Deep Learning Foundation Models for Low-Data Regimes from Classical Molecular Descriptors

Fast and accurate data-driven prediction of molecular properties is pivotal to scientific advancements across myriad chemical domains. Deep learning methods have recently garnered much attention, despite their inability to outperform classical machine learning methods when tested on practical, real-world benchmarks with limited training data. This study seeks to bridge this gap by introducing a new avenue for foundation model pretraining. We propose pretraining on low-noise, calculable molecular descriptors via supervised learning to obtain rich, highly transferable molecular representations. We demonstrate this strategy with CheMeleon, a O(10M) parameter foundation model that enables directed message-passing neural networks to finally exceed the performance of classical methods in the low-data regime. We evaluate on 58 benchmark data sets spanning a range of properties relevant to small-molecule drug discovery, sourced from the industry-led Polaris benchmarking initiative. Rigorous statistical comparisons show that CheMeleon outperforms classical baselines like Random Forest on molecular fingerprints and descriptors, as well as existing foundation models. We open-source the CheMeleon model and the pretraining framework to encourage adoption and extension of this pretraining strategy across chemical sciences.