mufu v1.0.0

MuFu, from CHM4930, fuses a frozen MoLFormer SMILES embedding, an ECFP4 fingerprint and a GINE graph encoder. For one SMILES it predicts EGFR and BACE1 pIC50 (with a conformal interval) and P(active), the 12 Tox21 assays and the 2 ClinTox tasks.

This is a 5-seed ensemble from commit 627f004, trained on data with sha256 b340ddf73f38e10f8b6a2948156417661a269ffc9a6d9b097b538c4690e1b041. pIC50 is the members' mean, and probabilities are averaged after each member's own calibration. The interval is μ ± q · max(σ, floor), where σ² adds the mean MC-Dropout variance and the variance between members, and q is fitted on validation.

The files only work as a set: each _mc_dropout.json is fitted to its checkpoint, and mufu_ensemble.json to exactly these checkpoints. Don't mix files across tags.

Test metrics

Task	Metric	Ensemble	Single model (mean ± std)	RF baseline
BACE1 pIC50	RMSE	0.795	0.827 ± 0.013	0.813
BACE1 pIC50	R²	0.610	0.578 ± 0.013	0.592
EGFR pIC50	RMSE	0.870	0.905 ± 0.015	0.877
EGFR pIC50	R²	0.510	0.470 ± 0.018	0.502
Tox21 (macro)	ROC AUC	0.751	0.735 ± 0.011	0.712
ClinTox (macro)	ROC AUC	0.861	0.822 ± 0.043	0.652

Single-model numbers are over the 5 seeds; RF is a random forest on ECFP4.

Usage

With the code and ai/src on the path (MolFormer is downloaded from ibm-research/MoLFormer-XL-both-10pct on first use):

from pathlib import Path

from huggingface_hub import snapshot_download

from encoders.smiles_encoder import SmilesEncoder
from mufu.ensemble import load_ensemble_params, load_members, member_paths

path = Path(snapshot_download("CHM4930/mufu", revision="v1.0.0"))
paths = member_paths(path, "mufu", range(5))
params = load_ensemble_params(paths, path, "mufu")
members = load_members(paths, SmilesEncoder(freeze=True))

Predict with mufu.ensemble.predict_with_ensemble_intervals, or serve it with MUFU_HF_REPO=CHM4930/mufu MUFU_MODEL_VERSION=v1.0.0 (see ai/README.md). For a single seed, use mufu.experiment.load_mufu_checkpoint.