Bilevel optimization densifies scarce labels to fix OOD molecular property prediction (ICML WS 2025)

CatAstro_Piyush · x · 2026-09-26

Accepted at GenBio ICML Workshop 2025 (arXiv:2506.11877), this work by Jina Kim, Jeffrey Willette, Bruno Andreis, and Sung Ju Hwang tackles a core drug-discovery problem: molecular prediction models generalize poorly to out-of-distribution compounds, and labeled data is scarce because experimental validation is costly. The authors propose a bilevel optimization approach that leverages unlabeled data to interpolate between in-distribution and OOD samples, teaching the model to generalize beyond the training set. They report significant gains on real-world datasets with heavy covariate shift, supported by t-SNE visualizations.

Original post →

More from Research

Research channel →