Concept bottlenecks improve control and clarity in molecular graph generation
Interpretable and Steerable Molecular Graph Generation using Concept Bottlenecks
Machine Learning
Summary
Creating new molecules with AI usually happens inside complex systems that are hard for people to understand. The authors introduced a way to make these systems easier to interpret and guide by integrating concept bottlenecks, which act like clear checkpoints representing understandable ideas about molecules. This lets users see and influence how new molecules are made, helping reduce unnecessary trial and error in drug discovery. Their approach also keeps the quality of generated molecules high while allowing more user control.
What this means in practice
- •For drug discovery teams: Generate molecular candidates with interpretable features that allow steering the process toward desirable drug properties.$Commercial implications: Enables creation of more controllable drug design tools that reduce costly trial and error in pharmaceutical development.
- •For materials scientists: Produce molecular structures with interpretable controls to explore new materials with targeted characteristics.
Authors
Sarabeshwar Balaji, Ritesh Kumar, Lav R. Varshney
Abstract
Deep graph generators (DGGs) largely lack interpretability, which is a significant limitation for one of their most crucial applications, drug discovery, where interpretable and transparent generation could provide greater control over the desirability of generated molecules, thereby reducing trial-and-error in candidate selection. Common approaches to interpretability, such as disentanglement, cannot guarantee that the learned representations are human-interpretable. To address this, we propose incorporating concept bottlenecks into DGGs, allowing users to interpret, debug, and steer the generation process. We show that this consistently improves distributional fidelity (lower FCD), with only modest, non-systematic effects on validity and uniqueness. We further explore Concept Controller, an optimization-based steering approach that has proven to be effective for steerable image generation. We evaluate our approach on a one-shot molecular graph generation framework, though the underlying ideas can be transferred to sequential graph generation as well.