Deleteriousness Predictions AlphaGenome Track Settings

AlphaGenome Variant Impact (AVI) score for all single-base substitutions

Configure track container: Variant Deleteriousness / Variant Impact Prediction Scores

Description

The "Prediction Scores" container track contains subtracks showing the results of variant impact prediction scores. Usually these are prediction algorithms that use protein features, conservation, nucleotide composition and similar signals to determine if a genome variant is pathogenic or not.

AlphaGenome - hg38

Interpretation: Scores are PHRED-scaled, so 10 marks the top 10 percent of the most impactful substitutions in the genome, 20 the top 1 percent and 30 the top 0.1 percent. Higher values are more likely to be damaging. The authors do not define a pathogenicity cutoff.

AlphaGenome shows the AlphaGenome Variant Impact (AVI) score from Google DeepMind, a machine-learning prediction that combines AlphaGenome regulatory predictions, AlphaMissense scores, sequence conservation and other functional annotations. Unlike most of the other scores in this container, it is not restricted to coding sequence: the track contains four bigWig subtracks (one per alternate allele) covering all 8.8 billion possible single-base substitutions in the genome, including introns and intergenic regulatory regions. The score has not yet been described in a peer-reviewed publication, and the data cannot be redistributed by UCSC.

BayesDel - Only hg19

BayesDel is a deleteriousness meta-score for coding and non-coding variants, single nucleotide variants, and small insertion/deletions. The range of the score is from -1.29334 to 0.75731. The higher the score, the more likely the variant is pathogenic.

MaxAF stands for maximum allele frequency. The old ACMG (American College of Medical Genetics and Genomics) rules utilize allele frequency to classify variants, so the "BayesDel without MaxAF" tracks were created to avoid double-dipping. However, new ACMG rules will not include allele frequency, so it is okay to use the "BayesDel with MaxAF" for variant classification in the future. For gene discovery research, it is better to use BayesDel with MaxAF.

For gene discovery research, a universal cutoff value (0.0692655 with MaxAF, -0.0570105 without MaxAF) was obtained by maximizing sensitivity and specificity in classifying ClinVar variants; Version 1 (build date 2017-08-24).

For clinical variant classification, Bayesdel thresholds have been calculated for a variant to reach various levels of evidence; please refer to Pejaver et al. 2022 for general application of these scores in clinical applications.

M-CAP - Only hg19

Interpretation: The authors define that at an M-CAP score > 0.025, 5% of pathogenic variants are misclassified as benign. 0.025 is the recommended cutoff.

The Mendelian Clinically Applicable Pathogenicity (M-CAP) score (Jagadeesh et al, Nat Genetics 2016) is a pathogenicity likelihood score that aims to misclassify no more than 5% of pathogenic variants while aggressively reducing the list of variants of uncertain significance. Much like allele frequency, M-CAP is readily interpreted; if it classifies a variant as benign, then that variant can be trusted to be benign with high confidence.

At an M-CAP score > 0.025, 5% of pathogenic variants are misclassified as benign. The score varies from 0.0 - 1.0, following a geometric distribution with a mean of 0.09.

MutScore - hg38/hg19

Interpretation: The authors defined the thresholds <0.140 for a variant to be benign, and > 0.730 for pathogenic with 95% confidence.

The within-gene clustering of pathogenic and benign DNA changes is an important feature of the human exome. MutScore score (Quinodoz, AJHG 2022) integrates qualitative features of DNA substitutions with new additional information derived from positional clustering. Variants of unknown significance that are scored as benign by other algorithms but located close to known pathogenic variants should be weighted more pathogenic by MutScore. The score ranges from 0.0-1.0, resembles a negative binomial distribution with a maximum ~0.05, depending on the nucleotide. MutScore was seen to outperform other scores by papers Porretta et al and Brock et al.

PrimateAI-3D - hg38/hg19

Interpretation: Scores range from 0 to 1, with higher values indicating greater predicted pathogenicity. The authors suggest a clinical threshold of 0.821 for distinguishing pathogenic from benign missense variants. 75% of all possible missense variants are classified as benign, 25% as pathogenic.

PrimateAI-3D (Gao et al, Science 2023) is a semi-supervised 3D convolutional neural network trained on 4.5 million benign missense variants from 233 primate species and common human variants. It operates on voxelized protein structures at 2 Å resolution (from AlphaFold or homology models) combined with multiple sequence alignments from 592 species. The track contains pre-computed scores for all 70.7 million possible single nucleotide missense variants. Pathogenic variants are shown in red, benign in blue. Items can be filtered by prediction and by percentile score.

PromoterAI - hg38

Interpretation: Scores range from -1 to 1. Positive scores indicate predicted disruption of promoter function, negative scores indicate the variant is tolerated.

PromoterAI predicts the impact of single nucleotide variants in gene promoter regions, scoring all possible substitutions within 500 bp of annotated transcription start sites. The track contains four bigWig subtracks (one per alternate allele) covering 39.5 million positions, plus a bigBed track for the 3.8% of positions where overlapping transcripts produce different scores.

ClinPred - hg38/hg19

Interpretation: Scores range from 0 to 1, with higher values indicating greater predicted likelihood of pathogenicity. The authors recommend a threshold of ≥ 0.5 to flag variants as likely disease-relevant.

ClinPred (Alirezaie et al, AJHG 2018) is a machine-learning predictor for nonsynonymous (missense) single-nucleotide variants. It combines existing pathogenicity scores with population allele frequency from gnomAD, and was trained on confidently annotated disease-causing and benign variants from ClinVar. The track contains four bigWig subtracks (one per alternate allele) with pre-computed scores for all possible human missense variants in the exome. Pathogenic variants are shown in red, benign in blue.

EVE - hg38

Interpretation: EVE scores range from 0 (benign) to 1 (pathogenic) and are normalized within each protein, so they are not directly comparable across proteins. A Class25 label assigns each variant to benign, uncertain, or pathogenic using a 25% uncertainty threshold.

EVE (Frazer et al, Nature 2021) is a deep generative model (a Bayesian variational autoencoder) trained per protein on evolutionary sequence alignments, without using clinical labels. The track shows scores for all possible missense substitutions in 2,949 disease-associated proteins as a heatmap (rows = amino acids, columns = protein positions), colored from benign (blue) through uncertain (white) to pathogenic (red).

popEVE - hg38

Interpretation: popEVE scores are a continuous, proteome-wide measure of deleteriousness and, unlike most missense scores, are calibrated to be comparable across genes; lower (more negative) scores are more deleterious. The authors define a high-confidence severe threshold at −5.056 and a moderate threshold at −4.617.

popEVE (Orenbuch et al, Nature Genetics 2025) builds on EVE and the ESM-1v protein language model, calibrating their scores against human population variation (UK Biobank) with a Gaussian process to place variants across the whole proteome on a single scale. The track shows scores for all single-nucleotide-reachable missense substitutions across roughly 18,000 proteins as a heatmap, colored on a global gradient from deleterious (red) to tolerated (blue).

To view the full description, click here.

All tracks in this collection (7)

Display mode: Reset to defaults

Type of graph:
Track height:	pixels (range: 8 to 128)
Data view scaling:	Always include zero:
Vertical viewing range:	min:	max: (range: 0 to 30)
Transform function:	Transform data points by:
Windowing function:		Smoothing window:	pixels
Negate values:
Draw y indicator lines:	at y = 0.0: at y =

Graph configuration help

List subtracks: only selected/visible all
	Dense Configure	Mutation: A	AlphaGenome AVI score: Mutation is A	(Download unavailable, see below)
	Dense Configure	Mutation: C	AlphaGenome AVI score: Mutation is C	(Download unavailable, see below)
	Dense Configure	Mutation: G	AlphaGenome AVI score: Mutation is G	(Download unavailable, see below)
	Dense Configure	Mutation: T	AlphaGenome AVI score: Mutation is T	(Download unavailable, see below)

Version: AlphaGenome Variant Impact score, Google DeepMind snapshot of August 27, 2026
Assembly: Human Dec. 2013 (GRCh38/hg38)

new Note: Released Sept. 10, 2026

Description

This track shows the AlphaGenome Variant Impact (AVI) score, a prediction from Google DeepMind of how likely a given substitution is to have an impact. Scores are provided for every possible single-base substitution at every position in the genome, about 8.8 billion in total. Unlike predictors that only cover protein-coding changes, the AVI score is defined genome wide, so it also scores variants in introns and in intergenic regulatory regions.

For an introduction to AlphaGenome and to the AVI score, see the AlphaGenome educational guides at Google DeepMind. The resource and the score are described in detail in the AlphaGenome Atlas preprint.

Display Conventions and Configuration

Because there are three possible substitutions at each position, the scores are split across four separate graphs, one for each alternate allele. The graph labelled "Mutation: A" shows the score for changing the reference base to an A, and so on. At any given position only three of the four graphs carry a real prediction; the fourth is the one matching the reference base, where no substitution exists, and it is shown as zero.

The source data gives two numbers per variant, a raw model score and a PHRED-scaled version of it. This track displays the PHRED-scaled score, which is easier to compare across variants: it is a rank in which 10 corresponds to the top 10 percent of the most impactful substitutions in the genome, 20 to the top 1 percent, and 30 to the top 0.1 percent. Higher values therefore indicate a variant predicted to be more damaging. Most substitutions score low, with a genome-wide median near 2.9, so the default view is scaled to 0 to 30 rather than to the observed maximum of about 89. Scores above the top of the range are drawn clipped at the top of the graph; the range can be changed on the track configuration page.

Individual scores are only shown when the view is zoomed in far enough. At wider zoom levels the graph shows the range of scores in each screen pixel rather than a single value.

Methods

The AVI score comes from a machine learning model that combines several kinds of evidence about a variant: AlphaGenome regulatory predictions, AlphaMissense pathogenicity scores for changes that fall in coding sequence, measures of evolutionary sequence conservation, and further functional annotations. Google DeepMind precomputed the score for all single-base substitutions in GRCh38 and released it as a tab-separated, tabix-indexed table of chromosome, position, reference allele, alternate allele, raw score and PHRED score. The method is described in the AlphaGenome Atlas preprint (see References below), which has been submitted for peer review but is not yet published. The authors do not define a pathogenicity cutoff, but recommend that users adopt genomic region- or application-aware thresholds, and prefer ranking variants by score over applying a hard cutoff where possible.

The scores can be downloaded from the AlphaGenome Atlas. At UCSC the table was split into one wig file per alternate allele and each was converted to bigWig. The PHRED column was copied through as text, so the values shown are exactly those in the source file. The commands are documented in the makeDoc, and the conversion program is in the kent source tree.

Data Access

Google DeepMind's terms do not allow UCSC to redistribute this dataset, so it is not available from the Table Browser, the Data Integrator, the Genome Browser API or our download server. The full set of precomputed scores can be obtained directly from the AlphaGenome Atlas, which provides the same tab-separated, tabix-indexed file that this track was built from. Scores for individual variants can also be read off the track in the browser by zooming in until single bases are visible.

Use of the AlphaGenome Services and their output is subject to the AlphaGenome Services Additional Terms of Service. In summary, the Services are available only to individuals and to non-commercial organizations such as universities, non-profit research institutes, educational bodies and government bodies, or for journalism, and their output may not be used for clinical purposes or relied on for medical advice.

Credits

Thanks to Dhavanthi Hariharan, Francesca Pietra, Tom Ward and Elisa Wong at Google DeepMind for providing this dataset and for their help with the track description.

References

Cheng J, Taylor KR, Nicolaisen L, Pan J, Bycroft C, Perino M, Ward T, Hawkes G, Covill LE, Weilert M et al. AlphaGenome Atlas: in silico mutagenesis of the entire human genome improves prioritization and interpretation of non-coding variants. Preprint, 2026. Submitted for peer review.