Missense mutations change a single amino acid in a protein and make up a large fraction of the variants observed in human populations, yet most have no established clinical significance. popEVE places missense variants on a single, proteome-wide spectrum of deleteriousness, so that variants in different genes can be compared directly. It is a deep generative model that combines cross-species evolutionary scores with human population variation: scores from EVE (an evolutionary variational autoencoder) and the ESM-1v protein language model are calibrated against allele observations in the UK Biobank using a Gaussian process, yielding a continuous, human-specific measure of variant impact. This track shows popEVE scores for single-nucleotide missense substitutions across roughly 18,000 human proteins.
Each entry spans one protein at its genomic locus. The heatmap columns correspond to individual amino acid positions in the protein, placed at the codon's genomic coordinate. The rows correspond to the 20 standard amino acids, ordered by amino acid class to match the MaveDB track. Each cell shows the popEVE score for substituting the wildtype amino acid at that position with the row amino acid. Because popEVE is distributed as a list of genomic single-nucleotide variants, only amino acid substitutions reachable by a single-nucleotide change are scored; cells for substitutions requiring more than one nucleotide change, for the wildtype amino acid, or for positions without a score (for example start codons) are left empty.
Unlike per-gene scores, popEVE is calibrated across the whole proteome, so cells are colored on a single global gradient keyed to the raw popEVE score (lower, more negative scores are more deleterious). The color is interpolated between the five anchors below: the published severe and moderate thresholds are fixed anchors, and the extremes saturate at the 0.5th and 99.5th percentiles of the proteome-wide score distribution.
| Color | popEVE score | Interpretation |
|---|---|---|
| ≤ −5.74 | Most deleterious (color saturates here) | |
| ≈ −5.056 | Severe threshold: high-confidence deleterious (99.99% likelihood of falling in the more deleterious distribution) | |
| ≈ −4.617 | Moderate threshold | |
| ≈ −3.5 | Near the proteome-wide median | |
| ≥ −2.29 | Most tolerated (color saturates here) |
Note: popEVE ranks deleteriousness to organismal fitness, weighted toward severe, often early-onset phenotypes, rather than classic clinical pathogenicity. Some well-known disease genes whose variants act mainly through loss of function or cause adult-onset conditions (for example BRCA1) may therefore show few or no cells in the severe range.
Hovering over a cell shows a summary of that substitution and the scores behind it, for example:
G1042→A
popEVE: -3.652 (tolerated)
EVE: 6.75
ESM1v: -8.779
popAdj: EVE -3.537 / ESM1v -3.767
gap: 0.277
| Field | Meaning |
|---|---|
| G1042→A | The substitution: the wildtype amino acid (G) at the protein position (1042) changed to the variant amino acid (A) represented by this row. |
| popEVE | The popEVE score and its severity class (severe / moderate / tolerated). This is the calibrated, proteome-wide score — comparable across genes, with lower (more negative) values more deleterious — and it determines the cell color. |
| EVE | The raw EVE evolutionary index used as an input to popEVE (a log-likelihood-ratio scale on which higher values are more deleterious). This is not the same as the 0–1 score shown in the separate EVE track. |
| ESM1v | The raw ESM-1v protein language model score used as the other input to popEVE, on which more negative values are more deleterious. |
| popAdj | The EVE and ESM-1v scores after popEVE's population-based recalibration onto the proteome-wide scale; the popEVE score is the combination of these two. |
| gap | The fraction of sequences with a gap at this position in the multiple sequence alignment (a per-position value; higher means sparser alignment coverage and a weaker evolutionary signal). |
For reverse-strand genes, protein positions are displayed left to right in genomic order (C-terminus to N-terminus on the screen), consistent with the standard genome browser orientation.
popEVE scores are computed per protein isoform, so a gene with more than one RefSeq isoform appears as several overlapping entries (one per protein accession) that may stack in the display. The track item name is the gene symbol; the specific RefSeq protein accession is available as a field in the Table Browser and details page.
popEVE is built on top of two unsupervised variant effect models: EVE, a Bayesian variational autoencoder trained per protein on multiple sequence alignments of evolutionary variation, and ESM-1v, a protein language model trained on the UniRef database. For each protein, the evolutionary scores from these models are treated as a prior, and a Gaussian process is used to model how the probability of observing a variant in the human population (the UK Biobank, ~500,000 exomes) depends on its evolutionary score. Population data are used only as a coarse "seen / not seen" signal rather than by allele frequency, which lets the resulting score rank the deleteriousness of variants across different proteins while leaving the within-gene ranking essentially unchanged. The inferred function rescales the evolutionary score per protein to account for the degree of constraint acting on that protein in the population, producing a calibrated, proteome-wide score. Low (more negative) scores indicate greater predicted deleteriousness. Following the authors, a high-confidence severe threshold is set at −5.056 and a moderate threshold at −4.617, each derived from a two-component Gaussian mixture fit. See Orenbuch et al. 2025 for full details.
The data were downloaded as a complete GRCh38 VCF (UKBB-trained popEVE, version 1.1) from https://pop.evemodel.org. The VCF lists, for every possible missense substitution, its genomic coordinates, the RefSeq protein accession and gene symbol, the popEVE score and component scores. Multiple codon changes encoding the same amino acid substitution carry identical popEVE scores and were deduplicated to one record per substitution. Records were converted to heatmap bigBed format with a custom Python script; full processing instructions are in the makedoc file, and the conversion scripts are available in our GitHub repository. The strand of each protein was taken from NCBI RefSeq. The global color gradient saturates at the 0.5th and 99.5th percentiles of the proteome-wide score distribution.
The data can be explored interactively in table format with the Table Browser or the Data Integrator and exported from there to spreadsheet or tab-separated tables. From scripts, the data can be accessed through our API, track=popEve.
For automated download and analysis, the genome annotation is stored in a bigBed file that can be downloaded from our download server. The file for this track is called popEve.bb. Individual regions or the whole genome annotation can be obtained using our tool bigBedToBed, which can be compiled from the source code or downloaded as a precompiled binary for your system. Instructions for downloading source code and binaries can be found here. The tool can also be used to obtain features within a given range, e.g. bigBedToBed http://hgdownload.soe.ucsc.edu/gbdb/$db/popEve/popEve.bb -chrom=chr17 -start=43000000 -end=43200000 stdout
The original annotation source data can be downloaded from https://pop.evemodel.org.
Thanks to Rose Orenbuch, Mafalda Dias, Jonathan Frazer, Debora S. Marks, and colleagues at Harvard Medical School, the Centre for Genomic Regulation, and collaborating institutions for developing popEVE and making the scores publicly available at pop.evemodel.org.
Orenbuch R, Shearer CA, Kollasch AW, Spinner AD, Hopf TA, van Niekerk L, Franceschi D, Dias M, Frazer J, Marks DS. Proteome-wide model for human disease genetics. Nat Genet. 2025 Dec;57(12):3165-3174. PMID: 41286104
Frazer J, Notin P, Dias M, Gomez A, Min JK, Brock K, Gal Y, Marks DS. Disease variant prediction with deep generative models of evolutionary data. Nature. 2021 Nov;599(7883):91-95. PMID: 34707284