qhjqhj00 ood-mol-opt-eval
Meta-learning Adaptive Deep Kernel Gaussian Processes for Molecular Property Prediction — Chen et al. (2022) (arXiv:2205.02708, 2022)
What this evaluates
Out-of-domain molecular property prediction and Bayesian optimization for molecular design. Tests transferability of learned representations to novel tasks.
Datasets
- Out-of-domain molecular design tasks — total ?; splits: test (-1)
Metrics
Top performing molecule property(primary) — range: other- Maximum property value found during Bayesian optimization over a fixed number of queries.
Input / output format
Input: Molecules from out-of-domain datasets. Feature representations extracted from models trained on FS-Mol.
Output: Selected molecules via Bayesian optimization with expected improvement acquisition function.
Scoring recipe
def score_ood_opt(features, dataset):
best_score = -inf
for _ in range(20):
init_mols = sample_worst_700(dataset, k=16)
score = bayesian_optimization(init_mols, features, acquisition='EI', batch_size=1)
best_score = max(best_score, score)
return best_score
Common pitfalls
- Uses expected improvement acquisition function with query-batch size 1.
- Starts from 16 randomly sampled molecules from the worst ~700 molecules in the dataset.
Evidence (verbatim from paper)
We repeat each BO experiment 20 times, each time starting from 16 randomly sampled molecules from the worst ~700 molecules within the dataset. Figure 4 shows that the ADKF-IFT representation enables fastest discovery of top performing molecules for the molecular docking, antibiotic discovery, and material design tasks.
Citation
@misc{chen2022metalearning,
title={Meta-learning Adaptive Deep Kernel Gaussian Processes for Molecular Property Prediction},
author={Chen et al. (2022)},
year={2022},
note={arXiv:2205.02708}
}
- arXiv: 2205.02708