Datasets:

bolinas-dna
/

evals_complex_traits

@@ -1,6 +1,7 @@
 ---
 license: apache-2.0
 tags:
   - genomics
   - dna
   - variant-effect-prediction
@@ -8,30 +9,34 @@ tags:
   - gwas
   - fine-mapping
 size_categories:
-  - 1K<n<10K
 ---
 # evals_complex_traits
-Variant-effect-prediction benchmark of UKBB fine-mapped complex-trait SNVs vs low-PIP SNVs, gene-matched 1:1 within consequence categories and MAF bins.
 ## Description
 | | |
 |---|---|
-| Positives | UKBB SuSiE+FINEMAP fine-mapped variants with PIP > 0.9 across 119 traits |
-| Negatives | PIP < 0.01 (and not null in any of the 119 traits), gene-matched 1:1 to positives |
-| Genome build | GRCh38 (lifted from hg19 via [`liftover`](https://github.com/konstantint/pyliftover)) |
 | Variant type | SNVs only |
 | Coordinates | 1-based (`pos` is 1-based; `ref`/`alt` are single bases) |
 ## Splits
-| Split | Variants (positives + matched negatives) | Chromosomes |
-|---|---:|---|
-| `train` | 1 128 (564 pos + 564 neg) | odd: 1, 3, …, X |
-| `test`  | 1 140 (570 pos + 570 neg) | even: 2, 4, …, Y |
-| **total** | **2 268** | |
 ## Columns
@@ -39,99 +44,95 @@ Variant-effect-prediction benchmark of UKBB fine-mapped complex-trait SNVs vs lo
 |---|---|---|
 | `chrom`, `pos`, `ref`, `alt` | str / int / str / str | Variant coordinates (1-based, GRCh38) |
 | `label` | bool | `True` for high-PIP positive, `False` for low-PIP matched negative |
-| `subset` | str | Consequence-group label for stratified eval (`distal`, `missense_variant`, `tss_proximal`, `3_prime_UTR_variant`, `non_coding_transcript_exon_variant`, `5_prime_UTR_variant`, `synonymous_variant`, `splicing`) |
-| `match_group` | int | Pos / neg pair ID (every group has 1 positive + 1 matched negative) |
 | `rsid` | str | dbSNP rsID (when available) |
 | `pip` | float | Maximum PIP across the 119 traits |
 | `traits` | str | Comma-separated list of traits with PIP > 0.9 (positives only) |
-| `MAF` | float | UKBB EUR minor allele frequency from the LD-score reference panel |
-| `ld_score` | float | UKBB EUR LD score (passthrough column; **not** used as a matching feature) |
-| `consequence`, `consequence_cre`, `consequence_final`, `consequence_group` | str | Ensembl VEP consequence + grouping used by the matcher |
-| `distance_tss_pc`, `distance_tss_nc` | int | Distance (0-based, half-open) to nearest protein-coding / non-protein-coding transcript TSS |
-| `distance_tss` | int | min(`distance_tss_pc`, `distance_tss_nc`); used by `consequence_group` recategorization |
-| `tss_closest_pc_gene_id`, `tss_closest_nc_gene_id`, `tss_closest_gene_id` | str | Ensembl gene IDs at those distances |
 | `distance_exon_pc`, `distance_exon_nc`, `distance_exon` | int | Same shape, for nearest exon |
 | `exon_closest_pc_gene_id`, `exon_closest_nc_gene_id`, `exon_closest_gene_id` | str | Same shape |
-| `distance_tss_pc_bin`, `distance_tss_nc_bin`, `distance_exon_pc_bin`, `MAF_bin` | str | Categorical bins used as exact-match keys during gene-matching |
 ## Per-subset retention
-| Subset | n_pos in dataset_all | matched (kept) | retention |
 |---|---:|---:|---:|
-| distal | 1 193 | 857 | 71.8 % |
-| missense_variant | 454 | 125 | 27.5 % |
-| tss_proximal | 244 | 56 | 23.0 % |
-| 3_prime_UTR_variant | 78 | 36 | 46.2 % |
-| non_coding_transcript_exon_variant | 75 | 28 | 37.3 % |
-| 5_prime_UTR_variant | 56 | 15 | 26.8 % |
-| synonymous_variant | 33 | 15 | 45.5 % |
-| splicing | 30 | 2 | 6.7 % |
-| **total** | **2 165** | **1 134** | **52.4 %** |
-52 % of input fine-mapped SNVs are retained. The biggest hits are:
-- **missense** — complex-trait missense positives skew to common-MAF GWAS hits while gnomAD missense negatives skew rare; even the per-subset 10-bin MAF scheme can only stretch so far.
-- **splicing** — splicing variants live next to a small set of specific exons; the gene-id match key starves neg supply.
-## Matching design (locked iter 33, [issue #156](https://github.com/Open-Athena/bolinas-dna/issues/156))
-Matching is exact on every categorical key, then Euclidean-nearest on the (RobustScaler-scaled) continuous features as a within-group tie-breaker. Without replacement, k=1.
 - **Continuous features**: `distance_tss_pc`, `distance_tss_nc`, `distance_exon_pc`, `distance_exon_nc`, `MAF`.
 - **Categorical features**:
   - `chrom`, `consequence_final`
-  - `tss_closest_pc_gene_id`, `tss_closest_nc_gene_id`, `exon_closest_pc_gene_id`, `exon_closest_nc_gene_id`
-  - `distance_tss_pc_bin`, `distance_tss_nc_bin` (`tss_proximal` only; edges `[0, 50, 100, 200, 500, 1000]`)
-  - `distance_exon_pc_bin` (`splicing` only; edges `[0, 5, 20, 30]`)
-  - `MAF_bin` — **per-subset tiered scheme** (`MAF_TIERED_V1`):
-    - `distal`, `tss_proximal`, `non_coding_transcript_exon_variant`: 20-bin (right-closed log-spaced toward low MAF; `MAF_BIN_EDGES_20`)
-    - `3_prime_UTR_variant`, `5_prime_UTR_variant`, `missense_variant`: 10-bin (`MAF_BIN_EDGES_10`)
-    - `synonymous_variant`, `splicing`, …: 5-bin (`MAF_BIN_EDGES_5`)
-`ld_score` is included as a passthrough column for downstream diagnostics but is **not** in the match key.
-**Matched-feature diagnostic** (iter 33): **zero Bonferroni-significant leaks** on the 5 matched features. `ld_score` (passthrough, **not** in the match key) still shows a Bonferroni-significant skew on `distal` (PA=0.376, p=3e-13 ★) — known dataset property, not a matching artefact.
-<details>
-<summary>Full per-(subset, feature) PA / p-value table</summary>
-Format: PA / p. `★` = Bonferroni-significant at α = 0.05 / 40 = 1.25e-3 (8 subsets × 5 matched features). Generated by [`snakemake/evals/scratch/leakage_v2.py`](https://github.com/Open-Athena/bolinas-dna/blob/6e34e07/snakemake/evals/scratch/leakage_v2.py).
 | subset | n | distance_tss_pc | distance_tss_nc | distance_exon_pc | distance_exon_nc | MAF |
-|---|---:|---|---|---|---|---|
-| distal | 857 | 0.470 / 9e-02 | 0.494 / 7e-01 | 0.468 / 7e-02 | 0.512 / 5e-01 | 0.538 / 3e-02 |
-| missense_variant | 125 | 0.504 / 1.0 | 0.488 / 9e-01 | 0.488 / 3e-01 | 0.536 / 3e-01 | 0.584 / 7e-02 |
-| tss_proximal | 56 | 0.482 / 9e-01 | 0.357 / 4e-02 | 0.500 / 1.0 | 0.347 / 6e-03 | 0.500 / 1.0 |
-| 3_prime_UTR_variant | 36 | 0.417 / 4e-01 | 0.333 / 7e-02 | 0.472 / 6e-01 | 0.500 / 1.0 | 0.500 / 1.0 |
-| non_coding_transcript_exon_variant | 28 | 0.321 / 9e-02 | 0.464 / 9e-01 | 0.429 / 6e-01 | — | 0.429 / 6e-01 |
-| synonymous_variant | 15 | 0.333 / 3e-01 | 0.533 / 1.0 | — | 0.467 / 1.0 | 0.667 / 3e-01 |
-| 5_prime_UTR_variant | 15 | 0.333 / 3e-01 | 0.467 / 1.0 | 0.467 / 1.0 | 0.667 / 2e-01 | 0.633 / 4e-01 |
-| splicing | 2 | 0.000 / 5e-01 | 0.000 / 5e-01 | 1.000 / 5e-01 | 0.500 / 1.0 | 0.500 / 1.0 |
 </details>
 ## Provenance
-Built by the [`bolinas-dna`](https://github.com/Open-Athena/bolinas-dna) eval pipeline at commit [`6e34e07`](https://github.com/Open-Athena/bolinas-dna/tree/6e34e07/snakemake/evals).
-- Curation pipeline: [`snakemake/evals/`](https://github.com/Open-Athena/bolinas-dna/tree/6e34e07/snakemake/evals)
-- Matching algorithm: [`src/bolinas/evals/matching.py`](https://github.com/Open-Athena/bolinas-dna/blob/6e34e07/src/bolinas/evals/matching.py)
-- Design discussion: [issue #156](https://github.com/Open-Athena/bolinas-dna/issues/156)
 The curation is a from-scratch reimplementation of the [TraitGym](https://github.com/songlab-cal/TraitGym) complex-traits pipeline.
-## Companion datasets
-- **[`bolinas-dna/evals_complex_traits_harness_255`](https://huggingface.co/datasets/bolinas-dna/evals_complex_traits_harness_255)** — same variants with 255 bp reference-genome windows materialized into `context` / `ref_completion` / `alt_completion` columns for direct use as eval-harness inputs.
 ## License
-Released under the same terms as its sources. UKBB summary-level data and the [Finucane lab fine-mapping release](https://huggingface.co/datasets/gonzalobenegas/finucane-ukbb-finemapping) are intended for non-commercial research; check upstream license if you plan to use commercially.
 ## Citation
-If you use this benchmark, please cite the upstream sources:
-- TraitGym — Benegas *et al.* 2025, [bioRxiv 2025.02.11.637758](https://www.biorxiv.org/content/10.1101/2025.02.11.637758v2) ([songlab-cal/TraitGym](https://github.com/songlab-cal/TraitGym))
 - UKBB fine-mapping — Wang *et al.* (Nat Commun 2021) and the [Finucane lab release](https://huggingface.co/datasets/gonzalobenegas/finucane-ukbb-finemapping)
 - LD scores — Bulik-Sullivan *et al.* (Nat Genet 2015)

 ---
 license: apache-2.0
 tags:
+  - biology
   - genomics
   - dna
   - variant-effect-prediction
   - gwas
   - fine-mapping
 size_categories:
+  - 10K<n<100K
 ---
 # evals_complex_traits
+Variant-effect-prediction benchmark of UKBB fine-mapped complex-trait SNVs vs
+low-PIP SNVs, 1:9 matched within consequence categories on `(chrom,
+consequence_final)` plus subset-targeted distance bins, with MAF entering as
+a continuous matching feature.
 ## Description
 | | |
 |---|---|
+| Positives | UKBB SuSiE+FINEMAP fine-mapped variants with max(PIP) > 0.9 across 119 traits |
+| Negatives | max(PIP) < 0.01 across 119 traits, 1:9 matched per positive |
+| Genome build | GRCh38 (lifted from hg19) |
 | Variant type | SNVs only |
 | Coordinates | 1-based (`pos` is 1-based; `ref`/`alt` are single bases) |
+| Matching ratio | 1:9 |
 ## Splits
+| Split | Variants (pos + 9·neg) | Positives | Chromosomes |
+|---|---:|---:|---|
+| `train` | 11,630 | 1,163 | odd: 1, 3, …, X |
+| `test` | 10,000 | 1,000 | even: 2, 4, …, Y |
+| **total** | **21,630** | **2,163** | |
 ## Columns
 |---|---|---|
 | `chrom`, `pos`, `ref`, `alt` | str / int / str / str | Variant coordinates (1-based, GRCh38) |
 | `label` | bool | `True` for high-PIP positive, `False` for low-PIP matched negative |
+| `subset` | str | Consequence-group label for stratified eval |
+| `match_group` | int | Integer ID grouping each positive with its 9 matched negatives |
 | `rsid` | str | dbSNP rsID (when available) |
 | `pip` | float | Maximum PIP across the 119 traits |
 | `traits` | str | Comma-separated list of traits with PIP > 0.9 (positives only) |
+| `MAF` | float | UKBB EUR minor allele frequency |
+| `ld_score` | float | UKBB EUR LD score (passthrough, **not** a matching feature) |
+| `consequence`, `consequence_cre`, `consequence_final`, `consequence_group` | str | Ensembl VEP consequence + grouping |
+| `distance_tss_pc`, `distance_tss_nc`, `distance_tss` | int | Distances to nearest protein-coding / non-protein-coding TSS (and min, used for `consequence_group` recategorization) |
+| `tss_closest_pc_gene_id`, `tss_closest_nc_gene_id`, `tss_closest_gene_id` | str | Ensembl gene IDs (passthrough — gene-id was *not* used in matching) |
 | `distance_exon_pc`, `distance_exon_nc`, `distance_exon` | int | Same shape, for nearest exon |
 | `exon_closest_pc_gene_id`, `exon_closest_nc_gene_id`, `exon_closest_gene_id` | str | Same shape |
+| `distance_tss_pc_bin`, `distance_exon_pc_bin` | str | Subset-prefixed bin labels used as exact-match keys; `BIN_NA` outside the binned subsets |
 ## Per-subset retention
+| Subset | n_pos in `dataset_all` | matched (kept) | retention |
 |---|---:|---:|---:|
+| `distal` | 1,193 | 1,193 | 100.0% |
+| `missense_variant` | 454 | 454 | 100.0% |
+| `tss_proximal` | 244 | 244 | 100.0% |
+| `3_prime_UTR_variant` | 78 | 78 | 100.0% |
+| `non_coding_transcript_exon_variant` | 75 | 75 | 100.0% |
+| `5_prime_UTR_variant` | 56 | 56 | 100.0% |
+| `synonymous_variant` | 33 | 33 | 100.0% |
+| `splicing` | 30 | 30 | 100.0% |
+| `mature_miRNA_variant` | 2 | 0 | 0.0% |
+| **total** | **2,165** | **2,163** | **99.9%** |
+## Matching design
+Matching is exact on every categorical key, then Euclidean-nearest on the
+(RobustScaler-scaled) continuous features as a within-group tie-breaker.
+Without replacement, k=9.
 - **Continuous features**: `distance_tss_pc`, `distance_tss_nc`, `distance_exon_pc`, `distance_exon_nc`, `MAF`.
 - **Categorical features**:
   - `chrom`, `consequence_final`
+  - `distance_tss_pc_bin` — `tss_proximal`: edges `[0, 100, 1000, ∞]`; `BIN_NA` elsewhere
+  - `distance_exon_pc_bin` —
+    - `tss_proximal`: edges `[0, 100, 1000, ∞]`
+    - `splicing`: edges `[0, 5, 30, ∞]`
+    - `BIN_NA` elsewhere
+Gene-ID columns are kept as passthrough metadata but **not** used as match
+keys.
+## Matched-feature AUPRC diagnostic
+Each continuous matching feature `f` is scored as a single-feature predictor
+within each subset: `{f}_auprc = max(AP(label, +f), AP(label, −f))`.
+**Baseline = 0.1 for 1:9 matching**.
+<details>
+<summary>Per-(subset, feature) AUPRC table</summary>
 | subset | n | distance_tss_pc | distance_tss_nc | distance_exon_pc | distance_exon_nc | MAF |
+|---|---:|---:|---:|---:|---:|---:|
+| `distal` | 1,193 | 0.101 | 0.102 | 0.109 | 0.105 | 0.101 |
+| `missense_variant` | 454 | 0.108 | 0.106 | 0.102 | 0.104 | 0.108 |
+| `tss_proximal` | 244 | 0.114 | 0.114 | 0.110 | 0.108 | 0.101 |
+| `3_prime_UTR_variant` | 78 | 0.119 | 0.116 | 0.108 | 0.108 | 0.114 |
+| `non_coding_transcript_exon_variant` | 75 | 0.110 | 0.112 | 0.124 | 0.100 | 0.106 |
+| `5_prime_UTR_variant` | 56 | 0.123 | 0.109 | 0.102 | 0.110 | 0.109 |
+| `synonymous_variant` | 33 | 0.120 | 0.106 | 0.107 | 0.111 | 0.107 |
+| `splicing` | 30 | 0.124 | 0.131 | 0.108 | 0.131 | 0.122 |
 </details>
 ## Provenance
+Built by the [`bolinas-dna`](https://github.com/Open-Athena/bolinas-dna) eval pipeline at commit
+[`main`](https://github.com/Open-Athena/bolinas-dna/tree/main/snakemake/evals).
+- Curation pipeline: [`snakemake/evals`](https://github.com/Open-Athena/bolinas-dna/tree/main/snakemake/evals)
+- Matching algorithm: [`src/bolinas/pipelines/evals/matching.py`](https://github.com/Open-Athena/bolinas-dna/blob/main/src/bolinas/pipelines/evals/matching.py)
+- Diagnostic helper: [`src/bolinas/pipelines/evals/matching_qc.py`](https://github.com/Open-Athena/bolinas-dna/blob/main/src/bolinas/pipelines/evals/matching_qc.py)
 The curation is a from-scratch reimplementation of the [TraitGym](https://github.com/songlab-cal/TraitGym) complex-traits pipeline.
 ## License
+Released under the same terms as its sources. UKBB summary-level data and
+the [Finucane lab fine-mapping release](https://huggingface.co/datasets/gonzalobenegas/finucane-ukbb-finemapping)
+are intended for non-commercial research; check upstream license if you plan
+to use commercially.
 ## Citation
+- TraitGym — Benegas *et al.* 2025, [bioRxiv 2025.02.11.637758](https://www.biorxiv.org/content/10.1101/2025.02.11.637758v2)
 - UKBB fine-mapping — Wang *et al.* (Nat Commun 2021) and the [Finucane lab release](https://huggingface.co/datasets/gonzalobenegas/finucane-ukbb-finemapping)
 - LD scores — Bulik-Sullivan *et al.* (Nat Genet 2015)