docs(dead-ends): DE-58 — censored (Tobit/AFT) CLint regression does not improve end-to-end Cmax#108
Merged
Conversation
…ot improve end-to-end Cmax
Deep-research candidate (Svensson et al. 2025). CLint training is 27.4% censored
(floor 3.0 / ceiling 150); shipped model treats the piles as exact. Controlled AFT
ablation (XGBoost survival:aft normal; AFT-exact vs AFT-censored isolates censoring):
Stage A: censoring moves CLint materially (median |Δlog10| 0.088, 52/107 >26%),
concentrated in low-CLint renal drugs pushed below the floor.
Stage B (end-to-end, one local stack):
engine AAFE +0.054 (worse), meta -0.0045 (noise vs CI hw ~0.42),
per-drug 28 better / 34 worse = net wash.
ml bit-identical across arms; AFT-exact meta 2.634 ~ shipped 2.622 (proxy valid).
Matches Svensson's own finding (censoring lifts uncertainty/NLL, not point accuracy)
and re-demonstrates §2 error-cancellation + §8/DE-42 + DE-43. Not integrated; headline
2.743 untouched (dev-state experiment, state-internal Δ).
- scripts/clint_censored_regression.py: reproducible gate (PYTHONPATH=src)
- data/validation/clint_censored_regression_2026-07-08.json: result artifact
- diagnosis.md §1: censored-regression bullet + external assay-reproducibility refs
jam-sudo
enabled auto-merge (squash)
July 9, 2026 01:53
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Gate-tests the censored-label regression candidate surfaced by the 2026-07-08 deep-research sweep (Svensson et al. 2025, AI in the Life Sciences 7:100128). Verdict: FAIL → DE-58. Headline 2.743 untouched (dev-state experiment, nothing integrated).
Motivation
The CLint training set (
clearance_hepatocyte_az.tab) is 27.4% censored (16.1% left-censored at the assay floor CLint=3.0; 11.3% right-censored at the ceiling 150), yet the shipped model (reg:squarederroron log10) treats those piles as exact point labels. Hypothesis: honest censoring recovers extreme-value information and improves the CLint input, hence Cmax.Method — controlled AFT ablation
Both arms are XGBoost
survival:aft(normal distribution = Tobit on log time), identical features (Morgan-2048 + 9 RDKit) and hyperparameters; only label bounds differ, isolating the censoring treatment:lower=upper=Y.(0, 3.0]; ceiling → right-censored[150, +inf).Stage A: per-drug ΔCLint on the 107-holdout. Stage B: feed shipped / AFT-exact / AFT-censored CLint through the full pipeline (predict → engine → meta) on one local stack; compare AAFE.
Result
Stage A — censoring moves CLint materially: median |Δlog10(CLint)| = 0.088 (~22%), 52/107 drugs move >26%, max 3.8× (isosorbide 6.5→1.8). Movers are predominantly low-CLint renally-cleared drugs pushed below the naive floor — exactly where hepatic CLint is not the Cmax-binding term.
Stage B — no end-to-end gain (dev-state, stack-internal Δ):
Why it fails
Svensson's own conclusion realised on our pipeline (censoring lifts uncertainty/NLL, not point accuracy), and a fresh controlled re-demonstration of three internal walls: §2 error-cancellation (better-motivated CLint breaks the joint balance → engine worse), §8/DE-42 (moved drugs are low-CLint renal cases where F binds Cmax, not CLint), DE-43 (fixed-weight meta damps the engine move, +0.054 → −0.0045). Per §4 the gate precedes integration → not integrated.
Changes
docs/research/dead-ends.md— DE-58 entrydocs/research/diagnosis.md— §1 censored-regression bullet + external assay-reproducibility refs (Fagerholm 2022; Bowman & Benet 2019)scripts/clint_censored_regression.py— reproducible gate (PYTHONPATH=src, deterministic seed 20260422)data/validation/clint_censored_regression_2026-07-08.json— result artifactSanitization guard passes (713 tracked files); ruff clean.