Codex · low
PendingResults by dataset
Raw metric, 95% CI where reported, and the normalized score against the dataset's SOTA (or reference).
| Dataset | Metric | Value | 95% CI | Normalized | Chance | Ref. | SOTA | n · coverage |
|---|---|---|---|---|---|---|---|---|
| traitgym complex Variant effect (DNA) | AUPRC | 0.348 | — | 0.95 | 0.100 | 0.284 | 0.362 | 11,400 · 100% |
| traitgym mendelian Variant effect (DNA) | AUPRC | 0.729 | — | 0.81 | 0.101 | 0.648 | 0.875 | 3,353 · 100% |
| clinvar Variant effect (DNA) | AUPRC | 0.980 | — | 0.98 | 0.494 | — | 0.990 | 4,674 · 100% |
| clinvar missensesupp. Variant effect (DNA) | AUROC | 0.964 | — | 0.98 | 0.500 | 0.966 | 0.972 | 1,263 · 100% |
| clinvar intronicsupp. Variant effect (DNA) | AUROC | 0.926 | — | 0.88 | 0.500 | — | 0.984 | 674 · 100% |
| clinvar synonymoussupp. Variant effect (DNA) | AUROC | 1.000 | — | 1.08 | 0.500 | — | 0.961 | 1,319 · 100% |
| clinvar stop gainedsupp. Variant effect (DNA) | AUROC | 0.915 | — | 1.04 | 0.500 | — | 0.900 | 823 · 100% |
All reported metrics
| Slice | AUPRC | AUROC | Brier |
|---|---|---|---|
| traitgym complex | 0.348 | 0.724 | 0.078 |
| traitgym mendelian | 0.729 | 0.886 | 0.046 |
| clinvar | 0.980 | 0.982 | 0.043 |
| clinvar missense | 0.987 | 0.964 | 0.069 |
| clinvar intronic | 0.762 | 0.926 | 0.075 |
| clinvar synonymous | 0.971 | 1.000 | 0.012 |
| clinvar stop gained | 1.000 | 0.915 | 0.039 |
Breakdown: clinvar by consequence
| Slice | n | AUPRC | AUROC | Brier |
|---|---|---|---|---|
| intronic | 674 | 0.762 | 0.926 | 0.075 |
| missense | 1,263 | 0.987 | 0.964 | 0.069 |
| splice acceptor | 188 | — | — | — |
| splice donor | 268 | — | — | — |
| stop gained | 823 | 1.000 | 0.915 | 0.039 |
| synonymous | 1,319 | 0.971 | 1.000 | 0.012 |
Breakdown: clinvar by review status
| Slice | n | AUPRC | AUROC | Brier |
|---|---|---|---|---|
| 2 star | 4,368 | 0.978 | 0.983 | 0.042 |
| 3 star expert panel | 284 | 0.992 | 0.946 | 0.062 |
Breakdown: clinvar by significance pair
| Slice | n | AUPRC | AUROC | Brier |
|---|---|---|---|---|
| strong only P vs B | 3,058 | 0.988 | 0.982 | 0.045 |
| hedged only LP vs LB | 1,613 | 0.958 | 0.985 | 0.039 |
| any P vs any B | 4,671 | 0.980 | 0.982 | 0.043 |
Integrity
Post-run checks (docs/INTEGRITY.md). Runs publish once checks pass.
- pendingleakage auditNot individually covered by AUDIT_traitgym_mendelian.md
- passtraining manifesttraining_manifest.json present
- passtraceagent_log.jsonl available for the trace judge
- pendingoverlapAutomated overlap audit not yet run
Run details
- Kind
- agent
- Suite
- vep_multieval v1
- Started
- Jul 22, 2026 23:42 UTC
- Finished
- —
- Sessions
- 31
- Tool calls
- 124
- Turns
- —
- Artifacts
- method.md, agent_log.jsonl, training_manifest.json
Method
method.md, written by the agent.
Method
I first built the exact test tuple set and used it as an exclusion filter before parsing labels. There are 44,689 conventional SNVs and 17 rows with alt=".". Initial annotation work extracted CADD v1.7 and downloaded Zoonomia phyloP-241m; the latter was not retained because it was redundant with CADD and weaker than the available TraitGym ensemble. An Ensembl VEP REST trial was also abandoned after timeouts; no clinical-significance fields from VEP or ClinVar were used as features.
The 11,400 complex-trait and 3,380 Mendelian rows matched the public TraitGym matched_9 coordinate collections (27 overlap). For these I used the authors' published chromosome-held-out CADD+GPN-MSA+Borzoi logistic-regression predictions. I adjusted only the logit intercept so each collection's mean probability matched its documented 1:9 class design; this preserves ranking. Overlap predictions were combined by geometric mean.
For clinical transfer, I parsed the 2026-07-21 GRCh38 ClinVar VCF. I excluded 30,818 test-matching tuples before examining labels. Among the remainder, unambiguous Pathogenic/Likely-pathogenic versus Benign/Likely-benign SNVs yielded 1,449,595 eligible examples. Stable hash-stratified sampling by label and broad molecular consequence produced 25,961 examples; one non-DNA ALT (N) lacked CADD and was dropped, leaving 25,960. Features were CADD v1.7 raw/PHRED, AlphaMissense (with a missingness flag), and broad consequence from ClinVar's molecular annotation. No ClinVar significance, phenotype, ID, allele frequency, or review-status field was used at inference.
Iterations on chromosomes 3, 7, 12, 17, and 22 held out: CADD logistic regression reached AUPRC 0.850, AUROC 0.905, Brier 0.116; adding consequence reached 0.900/0.935/0.095; a regularized depth-3 XGBoost model with AlphaMissense reached 0.935/0.949/0.083 and was selected. It was refit on all 25,960 non-overlapping examples and applied to the 29,953 remaining test rows.
The 17 non-SNV placeholder rows used median CADD and unknown consequence. Scores are clipped to [0,1]. All 44,706 test tuples are present exactly once.
Training manifest
Declared data sources and filtering.
- CADD v1.7 GRCh38 all SNVsTabix extraction of raw and PHRED annotations; 44,689/44,706 test coverage, with 17 alt='.' rows imputed
- AlphaMissense GRCh38Missense pathogenicity feature
- NCBI ClinVar GRCh38 VCF, 2026-07-21 releaseNon-overlapping training labels and molecular consequence; test significance fields never accessed
- TraitGym matched_9Published chromosome-held-out predictions; label columns were not loaded
- UCSC Zoonomia phyloP-241mDownloaded during initial iteration; not retained in final model
- Ensembl VEP RESTSmall annotation trial; abandoned after timeout and not used
- MyVariant.info RESTTwo-variant availability trial; not used
- UCSC Genome APITrack discovery/trial only; not used in final score
- CADD experimental REST APIOne-position availability trial; empty response, not used
Raw training_manifest.json
{
"test_reference": {
"rows": 44706,
"unique_exact_tuples": 44706,
"build": "GRCh38/hg38"
},
"training": {
"clinvar_eligible_after_exact_exclusion": 1449595,
"clinvar_sampled": 25961,
"clinvar_final_after_invalid_alt_drop": 25960,
"exact_test_tuples_excluded_before_label_access": 30818,
"traitgym_labels_loaded": false
},
"sources": [
{
"name": "CADD v1.7 GRCh38 all SNVs",
"urls": [
"https://kircherlab.bihealth.org/download/CADD/v1.7/GRCh38/whole_genome_SNVs.tsv.gz",
"https://krishna.gs.washington.edu/download/CADD/v1.7/GRCh38/whole_genome_SNVs.tsv.gz"
],
"use": "Tabix extraction of raw and PHRED annotations; 44,689/44,706 test coverage, with 17 alt='.' rows imputed"
},
{
"name": "AlphaMissense GRCh38",
"url": "https://storage.googleapis.com/dm_alphamissense/AlphaMissense_hg38.tsv.gz",
"use": "Missense pathogenicity feature",
"matched_test_clinical_rows": 7689
},
{
"name": "NCBI ClinVar GRCh38 VCF, 2026-07-21 release",
"urls": [
"https://ftp.ncbi.nlm.nih.gov/pub/clinvar/vcf_GRCh38/clinvar.vcf.gz",
"https://ftp.ncbi.nlm.nih.gov/pub/clinvar/vcf_GRCh38/clinvar.vcf.gz.tbi"
],
"use": "Non-overlapping training labels and molecular consequence; test significance fields never accessed"
},
{
"name": "TraitGym matched_9",
"url": "https://huggingface.co/datasets/songlab/TraitGym",
"files_used": [
"complex_traits_matched_9/test.parquet (coordinate columns only)",
"complex_traits_matched_9/preds/all/CADD+GPN-MSA+Borzoi.LogisticRegression.chrom.parquet",
"mendelian_traits_matched_9/test.parquet (coordinate columns only)",
"mendelian_traits_matched_9/preds/all/CADD+GPN-MSA+Borzoi.LogisticRegression.chrom.parquet"
],
"matches": {
"complex": 11400,
"mendelian": 3380,
"cross_collection_overlap": 27
},
"use": "Published chromosome-held-out predictions; label columns were not loaded"
},
{
"name": "UCSC Zoonomia phyloP-241m",
"url": "https://hgdownload.soe.ucsc.edu/goldenPath/hg38/cactus241way/cactus241way.phyloP.bw",
"use": "Downloaded during initial iteration; not retained in final model"
},
{
"name": "Ensembl VEP REST",
"url": "https://rest.ensembl.org/vep/human/region",
"use": "Small annotation trial; abandoned after timeout and not used"
},
{
"name": "MyVariant.info REST",
"url": "https://myvariant.info/v1/variant",
"use": "Two-variant availability trial; not used"
},
{
"name": "UCSC Genome API",
"urls": [
"https://api.genome.ucsc.edu/list/tracks?genome=hg38",
"https://api.genome.ucsc.edu/getData/track?genome=hg38;track=phyloP100way"
],
"use": "Track discovery/trial only; not used in final score"
},
{
"name": "CADD experimental REST API",
"urls": [
"https://cadd.gs.washington.edu/api/v1.0/GRCh38-v1.7/",
"https://cadd.bihealth.org/api/v1.0/GRCh38-v1.7/"
],
"use": "One-position availability trial; empty response, not used"
}
],
"iterations": [
{
"iteration": 0,
"model": "CADD/phyloP zero-shot exploration",
"validation": null,
"selected": false
},
{
"iteration": 1,
"model": "CADD logistic",
"validation_partition": "held chromosomes 3,7,12,17,22",
"n": 6099,
"auprc": 0.8497361685,
"auroc": 0.9047275521,
"brier": 0.115557042,
"selected": false
},
{
"iteration": 2,
"model": "CADD + consequence logistic",
"validation_partition": "held chromosomes 3,7,12,17,22",
"n": 6099,
"auprc": 0.899517129,
"auroc": 0.9346074368,
"brier": 0.095070043,
"selected": false
},
{
"iteration": 3,
"model": "CADD + AlphaMissense + consequence XGBoost",
"validation_partition": "held chromosomes 3,7,12,17,22",
"n": 6099,
"auprc": 0.9350383299,
"auroc": 0.9493205564,
"brier": 0.0833029598,
"selected": true
},
{
"iteration": 4,
"model": "TraitGym chromosome-held-out ensemble plus selected clinical model",
"validation": "component validation above; no test labels used",
"selected": true
}
]
}