Applying Omnii-1 across the genetic architecture of human disease
Identifying disease-causing DNA variants and understanding their effects remains a central problem in human genetics. We use Omnii-1, our multimodal genome language model, to prioritize candidate causal variants in unsolved rare disease cases and to study the effects of rare and common variants in Parkinson's disease.
Highlights
- New leads for unsolved rare disease cases.
Omnii-1nominates candidate disease-causing variants in four of five children whose rare disease had no known genetic cause, and for one, proposes how the variant disrupts the gene. - Finding harmful variants in Parkinson's disease (PD). Across nearly 15,000 PD patients,
Omnii-1finds rare variants linked to earlier disease onset better than the best existing tool, and identifies the strongest single variant five times earlier in its ranking than any other tested model. - Fewer experiments to find the variant behind a disease signal.
Omnii-1ranks a functional variant within its first four picks in 90% of genome-wide association study (GWAS) credible sets, fewer selections than any other method tested. - Designing testable hypotheses from GWAS signals. Across 153 PD GWAS signals,
Omnii-1proposes both a candidate variant and its molecular mechanism, guiding experimental follow-up.
Explaining how genetic variation leads to disease means separating the few variants that disrupt biological function from the many benign ones that do not. This task is especially hard for variants in non-coding DNA, which makes up the large majority of the genome. For a coding variant, the genetic code tells us how the protein changes: a variant that introduces a premature stop codon, for example, can be recognized as damaging from sequence alone. Non-coding variants, by contrast, often act by altering when, where, or how strongly a gene is expressed. This makes them harder to interpret both experimentally, since testing them requires knowing which gene they regulate and in which cell type, and computationally, since existing tools have less predictive power for non-coding variants and offer little biological understanding of how they act.
Variant frequency, variant effect size, and disease prevalence are related properties of genetic disease, and each combination of them calls for its own study designs and statistical tools (see The genetic architecture of human disease). Yet these approaches share a common bottleneck: predicting from sequence which variants disrupt function, how they do so, and how that disruption connects to disease.
The genetic architecture of human disease
In Mendelian disease, one or two rare variants explain a patient's diagnosis, and the challenge is to pinpoint them. In complex disease, common variants each have small effects, and the challenge is to identify the causal variant among the many correlated candidates at each GWAS signal. Rare variants in complex disease fall in between: each is too rare to test alone, so the challenge is to decide which ones to analyze together.
Omnii-1 is a multimodal genome language model trained on trillions of tokens across omics and other biological modalities. Its variant effect predictions cover both sequence-level constraint and molecular readouts, so it can prioritize the variants most likely to disrupt function and also predict the molecular mechanism of their effects.
We apply Omnii-1 across the full genetic architecture of human disease, from Mendelian to complex disease, in the latter case using Parkinson's disease (PD) to examine both rare and common variation (Figure 1):
- Mendelian disease: closing the molecular diagnosis gap. In collaboration with the Manton Center for Orphan Disease Research at Boston Children's Hospital, we studied five children with congenital myopathy who had been sequenced extensively but still lacked a clear genetic cause for their disease.
Omnii-1was integrated into existing diagnostic workflows and nominated candidate disease-causing variants in four of the five cases. For one patient,Omnii-1proposed a candidate RYR1 variant in a difficult-to-characterize region of the non-coding genome, and predicted that it creates a new splice site, a mechanism that can be directly tested in the patient's RNA. - Rare variants in complex disease: finding the variants that matter. In whole-genome sequencing (WGS) data from nearly 15,000 PD patients, we asked
Omnii-1and five other prediction tools to nominate the variants associated with earlier age of onset, beyond those that obviously disrupt a gene.Omnii-1outperformed the best existing tool (REVEL2) while also scoring the non-coding variants that REVEL cannot, and ranked a key disease-causing variant far earlier than other models. Of all the models tested, it was also the most complementary to REVEL when the two were combined. - Common variants in complex disease: from benchmarks to PD GWAS. Genome-wide association studies point to regions of the genome linked to disease, but each region holds many variants that are statistically indistinguishable. In regions where experiments had already identified the functional variant,
Omnii-1ranked one within its top four picks in 90% of credible sets, ahead of every other method tested. We then appliedOmnii-1variant scoring prospectively to the largest PD GWAS to date. For 153 reported signals,Omnii-1nominated a candidate variant and proposed a mechanism for it in brain and monocytes, suggesting specific experiments for follow-up.
Scoring variants with Omnii-1
Omnii-1 scores a variant by comparing its predictions for the reference and alternate sequence in a region around the variant, and returns two scores (Figure 2). Omnii-1 DNA uses sequence alone and flags changes that break patterns the model has learned, in both coding and non-coding DNA. The Omnii-1 track score uses the molecular signals Omnii-1 predicts in a chosen biosampleaa The tissue or cell type the predictions are made for, such as brain or monocytes., such as chromatin accessibility, histone marks and RNA coverage, so it can also suggest how a variant acts and in which cell type. We report each score as a percentile against population variants from the same kind of region, where lower means more extreme compared with that background.
How the scores and percentiles are calculated
DNA score. Omnii-1 reads the sequence around the variant twice, once with the reference base and once with the alternate, and measures how much less the alternate fits its surrounding sequence context, based on what the model has learned across genomes. A strongly negative score means the change disrupts that context, much of which reflects evolutionary constraint.
Track score. Omnii-1 predicts how a variant changes molecular signals in the chosen biosample, such as chromatin accessibility, histone marks, transcription and splicing. The track score combines these predicted changes into one number that measures overall disruption, while the individual tracks show which signals change and in which direction.
Percentiles. We rank each score against single-nucleotide variants from gnomAD, scored the same way and in the same biosample, separately for common variants (above 5% frequency) and rare ones (seen once). A variant is compared only with background variants from the same kind of region, such as coding, intronic or intergenic, and with the same type of base change. Its percentile is the share of those variants whose score is at least as extreme, so lower means more unusual.
Closing the molecular diagnosis gap in rare disease
In collaboration with Catherine Brownstein, Alan Beggs, Katherine Anderson, and Casie Genetti of the Manton Center for Orphan Disease Research at Boston Children's Hospital.
When seeking the genetic cause of rare or Mendelian disease, the goal is to find the one or two high-impact variants that explain a patient's diagnosis. Standard approaches usually encompass DNA sequencing of patients and, when possible, additional sequencing of family members to aid in patient variant interpretation. Finding the precise molecular cause of disease is critical in several ways. It ends what is often a long diagnostic search, establishes the pattern of inheritance and the risk of recurrence in future children, can inform prognosis and gene-specific clinical care, and determines eligibility for gene-targeted trials and therapies.
Typically, clinical laboratories classify candidate variants using guidelines from the American College of Medical Genetics and Genomics (ACMG)3, which weigh evidence such as population frequency, inheritance, functional data, and computational variant effect predictions. This framework leaves two major gaps. First, many variants lack enough evidence to be called either pathogenic or benign and are reported as variants of uncertain significance (VUS). Second, these guidelines are applied mostly to coding variants, but causal variants may also lie in non-coding DNA.
Omnii-1 nominates strong candidate variants for an unsolved case of congenital myopathy
To further understand Omnii-1's capabilities to address these gaps, we utilized data from five unsolved cases of congenital myopathy obtained in collaboration with the Boston Children's Hospital Manton Center for Orphan Disease Research. Congenital myopathies are inherited muscle disorders that typically cause low muscle tone and weakness from birth or early childhood. We started with one trio: a girl with congenital fiber type disproportion (CFTD) and her unaffected parents. CFTD is diagnosed by muscle biopsy, in which one type of muscle fiber (type 1) is consistently smaller than the other. It can be caused by variants in several different genes, and many patients remain without a genetic diagnosis4.
All three trio members had whole-genome sequencing (long and short read) and whole-exome sequencing (Figure 3A). Having sequencing data on both parents meant we could assign every variant in the child to her mother or her father, screen for de novobb A variant present in the child but in neither parent. variants, and use the long reads to place nearby variants on the same or opposite copies of a gene. We began with the genes already known to cause CFTD (TPM3, RYR1, ACTA1, MYH7, SELENON, TPM2 and LMNA)4. Across the three available sequencing platforms, 34 rare variants were called in the probandcc The affected family member through whom the family came to attention, here the daughter. in this gene set, among which 10 were defined as well supported by the sequencing data (see Methods). Among these 10 candidates, the most compelling evidence emerged in RYR1, which is associated with autosomal recessive CFTD, in which damaging variants on both inherited copies of a gene constitute a genetic diagnosis (Figure 3B).
Methods
Search space
The genes to search were fixed from the diagnosis alone, before any variant was examined. For CFTD these were the 7 genes listed by Clarke 20114: TPM3, RYR1, ACTA1, MYH7, SELENON, TPM2 and LMNA. For nemaline myopathy they were the 13 genes listed by Sewry et al. 20195. Each gene's region spanned all of its GENCODE v47 transcripts, extended by 20 kb upstream and 10 kb downstream relative to the gene's strand, so that promoters and nearby regulatory sequence were searched along with the gene body.
Variant screening across platforms
All calls in these regions, from every platform a person had, were merged into a single table per person. Small variants (single-nucleotide changes, and insertions and deletions under 50 bp) were matched across platforms on chromosome, position and alleles after each allele was reduced to its minimal representation. Structural variants and copy-number changes (50 bp or larger) were merged across callers into a single event at 50% reciprocal overlap; insertions were merged if they lay within 500 bp of each other and were of similar length. A platform was counted as having found a variant only if its call passed the caller's own filters. Otherwise the table recorded whether that platform saw the reference allele at depth, did not cover the site (for example, outside the exome capture), or had no data.
Defining rare variants
A variant was considered rare if its allele frequency was at most 1% in both gnomAD v46 and CoLoRSdb v1.2.0. For gnomAD, the frequency used was the highest across all genetic ancestry groups, in both the exome and genome call sets. A group counted only if it had at least 1,000 alleles at the site, so that a small group at a poorly covered site could not make a variant appear common from one or two alleles. Structural and copy-number changes were compared against gnomAD-SV v4 and gnomAD's exome copy-number calls. CoLoRSdb, a set of 1,427 long-read genomes, was used with the same 1,000-allele minimum. It captures variation that short reads miss, particularly insertions and repeat-length changes, which can appear rare in gnomAD only because gnomAD cannot detect them. A variant absent from a reference set, or one where no group reached 1,000 alleles, was retained as rare, and the reason was recorded.
Call support
Low-confidence calls were flagged rather than removed, so every rare variant remained available for review. A rare variant was considered well supported if two or more platforms found it, or if the single platform that found it reported a genotype quality of at least 20. Structural variants followed the same rule using the caller's quality score, or at least 5 supporting reads for pbsv.
Inheritance, de novo variants and phasing
For each of the proband's rare variants, each parent's calls were examined on every platform that parent had. A parent was classed as a carrier if any platform called the variant, or if at least 3 reads, making up at least 15% of the depth, carried it without a call. A parent was classed as reference if no platform carried the variant and at least one platform observed only the reference allele at sufficient depth: at least 10 long reads or 15 short reads, with no more than one read carrying the variant. Otherwise the parent was classed as uncertain. A variant was then labeled maternal or paternal if only that parent carried it, inherited from both parents if both did, de novo if both parents were reference, and uncertain otherwise. Whether two variants in the same gene lay on the same copy (cis) or opposite copies (trans) was read from the proband's own phased long-read calls: two variants in the same phase block lay in trans if they were on different haplotypes and in cis if they were on the same haplotype. Where the parental calls could not place a variant, its long-read haplotype assigned it to a parent. This was done only within a phase block where at least two parentally placed variants agreed on that parent and none disagreed.
The maternal variant is a missense change, p.Pro3208Leu, already listed in ClinVar, where five of six submissions classify it as pathogenic or likely pathogenic and one as a VUS. Omnii-1 DNA places it among the lowest-scoring 1.3% of rare variants genome-wide. Relative to other RYR1 missense variants in ClinVar, it scores lower than every benign or likely benign variant and lower than most pathogenic or likely pathogenic ones (Figure 4A). Omnii-1 DNA is therefore consistent with a pathogenic label for this variant, which constitutes a strong first candidate RYR1 hit.
There are two paternal candidate variants, both of which lack entries in ClinVar and lie deep in introns. Genome-wide, Omnii-1 DNA places c.801-724C>A in the lowest 9.7% of rare variants and c.6275-651G>A in the lowest 15.7%. Within RYR1, only 2.0% of the 754 rare intronic variants that ClinVar classifies as benign or likely benign score as low as c.801-724C>A, and 8.2% as low as c.6275-651G>A (Figure 4B).
Because both paternal candidates are non-coding, we scored them with Omnii-1 tracks in skeletal muscle, the tissue affected in CFTD. Of the two, c.801-724C>A stood out. It sits 724 bases from the nearest exon in intron 9, has never been seen in gnomAD, and is predicted to create a new splice site, raising splice site usage at a neighboring base from 0.003 to 0.18 (Figure 5). A new splice site here can bring part of intron 9 into the mRNA, which can disrupt the protein and disable that copy of the gene. This can be tested directly in the patient's RNA. The sequence itself supports the prediction, since the change turns GAG|GTCAGG into GAG|GTAAGG, a close match to the signal cells use to mark the end of an exon. We therefore propose c.801-724C>A as a strong candidate for the second RYR1 hit, on the paternal copy and in transdd On the opposite copy of the gene, so that both copies carry a damaging variant. with the maternal missense variant.
A workbench for rare disease molecular diagnosis
We then applied the same approach to the remaining four families, each with its own mix of available data from family members and sequencing platforms. We present the results in the Omnii-1 Rare Disease Workbench, a flexible framework for bringing Omnii-1 scoring into rare disease molecular diagnosis. For every family, the workbench:
- Screens every variant the available data can show, from single-nucleotide variants (SNVs) to structural variants and copy-number changes, across whichever sequencing technologies a family has.
- Places each variant on the mother's or the father's copy, using the parents where they were sequenced and long-read phasing where they were not.
- Scores each variant with
Omnii-1DNA and withOmnii-1tracks in the relevant tissue, and ranks each score two ways: against rare variants across the genome, and against the gene's own ClinVar variants. - Ends with a hypothesis that grades the evidence for each copy of the gene.
Across all five families, the workbench points to a candidate diagnosis in four: NEB in the three families with nemaline myopathy, and RYR1 in the trio described above. In the fifth, a second family with CFTD, no candidate has yet emerged.
Through these explorations, we demonstrate the compelling utility of Omnii-1 in rare disease molecular diagnosis. For a patient whose disease had no known genetic cause, Omnii-1 helped assemble a candidate biallelic diagnosis: it added evidence for a maternal missense variant, and, on the paternal copy, singled out a deep intronic variant that standard interpretation would likely pass over. It also predicted how this variant acts, by creating a new splice site in skeletal muscle, which can be tested directly in the patient's RNA. Like the other candidates, this variant remains to be confirmed experimentally.
Finding disruptive rare variants beyond loss-of-function
Rare variants may also contribute to common (complex) disease, though usually with smaller effects than in Mendelian disease. GWAS, which tests common variants one at a time, largely excludes them, because each rare variant appears in too few people to test on its own. Instead, rare variants are often grouped by gene and tested in aggregate via gene burden testsee A test that pools the rare variants in a gene or gene set (as done here) and asks whether their carriers, as a group, differ from non-carriers, here in age at onset..
Parkinson's age of onset is a useful readout for finding deleterious rare variants
Gene burden tests typically start from predicted loss-of-function (pLoF) variantsff Nonsense, frameshift and essential splice variants., which are defined only by how they change the protein, not by a model's prediction of how damaging they are7,8. The open question is what to aggregate beyond them. Most rare variants in a gene are missense, synonymous, or non-coding, and only some of them disrupt function. Adding the right ones strengthens a burden test, while adding neutral ones dilutes it. In practice, burden tests rarely extend beyond missense variants predicted to be damaging. Most are built on exome sequencing, which captures little non-coding sequence, and few tools can reliably say which non-coding variants matter. Choosing which variants to include alongside pLoF, both coding and non-coding, is exactly the task a variant effect predictor should help with.
To evaluate this, we designed a test that leans on a well-established observation: damaging variants in canonical PD genes are enriched among patients with earlier onset9,10,11. We reasoned that if a predictor is truly identifying disruptive variants, carriers of its top-ranked variants in these genes should likewise develop PD earlier. We tested this in a WGS cohort of 14,678 European-ancestry PD cases from the Global Parkinson's Genetics Program (GP2).
Positive and negative control variant sets
In this cohort, we observed that carriers of rare (maximum gnomAD population allele frequency below 1%) pLoF variants in 14 monogenic PD genes12 develop the disease 7.5 years earlier on average (p = 3 × 10−24), or 4.5 years (p = 1 × 10−4) for pLoF SNVs, the class most predictors score. We also observed that rare synonymous variants in this same gene set, which are not expected to disrupt function, show no association (p = 0.61). These positive and negative control results validate our general approach and gene burden testing pipeline.
Methods
Cohort
We analyzed whole-genome sequencing data from GP2 release 12. The analysis set comprised 14,678 European-ancestry PD cases with a recorded age at onset.
Genes and regions
We analyzed 14 genes previously implicated in monogenic forms of PD with "high" or "very high" confidence12: ATP13A2, DNAJC6, FBXO7, GBA1, LRRK2, PARK7, PINK1, PLA2G6, POLG, PRKN, SNCA, SYNJ1, VPS13C, and VPS35. Regions were defined by MANE Select transcript boundaries without padding, so every analyzed variant lies within a MANE Select transcript.
Annotation
Variants were annotated with Ensembl VEP v113 against the merged GRCh38 cache (Ensembl 113 / GENCODE 47), using --per_gene with MANE Select transcripts prioritized.
Variant classes
pLoF variants were defined by VEP consequence alone: transcript ablation, splice acceptor, splice donor, stop gained, and frameshift variants. Synonymous variants were defined by the synonymous_variant consequence.
Rarity filters
Variants were classified as rare if their maximum population allele frequency in gnomAD was below 1% or if they were absent from gnomAD. A within-cohort allele frequency ceiling of 5% was also applied, matching the --freqUpper threshold in rvtests13.
Burden model
Association with age at onset was tested using the rvtests CMC Wald test (cmcWald), which collapses the variants in each set into a single carrier indicator per person. The model was age at onset ~ carrier + sex + PC1–10, where PC1–10 are the first ten genetic principal components (see Principal components). SKAT-O was run alongside as a secondary test. Coefficients are shifts in years, with negative values indicating earlier onset among carriers. Downstream analyses refit the same model in R, which is algebraically equivalent to cmcWald, and every refit was checked against the rvtests output before further analysis.
Principal components
Principal components were estimated within ancestry, from genome-wide common variation rather than from the target genes. Autosomal variants were retained at minor allele frequency (MAF) ≥ 5% and genotype missingness ≤ 5%, restricted to biallelic sites, and three long-range linkage disequilibrium (LD) regions were excluded outright: the MHC (chr6:25-34 Mb), the 8p23.1 inversion (chr8:7-13 Mb) and the 17q21.31 / MAPT inversion (chr17:44-47 Mb). These are large correlated blocks that LD pruning alone does not break up, and a leading PC that tracks one of them describes a haplotype rather than ancestry. Excluding 17q21.31 also matters substantively here, since a PC tracking the MAPT H1/H2 haplotype would be a poor covariate at a locus itself associated with PD. The remaining variants were LD-pruned with plink2 --indep-pairwise 200kb 0.1, i.e. a 200 kb window and an r2 ceiling of 0.1. The window is specified in physical distance rather than variant count because at whole-genome density a 50-variant window spans only 10-15 kb, far shorter than European LD blocks. Principal components were then computed on the pruned set with plink2 --pca 10 approx, the randomized algorithm recommended above 5,000 samples.
Omnii-1 and REVEL are complementary tools for finding deleterious variants beyond pLoF
Having validated our approach and pipeline (see Positive and negative control variant sets), we then sought to evaluate Omnii-1 DNA against other commonly used in silico predictors in identifying the non-pLoF variants to add to a gene burden test. We removed all pLoF variants from the candidate pool, leaving the rare missense, synonymous, and non-coding SNVs in the 14 genes. Each predictor then ranked its top 100 variants, and we tested whether carriers of those variants develop PD earlier. pLoF carrier status was included as a covariate rather than added to each set, so each result measures only what a predictor contributes beyond pLoF. Carriers of the top 100 variants from Omnii-1 DNA develop PD 4.4 years earlier (p = 5 × 10−10), surpassing REVEL (−3.4 years, p = 3 × 10−8). The two differ in what they contribute: REVEL can score only missense variants, while 11 of Omnii-1's 100 variants are non-coding. AlphaGenome Variant Impact (AVI)14 (−2.8 years, p = 2 × 10−5) and Evo 215 (−1.8 years, p = 8 × 10−3) added weaker signal, and CADD16 (p = 0.05) and Borzoi17 (p = 0.75) added none detectable (Figure 6).
Methods
LoF-plus design
After the pLoF controls, all pLoF variants were removed from the candidate pool, leaving the rare missense, synonymous and non-coding SNVs in the 14 genes (124,432 variants). Each predictor ranked this same pool and its 100 top-ranked variants were taken as its selection. Those 100 variants, and only those, were collapsed into a single carrier indicator and tested, without adding the pLoF set to them. Instead, pLoF carrier status entered the model as a covariate, so each result measures what a predictor contributes beyond pLoF rather than the pLoF signal showing through it.
Predictors compared
Omnii-1 (DNA score); Evo 2 7B15; Borzoi17 and AVI14, both regulatory-track models; CADD v1.7 (PHRED)16; and REVEL v1.3. CADD and REVEL were applied through their VEP plugins. REVEL is missense-only by construction, so it could rank only the 1,587 missense variants in the pool, while the other predictors could rank the full pool.
Selection and ties
Variants were ranked within the eligible pool, and ties were broken by a seeded shuffle rather than by position in the file, so that a tied block is not drawn preferentially from the earliest gene.
The choice of 100 variants is arbitrary, and a single cutoff hides where in each predictor's ranking the signal comes from. We therefore tracked the cumulative burden of each predictor's selections. Walking down each ranking one variant at a time, we summed the person-years of earlier onsetgg A variant's contribution is its own effect on age of onset multiplied by the number of people carrying it. attributable to each variant's carriers (Figure 7).
Omnii-1's curve drops earliest and most steeply, and the predictors otherwise reach a given level of burden in a broadly consistent order: REVEL, AVI, Evo 2, CADD, and Borzoi last. All six predictors eventually select PRKN p.R275W, and in each case it is the single largest step in their curve, but Omnii-1 reaches it at rank 22 where Evo 2 needs 121, REVEL 257, Borzoi 333, AVI 351, and CADD 437. This variant is a positive control, as it is classified as pathogenic in ClinVar for early-onset PD, but it is not a pLoF variant, so its addition to gene burden tests is dependent on how well it can be classified as deleterious via in silico tools. Omnii-1 ranks it well ahead of every other model, including REVEL, which was trained specifically to separate disease-causing missense variants from rare neutral ones2.
Because REVEL is optimized for missense variants, while Omnii-1 also ranks non-coding sequence that REVEL cannot score, we reasoned that the two models might be complementary. We therefore combined their rankings into an ensemble, and did the same for each of the other predictors (Figure 8). Omnii-1 DNA + REVEL takes the lead early and holds it continuously from rank 44 to rank 256. This early lead is meaningful because it shows this ensemble placing the most damaging variants first.
Taken together, this analysis suggests that among the predictors able to score variants beyond missense (Omnii-1 DNA, AVI, Evo 2, Borzoi, and CADD), Omnii-1 DNA is the most informative for choosing which variants to add to a burden test. It recovers the largest shift in age at onset at 100 variants, accumulates burden earliest, and gives the most durable gain when fused with REVEL.
Methods
Cumulative burden
Each variant in a predictor's ranking was fitted on its own with the same model and covariates as the LoF-plus test, including pLoF carrier status, giving a per-variant effect βi (years) and carrier count ni. The cumulative burden at rank N is the sum of βi × ni over the top N variants, in person-years of earlier onset. Weighting by carrier count stabilizes the measure, since most top-ranked variants are singletons with highly variable individual estimates. This is a descriptive accumulation of marginal effects, not a joint model: a carrier of more than one selected variant is counted once per variant, so the shape and ordering of the curves are more informative than their absolute values.
Ensembling by reciprocal rank fusion
Each predictor independently ranks every eligible variant from most to least damaging, and a variant's fused score is the sum over the predictors in the ensemble of 1 / (k + rankm(v)), with k = 60 following Cormack, Clarke and Buettcher18. Variants are then ordered by descending fused score and that order is used exactly as a single predictor's ranking would be, so a fused arm and a single arm are selected, tested and accumulated identically. Five arms are reported here, each pairing REVEL with one of the other predictors.
Omnii-1 narrows GWAS signals to functional variants
In complex disease, risk is spread across many variants, mostly common ones of small effect. Because nearby variants are inherited together (via linkage disequilibrium, or LD), each GWAS signal points to a set of correlated candidates rather than a single variant. Fine-mapping narrows these to a credible sethh The smallest set of variants that contains the causal variant with 95% probability. and gives each variant a posterior inclusion probability (PIP)ii The probability that the variant is causal.19, but identifying which variants change molecular function requires experimental testing.
Omnii-1 finds functional GWAS variants with fewer experiments
Testing every variant in a credible set is slow and expensive, so testing them in a reliable order can help find the functional variant sooner. We asked whether Omnii-1 can provide that order, using public studies that measured the molecular effects of fine-mapped GWAS variants. Two massively parallel reporter assays (MPRAs; Siraj et al.20, Zhu et al.21) measure how each allele changes regulatory activity, and ADASTRA (Abramov et al.22) catalogs how it changes transcription factor binding. In each of nearly 1,800 credible sets, we ranked the measured variants with Omnii-1 and compared that order with five widely used variant scores (Borzoi17, AVI14, Evo 215, GPN-Star23 and CADD16) and two baselines, PIP and distance to the nearest transcription start site.
Omnii-1's first pick was functional in about half of reporter assay sets (53%) and two thirds of binding sets (64%), the highest rate of any method tested (Figure 9). To reach a functional variantjj A variant that a reporter or binding assay showed changes molecular activity. in 90% of sets, a lab following Omnii-1's ranking would test four variants per set, against five for Borzoi and seven for AVI, so Borzoi and AVI would need 25% and 75% more experiments. We highlight four selected credible sets in which Omnii-1 alone ranks the functional variant first (Figure 10).
Even though a functional effect in an assay does not prove that a variant causes disease, narrowing each credible set to a few strong candidates lets a lab test many more loci or study the most promising ones in depth. Omnii-1 works from the DNA sequence and a chosen cell type, so it can rank variants at any GWAS locus, including the many that have never been tested. We next apply it to PD, where most loci have no experimental ground truth, to nominate candidate functional variants and the mechanisms through which they may act.
Benchmark details
Data
We used SuSiE 95% credible sets of SNVs from UK Biobank and BioBank Japan fine-mapping24 (Finucane lab release). We counted a variant as functional if the original study reported that its alleles differ in regulatory activity in SK-N-SH cells (Siraj et al. 202620) or LX-2 cells (Zhu et al. 202621), or if ADASTRA22 reported allele-specific binding at a 5% false discovery rate (FDR) for a transcription factor with a known motif other than CTCF. Variants whose binding change is explained by a change in the factor's own motif were counted as not functional, so that the labels test more than a motif scan. Each credible set needed at least one functional and one non-functional measured variant, and at least one of its functional variants had to be backed by an expression quantitative trait locus (eQTL) or a regulatory element. Within each set, only experimentally measured variants were ranked. This gives 756 sets at 388 loci (4,356 variants) for the reporter assays and 1,031 sets at 534 loci (18,999 variants) for binding.
Scoring
Omnii-1 scores each variant in the assay's cell type, or in K562 for ADASTRA, which pools many cell types. The other methods use their standard scores (Borzoi with Flashzoi weights25, AVI, Evo 2, GPN-Star and CADD) and are not matched to cell type.
Metrics
Top-k is the share of credible sets with a functional variant among a method's k highest-ranked variants, with tied variants sharing the credit. Within each study, credible sets that share a variant form one locus, and each locus counts once. Reporter-assay results average the two studies, and the random line is the exact expected value for each set.
The Omnii-1 GWAS Workbench displays candidate variants and mechanisms in Parkinson's disease
We applied Omnii-1 to the largest PD GWAS to date from GP2, which reports 157 independent signals across 134 loci in a European ancestry cohort26. We ran Omnii-1 tracks in brain and in monocytes to propose a mechanism for the top candidates in each. Monocytes were chosen because of previous work demonstrating the impact of PD risk alleles on monocyte gene expression27,28.
We present results for 153 of the 157 signals, in 132 loci, in the Omnii-1 GWAS Workbench (the other four could not be fine-mapped, see methods in the Workbench 'Overview' tab). The workbench shows each signal's credible set and Omnii-1 scores on one genomic axis, with every track change expressed as the effect of the risk allele. In each biosample, the workbench nominates a variant per signal from Omnii-1's predicted tracks and pairs it with the regulatory annotation and a hypothesis for its mechanism (Figure 11).
Omnii-1 nominates a variant and mechanism at the NUCKS1/RAB29 PD risk locus
To illustrate these features, we highlight a well-established PD-associated signal in a region harboring the NUCKS1 and RAB29 genes (Workbench Locus 7, Signal 11; Figure 12). Here the GWAS lead rs708723 is almost perfectly correlated with three nearby variants, so fine-mapping cannot tell the four apart. Omnii-1 tracks rank one of them, rs823114, first in both brain and monocytes, well ahead of the lead (Figure 12A). Omnii-1 predicts that the rs823114 risk allele lowers active histone marks, chromatin accessibility, and NUCKS1 RNA in monocytes, consistent with a weaker regulatory element (Figure 12B). Brain predictions point the same way, with a smaller decrease in NUCKS1 RNA. Existing annotation agrees, placing rs823114 in an ENCODE promoter-like element 222 bp upstream of the NUCKS1 transcription start site.
Independent data support the prediction, as NUCKS1 expression decreases with the rs823114 risk allele (rs823114-A) in GTEx whole blood29 and brain tissues30, and the PD GWAS signal at this locus colocalizes with a monocyte expression quantitative trait locus (eQTL) for NUCKS1 (posterior probability of a shared causal variant, PP4 = 0.68)31. In neural progenitor cells derived from induced pluripotent stem cells (iPSCs), the haplotype carrying rs823114-A lowers NUCKS1 expression, matching the Omnii-1 prediction, and also reduces CTCF binding at the NUCKS1 promoter30.
Together, these results show that Omnii-1 can turn the genetic signals of complex disease into candidate variants with predicted mechanisms in specific cell types. Applied prospectively to PD GWAS, where the functional variant at most loci is unknown, the same scores narrow most signals to a nominated variant and a hypothesis for how it acts. We are now extending Omnii-1 to more specific cell types, such as neuronal subtypes and microglial states, so that each nomination can be made in the cell type where the risk allele is most likely to have its effect.
Closing
Here, we applied Omnii-1 across the full genetic architecture of disease, spanning Mendelian and complex disease and, within PD, rare and common variation. In Mendelian disease, we showed that Omnii-1 can help close the molecular diagnosis gap, supporting a missense variant on one copy of RYR1 and nominating a deep intronic variant on the other, along with a predicted splicing mechanism that can be tested in the patient's RNA. For rare variants in complex disease, Omnii-1 DNA outperforms the best supervised missense predictor at choosing variants for a gene burden test, extends that choice into non-coding sequence, and ranks a known disease-causing variant far earlier than other models. For common variants, the Omnii-1 track score matches or outperforms other methods in ranking functional variants when a ground truth is available, and it can nominate causal variants and cell-type-specific hypotheses when applied prospectively to GWAS data. Omnii-1 fits readily into diverse genetic study designs and statistical tools, bridging the gap from variant to mechanism.
Partner with us
Partners can contribute genomic and phenotypic data to help discover disease-associated variants, identify causal mechanisms, improve genetic risk prediction, and uncover potential therapeutic targets. We are particularly interested in large patient cohorts and unresolved rare diseases where existing methods have fallen short.
Acknowledgements
We thank the patients and families from the Manton Center congenital myopathy cohort, whose contribution of genomic and clinical information made this research possible.
Data used in the preparation of this blog were obtained from the Global Parkinson's Genetics Program (GP2). GP2 is funded by the Aligning Science Against Parkinson's (ASAP) initiative and implemented by The Michael J. Fox Foundation for Parkinson's Research (www.gp2.org). For a complete list of GP2 members, see www.gp2.org.
Citation
@misc{omnii_genetics_2026,
title = {Applying Omnii-1 across the genetic architecture of human disease},
year = {2026},
month = {Oct},
url = {https://www.radicalnumerics.ai/blog/omnii-genetics},
organization = {Radical Numerics}
}References
1.Manolio, T.A., Collins, F.S., Cox, N.J., et al. "Finding the missing heritability of complex diseases." Nature 461, 747–753 (2009).
2.Ioannidis, N.M., Rothstein, J.H., Pejaver, V., et al. "REVEL: An Ensemble Method for Predicting the Pathogenicity of Rare Missense Variants." American Journal of Human Genetics 99(4), 877–885 (2016).
3.Richards, S., Aziz, N., Bale, S., et al. "Standards and guidelines for the interpretation of sequence variants: a joint consensus recommendation of the American College of Medical Genetics and Genomics and the Association for Molecular Pathology." Genetics in Medicine 17(5), 405–424 (2015).
4.Clarke, N.F. "Congenital fiber-type disproportion." Seminars in Pediatric Neurology 18(4), 264–271 (2011).
5.Sewry, C.A., Laitila, J.M., Wallgren-Pettersson, C. "Nemaline myopathies: a current view." Journal of Muscle Research and Cell Motility 40(2), 111–126 (2019).
6.Guez, J., Goodrich, J.K., Moldovan, M.A., et al. "Integrating 730,947 exome sequences with clinical literature improves gene discovery." medRxiv (2026).
7.Karczewski, K.J., Francioli, L.C., Tiao, G., et al. "The mutational constraint spectrum quantified from variation in 141,456 humans." Nature 581(7809), 434–443 (2020).
8.Backman, J.D., Li, A.H., Marcketta, A., et al. "Exome sequencing and analysis of 454,787 UK Biobank participants." Nature 599(7886), 628–634 (2021).
9.Andrews, S.V., Kukkle, P.L., Menon, R., et al. "The Genetic Drivers of Juvenile, Young, and Early-Onset Parkinson's Disease in India." Movement Disorders 39(2), 339–349 (2024).
10.Zhao, Y., Qin, L., Pan, H., et al. "The role of genetics in Parkinson's disease: a large cohort study in Chinese mainland population." Brain 143(7), 2220–2234 (2020).
11.Trinh, J., Lohmann, K., Baumann, H., et al. "Utility and implications of exome sequencing in early-onset Parkinson's disease." Movement Disorders 34(1), 133–137 (2019).
12.Blauwendraat, C., Nalls, M.A., Singleton, A.B. "The genetic architecture of Parkinson's disease." Lancet Neurology 19(2), 170–178 (2020).
13.Zhan, X., Hu, Y., Li, B., et al. "RVTESTS: an efficient and comprehensive tool for rare variant association analysis using sequence data." Bioinformatics 32(9), 1423–1426 (2016).
14.Cheng, J., Taylor, K.R., Nicolaisen, L., et al. "AlphaGenome Atlas: in silico mutagenesis of the entire human genome improves prioritization and interpretation of non-coding variants." medRxiv (2026).
15.Brixi, G., Durrant, M.G., Ku, J., et al. "Genome modelling and design across all domains of life with Evo 2." Nature 652, 1349–1361 (2026).
16.Schubach, M., Maass, T., Nazaretyan, L., et al. "CADD v1.7: using protein language models, regulatory CNNs and other nucleotide-level scores to improve genome-wide variant predictions." Nucleic Acids Research 52(D1), D1143–D1154 (2024).
17.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).
18.Cormack, G.V., Clarke, C.L.A., Buettcher, S. "Reciprocal rank fusion outperforms Condorcet and individual rank learning methods." Proceedings of the 32nd International ACM SIGIR Conference, 758–759 (2009).
19.Wang, G., Sarkar, A., Carbonetto, P., et al. "A simple new approach to variable selection in regression, with application to genetic fine mapping." Journal of the Royal Statistical Society Series B 82(5), 1273–1300 (2020).
20.Siraj, L., Castro, R.I., Dewey, H.B., et al. "Functional dissection of complex trait variants at single-nucleotide resolution." Nature (2026).
21.Zhu, B., He, N., Xiao, Y., et al. "Integrative analyses elucidate transcriptional regulatory functions of risk alleles for metabolic liver disease." Nature Genetics 58(6), 1280–1294 (2026).
22.Abramov, S., Boytsov, A., Bykova, D., et al. "Landscape of allele-specific transcription factor binding in the human genome." Nature Communications 12(1), 2751 (2021).
23.Ye, C., Benegas, G., Albors, C., et al. "Predicting genome-wide functional constraints with GPN-Star." Nature (2026).
24.Kanai, M., Ulirsch, J.C., Karjalainen, J., et al. "Insights from complex trait fine-mapping across diverse populations." medRxiv (2021).
25.Hingerl, J.C., Karollus, A., Gagneur, J. "Flashzoi: an enhanced Borzoi for accelerated genomic analysis." Bioinformatics 41(9), btaf467 (2025).
26.Leonard, H.L., Global Parkinson's Genetics Program (GP2). "Novel Parkinson's Disease Genetic Risk Factors Within and Across European Populations." medRxiv (2025).
27.Navarro, E., Udine, E., Lopes, K.P., et al. "Dysregulation of mitochondrial and proteolysosomal genes in Parkinson's disease myeloid cells." Nature Aging 1(9), 850–863 (2021).
28.Mattei, D., Brophy, E., Rosen, M., et al. "Integrated Stress Response Signatures Drive Monocyte Dysfunction in GBA1- and LRRK2-Linked Parkinson's Disease." medRxiv (2025).
29.GTEx Consortium. "The GTEx Consortium atlas of genetic regulatory effects across human tissues." Science 369(6509), 1318–1330 (2020).
30.Oh, S., Shao, J., Mitra, J., et al. "Enhancer release and retargeting activates disease-susceptibility genes." Nature 595(7869), 735–740 (2021).
31.Lopes, K.P., Snijders, G.J.L., Humphrey, J., et al. "Genetic analysis of the human microglial transcriptome across brain regions, aging and disease pathologies." Nature Genetics 54(1), 4–17 (2022).
