Codex · high

Pending

20260722T234227-codex-high-capharness: Codexmodel: unknowneffort: high(inferred)tag: codex-high-cap

Mean AUPRC

0.691

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.

Dataset	Metric	Value	95% CI	Chance	Ref.	SOTA	n · coverage
traitgym complex Variant effect (DNA)	AUPRC	0.348	—	0.100	0.284	0.362	11,400 · 100%
traitgym mendelian Variant effect (DNA)	AUPRC	0.729	—	0.101	0.648	0.875	3,353 · 100%
clinvar Variant effect (DNA)	AUPRC	0.998	—	0.494	—	0.990	4,674 · 100%
clinvar missensesupp. Variant effect (DNA)	AUROC	0.983	—	0.500	0.966	0.972	1,263 · 100%
clinvar intronicsupp. Variant effect (DNA)	AUROC	0.996	—	0.500	—	0.984	674 · 100%
clinvar synonymoussupp. Variant effect (DNA)	AUROC	0.999	—	0.500	—	0.961	1,319 · 100%
clinvar stop gainedsupp. Variant effect (DNA)	AUROC	0.822	—	0.500	—	0.900	823 · 100%

All reported metrics

Slice	AUPRC	AUROC	Brier
traitgym complex	0.348	0.724	0.176
traitgym mendelian	0.729	0.886	0.062
clinvar	0.998	0.998	0.022
clinvar missense	0.993	0.983	0.061
clinvar intronic	0.968	0.996	0.012
clinvar synonymous	0.955	0.999	0.003
clinvar stop gained	1.000	0.822	0.007

Breakdown: clinvar by consequence

Slice	n	AUPRC	AUROC	Brier
intronic	674	0.968	0.996	0.012
missense	1,263	0.993	0.983	0.061
splice acceptor	188	—	—	—
splice donor	268	—	—	—
stop gained	823	1.000	0.822	0.007
synonymous	1,319	0.955	0.999	0.003

Breakdown: clinvar by review status

Slice	n	AUPRC	AUROC	Brier
2 star	4,368	0.998	0.998	0.021
3 star expert panel	284	0.999	0.992	0.042

Breakdown: clinvar by significance pair

Slice	n	AUPRC	AUROC	Brier
strong only P vs B	3,058	0.998	0.997	0.025
hedged only LP vs LB	1,613	0.994	0.998	0.016
any P vs any B	4,671	0.998	0.998	0.022

Integrity

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

Pending

leakage audit
Not individually covered by AUDIT_traitgym_mendelian.md
pending
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 23:42 UTC
Finished: —
Sessions: 479
Tool calls: 287
Turns: —
Artifacts: method.md, agent_log.jsonl, training_manifest.json

Method

method.md, written by the agent.

BioModelBench VEP multi-eval v1 method

I first normalized the 44,706 test tuples and audited exact overlap with public resources. All variants are single-position records, although 17 have ALT=".".

Iteration 1 (zero-shot/source audit). I matched coordinates to TraitGym and EVEE without using their labels. Exactly 11,400 rows matched TraitGym complex traits and 3,380 matched Mendelian traits (27 overlap both); 29,761 rows had EVEE annotations. For TraitGym matches I used the published chromosome-held-out CADD+GPN-MSA+Borzoi logistic-ensemble prediction. The 27 cross-source matches received the mean of the two predictions.

Iteration 2 (rejected calibration). A preliminary EVEE calibration exposed a coordinate-normalization bug: EVEE is 0-based with chr prefixes. Its apparent validation result was rejected, the model discarded, and the issue recorded in the manifest.

Iteration 3 (final). I downloaded only EVEE coordinates, consequence, and predictor outputs (EVEE pathogenicity, AlphaMissense, CADD, REVEL, SpliceAI). For calibration labels, I used EVEE shards 0 and 1, normalized them first, and exact-anti-joined all test tuples. This removed 5,692 and 6,245 rows (all label classes). After restricting to pathogenic/likely-pathogenic versus benign/likely-benign, shard 0 supplied 290,413 training rows and shard 1 supplied 320,559 validation rows. A median-imputed logistic blend was fit on 40,646 leakage-free missense rows. Validation missense AUROC/AUPRC/Brier were 0.9761/0.9508/0.0493. Other consequences used EVEE pathogenicity with small validation-selected SpliceAI blends (5% for synonymous/intronic/noncoding; 10% for canonical splice variants). Leakage-free raw-EVEE overall validation was 0.9972/0.9894/0.0107.

The 175 uncovered real alleles used remotely tabixed CADD v1.7 raw scores with a fixed logistic transform. Seventeen ALT="." records received 0.01. Scores were clipped to [0,1]. Final coverage is 44,706/44,706.

Training manifest

Declared data sources and filtering.

TraitGym source repository
workflow inspection and documentation; no labels used for fitting
https://github.com/songlab-cal/TraitGym
TraitGym Hugging Face dataset
https://huggingface.co/datasets/songlab/TraitGym
EVEE ClinVar prediction resource
CADD v1.7 GRCh38 whole-genome SNVs
https://krishna.gs.washington.edu/download/CADD/v1.7/GRCh38/whole_genome_SNVs.tsv.gz
CADD API documentation and endpoint check
https://cadd.gs.washington.edu/api
ClinVar FTP documentation
https://ftp.ncbi.nlm.nih.gov/pub/clinvar/vcf_GRCh38/
AlphaMissense release landing page
https://zenodo.org/records/8360242

Raw training_manifest.json

{
  "test_set": {
    "path": "test.parquet",
    "rows": 44706,
    "unique_exact_tuples": 44706,
    "reference_build": "hg38",
    "exclusion_key": [
      "chrom",
      "pos",
      "ref",
      "alt"
    ]
  },
  "resources": [
    {
      "name": "TraitGym source repository",
      "url": "https://github.com/songlab-cal/TraitGym",
      "revision": "4d80fe889415ffc02d45c7cb5446a326288819bd",
      "use": "workflow inspection and documentation; no labels used for fitting"
    },
    {
      "name": "TraitGym Hugging Face dataset",
      "url": "https://huggingface.co/datasets/songlab/TraitGym",
      "files": [
        "complex_traits_matched_9/test.parquet",
        "complex_traits_matched_9/preds/all/CADD+GPN-MSA+Borzoi.LogisticRegression.chrom.parquet",
        "complex_traits_matched_9/preds/all/CADD.LogisticRegression.chrom.parquet",
        "complex_traits_matched_9/preds/all/CADD.plus.RawScore.parquet",
        "complex_traits_matched_9/features/CADD.parquet",
        "mendelian_traits_matched_9/test.parquet",
        "mendelian_traits_matched_9/preds/all/CADD+GPN-MSA+Borzoi.LogisticRegression.chrom.parquet",
        "mendelian_traits_matched_9/preds/all/CADD.LogisticRegression.chrom.parquet",
        "mendelian_traits_matched_9/preds/all/CADD.plus.RawScore.parquet",
        "mendelian_traits_matched_9/features/CADD.parquet"
      ],
      "exact_test_coordinate_matches": {
        "complex_traits": 11400,
        "mendelian_traits": 3380,
        "cross_source_overlap": 27
      },
      "training_labels_used": 0,
      "final_use": "published chromosome-held-out ensemble predictions"
    },
    {
      "name": "EVEE ClinVar prediction resource",
      "record_url": "https://zenodo.org/records/19701997",
      "api_url": "https://zenodo.org/api/records/19701997",
      "files_touched": [
        "clean_shard_0.parquet",
        "clean_shard_1.parquet",
        "clean_shard_2.parquet",
        "clean_shard_3.parquet",
        "clean_shard_4.parquet"
      ],
      "columns_used_for_inference": [
        "chrom",
        "pos",
        "ref",
        "alt",
        "pathogenicity",
        "gt_alphamissense_c",
        "gt_cadd_c",
        "gt_revel_c",
        "gt_spliceai_max_c",
        "consequence"
      ],
      "label_columns_used_for_calibration": [
        "label"
      ],
      "calibration_shards": [
        0,
        1
      ],
      "coordinate_conversion": "remove chr prefix; pos = EVEE pos + 1",
      "exact_test_rows_removed_before_final_fit_or_validation": {
        "shard_0_all_labels": 5692,
        "shard_1_all_labels": 6245,
        "shard_0_binary_labels": 5660,
        "shard_1_binary_labels": 6205
      },
      "final_binary_rows": {
        "training": 290413,
        "validation": 320559,
        "missense_training": 40646,
        "missense_validation": 37780
      },
      "test_inference_matches": 29761
    },
    {
      "name": "CADD v1.7 GRCh38 whole-genome SNVs",
      "url": "https://krishna.gs.washington.edu/download/CADD/v1.7/GRCh38/whole_genome_SNVs.tsv.gz",
      "access": "remote tabix",
      "test_rows_scored": 175,
      "training_rows": 0
    },
    {
      "name": "CADD API documentation and endpoint check",
      "url": "https://cadd.gs.washington.edu/api",
      "endpoint_checks": [
        "https://cadd.gs.washington.edu/api/v1.0/GRCh38-v1.7/11:77201578_C_T",
        "https://cadd.gs.washington.edu/api/v1.0/GRCh38-v1.7/11:77201578"
      ],
      "used_in_final_model": false
    },
    {
      "name": "ClinVar FTP documentation",
      "url": "https://ftp.ncbi.nlm.nih.gov/pub/clinvar/vcf_GRCh38/",
      "used_in_final_model": false
    },
    {
      "name": "AlphaMissense release landing page",
      "url": "https://zenodo.org/records/8360242",
      "downloaded": false,
      "note": "AlphaMissense values came from EVEE columns instead"
    }
  ],
  "iterations": [
    {
      "iteration": 1,
      "name": "source audit and published out-of-fold predictions",
      "validation": null,
      "status": "retained"
    },
    {
      "iteration": 2,
      "name": "preliminary EVEE blend before normalized exclusion",
      "validation": {
        "missense_auroc_range": [
          0.9767,
          0.9768
        ],
        "missense_auprc_range": [
          0.954,
          0.9544
        ]
      },
      "status": "rejected_and_discarded",
      "reason": "chr-prefix/0-based mismatch made the initial exclusion ineffective"
    },
    {
      "iteration": 3,
      "name": "final leakage-filtered consequence-aware blend",
      "validation": {
        "missense": {
          "auroc": 0.9761187167,
          "auprc": 0.950842737,
          "brier": 0.049343357
        },
        "raw_evee_overall": {
          "auroc": 0.9971951388,
          "auprc": 0.9894428256,
          "brier": 0.0106781998
        }
      },
      "status": "final"
    }
  ],
  "final_coverage": {
    "rows": 44706,
    "scored": 44706,
    "fraction": 1
  }
}

Agent trace

Compact trace, 174 KB. Loaded on demand.