{
  "schema_version": 1,
  "research_line": "rna_drug_response_controls",
  "status": "exploratory development checks on previously viewed cells; independent evaluation not completed",
  "freeze": {
    "commit": "ce5a4fb",
    "frozen_utc": "2026-10-06T11:38:50Z",
    "before_scoring": true,
    "result_date": "2026-10-06"
  },
  "counts": {
    "development_cells": 44,
    "overlapping_splits": 20,
    "locked_cells_evaluated": 0
  },
  "measurement": {
    "metric": "MAE and RMSE in pGI50; lower is better",
    "features": "CellMiner RNA, log2(FPKM + 1)",
    "comparison": "The same previously viewed development cells and overlapping splits are reused."
  },
  "summary": {
    "models": {
      "drug_mean": {
        "MAE": {
          "mean": 0.6079409698275863,
          "median": 0.6029837592364531,
          "min": 0.49925697865353047,
          "max": 0.6800223214285714
        },
        "RMSE": {
          "mean": 0.7979660442002847,
          "median": 0.8076867020516963,
          "min": 0.6574408665693837,
          "max": 0.871645145194509
        }
      },
      "tissue_fixed": {
        "MAE": {
          "mean": 0.5946546561906478,
          "median": 0.6009306718509309,
          "min": 0.44086162620220515,
          "max": 0.6662877813625451
        },
        "RMSE": {
          "mean": 0.7839902962858876,
          "median": 0.8026572914275383,
          "min": 0.6009375367736373,
          "max": 0.8608311076184902
        }
      },
      "tissue_cv": {
        "MAE": {
          "mean": 0.6101590908930449,
          "median": 0.6192447925211167,
          "min": 0.47951247085001536,
          "max": 0.6704106977513228
        },
        "RMSE": {
          "mean": 0.8005956755666523,
          "median": 0.8215966180812446,
          "min": 0.6372255192373993,
          "max": 0.8915387134439573
        }
      },
      "molecular": {
        "MAE": {
          "mean": 0.5509287944226826,
          "median": 0.5588387287520505,
          "min": 0.40197166758700253,
          "max": 0.6622713418025448
        },
        "RMSE": {
          "mean": 0.7414008580304515,
          "median": 0.7335622639259162,
          "min": 0.5479722863748202,
          "max": 0.8602295268552776
        }
      },
      "rna_permuted": {
        "MAE": {
          "mean": 0.6076619522350397,
          "median": 0.6122887892708979,
          "min": 0.5121838895374287,
          "max": 0.6872988650043131
        },
        "RMSE": {
          "mean": 0.8083265038877617,
          "median": 0.8075397670691279,
          "min": 0.6973813346969855,
          "max": 0.9350239738848161
        }
      }
    },
    "delta_MAE": {
      "molecular_vs_drug_mean": {
        "mean": -0.05701217540490358,
        "median": -0.05835274387295791,
        "min": -0.11428517418928064,
        "max": 0.016677605891928127,
        "wins_negative": 19,
        "n_splits": 20
      },
      "molecular_vs_tissue_fixed": {
        "mean": -0.043725861767965134,
        "median": -0.05190889086407485,
        "min": -0.101035457960641,
        "max": 0.033626676140052325,
        "wins_negative": 18,
        "n_splits": 20
      },
      "molecular_vs_tissue_cv": {
        "mean": -0.0592302964703622,
        "median": -0.06606382183084542,
        "min": -0.10401538465950705,
        "max": 0.026072282902291577,
        "wins_negative": 19,
        "n_splits": 20
      },
      "molecular_vs_rna_permuted": {
        "mean": -0.056733157812357084,
        "median": -0.057305030486233255,
        "min": -0.11961772162475032,
        "max": 0.03965593473032736,
        "wins_negative": 19,
        "n_splits": 20
      },
      "tissue_cv_vs_tissue_fixed": {
        "mean": 0.015504434702397069,
        "median": 0.0183374478917947,
        "min": -0.013537131150802573,
        "max": 0.04550918920382452,
        "wins_negative": 2,
        "n_splits": 20
      },
      "rna_permuted_vs_drug_mean": {
        "mean": -0.00027901759254649885,
        "median": -0.00301281751499477,
        "min": -0.05740691435635403,
        "max": 0.06559413700928873,
        "wins_negative": 12,
        "n_splits": 20
      }
    },
    "delta_sign": "first named model MAE minus second; negative improves first"
  },
  "method": {
    "tissue_fixed": "shrinkage fixed at 10",
    "tissue_cv": "training-only cross-validation selects among four candidates",
    "molecular": "training-only cross-validation selects among six PCA/Ridge candidates",
    "rna_permuted": "One fixed whole-RNA-row shuffle within each training, validation, calibration and evaluation partition; no crossing partitions.",
    "control_limit": "The shuffle also breaks tissue information; it is not a causal or isolated incremental RNA-effect estimate."
  },
  "timings": {
    "elapsed_seconds": 17.236675249878317,
    "includes": "cached input parsing and 20-split evaluation",
    "excludes": "download, plotting and GitHub publication",
    "threads": 1
  },
  "verification": {
    "tests_passed": 21,
    "tests_skipped": 0,
    "scope": "local execution, not CI",
    "new_response_or_rna_downloads": 0
  },
  "sources": [
    {
      "title": "CellMiner 2.15 RNA-seq composite expression",
      "url": "https://discover.nci.nih.gov/cellminer/download/processeddataset/nci60_RNA__RNA_seq_composite_expression.zip",
      "terms_url": "https://discover.nci.nih.gov/disclaimer.jsp",
      "use": "original matrix is not redistributed here"
    }
  ],
  "limitations": [
    "previously viewed data and overlapping development splits; no independent final test",
    "split wins are not independent trials; no significance test or p-value",
    "one fixed negative control is diagnostic, not a permutation null distribution",
    "RNA shuffle also breaks tissue information; no causal or incremental RNA effect claim",
    "four tissue and six molecular candidates do not equalize model class complexity",
    "imputation, variance selection, scaling and PCA fit only each training subset",
    "inner RNA permutations stay inside inner train/validation separately",
    "known lineage representatives only; unknown related lineages may remain",
    "GI50 censoring and aggregate measurement conditions unresolved; no clinical efficacy claim",
    "source FPKM preprocessing not independently reproduced; interval coverage descriptive",
    "60-second cap is soft and does not interrupt parsing or an active linear algebra call"
  ]
}