OpenADMET PXR Induction Blind Challenge — v43 Final Blend

TL;DR

• v31 is itself a cascade: v7 → v22 → v26 → v28 → v29 → v31, blending five Chemprop multitask models (T1, T1v5, T1v2, T13, T14), an AutoGluon-CheMeleon tabular model (T2), and two LightGBM baselines (v3, v4) on FCFP4-count + Mordred + CheMeleon-2048d features.
• T1v5 is a 5-head Chemprop D-MPNN multitask model trained jointly on the OpenADMET DRC, the counter-screen, NCATS qHTS PXR (AID 1346982 / 1346985), Tox21 SR-ARE, and ChEMBL NR1I2.
• Final calibration is honest per-fold isotonic regression — fit on out-of-fold predictions only, leave-one-fold-out. Isotonic fit on full training OOFs over-predicts improvement by ~0.01 RAE; honest fit was the gate that prevented submitting overfit garbage.

Repository layout

.
├── README.md                          # this card
├── docs/
│   ├── METHODOLOGY_REPORT.md          # required contest methodology report (source)
│   └── METHODOLOGY_REPORT.pdf         # rendered PDF
├── code/
│   ├── requirements.txt               # pinned deps
│   ├── featurization/                 # CheMeleon, Mordred, FCFP4 wrappers
│   ├── baseline/                      # v2 XGBoost, v3/v4 LightGBM
│   ├── multitask/                     # T1, T1v5 (Chemprop), T2 (AutoGluon)
│   ├── ensemble/                      # v26, v31, v43 cascade blends
│   ├── submit/                        # Gradio API submission script
│   └── utils/                         # validation utilities
├── models/
│   ├── T1_chemprop/                   # 5-seed multitask Chemprop (12 MB)
│   ├── T1v5_chemprop/                 # 5-fold + pretrained 5-head Chemprop (9.5 MB)
│   └── T2_autogluon_chemeleon/        # AutoGluon-tabular on CheMeleon emb (1.6 GB)
├── data/
│   ├── raw/                           # OpenADMET-provided CSVs (also at openadmet/pxr-challenge-train-test)
│   ├── parent_clusters_fcfp4.csv      # FCFP4-NN parent-cluster assignment (LOCO grouping)
│   └── oof_predictions/               # per-track OOF + test predictions (inputs to v43 blend)
└── submission/
    ├── v43_final.parquet              # 513-row final submission
    ├── v43_final.csv
    └── v43_final.json                 # manifest with CV metrics + LB outcome

Reproducing v43 from the OOF predictions

Reproducing the trained models from scratch

1. Pull the OpenADMET data: data/raw/pxr-challenge_TRAIN.csv, pxr-challenge_TEST_BLINDED.csv, the counter-screen and single-concentration files. (Also at openadmet/pxr-challenge-train-test.)
2. Generate features: python code/featurization/featurize_mordred.py and code/featurization/featurize_chemeleon.py.
3. Train baselines: python code/baseline/v3_v4_lgbm.py → produces v3 (Butina-LOCO LGBM) and v4 (kitchen-sink LGBM) OOFs.
4. Train multitask models: code/multitask/T1_chemprop_external_pretrain.py, T1v5_chemprop_chembl_nr1i2.py, T2_autogluon_chemeleon.py.
5. Run cascade: code/ensemble/v26_cascade_blend.py → v31_v29_t2_blend.py → v43_final_blend.py.

Final-leaderboard metrics (live leaderboard, partial test set)

What worked

1. Diversity over scale. A 300-parameter Chemprop blended with a 4,900-feature LightGBM beat single models 10× larger. The error modes of D-MPNN graph models and gradient-boosted trees on descriptors are decorrelated; combining them captures variance neither covers alone.
2. Honest per-fold isotonic. In-sample isotonic fit on full OOFs is optimistic by ~0.009 RAE; chained iso-on-iso compounds at ~0.01 per step. Per-fold (leave-one-fold-out) isotonic was the only protocol that produced CV→LB shifts consistent with the actual leaderboard.
3. Parent-cluster LOCO. Random folds leak parent-series into validation. The test set is 513 analogs of ~89 parent compounds; FCFP4-NN-based grouping reproduces that structure for CV.
4. External multitask aux. ChEMBL NR1I2 (907 compounds, T1v5 head 5) and NCATS qHTS PXR (AID 1346982 / 1346985, ~10K compounds, T1 heads 2-3) added measurable lift even though individual heads had Spearman ≤ 0.4 against the primary task. Multitask transfer works through shared representations, not direct correlation.

What did not work (in case it saves you time)

• NN-transfer of any biological-assay readout from train compounds to test. This caused a catastrophic LB regression (v50: OOF 0.4536 → LB 0.658, rank 87). Test compounds include "SAR misses" — high structural similarity to active train compounds but biologically inactive. Nearest-neighbor copying of the train assay readout assigns high activity to inactive misses and overshoots pEC50.
• TabPFN on CheMeleon. Spearman 0.93 against v43 → no orthogonal signal. Foundation-model embeddings dominate the representation regardless of the tabular learner on top.
• ChemBERTa, GIN-ACtriplet, UniMolV2-310M, MaskMol, ImageNet ViT on molecular images. All converged to Spearman > 0.88 against v43. The 2D-graph representation space is saturated for this task and this data size.
• Tail-weighted loss / quantile regression heads. Cannot fix the data ceiling: only 11 training compounds have pEC50 ≥ 6.5.

License

Citation

Acknowledgements