{
  "tier_rows": 213,
  "metadata_sha256": "dc15911dee99278033ecd3f1cadd67ba34490d8040a30f30c3b41e0b4f38e9ba",
  "model": "equal-weight ensemble",
  "members": [
    "ridge",
    "svr",
    "xgboost"
  ],
  "seed": 0,
  "block_order": [
    "sequence",
    "smiles",
    "molecule_annotation",
    "study_context",
    "termini",
    "physchem"
  ],
  "block_slices": {
    "sequence": [
      0,
      1280
    ],
    "smiles": [
      1280,
      2048
    ],
    "molecule_annotation": [
      2048,
      2057
    ],
    "study_context": [
      2057,
      2090
    ],
    "termini": [
      2090,
      2103
    ],
    "physchem": [
      2103,
      2119
    ]
  },
  "training_composition": {
    "species": {
      "mouse": 34,
      "rat": 179
    },
    "route": {
      "iv": 213
    }
  },
  "application_caveats": [
    "Route of administration is a model input. A prediction is route-specific and must not be read as a prediction for another route: IV and SC terminal half-lives do not share an absolute scale, because an SC terminal slope can report absorption rather than elimination (flip-flop kinetics). Transferring across the two is offset by roughly half a log unit - see evaluate_route_flipflop.py and figures/tabs/table33.",
    "Per-route performance is very uneven; see figures/tabs/table34_muriverse_halflife_performance_by_route.csv. The pooled score is carried by IV. Routes with few training rows rank at or below chance and the pooled number does not transfer to them."
  ],
  "models": {
    "shipped": {
      "features": "annotation + context + termini + physchem",
      "blocks": [
        "molecule_annotation",
        "study_context",
        "termini",
        "physchem"
      ],
      "dims": 71,
      "cluster_pmid_spearman": 0.5736096354933952,
      "n_train": 213
    },
    "structural": {
      "features": "physchem + study context + termini",
      "blocks": [
        "physchem",
        "study_context",
        "termini"
      ],
      "dims": 62,
      "cluster_pmid_spearman": 0.5865509888466302,
      "n_train": 213
    }
  },
  "not_included": "structural_embedding (SMILES + physchem + study context) needs PeptideCLM and scored worse; it is in the GitHub repository",
  "metadata_sha256_note": "hash of MuriVerse/rodent/strict/halflife_meta_with_split.csv at the release this model was fitted on",
  "applicability": {
    "physchem_range": {
      "mw": [
        301.29900000000004,
        5732.5089999999855
      ],
      "clogp": [
        -33.85392000000005,
        2.6673700000000107
      ],
      "molmr": [
        70.22570000000003,
        1435.4526000000078
      ],
      "tpsa": [
        150.03,
        2536.940000000001
      ],
      "labute_asa": [
        120.63192558805426,
        2316.48058143763
      ],
      "hbd": [
        4.0,
        90.0
      ],
      "hba": [
        5.0,
        87.0
      ],
      "rotatable_bonds": [
        6.0,
        194.0
      ],
      "rings": [
        0.0,
        13.0
      ],
      "aromatic_rings": [
        0.0,
        12.0
      ],
      "frac_csp3": [
        0.34615384615384615,
        0.7475728155339806
      ],
      "heavy_atoms": [
        21.0,
        399.0
      ],
      "amide_bonds": [
        2.0,
        58.0
      ],
      "n_basic": [
        0.0,
        21.0
      ],
      "n_acidic": [
        0.0,
        12.0
      ],
      "net_charge_ph74": [
        -6.0,
        17.0
      ]
    },
    "note": "min and max of each physicochemical descriptor over the 213 intravenous training molecules"
  }
}