Ark-ADMET: released CatBoost and Chemprop checkpoints for the TDC ADMET benchmark

This repository contains one five-seed CatBoost model set for each of the 22 Therapeutics Data Commons (TDC) ADMET endpoints, plus a separate five-seed Chemprop model set for Lipophilicity_AstraZeneca.

Models were evaluated using the official TDC splits and endpoint-specific metrics. CatBoost feature and hyperparameter selection used an internal scaffold split derived from the seed 1 training partition. The Chemprop model is an independent submission: its predictions are neither used as features by nor averaged with the CatBoost models.

Overview

• 22 CatBoost model sets, one directory per TDC endpoint, using normalized lower-case directory names.
• 5 official TDC splits per endpoint (seed1 to seed5). Each catboost.cbm is fitted on the training partition from the corresponding split.
• CatBoost prediction is the mean over the five seeds.
  • Classification: the score is the positive-class probability.
  • Regression: the score is on the original label scale (the target scaler is inverted using y_scaler.json).
• Optional feature blocks (PaDEL 3D descriptors, GIN embeddings, MiniMol embeddings) are used only by the CatBoost endpoints that kept them; see the endpoint table.
• match.json is an audit log from the CatBoost refit only: every saved test prediction matched the reported vector with a maximum absolute difference of 0 at atol = 1e-5. It is not needed to load or run a model.
• Chemprop lives under chemprop/lipophilicity_astrazeneca/ and does not replace the CatBoost directory lipophilicity_astrazeneca/.

Quick start

Install the dependencies from requirements.txt:

Package	Version
Python	3.11.16
catboost	1.2.10
numpy	1.26.4
rdkit	2026.03.6
huggingface_hub	0.36.2

Download a snapshot and load an endpoint that uses only the fingerprint block:

from huggingface_hub import snapshot_download
from predict import EndpointModel

repo_dir = snapshot_download("Eonark/Ark-ADMET")
model = EndpointModel.from_pretrained(repo_dir, "pgp_broccatelli")
print(model.predict(["CCO", "c1ccccc1"]))

Endpoints that kept extra blocks need those arrays supplied by the caller:

from predict import EndpointModel

model = EndpointModel.from_pretrained(repo_dir, "herg")

# padel:   (n, 431) in feature_spec["padel_3d"]["columns"] order,
#          or (n, n_3d) already sliced to the endpoint's kept columns
# gin:     (n, 300) from the frozen GIN encoder described below
# minimol: (n, 512) from MiniMol v1; required when blend_weight > 0
#          or minimol_concat is true
pred = model.predict(smiles, padel=padel, gin=gin, minimol=minimol)

The MiniMol blend head is stored as minimol_head.npz, which is the preferred format and does not require scikit-learn. minimol_head.joblib is kept as the original fitted object.

Endpoints

The directory name is the second argument to EndpointModel.from_pretrained. The metric is the official TDC ADMET metric for that endpoint. Columns 3D, GIN, MiniMol concat, and Blend show which optional feature blocks the endpoint uses (3D is the number of PaDEL 3D columns kept; Blend is blend_weight).

Directory	TDC name	Task	Metric	3D	GIN	MiniMol concat	Blend
dili	DILI	classification	ROC-AUC	431			0.1
hia_hou	HIA_Hou	classification	ROC-AUC	0	yes		0
bioavailability_ma	Bioavailability_Ma	classification	ROC-AUC	431	yes	yes	0.1
herg	hERG	classification	ROC-AUC	200	yes		0.2
half_life_obach	Half_Life_Obach	regression	Spearman	431			0.8
cyp2d6_substrate_carbonmangels	CYP2D6_Substrate_CarbonMangels	classification	PR-AUC	100	yes		0
cyp2c9_substrate_carbonmangels	CYP2C9_Substrate_CarbonMangels	classification	PR-AUC	100			0
cyp3a4_substrate_carbonmangels	CYP3A4_Substrate_CarbonMangels	classification	ROC-AUC	431			0.2
caco2_wang	Caco2_Wang	regression	MAE	431			0
clearance_microsome_az	Clearance_Microsome_AZ	regression	Spearman	431	yes	yes	0
vdss_lombardo	VDss_Lombardo	regression	Spearman	431		yes	0
clearance_hepatocyte_az	Clearance_Hepatocyte_AZ	regression	Spearman	431			0
pgp_broccatelli	Pgp_Broccatelli	classification	ROC-AUC	0			0
bbb_martins	BBB_Martins	classification	ROC-AUC	431			0
ppbr_az	PPBR_AZ	regression	MAE	25	yes		0.1
lipophilicity_astrazeneca	Lipophilicity_AstraZeneca	regression	MAE	431			0.3
ames	AMES	classification	ROC-AUC	0		yes	0.2
ld50_zhu	LD50_Zhu	regression	MAE	0			0.2
solubility_aqsoldb	Solubility_AqSolDB	regression	MAE	50			0.2
cyp2c9_veith	CYP2C9_Veith	classification	PR-AUC	0	yes		0.2
cyp3a4_veith	CYP3A4_Veith	classification	PR-AUC	200	yes		0.1
cyp2d6_veith	CYP2D6_Veith	classification	PR-AUC	431	yes	yes	0

Features

Fingerprint block (all endpoints)

Every model starts with the same 2563 fingerprint columns, in this order:

1. Morgan count, radius 2, 1024 bins (GetHashedMorganFingerprint)
2. Avalon count, 1024 bins (GetAvalonCountFP)
3. ErG, 315
4. 200 RDKit descriptors, named in feature_spec.json

predict computes this block directly from SMILES. Invalid SMILES are represented internally as missing features; callers should validate inputs and reject such rows rather than interpret the resulting predictions.

Optional blocks appended after the fingerprint

Block	Used when	Width	Supplied via
PaDEL 3D	n_3d > 0	n_3d	padel
GIN	gin is true	300	gin
MiniMol	minimol_concat is true	512	minimol

PaDEL 3D (431 columns).

• Only the 3D descriptor group is used (PaDEL flags -3d -retain3d -retainorder).
• Conformers come from RDKit ETKDGv3 with randomSeed=1, hydrogens kept, then MMFF optimization with maxIters=200.
• If the first embedding fails, it is retried with useRandomCoords=True. A molecule that still fails is omitted, and its 3D descriptors stay missing.
• Some endpoints keep a truncated, ranked subset of the 431 columns. The kept names and their indices into the 431-column list are stored in feature_spec.json under endpoints.<name>.padel_columns and padel_index. An endpoint with n_3d = 431 uses the full list in that order.

GIN (300 dimensions).

• Encoder: dgllife.model.load_pretrained("gin_supervised_masking").
• Graphs: mol_to_bigraph(add_self_loop=True, canonical_atom_order=False) with PretrainAtomFeaturizer and PretrainBondFeaturizer.
• Readout: dgl AvgPooling.
• The encoder weights are not bundled in this repository.

MiniMol (512 dimensions).

• Embeddings come from public MiniMol v1.
• The encoder weights are not bundled in this repository.

MiniMol blend head

Where blend_weight is above 0, each seed also has a small head on the MiniMol vector: StandardScaler, one hidden layer of 128 ReLU units, then a linear output. The per-seed prediction is

(1 - alpha) * CatBoost + alpha * head

where alpha is the endpoint's blend_weight. If a row's MiniMol vector is not finite, that row keeps the CatBoost prediction unchanged.

Intended use and limitations

These checkpoints are intended for research, benchmark reproduction, and method comparison. They have not been validated for clinical, diagnostic, regulatory, or safety-critical use. Predictions may be unreliable for molecules and assay conditions outside the chemical and experimental domains represented by the TDC datasets.

The repository contains fitted parameters and inference helpers, not a complete ADMET decision system. Users remain responsible for input validation, applicability-domain assessment, uncertainty analysis, and experimental confirmation.

• Unfeaturizable SMILES. Two solubility-test molecules have no MiniMol vector and fall back to CatBoost: CC1=CC=C[NH+2]([O-])[CH-]1 and O=C(O)C1=C[NH+2]([O-])[CH-]C=C1.
• CYP2D6 substrate class weighting. CYP2D6_Substrate_CarbonMangels was fitted with scale_pos_weight = 3. This weight is already baked into its trees; do not apply it again.
• External encoders. GIN and MiniMol embeddings must be computed by the caller with the specifications above. Mismatched encoder versions or featurization settings will change predictions.

Separate Chemprop submission (Lipophilicity_AstraZeneca)

Directory: chemprop/lipophilicity_astrazeneca/.

This is the independent Chemprop 2.3.1 message-passing model reported alongside the CatBoost family. For each of the five official TDC splits, the checkpoint with minimum validation loss was retained; this is the only use of the official validation partition in this work. The released prediction is the mean of the five checkpoint outputs. Hyperparameters: message-passing depth 3, feed-forward depth 2, hidden dimension 300, dropout 0, up to 50 epochs, and molecular-graph input only (no extra RDKit descriptors).

Item	Value
Task / metric	regression / MAE
Five-seed test MAE	0.412 ± 0.012 (unrounded mean ± sample SD)
Relation to CatBoost	separate submission; not blended
CatBoost counterpart	still in lipophilicity_astrazeneca/

The root requirements.txt covers CatBoost inference. Chemprop checkpoints require the packages in chemprop/requirements.txt. Load the five seed{1-5}/best.pt checkpoints with Chemprop 2.3.1 and average their predictions. See chemprop/lipophilicity_astrazeneca/config.json for the public hyperparameter record.

Repository layout

<endpoint>/config.json                      # CatBoost endpoints
<endpoint>/match.json                       # audit log only
<endpoint>/seed{1-5}/catboost.cbm
<endpoint>/seed{1-5}/y_scaler.json          # regression only
<endpoint>/seed{1-5}/minimol_head.npz       # blend_weight > 0 (preferred)
<endpoint>/seed{1-5}/minimol_head.joblib    # same head, original object
chemprop/lipophilicity_astrazeneca/config.json
chemprop/lipophilicity_astrazeneca/seed{1-5}/best.pt
chemprop/lipophilicity_astrazeneca/seed{1-5}/metrics.json
chemprop/requirements.txt
feature_spec.json
versions.json
predict.py                                  # CatBoost inference helper
requirements.txt

License

Released under the Apache License 2.0. Copyright 2026 ADMET endpoint models contributors.

Third-party fingerprints, PaDEL, GIN, and MiniMol remain under their own licenses. Their pretrained weights are not included in this repository.