Learning molecular representation is a critical step in molecular machine learning that significantly influences modeling success, particularly in data-scarce situations. The concept of broadly pre-training neural networks has advanced fields such as computer vision, natural language processing, and protein engineering. However, similar approaches for small organic molecules have not achieved comparable success. In this work, we introduce a novel pre-training strategy, substrate scope contrastive learning, which learns atomic representations tailored to chemical reactivity. This method considers the grouping of substrates and their yields in published substrate scope tables as a measure of their similarity or dissimilarity in terms of chemical reactivity. We focus on 20,798 aryl halides in the CAS Content Collection spanning thousands of publications to learn a representation of aryl halide reactivity. We validate our pre-training approach through both intuitive visualizations and comparisons to traditional reactivity descriptors and physical organic chemistry principles. The versatility of these embeddings is further evidenced in their application to yield prediction, regioselectivity prediction, and the diverse selection of new substrates. This work not only presents a chemistry-tailored neural network pre-training strategy to learn reactivity-aligned atomic representations, but also marks a first-of-its-kind approach to benefit from the human bias in substrate scope design.

Capturing the microscopic interactions that determine molecular reactivity poses a challenge across the physical sciences. Even a basic understanding of the underlying reaction mechanisms can substantially accelerate materials and compound design, including the development of new catalysts or drugs. Given the difficulties routinely faced by both experimental and theoretical investigations that aim to improve our mechanistic understanding of a reaction, recent advances have focused on data-driven routes to derive structure-property relationships directly from high-throughput screens. However, even these high-quality, high-volume data are noisy, sparse and biased -- placing them in a regime where machine-learning is extremely challenging. Here we show that a statistical approach based on deep filtering of nonlinear feature networks results in physicochemical models that are more robust, transparent and generalize better than standard machine-learning architectures. Using diligent descriptor design and data post-processing, we exemplify the approach using both literature and fresh data on asymmetric catalytic hydrogenation, Palladium-catalyzed cross-coupling reactions, and drug-drug synergy. We illustrate how the sparse models uncovered by the filtering help us formulate physicochemical reaction ``pharmacophores'', investigate experimental bias and derive strategies for mechanism detection and classification.

We demonstrate that the chemical-feature model described in our original paper is distinguishable from the nongeneralizable models introduced by Chuang and Keiser. Furthermore, the chemical-feature model significantly outperforms these models in out-of-sample predictions, justifying the use of chemical featurization from which machine learning models can extract meaningful patterns in the dataset, as originally described. In Ahneman et al. (1), we showed that a random forest (RF) algorithm built using computationally derived chemical descriptors for the components of a Pd-catalyzed C–N cross-coupling reaction (aryl halide, ligand, base, and potentially inhibitory isoxazole additive) could identify predictive and meaningful relationships in a multidimensional chemical dataset comprising 4608 reactions. Chuang and Keiser (2) built alternative models using random barcode features ("straw" models), wherein the chemical descriptors are replaced with random numbers selected from a standard normal distribution. One-hot encoded features, wherein each reagent acts as a categorical descriptor and is marked as absent or present, were also evaluated.