fix(renal): account for passive tubular reabsorption

data/validation/renal_reabsorption_2026-09-22.json

@@ -0,0 +1,174 @@
+{
+  "date": "2026-09-22",
+  "change": "renal CL: filtration-only -> filtration + passive tubular reabsorption",
+  "model": "CL_renal = fup*UF*(GFR+PS)/(UF+PS), PS = peff*TSA",
+  "source": {
+    "ref": "Scotcher et al. 2016, Eur J Pharm Sci 94:59-71",
+    "doi": "10.1016/j.ejps.2016.03.018",
+    "note": "one-region reduction of the published 5-compartment model; verified algebraically identical to fup*GFR*(1-F_reabs), F_reabs=F'(1-UF/GFR), F'=PS/(PS+UF)"
+  },
+  "constants": {
+    "GFR_L_PER_H": 7.5,
+    "URINE_FLOW_L_PER_H": 0.06,
+    "TUBULAR_SURFACE_AREA_CM2": 57.8,
+    "tsa_provenance": "back-calculated from the paper's F'=0.5 anchor (Caco-2 Papp 14.8e-6 cm/s) through adme._CACO2_TO_INVIVO_OFFSET; effective lumped area, NOT fitted to Cmax loss (Invariant #8)"
+  },
+  "renal_cl_vs_clinical": {
+    "caffeine": {
+      "fup": 0.7009,
+      "peff": 9.105,
+      "clinical_cl_renal_l_h": 0.1,
+      "old_l_h": 5.257,
+      "new_l_h": 1.296,
+      "old_fold": 52.57,
+      "new_fold": 12.96
+    },
+    "acetaminophen": {
+      "fup": 0.7585,
+      "peff": 5.303,
+      "clinical_cl_renal_l_h": 0.6,
+      "old_l_h": 5.688,
+      "new_l_h": 2.033,
+      "old_fold": 9.48,
+      "new_fold": 3.39
+    },
+    "antipyrine": {
+      "fup": 0.5113,
+      "peff": 11.448,
+      "clinical_cl_renal_l_h": 0.15,
+      "old_l_h": 3.835,
+      "new_l_h": 0.796,
+      "old_fold": 25.56,
+      "new_fold": 5.31
+    },
+    "theophylline": {
+      "fup": 0.5796,
+      "peff": 6.569,
+      "clinical_cl_renal_l_h": 0.4,
+      "old_l_h": 4.347,
+      "new_l_h": 1.35,
+      "old_fold": 10.87,
+      "new_fold": 3.38
+    },
+    "midazolam": {
+      "fup": 0.0318,
+      "peff": 7.563,
+      "clinical_cl_renal_l_h": 0.02,
+      "old_l_h": 0.239,
+      "new_l_h": 0.067,
+      "old_fold": 11.94,
+      "new_fold": 3.36
+    },
+    "atenolol": {
+      "fup": 0.3394,
+      "peff": 0.812,
+      "clinical_cl_renal_l_h": 8.0,
+      "old_l_h": 2.546,
+      "new_l_h": 1.991,
+      "old_fold": 0.32,
+      "new_fold": 0.25
+    },
+    "gabapentin": {
+      "fup": 0.946,
+      "peff": 4.819,
+      "clinical_cl_renal_l_h": 6.0,
+      "old_l_h": 7.095,
+      "new_l_h": 2.691,
+      "old_fold": 1.18,
+      "new_fold": 0.45
+    },
+    "metformin": {
+      "fup": 0.8573,
+      "peff": 0.229,
+      "clinical_cl_renal_l_h": 31.0,
+      "old_l_h": 6.43,
+      "new_l_h": 5.961,
+      "old_fold": 0.21,
+      "new_fold": 0.19
+    },
+    "lisinopril": {
+      "fup": 0.1461,
+      "peff": 0.605,
+      "clinical_cl_renal_l_h": 6.0,
+      "old_l_h": 1.096,
+      "new_l_h": 0.907,
+      "old_fold": 0.18,
+      "new_fold": 0.15
+    }
+  },
+  "holdout_107_public_clone": {
+    "note": "local stack reproduces the committed cache bit-identically",
+    "baseline": {
+      "meta": {
+        "aafe": 2.7428,
+        "pct_2fold": 44.9,
+        "pct_3fold": 63.6
+      },
+      "engine": {
+        "aafe": 4.2779,
+        "pct_2fold": 27.1,
+        "pct_3fold": 43.9
+      },
+      "ml": {
+        "aafe": 2.9981,
+        "pct_2fold": 43.0,
+        "pct_3fold": 58.9
+      }
+    },
+    "renal": {
+      "meta": {
+        "aafe": 2.7416,
+        "pct_2fold": 44.9,
+        "pct_3fold": 63.6
+      },
+      "engine": {
+        "aafe": 4.2548,
+        "pct_2fold": 27.1,
+        "pct_3fold": 43.0
+      },
+      "ml": {
+        "aafe": 2.9981,
+        "pct_2fold": 43.0,
+        "pct_3fold": 58.9
+      }
+    },
+    "delta_meta": -0.0012,
+    "delta_engine": -0.0232,
+    "in_domain_n": 81,
+    "in_domain_meta": [
+      2.7906,
+      2.7894
+    ],
+    "n_engine_cmax_moved": 37,
+    "max_move": "<=4%"
+  },
+  "pkdb_curve_fold_error": {
+    "note": "open-licence PK-DB timecourses, old vs new renal on the same canonical engine",
+    "caffeine_oral": [
+      28.04,
+      24.91
+    ],
+    "acetaminophen_iv": [
+      32.67,
+      29.2
+    ],
+    "acetaminophen_oral": [
+      22.45,
+      21.36
+    ],
+    "midazolam_iv": [
+      1.96,
+      1.95
+    ],
+    "midazolam_oral": [
+      2.01,
+      2.01
+    ]
+  },
+  "not_modelled": [
+    "active tubular secretion",
+    "pH/ionisation-dependent reabsorption",
+    "peff uncertainty in the returned cv"
+  ],
+  "cache_regen": "required on the canonical Linux stack; delta is within the +/-0.020 pin so the existing cache still passes"
+}

docs/research/experiment-log.md

@@ -1,5 +1,5 @@
 ---
-last_updated: 2026-07-03
+last_updated: 2026-09-22
 parent: ../../README.md
 charter: Chronological log of Sisyphus experiments (successes, negatives, infrastructure). Latest first.
 ---
@@ -10,6 +10,44 @@ Reverse-chronological. The project README carries only the **current** headline
 
 ---
 
+## 2026-09-22 — Passive tubular reabsorption: renal CL is no longer filtration-only (correctness; headline-neutral)
+
+`_estimate_renal_clearance` was `CL_renal = GFR·fup` for every drug — the zero-permeability
+limit, i.e. everything filtered is excreted and nothing is reabsorbed. That over-predicts renal
+CL by ~an order of magnitude for permeable, high-fup drugs, which are filtered freely and then
+almost entirely reabsorbed. Replaced with the one-compartment (well-stirred) tubule balance
+
+    CL_renal = fup · UF · (GFR + PS) / (UF + PS),    PS = peff · TSA
+
+reusing the already-predicted `peff`; `UF` = 1 mL/min urine flow. This is algebraically identical
+(verified to machine precision) to the Scotcher et al. 2016 form `fup·GFR·(1−F_reabs)` with
+`F_reabs = F'(1−UF/GFR)`, `F' = PS/(PS+UF)`, reduced to a single tubular region
+(Eur J Pharm Sci 94:59-71, doi:10.1016/j.ejps.2016.03.018). Limits are exact: `PS→0` recovers
+the old `fup·GFR`, `PS→∞` gives `fup·UF`. `TSA = 57.8 cm²` is back-calculated from that paper's
+own model calibration point (F′=0.5 at Caco-2 Papp 14.8e-6 cm/s) through the existing
+`_CACO2_TO_INVIVO_OFFSET`; it is an *effective* lumped area (a well-stirred tubule over-reabsorbs
+vs the 5-region plug-flow structure, and the lumping difference lands here). **Anchored to the
+published renal model only — never fitted to Cmax loss (Invariant #8).**
+
+- **Renal CL vs clinical** (fold error, old → new): caffeine 52.6→13.0, antipyrine 25.6→5.3,
+  theophylline 10.9→3.4, paracetamol 9.5→3.4, midazolam 11.9→3.4. Secretion-dominant drugs
+  (metformin, atenolol, lisinopril) are unchanged-and-still-under-predicted — active secretion
+  is explicitly out of scope and now documented as such.
+- **Headline: neutral.** 107-holdout, public-clone, local stack (reproduces the committed cache
+  bit-identically): Meta 2.7428 → **2.7416** (−0.0012), Engine 4.2779 → 4.2548 (−0.0232), ML
+  unchanged. Meta %2-fold and %3-fold are unchanged; Engine %2-fold is unchanged and %3-fold
+  moves 43.9% → 43.0%. 37/107 engine Cmax values move, all ≤4%. Consistent
+  with DE-43 (the fixed-weight meta damps engine moves) — renal is a minor elimination route for
+  most holdout drugs. Ships on correctness, not on the number.
+- **PK-DB curve check** (open-licence timecourses, old vs new on the same canonical engine):
+  curve fold error improves caffeine 28.0→24.9, paracetamol IV 32.7→29.2, oral 22.5→21.4;
+  midazolam unchanged (1.96→1.95). Directionally right but small — the dominant residual on
+  those two curves is the CLint assay-floor artifact, not renal (DE-58).
+
+The committed `4track_holdout_predictions.json` is **not** regenerated here (macOS is not the
+canonical numerics stack); the Δ is within the ±0.020 pin so `test_cached_holdout_aafe_is_2p743`
+still passes against the existing cache. A canonical Linux regen is the follow-up.
+
 ## 2026-07-07 (cont.) — N50' clean re-curation: infeasible from repo data (pool=0), deferred to human-led curation
 
 Scoped the clean N50' secondary-holdout curation (the unbiased-generalization instrument the

src/sisyphus/engine/flux.py

@@ -315,7 +315,7 @@ def apply(
             renal_cl = params.drug_param("renal_clearance")
             if renal_cl <= 0:
                 return
-            # RBP-2: renal_cl = GFR·fup is plasma-basis → filter PLASMA A/(V·Kp),
+            # RBP-2: renal_cl (fup-scaled, ivive.py) is plasma-basis → filter PLASMA A/(V·Kp),
             # not blood. The convective edge supplies flow limitation.
             v = params.node_param(self.source_name, "volume")
             kp = params.drug_kp(self.source_name)

src/sisyphus/mipd/covariates.py

@@ -1,10 +1,14 @@
 """Patient covariates that deterministically individualize the engine prior.
 
 v1: renal function only — a measured creatinine clearance (CrCl). CrCl scales
-the drug's renal (glomerular-filtration) clearance: the engine's reference renal
-model is ``CL_renal = GFR*fup`` with GFR = 7.5 L/h (~125 mL/min), so an
-individual's renal CL is scaled by ``CrCl / 125``. Weight/age covariates (via
-sbi.physiology_generator) are a documented future extension — see the design spec
+the drug's renal clearance by ``CrCl / 125`` (reference GFR = 7.5 L/h ~= 125
+mL/min). The engine's reference renal model is filtration with passive tubular
+reabsorption, ``CL_renal = fup*UF*(GFR+PS)/(UF+PS)`` (predict/ivive.py). The
+linear CrCl scaling is exact in the filtration-dominated limit (small PS) and
+over-scales highly reabsorbed drugs, whose CL_renal (-> fup*UF) is nearly
+GFR-independent — but their renal CL is then negligible in absolute terms.
+Weight/age covariates (via sbi.physiology_generator) are a documented future
+extension — see the design spec
 docs/_internal/specs/2026-06-11-mipd-crcl-renal-individualization-design.md.
 """
 from __future__ import annotations
@@ -124,7 +128,8 @@ def warnings(self) -> tuple[str, ...]:
         if self.crcl_ml_min is not None and not (5.0 <= self.crcl_ml_min <= 200.0):
             w.append(
                 f"crcl:extreme:{self.crcl_ml_min}: the engine renal model is "
-                "glomerular-filtration-only and least reliable outside [5, 200] mL/min"
+                "filtration + passive reabsorption only (no active secretion) and least "
+                "reliable outside [5, 200] mL/min"
             )
         if self.body_weight_kg is not None and not (2.0 <= self.body_weight_kg <= 250.0):
             w.append(

src/sisyphus/predict/ivive.py

@@ -42,6 +42,26 @@
 # GFR: glomerular filtration rate (L/h)
 _GFR_L_PER_H = 7.5  # ~125 mL/min = 7.5 L/h
 
+# Urine flow at the end of the nephron (L/h). Scotcher et al. 2016,
+# Eur J Pharm Sci 94:59-71 (doi:10.1016/j.ejps.2016.03.018) assume 1 mL/min.
+_URINE_FLOW_L_PER_H = 0.06  # 1 mL/min
+
+# Effective lumped tubular surface area (cm^2) for scaling peff -> PS.
+# Back-calculated from the Scotcher 2016 mechanistic-model calibration point:
+# their 5-compartment model reaches F' = 0.5 (half of urine/plasma equilibrium)
+# at Caco-2 Papp = 14.8e-6 cm/s under a pH 6.5/7.4 gradient. In the
+# one-compartment reduction used below F' = PS/(PS + UF), so F' = 0.5 means
+# PS = UF = 0.06 L/h. Sisyphus `peff` is in-vivo-calibrated Caco-2
+# (peff = 10**1.29 * Papp ~= 19.5x, adme._CACO2_TO_INVIVO_OFFSET), putting the
+# half-reabsorption point at peff = 2.886e-4 cm/s, hence
+#   TSA = 0.06 L/h * 1000 cm3/L / (2.886e-4 cm/s * 3600 s/h) = 57.8 cm^2.
+# This is an EFFECTIVE area, well below the geometric proximal-tubule area: a
+# well-stirred tubule sees fully concentrated filtrate everywhere and so
+# over-reabsorbs relative to Scotcher's 5-region plug-flow-with-water-removal
+# structure, and the lumping difference is absorbed here. Anchored to the
+# published renal model ONLY — never fitted to Sisyphus Cmax loss (Invariant #8).
+_TUBULAR_SURFACE_AREA_CM2 = 57.8
+
 # ---------------------------------------------------------------------------
 # Enzyme abundances (pmol, total organ) — from reference_man.yaml
 # ---------------------------------------------------------------------------
@@ -539,23 +559,46 @@ def _compute_all_kp(
 # ---------------------------------------------------------------------------
 
 
-def _estimate_renal_clearance(fup: Distribution, profile: MolecularProfile) -> Distribution:
-    """Estimate renal clearance from GFR and fraction unbound.
+def _estimate_renal_clearance(fup: Distribution, peff: Distribution) -> Distribution:
+    """Estimate renal clearance from filtration and passive tubular reabsorption.
+
+    One-compartment (well-stirred) tubule mass balance: drug enters by
+    glomerular filtration (``fup * GFR``), leaves in urine (``UF * C_tubule``),
+    and exchanges passively across the tubular wall in both directions
+    (``PS * (C_tubule - fup * C_plasma)``). At steady state this gives
+
+        CL_renal = fup * UF * (GFR + PS) / (UF + PS)
 
-    For drugs undergoing glomerular filtration without active secretion:
-        CL_renal = GFR * fup
+    algebraically identical to the Scotcher et al. 2016 formulation
+    ``CL = fup * GFR * (1 - F_reabs)``, ``F_reabs = F' * (1 - UF/GFR)``,
+    ``F' = PS / (PS + UF)``, reduced to a single tubular region.
 
-    Active secretion/reabsorption would require transporter data.
+    Limits: ``PS -> 0`` recovers ``fup * GFR`` (impermeable; filtration only),
+    ``PS -> inf`` gives ``fup * UF`` (equilibrates with plasma, so the drug
+    leaves only as fast as water does). The previous model was the ``PS = 0``
+    limit applied to *every* drug, which over-predicted renal clearance for
+    permeable, high-fup compounds (e.g. caffeine, paracetamol) by ~an order of
+    magnitude — they are filtered freely but almost entirely reabsorbed.
+
+    NOT modelled: active secretion (so secretion-dominant drugs are still
+    under-predicted), pH/ionisation-dependent reabsorption, and peff
+    uncertainty — the returned cv tracks fup only.
 
     Args:
         fup: Fraction unbound in plasma.
-        profile: Molecular profile (reserved for future use, e.g. charge-based
-            reabsorption estimates).
+        peff: Effective permeability (x10^-4 cm/s).
 
     Returns:
         Renal clearance (L/h) as Distribution.
     """
-    cl_renal = _GFR_L_PER_H * fup.mean
+    # peff (x10^-4 cm/s) * TSA (cm^2) -> L/h
+    ps_l_per_h = peff.mean * 1e-4 * 3600.0 * _TUBULAR_SURFACE_AREA_CM2 / 1000.0
+    cl_renal = (
+        fup.mean
+        * _URINE_FLOW_L_PER_H
+        * (_GFR_L_PER_H + ps_l_per_h)
+        / (_URINE_FLOW_L_PER_H + ps_l_per_h)
+    )
     return Distribution(mean=cl_renal, cv=fup.cv)
 
 
@@ -748,7 +791,7 @@ def build_drug_on_graph(
     particle_radius = _estimate_particle_radius(adme.solubility)
 
     # Estimate renal clearance
-    renal_cl = _estimate_renal_clearance(adme.fup, profile)
+    renal_cl = _estimate_renal_clearance(adme.fup, adme.peff)
 
     # Set administration node based on route
     if route == "oral":