fix(pk): terminal t½ window selection; delete two dead code paths#100
Conversation
…ad code paths Three verified minor review items, all headline-neutral (Cmax/AUC unchanged). 1. pk/nca.py — terminal_half_life fit the log-linear slope over ALL post-Cmax points, blending the steep distribution (α) phase into λz and biasing t½ short for multi-compartment disposition. Now selects the terminal window by the adjusted-R² "best fit" rule (Phoenix/WinNonlin): among last-k-point windows (k≥3), maximize adjusted R², preferring more points on a tie. A mono-exponential curve still uses the whole tail (unchanged). t½ is a secondary endpoint (does not feed the headline Cmax); its one numeric consumer is MIPD steady-state dose sizing. Added a bi-exponential regression test that the old all-points fit fails (returns ln2/α instead of ln2/β). 2. predict/drugbank.py — delete dead get_ugt_enzymes (+ its _ugt_substrates store, the UGT populate block, and the now-unused _UGT_NORMALIZATION map). Zero callers; UGT fm is sourced from the JSON substrate registry (non_cyp_substrates), not DrugBank. 3. predict/chemistry.py — delete dead _load_pka_models / _predict_pka_from_model / _PKA_MODEL_DIR (the XGBoost pKa loader). Never wired into compute_profile (which uses DrugBank pKa + the SMARTS _estimate_pka_type fallback); the model was tested and reverted (worsened AAFE). Left intact (verified but out of scope): the graph merge_overlay stub + sc/tumor overlay YAMLs (intentional Phase-3 extension seam), and RBP reset (real but on the headline benchmark path → needs a canonical regen, tracked separately). Tests: test_endpoints (12) + test_drugbank + test_drugbank_integration + test_features_chemistry (61) pass; ruff clean.
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d531cb7b61
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| y_hat = slope * t_w + intercept | ||
| ss_res = float(np.sum((y_w - y_hat) ** 2)) | ||
| ss_tot = float(np.sum((y_w - y_w.mean()) ** 2)) | ||
| r2 = 1.0 if ss_tot <= 0.0 else 1.0 - ss_res / ss_tot |
There was a problem hiding this comment.
When the last three post-Cmax samples are an identical positive floor (for example an assay/numerical LLOQ tail like [10, 5, 2, 1e-9, 1e-9, 1e-9]), ss_tot <= 0.0 makes that flat window look like a perfect adjusted-R² fit; np.polyfit can still return a tiny negative roundoff slope, so the new selector picks the plateau and returns an enormous t½ (~6e14 h) instead of using the preceding declining tail or returning None. Flat/no-variance windows should be rejected (or require a meaningful negative slope) rather than scored as perfect terminal decline.
Useful? React with 👍 / 👎.
1 participant
Three verified minor review items, all headline-neutral (Cmax/AUC unchanged). Each verified before acting — two other "minor" claims (solubility mol/L-vs-mg/mL, an RBP asymmetry severity) were checked in the same pass and handled differently (see below).
1.
pk/nca.py— terminal t½ fit on a selected terminal window (real bug)terminal_half_liferegressed the log-linear slope over all post-Cmax points, blending the steep distribution (α) phase into λz → t½ biased short for multi-compartment disposition. Now selects the terminal window by the adjusted-R² "best fit" rule (Phoenix/WinNonlin): among last-k-point windows (k≥3), maximize adjusted R², preferring more points on a tie. A mono-exponential curve still uses the whole tail (behavior unchanged).t½ is a secondary endpoint (does not feed the headline Cmax); its one numeric consumer is MIPD steady-state dose sizing. Added a bi-exponential regression test the old all-points fit fails (returns ≈ln2/α instead of ln2/β).
2.
predict/drugbank.py— delete deadget_ugt_enzymesZero callers; superseded by the JSON substrate registry (
non_cyp_substrates). Also removed its now-dead plumbing:_ugt_substratesstore, the UGT populate block, and the unused_UGT_NORMALIZATIONmap.3.
predict/chemistry.py— delete dead XGBoost pKa loader_load_pka_models/_predict_pka_from_model/_PKA_MODEL_DIRwere never wired intocompute_profile(which uses DrugBank pKa + the SMARTS_estimate_pka_typefallback). The model was tested and reverted (worsened AAFE).Left intact (verified, out of scope)
merge_overlaystub +sc_overlay.yaml/tumor_overlay.yaml— dead today, but an intentional, documented Phase-3 extension seam (matched stub + fixture pair). Not cruft; kept.Verification
test_endpoints(12, incl. the new bi-exponential test) +test_drugbank+test_drugbank_integration+test_features_chemistry(61) pass; ruff clean; no residual references to the deleted symbols.