Monroe is a molecular foundation model designed for property prediction in low-data regimes. It combines large-scale multi-task pretraining on quantum-chemical and bioactivity data with explicit stereochemical representations and conformer denoising, and uses TabPFN for downstream prediction from frozen molecular embeddings. I will present the main ideas behind Monroe, its performance across molecular property prediction benchmarks, and what we learned about combining molecular foundation models with tabular inference.

About the speaker: Błażej Banaszewski is a researcher at the Max Planck Institute of Biochemistry working at the intersection of machine learning and molecular science. His research focuses on foundation models for molecular representation, property prediction, and biomolecular design.