged in others, as in the gene editing panel of Figure 1. This kind of control would make gene and cell therapies more precise, and could counteract disease-causing variants in the regulatory genome itself. Accurate variant effect prediction is what makes this possible, because a model that can score the effect of a change can also search for the change that has the effect we want.
We also built Omnii-1 to take on new tasks and data types after pre-training. Our posts on cancer diagnostics and cancer vaccines describe two cases where we have done this.
Closing
In this post, we evaluated Omnii-1 on three problems in reading and interpreting the regulatory genome. In track reconstruction, the model predicts held-out assay data in a cell-type-specific way, including DNA methylation, which none of the sequence-to-omics models we compare against predict. In variant effect prediction, Omnii-1 identifies the variants that alter gene expression, splicing and chromatin accessibility more accurately than the leading published models on nearly every benchmark. In enhancer-to-gene linking, the model's largest gains come from long-range interactions, where an enhancer skips over other genes to reach its target.
Most of the genetic signal behind rare and common disease lies in the regulatory genome. Our disease genetics post applies Omnii-1 to variants in rare and complex disease.
If you're looking to help build general biological intelligence, we'd like to hear from you.
Citation
@misc{omnii_genome_2026,
title = {A unified model for generative multiomics with Omnii-1},
year = {2026},
month = {Oct},
url = {https://www.radicalnumerics.ai/blog/omnii-genome},
organization = {Radical Numerics}
}References
1. Nguyen, E., Poli, M., et al. "HyenaDNA: Long-Range Genomic Sequence Modeling at Single Nucleotide Resolution." Advances in Neural Information Processing Systems (NeurIPS) 36 (2023).
2. Nguyen, E., Poli, M., Durrant, M.G., et al. "Sequence modeling and design from molecular to genome scale with Evo." Science 386, eado9336 (2024).
3. Brixi, G., Durrant, M.G., Ku, J., et al. "Genome modelling and design across all domains of life with Evo 2." Nature (2026).
4. Radical Numerics. "A new frontier in generative genomics with Omnii." Radical Numerics blog (2026).
5. Avsec, Ž., Latysheva, N., Cheng, J., et al. "Advancing regulatory variant effect prediction with AlphaGenome." Nature 649, 1206–1218 (2026).
6. Linder, J., Srivastava, D., Yuan, H., et al. "Predicting RNA-seq coverage from DNA sequence as a unifying model of gene regulation." Nature Genetics 57, 949–961 (2025).
7. Pampari, A., Shcherbina, A., Kvon, E.Z., et al. "ChromBPNet: bias factorized, base-resolution deep learning models of chromatin accessibility reveal cis-regulatory sequence syntax, transcription factor footprints and regulatory variants." bioRxiv (2024).
8. Greenberg, M.V.C. and Bourc'his, D. "The diverse roles of DNA methylation in mammalian development and disease." Nature Reviews Molecular Cell Biology 20(10), 590–607 (2019).
9. Liu, M.C., Oxnard, G.R., Klein, E.A., et al. "Sensitive and specific multi-cancer detection and localization using methylation signatures in cell-free DNA." Annals of Oncology 31(6), 745–759 (2020).
10. Zeng, T. & Li, Y.I. "Predicting RNA splicing from DNA sequence using Pangolin." Genome Biology 23, 103 (2022).
11. The GTEx Consortium. "The GTEx Consortium atlas of genetic regulatory effects across human tissues." Science 369(6509), 1318–1330 (2020).
12. Kerimov, N., Hayhurst, J.D., Peikova, K., et al. "A compendium of uniformly processed human gene expression and splicing quantitative trait loci." Nature Genetics 53, 1290–1299 (2021).
13. Hingerl, J.C., Karollus, A., Gagneur, J. "Flashzoi: an enhanced Borzoi for accelerated genomic analysis." Bioinformatics 41(9), btaf467 (2025).
14. Alasoo, K., Rodrigues, J., Danesh, J., Freitag, D.F., Paul, D.S., Gaffney, D.J. "Genetic effects on promoter usage are highly context-specific and contribute to complex traits." eLife 8, e41673 (2019).
15. DeGorter, M.K., et al. "Transcriptomics and chromatin accessibility in multiple African population samples." bioRxiv (2023).
16. Meuleman, W., Muratov, A., Rynes, E., et al. "Index and biological spectrum of human DNase I hypersensitive sites." Nature 584, 244–251 (2020).
17. Gschwind, A.R., Mualim, K.S., Karbalayghareh, A., et al. "An encyclopedia of enhancer-gene regulatory interactions in the human genome." bioRxiv (2023).
18. Fulco, C.P., Nasser, J., Jones, T.R., et al. "Activity-by-contact model of enhancer–promoter regulation from thousands of CRISPR perturbations." Nature Genetics 51, 1664–1669 (2019).
