How neural networks learn the shape of molecules?

Neural networks need to understand molecules well for accurate predictions. A research team tested a different approach: instead of letting the network mimic chemical structure on its own, they explicitly taught it the central core, called the Bemis-Murcko scaffold. Think of the scaffold as a bicycle frame: it's the central part where various components attach. The network learned to group molecules with similar scaffolds, making predictions more precise. The researchers also tried two different ways to organize representation spaces: Euclidean (like the 3D space we know) and hyperbolic (where space expands infinitely). The hyperbolic method worked better for organizing molecules, but neither was consistently faster or more accurate in all cases. Explicit teaching of structure thus seems to yield reliable results, though benefits always depend on the type of chemical property to be predicted.

In the field of machine learning applied to chemistry, molecular representations are learned using unsupervised learning techniques. This means neural networks analyze large amounts of chemical data without explicit labels, seeking to capture hidden relationships. The concept of the Bemis-Murcko scaffold identifies the central structure of a molecule, removing side groups and secondary chains. This structural framework is crucial because it determines many of the molecule's chemical properties. The geometric spaces used to represent molecules can follow Euclidean or hyperbolic geometries. In Euclidean geometry, distances follow traditional geometric rules. In hyperbolic geometry, space has negative curvature, allowing complex hierarchies to be represented more efficiently. Contrastive functions, such as those used in this study, push the network to keep similar molecules close in the representation space. The hyperbolic approach seems particularly suitable for hierarchical data, such as relationships between chemical scaffolds. However, specific benefits heavily depend on the target task and the quality of available data.