Search papers, labs, and topics across Lattice.
This paper introduces Monroe, a molecular foundation model (MFM) that enhances bioassay activity prediction by leveraging a dataset of over 81 million molecules and innovative training techniques. Key advancements include improved graph representations for stereochemistry, novel training losses, and a prior-data-fitted model for in-context predictions, leading to superior performance on established benchmarks. Monroe not only matches or exceeds existing MFMs but also significantly enhances the performance of leading models like MiniMol and CheMeleon through its downstream adaptation strategy.
Monroe achieves state-of-the-art performance in molecular activity prediction, significantly outperforming existing models while enhancing others through innovative training strategies.
Bioassay activity prediction is often data-limited because drug-discovery datasets rely on time-consuming and expensive wet-lab experiments for data generation and evaluation. This challenge has inspired recent research into molecular foundation models (MFMs), which aim to encode general-purpose chemical knowledge into molecular representations that generalize well in data-constrained scenarios. This paper presents Monroe, a new MFM with several innovations over the existing state of the art: increased scale allowing pre-training on over 81 million molecules from the PM6 quantum chemistry dataset; improved graph representation of stereochemistry; improved training losses including conformer denoising and embedding decorrelation; improved multi-task learning; and the use of a prior-data-fitted model (TabPFN) for downstream in-context prediction. Our evaluations use a principled pairwise comparison framework that measures statistically significant performance differences. Across established Polaris benchmarks, Monroe matches or exceeds existing MFMs, while on activity cliff benchmarks, designed to assess utility for molecular discovery, it achieves significant improvements over prior methods. Finally, ablation and transfer experiments show that PFN-based downstream predictors also substantially improve two leading existing models, MiniMol and CheMeleon, yielding new state-of-the-art variants we call MiniMol_PFN and CheMeleon_PFN, suggesting that our downstream adaptation strategy generalizes beyond Monroe. Source code is at github.com/blazejba/monroe.