README.md

---
license: mit
---

# evoeval

A collection of genomic **variant-effect evaluation** datasets. Each dataset is a table of
variants keyed on GRCh38 coordinates with a label and annotation columns, stored as Parquet.
This repo holds the **base (unscored) variant tables** — model delta scores are computed
downstream, not shipped here.

## Standard column schema

Every parquet has these coordinate columns with standardized types:

| Column | Type | Description |
|---|---|---|
| `chrom` | str | Chromosome, GRCh38, always `chr`-prefixed (e.g. `chr1`, `chrX`) |
| `pos` | int64 | 1-based VCF position |
| `ref` | str | Reference allele |
| `alt` | str | Alternate allele |

Most datasets are also annotated with:

| Column | Type | Description |
|---|---|---|
| `genomic_element` | str | Coarse location: `CDS` / `splice_site` / `5UTR` / `3UTR` / `intron` / `ncRNA` / `intergenic` |
| `consequence` | str | Most-severe SnpEff SO term from MANE Select transcripts (except where a dataset carries its own scheme — see per-dataset notes) |
| `variant_type` | str | `SNV` / `insertion` / `deletion` / `MNV` (derived from ref/alt lengths) |

`omim`, `gnomad_balanced`, and the TraitGym sets carry their own pre-annotation scheme for
`consequence` / `genomic_element` — see the individual sections.

---

## Dataset index

| File | Task | n (total) | n pos | n neg |
|---|---|---|---|---|
| `clinvar/clinvar.parquet` | P/LP vs B/LB | 200,036 | 66,987 | 133,049 |
| `splicevar/splicevar.parquet` | Splice-altering vs Normal | 11,978 | 7,147 | 4,831 |
| `gpn_star/cosmic.parquet` | Somatic cancer mutations vs neutral | 18,903 | 183 | 18,720 |
| `gpn_star/omim.parquet` | Disease-associated vs common variants | 2,640,672 | 406 | 2,640,266 |
| `gpn_star/gnomad_balanced.parquet` | Balanced variant benchmark | 11,992,284 | 5,996,146 | 5,996,138 |
| `traitgym/complex_traits.parquet` | GWAS causal (non-coding) vs matched | 11,400 | 1,140 | 10,260 |
| `traitgym/mendelian_traits.parquet` | Mendelian disease causal vs matched | 3,380 | 338 | 3,042 |

---

## ClinVar (`clinvar/`)

**Source:** ClinVar (GRCh38). Combined (un-split) variant table.
**Eval task:** Distinguish Pathogenic/Likely Pathogenic from Benign/Likely Benign variants.
**Positive class:** `ClinSigSimple = 1` (P/LP)
**Negative class:** `ClinSigSimple = 0` (B/LB)
**Class balance:** 66,987 positive : 133,049 negative

### Columns

| Column | Type | Notes |
|---|---|---|
| `chrom` | str | chr-prefixed |
| `pos` | int64 | |
| `ref` | str | |
| `alt` | str | |
| `ClinSigSimple` | int64 | **Label**: 1 = P/LP, 0 = B/LB |
| `ClinicalSignificance` | str | Raw ClinVar significance text (5 distinct values) |
| `ReviewStatus` | str | ClinVar submission confidence tier (3 distinct values) |
| `NumberSubmitters` | int64 | Number of submitting labs |
| `GeneSymbol` | str | Gene symbol |
| `VariationID` | int64 | ClinVar variant identifier |
| `feature_lvl2` | str | Author-provided element annotation |
| `genomic_element` | str | SnpEff coarse element |
| `consequence` | str | SnpEff most-severe SO term (24 distinct values) |
| `variant_type` | str | SNV / insertion / deletion / MNV |

### Stratification

- **`genomic_element`** — CDS (124,449), intron (43,273), splice_site (16,910), 3UTR (7,488), intergenic (6,064), 5UTR (1,852)
- **`variant_type`** — SNV (173,657), deletion (16,975), insertion (9,114), MNV (290)
- **`consequence`** — 24 SnpEff SO strata

---

## SpliceVarDB (`splicevar/`)

**Source:** SpliceVarDB. Combined (un-split) variant table.
**Eval task:** Distinguish experimentally confirmed splice-altering variants from normal (non-altering) variants.
**Positive class:** `classification = "Splice-altering"`
**Negative class:** `classification = "Normal"`
**Class balance:** 7,147 positive : 4,831 negative. All variants are SNVs.

### Columns

| Column | Type | Notes |
|---|---|---|
| `chrom` | str | chr-prefixed |
| `pos` | int64 | Genomic VCF position |
| `ref` | str | |
| `alt` | str | |
| `classification` | str | **Label**: "Splice-altering" or "Normal" |
| `gene` | str | Gene symbol |
| `hgvs` | str | HGVS notation |
| `method` | str | Experimental assay (RNA-Seq, MFASS, MaPSy, Minigene Assay, RT-PCR, …) |
| `location` | str | "Intronic", "Exonic", or "Exonic,Intronic" |
| `is_coding` | bool | Whether the variant falls in coding sequence |
| `genomic_element` | str | SnpEff coarse element |
| `consequence` | str | SnpEff most-severe SO term |
| `variant_type` | str | All SNV |

### Stratification

- **`method`** — RNA-Seq (6,117), MFASS (5,022), MaPSy (326), Minigene Assay (296), RT-PCR (119), Unspecified (45), plus a few combined-assay values
- **`location`** — Intronic (7,407), Exonic (4,338), Exonic,Intronic (233)
- **`is_coding`** — True (8,201) / False (3,777)

---

## GPN-star benchmarks (`gpn_star/`)

**Source:** [songlab/gpn-star](https://huggingface.co/collections/songlab/gpn-star) on HuggingFace.
Pre-computed model scores (GPN-MSA, CADD, phyloP, phastCons, ESM-1b, NT) have been stripped;
these are the base variant tables.

---

### `cosmic.parquet` — COSMIC somatic mutations vs neutral variants

**Eval task:** Distinguish COSMIC-catalogued somatic cancer mutations from common neutral variants.
**Positive class:** `label = True` — COSMIC-catalogued somatic cancer mutation
**Negative class:** `label = False` — common benign missense variant (gnomAD, AF > 5%)
**Class balance:** 183 positive : 18,720 negative (102:1)

| Column | Type | Notes |
|---|---|---|
| `chrom` | str | chr-prefixed |
| `pos` | int64 | |
| `ref` | str | |
| `alt` | str | |
| `label` | bool | **Label**: True = COSMIC cancer mutation |
| `consequence` | str | SnpEff most-severe SO term |
| `genomic_element` | str | SnpEff coarse element |
| `variant_type` | str | All SNV |

> **Important caveat:** 180 of 183 positives (98%) are `missense_variant` in `CDS`. All other
> genomic_element/consequence strata have 0 or 3 positives. Stratified AUCs will mostly return
> None; the overall AUC is the meaningful metric here.

### Stratification

- **`genomic_element`** — only `CDS` (180 pos) and `intergenic` (3 pos) have any positives.
- **`consequence`** — only `missense_variant` (180 pos) and `intergenic_region` (3 pos) have positives.

---

### `omim.parquet` — OMIM disease-associated variants vs common variants

**Eval task:** Distinguish OMIM-listed regulatory disease variants from common population variants.
**Positive class:** `label = True` — OMIM-curated pathogenic regulatory variant
**Negative class:** `label = False` — common variant (gnomAD, AF > 5%)
**Class balance:** 406 positive : 2,640,266 negative (6,500:1)

| Column | Type | Notes |
|---|---|---|
| `chrom` | str | chr-prefixed |
| `pos` | int64 | |
| `ref` | str | |
| `alt` | str | |
| `label` | bool | **Label**: True = OMIM disease-associated |
| `consequence` | str | **Custom OMIM categories** (not SnpEff SO terms) |
| `variant_type` | str | All SNV |
| `genomic_element` | str | Mapped from custom consequence (see below) |

**Custom consequence values and their genomic_element mapping:**

| consequence | genomic_element | n total | n pos |
|---|---|---|---|
| `Enhancer` | `intergenic` | 2,357,316 | 37 |
| `ncRNA` | `ncRNA` | 137,330 | 60 |
| `3' UTR` | `3UTR` | 68,413 | 38 |
| `Promoter` | `intergenic` | 62,575 | 130 |
| `5' UTR` | `5UTR` | 15,030 | 133 |
| `MicroRNA Gene` | `ncRNA` | 5 | 5 |
| `Imprinting Control Region` | `intergenic` | 3 | 3 |

> All variants are **non-coding regulatory** elements. The `Enhancer` category dominates (89% of
> rows) with very sparse positives (37 of 2.36M).

### Stratification

- **`genomic_element`** — 4 strata, all ≥20 rows with both classes: intergenic (170 pos), ncRNA (65), 3UTR (38), 5UTR (133).
- **`consequence`** — 5 valid strata: Enhancer (37), ncRNA (60), 3' UTR (38), Promoter (130), 5' UTR (133).

---

### `gnomad_balanced.parquet` — gnomAD balanced benchmark

**Eval task:** General variant-effect prediction benchmark using gnomAD population variants.
**Positive class:** `label = True` — rare variant (singleton in gnomAD, under purifying selection)
**Negative class:** `label = False` — common variant (AF > 5% in gnomAD, likely neutral)
**Class balance:** 5,996,146 : 5,996,138 (perfectly balanced, 50:50)

> **Label semantics (GPN-star paper, Benegas et al. 2024):** True = gnomAD singleton; False =
> common variant (AF > 5%). Purifying selection keeps deleterious variants rare, so rare variants
> are enriched for functional/deleterious effect. A model with signal assigns more negative delta
> scores to singletons than to common variants.

| Column | Type | Notes |
|---|---|---|
| `chrom` | str | chr-prefixed |
| `pos` | int64 | |
| `ref` | str | |
| `alt` | str | |
| `label` | bool | **Label**: True/False (see note above) |
| `consequence` | str | Simplified consequence terms (see below) |
| `variant_type` | str | All SNV |
| `genomic_element` | str | Mapped from simplified consequence |

**Simplified consequence values and their genomic_element mapping (top strata):**

| consequence | genomic_element | n total | n pos | n neg |
|---|---|---|---|---|
| `intron` | `intron` | 6,481,608 | 3,240,804 | 3,240,804 |
| `intergenic` | `intergenic` | 4,714,558 | 2,357,279 | 2,357,279 |
| `non_coding_transcript_exon` | `ncRNA` | 274,540 | 137,270 | 137,270 |
| `3_prime_UTR` | `3UTR` | 136,750 | 68,375 | 68,375 |
| `downstream_gene` | `intergenic` | 129,544 | 64,772 | 64,772 |
| `upstream_gene` | `intergenic` | 124,890 | 62,445 | 62,445 |
| `synonymous` | `CDS` | 37,948 | 18,974 | 18,974 |
| `missense` | `CDS` | 37,452 | 18,726 | 18,726 |
| `5_prime_UTR` | `5UTR` | 29,794 | 14,897 | 14,897 |
| `splice_region` | `splice_site` | 24,170 | 12,085 | 12,085 |

> Consequence terms are simplified (not standard SO format): `intron` not `intron_variant`,
> `missense` not `missense_variant`. Perfectly balanced within every stratum (pos/neg equal or off by 1).

### Stratification

- **`genomic_element`** — 7 strata, all perfectly balanced and ≥20 rows.
- **`consequence`** — 13 valid strata, all perfectly balanced.

---

## TraitGym (`traitgym/`)

**Source:** [songlab/TraitGym](https://huggingface.co/datasets/songlab/TraitGym) on HuggingFace.
**Eval task:** Distinguish causal variants from matched controls selected within the same functional
category with similar MAF and LD score — the model cannot win on consequence or frequency bias alone.
**Positive class:** `label = True` (causal variant)
**Negative class:** `label = False` (matched control)

**Annotation:** Pre-annotated by the TraitGym authors using ENCODE regulatory-element categories
(`dELS`, `pELS`, `PLS`, …) combined with SO terms — these are kept as-is (SnpEff annotation is
deliberately not applied, as it would overwrite the informative ENCODE categories).

**Consequence categories used:**

| Term | Meaning |
|---|---|
| `PLS` | Promoter-like signature (ENCODE cRE) |
| `pELS` | Proximal enhancer-like signature |
| `dELS` | Distal enhancer-like signature |
| `pELS_flank`, `dELS_flank` | Flanking regions of ELS elements |
| `intron_variant`, `intergenic_variant`, `non_coding_transcript_exon_variant`, `3_prime_UTR_variant`, `5_prime_UTR_variant`, `upstream_gene_variant` | SO terms |

---

### `complex_traits.parquet` — GWAS fine-mapped non-coding variants

**Eval task:** GWAS fine-mapped causal non-coding variants vs matched controls.
**Class balance:** 1,140 causal : 10,260 controls (1:9)

| Column | Type | Notes |
|---|---|---|
| `chrom` | str | chr-prefixed |
| `pos` | int64 | |
| `ref` | str | |
| `alt` | str | |
| `pip` | float64 | Posterior inclusion probability (causal prior) |
| `trait` | str | GWAS trait name (empty string `""` for controls) |
| `label` | bool | **Label**: True = causal |
| `maf` | float64 | Minor allele frequency |
| `ld_score` | float64 | LD score |
| `consequence` | str | ENCODE/SO category (see above) |
| `tss_dist` | int64 | Distance to nearest transcription start site |
| `match_group` | str | Matching group ID (each causal paired with controls) |

> `trait` is empty string for all controls — per-trait stratification would create causal-only
> strata, so it is excluded from stratification.

### Stratification

- **`consequence`** — 15 valid strata: dELS (314 pos), intron_variant (190), dELS_flank (167), intergenic_variant (103), pELS (101), … Each stratum keeps the 1:9 pos:neg ratio by design.

---

### `mendelian_traits.parquet` — Mendelian disease regulatory variants

**Eval task:** Mendelian disease regulatory causal variants vs matched controls (promoters, UTRs, ncRNA from OMIM).
**Class balance:** 338 causal : 3,042 controls (1:9)

| Column | Type | Notes |
|---|---|---|
| `chrom` | str | chr-prefixed |
| `pos` | int64 | |
| `ref` | str | |
| `alt` | str | |
| `OMIM` | str | OMIM gene/disease ID (NaN for controls) |
| `consequence` | str | ENCODE/SO category (see above) |
| `label` | bool | **Label**: True = disease-causing |
| `tss_dist` | int64 | Distance to nearest TSS |
| `match_group` | str | Matching group ID |

### Stratification

- **`consequence`** — 11 valid strata: 5_prime_UTR_variant (114 pos), non_coding_transcript_exon_variant (71), PLS (63), 3_prime_UTR_variant (29), upstream_gene_variant (21), … Smaller counts — some strata will have low power.