Overview
| Name | popEVE · aliases: popEVE |
|---|---|
| Full name | The paper gives no expanded name |
| Ambiguous names (not auto-linked) | EVE |
| Family | Not a family entry. The predecessor EVE has no entity record, so no family entry |
| Organization | Harvard Medical School (Blavatnik Institute), Centre for Genomic Regulation (CRG), Barcelona |
| Organization type | Academic lab |
| Country | United States, Spain |
| Model class | Missense pathogenicity prediction (evolution + population constraint) |
| Domain | Genomics & RNA (also: Proteins) |
| First release | · paper revision Basis: medRxiv v1 (10.1101/2023.11.27.23299062) Note: The site said only "preprint 2023"; v1 carried a different title from the retitled v2 (2025-03) source |
| Other milestones | |
| Latest update | † partly verified · journal Basis: Journal publication Journal publication (DOI 10.1038/s41588-025-02400-1); medRxiv v2 retitled 2025-03; last code commit 2024-07-01 source The most recent sourced event, not a claim that nothing changed since. |
| Status | Active · Journal publication within 12 months; no code commits since 2024-07 |
| Openness | Code: Open · GPL-3.0 Weights: Unclear Code is open with an online per-gene portal; whether weights are released separately is unclear |
| Benchmarks reported | 1 counted (modified protocol 1) |
| Wet-lab validation | None recorded |
| Unverified fields | † latest_update, release_milestones.journal_online, open_status.weights |
| As of | 2026-11-16 |
What it does
Adds a protein language model and population constraint (a public allele-frequency resource and a large cohort) on top of an evolutionary variational model, then calibrates scores proteome-wide with a Gaussian process so variants in different genes sit on one scale - built for ranking which variant in a patient genome is most likely causal. Like AlphaMissense it predicts missense effects, but it emphasises cross-gene calibration and rare-disease cohorts rather than a precomputed proteome-wide catalogue.
Benchmarks reported
Every row is one pairing from the survey's relation file; a score belongs to the benchmark page's ranking, where it is shown with metric, eval set, source type and date. 1 of 1 pairings are counted.
| Benchmark | Relation | Eval set / protocol | Source type | Note and evidence |
|---|---|---|---|---|
| 11 ClinVar hold-out protocol | Developer report, modified protocol | — | Self-reported | An expert-curated subset from earlier years · evidence |
Other evaluations (outside the 44)
- Severe developmental disorder cohort (Own task set) Self-reported: More than 31,000 families; the developers report the causal variant ranked first in most cases, hundreds of genes and over a hundred new candidates, with some associations reported as independently confirmed in press coverage
- Cross-gene comparability assessment (Own task set) Self-reported: The paper argues the public held-out protocol lacks resolution between leading models, citing rank changes between the two years
Developers report: once scores are calibrated proteome-wide, variants in different genes can be ranked directly, putting the causal variant first in most cases of a large developmental-disorder cohort.
Independent evidence: No independent evidence found. No third-party evaluation on a site benchmark was found; external confirmation of candidate genes concerns gene associations, not model evaluation
Training data
Group in Figure 7: Public data only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Yes | Evolutionary MSAs, a public allele-frequency resource, constraint from a large population cohort, and public DMS data source |
| b. Private or licensed data | No | source |
| c. Own wet-lab data used for training | No | |
| d. Wet lab used only for validation | No | No own wet lab; new candidate genes were confirmed by external studies source |
Wet-lab validation
None. The developers report no wet-lab validation.
Background
- Organization: An academic project: an institute at a US medical school working with a genomics research centre in Barcelona on evolutionary sequence models and variant-effect prediction, with a UK genome institute as a partner. Not a company project.
- Funding: Funding from a technology philanthropy's neurodegeneration network, awards from two US federal science agencies and a Catalan government research-centre programme (per acknowledgements and press release; the list may be incomplete).
- Products and partnerships: Builds on the group's earlier evolutionary variational model; scores are being integrated into two public variant-annotation resources, with collaborations covering two children's-hospital rare-disease networks and a national genomics programme. Clinical deployment still needs safety and accuracy validation.
- Related models: AlphaMissense
Reading notes
- On the ClinVar protocol only a qualitative "comparable" claim appears and the text gives no AUC, so it cannot be read as beating AlphaMissense.
- Variant labels and predictors can contaminate each other, so the protocol's numbers need gene-level assessment alongside.
- The cohort results are closer to real use, but gene counts and hit definitions appear in several framings, so always state the denominator.
- CAGI does not appear in the full text, so the old candidate pairing does not hold.
Caveats
- Per-model AUC values sit only in extended data figures, not the main text.
- The cohort gene count has two framings and press coverage differs from the paper.
- The journal online date comes from a press release, not the article header; partly verified.
- No code commits since 2024-07, and clinical deployment still needs validation.
Sources
- Journal paper 2025 (journal)
- Journal full text (public archive) (journal)
- Preprint v1 (preprint)
- Code repository (repository)
- Institutional press release (press)
Corrections
2 corrections
Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.