Codex · low

Clean
20260722T065453-codex-lowharness: Codexmodel: unknowneffort: low(inferred)tag: codex-low
Mean AUPRC
0.690
task headline
Wall clock
—
of 10 h budget
Compute
1xH100
Agent cost
—
summed from trace
Started
Jul 22, 2026

Results by dataset

Raw metric and 95% CI where reported, next to the dataset's chance level, reference and SOTA.

DatasetMetricValue95% CIChanceRef.SOTAn · coverage
traitgym complex
Variant effect (DNA)
AUPRC0.348—0.1000.2840.36211,400 · 100%
traitgym mendelian
Variant effect (DNA)
AUPRC0.729—0.1010.6480.8753,353 · 100%
clinvar
Variant effect (DNA)
AUPRC0.992—0.494—0.99029,953 · 100%
clinvar missensesupp.
Variant effect (DNA)
AUROC0.955—0.5000.9660.9727,861 · 100%
clinvar intronicsupp.
Variant effect (DNA)
AUROC0.975—0.500—0.9844,368 · 100%
clinvar synonymoussupp.
Variant effect (DNA)
AUROC0.993—0.500—0.9618,155 · 100%
clinvar stop gainedsupp.
Variant effect (DNA)
AUROC0.823—0.500—0.9005,762 · 100%
All reported metrics
SliceAUPRCAUROCBrier
traitgym complex0.3480.7240.078
traitgym mendelian0.7290.8860.048
clinvar0.9920.9920.038
clinvar missense0.9820.9550.076
clinvar intronic0.8800.9750.041
clinvar synonymous0.9050.9930.031
clinvar stop gained1.0000.8230.003
Breakdown: clinvar by consequence
SlicenAUPRCAUROCBrier
3 prime utr4450.0380.9440.023
5 prime utr1700.5560.8340.106
initiator codon1630.9970.8680.066
intronic4,3680.8800.9750.041
missense7,8610.9820.9550.076
other1330.5870.8510.105
splice acceptor1,2311.0000.9390.010
splice donor1,6491.0000.9180.009
stop gained5,7621.0000.8230.003
synonymous8,1550.9050.9930.031
Breakdown: clinvar by review status
SlicenAUPRCAUROCBrier
2 star28,0580.9920.9920.038
3 star expert panel1,7640.9970.9860.037
unknown1240.9930.9950.029
Breakdown: clinvar by significance pair
SlicenAUPRCAUROCBrier
strong only P vs B20,1450.9950.9910.036
hedged only LP vs LB9,7900.9820.9910.042
any P vs any B29,9350.9920.9920.038

Integrity

Post-run checks (docs/INTEGRITY.md). Runs publish once checks pass.

CleanManual audit in tasks/vep_multieval/AUDIT_traitgym_mendelian.md
  • leakage audit
    no test-label leakage (AUDIT_traitgym_mendelian.md, 2026-07-23)
    pass
  • training manifest
    training_manifest.json present
    pass
  • trace
    agent_log.jsonl available for the trace judge
    pass
  • overlap
    Automated overlap audit not yet run
    pending

Run details

Kind
agent
Suite
vep_multieval v1
Started
Jul 22, 2026 06:54 UTC
Finished
—
Sessions
11
Tool calls
54
Turns
—
Artifacts
method.md, agent_log.jsonl, training_manifest.json

Method

method.md, written by the agent.

Method

I first checked normalization and uniqueness: the file contains 44,706 unique hg38 SNVs (22 autosomes plus X; no indels or duplicate tuples). No labeled training set was compiled, and no test labels were used.

Iterations were: (1) MyVariant batch lookup, rejected because CADD/dbNSFP coverage was poor; (2) sparse remote-tabix CADD v1.7 retrieval, confirmed exact allele scoring but rejected for full use because request latency was high; (3) exact joins to TraitGym's public matched and all-candidate files, extracting only coordinates, published predictions, CADD, and GPN-MSA columns; (4) sparse extraction of EVEE ClinVar predictions, rejected for the final model because its supervised ClinVar probe could overlap test variants; and (5) final Ensembl VEP batch annotation with CADD v1.7, AlphaMissense, SpliceAI, SIFT, PolyPhen, and consequence. The VEP parser intentionally discarded colocated ClinVar significance and all other label fields.

For tuples in TraitGym matched sets, I used its released chromosome-held-out CADD + GPN-MSA + Borzoi logistic-ensemble prediction. The 27 tuples shared by complex and Mendelian sets were averaged. Because TraitGym fit logistic regression with balanced class weights while matched_9 contains one positive per nine controls, I applied a prior-odds correction from 1:1 to 1:9; this preserves rankings.

All remaining rows received a consequence-aware clinical score. CADD PHRED was transformed by a fixed sigmoid centered at 15. Missense scores blend calibrated AlphaMissense (70%), CADD (20%), and available SIFT/PolyPhen (5% each). Core splice variants blend SpliceAI (65%) and CADD (35%); synonymous/intronic/splice-region variants take the stronger splice-aware signal. UTR variants use 80% CADD and 20% SpliceAI; other consequences use transformed CADD. Missing optional annotations are reweighted rather than imputed from labels. Scores are clipped to [0,1].

Validation selection used the published TraitGym chromosome-weighted AUPRCs supplied with the task (0.362 complex; 0.648 Mendelian); no local label-based validation was performed.

Training manifest

Declared data sources and filtering.

Raw training_manifest.json
{
  "schema_version": "1.0",
  "test_set": {
    "path": "/task/test.parquet",
    "build": "GRCh38/hg38",
    "rows": 44706,
    "unique_exact_tuples": 44706,
    "exact_duplicate_rows": 0
  },
  "labeled_training": {
    "compiled": false,
    "rows_before_test_exclusion": 0,
    "exact_test_overlaps_excluded": 0,
    "rows_after_test_exclusion": 0,
    "note": "No labels were parsed or used for fitting/calibration. Released out-of-chromosome predictions and label-independent annotations were used as features/scores."
  },
  "sources": [
    {
      "name": "TraitGym code",
      "url": "https://github.com/songlab-cal/TraitGym",
      "use": "Inspected scoring workflow and confirmed class_weight=balanced."
    },
    {
      "name": "TraitGym Hugging Face dataset",
      "url": "https://huggingface.co/datasets/songlab/TraitGym",
      "files_touched": [
        "complex_traits_matched_9/test.parquet",
        "complex_traits_matched_9/preds/all/CADD+GPN-MSA+Borzoi.LogisticRegression.chrom.parquet",
        "complex_traits_matched_9/features/{CADD,GPN-MSA_absLLR,Borzoi_L2_L2}.parquet",
        "mendelian_traits_matched_9/test.parquet",
        "mendelian_traits_matched_9/preds/all/CADD+GPN-MSA+Borzoi.LogisticRegression.chrom.parquet",
        "mendelian_traits_matched_9/features/{CADD,GPN-MSA_absLLR,Borzoi_L2_L2}.parquet",
        "complex_traits_v22_matched_9/test.parquet",
        "complex_traits_all/{test.parquet,features/CADDv1.7.RawScore.parquet,features/GPN-MSA_absLLR.parquet}",
        "mendelian_traits_all/{test.parquet,features/CADDv1.7.RawScore.parquet,features/GPN-MSA_absLLR.parquet}"
      ],
      "dedup": {
        "matched_rows_before_cross_dataset_dedup": 14780,
        "shared_exact_tuples": 27,
        "unique_exact_test_tuples_used": 14753
      },
      "labels_used": false
    },
    {
      "name": "Ensembl VEP REST",
      "url": "https://rest.ensembl.org/vep/homo_sapiens/region",
      "annotations": [
        "CADD",
        "AlphaMissense",
        "SpliceAI",
        "SIFT",
        "PolyPhen",
        "consequence"
      ],
      "input_rows": 44706,
      "unique_input_tuples": 44706,
      "output_rows": 44706,
      "cadd_coverage": 44706,
      "alphamissense_coverage": 7308,
      "spliceai_coverage": 34508,
      "clinical_significance_fields_parsed": false
    },
    {
      "name": "CADD v1.7 GRCh38 prescored SNVs",
      "urls": [
        "https://kircherlab.bihealth.org/download/CADD/v1.7/GRCh38/whole_genome_SNVs.tsv.gz",
        "https://kircherlab.bihealth.org/download/CADD/v1.7/GRCh38/whole_genome_SNVs.tsv.gz.tbi",
        "https://krishna.gs.washington.edu/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.tbi"
      ],
      "use": "Remote-tabix feasibility pilot; final CADD values came through Ensembl VEP."
    },
    {
      "name": "MyVariant.info",
      "urls": [
        "https://myvariant.info/v1/variant",
        "https://myvariant.info/v1/query"
      ],
      "use": "Coverage pilot only; rejected and unused in final scoring."
    },
    {
      "name": "EVEE ClinVar prediction resource",
      "urls": [
        "https://zenodo.org/records/19701997/files/manifest.json",
        "https://zenodo.org/records/19701997/files/clean_shard_0.parquet",
        "https://zenodo.org/records/19701997/files/clean_shard_1.parquet",
        "https://zenodo.org/records/19701997/files/clean_shard_2.parquet",
        "https://zenodo.org/records/19701997/files/clean_shard_3.parquet"
      ],
      "rows_annotation_columns_touched": 3372596,
      "exact_test_matches": 1604,
      "labels_read": false,
      "used_in_final": false,
      "reason_rejected": "Conservative avoidance of possible overlap with the supervised ClinVar probe."
    },
    {
      "name": "EVEE variant viewer code",
      "url": "https://github.com/goodfire-ai/variant-viewer",
      "use": "Repository access attempted during schema discovery; no final dependency."
    },
    {
      "name": "AlphaMissense release documentation",
      "urls": [
        "https://storage.googleapis.com/dm_alphamissense/README.pdf",
        "https://zenodo.org/records/8360242"
      ],
      "use": "Documentation consulted; scores were returned by Ensembl VEP."
    }
  ],
  "iterations": [
    {
      "iteration": 1,
      "method": "Input normalization/uniqueness audit",
      "status": "completed",
      "coverage": 1,
      "validation_score": null
    },
    {
      "iteration": 2,
      "method": "MyVariant CADD/dbNSFP lookup",
      "status": "rejected_low_coverage",
      "coverage": null,
      "validation_score": null
    },
    {
      "iteration": 3,
      "method": "Sparse remote-tabix CADD v1.7",
      "status": "rejected_high_latency",
      "coverage": null,
      "validation_score": null
    },
    {
      "iteration": 4,
      "method": "TraitGym released chromosome-held-out ensemble exact join",
      "status": "used",
      "rows": 14753,
      "validation_score": null,
      "published_reference_auprc": {
        "complex_traits": 0.362,
        "mendelian_traits": 0.648
      }
    },
    {
      "iteration": 5,
      "method": "TraitGym all-candidate CADD/GPN sparse extraction",
      "status": "intermediate_fallback_only",
      "unique_test_matches": 16981,
      "validation_score": null
    },
    {
      "iteration": 6,
      "method": "EVEE sparse prediction lookup",
      "status": "rejected_leakage_conservatism",
      "exact_test_matches": 1604,
      "validation_score": null
    },
    {
      "iteration": 7,
      "method": "Final TraitGym ensemble plus VEP consequence-aware clinical ensemble",
      "status": "final",
      "coverage": 1,
      "validation_score": null
    }
  ]
}

Agent trace

Compact trace, 53 KB. Loaded on demand.