What happened
Dr. Andrea Mastropietro and Prof. Dr. Jürgen Bajorath from the University of Bonn and Lamarr Institute published research in Cell Reports Physical Science, introducing DiffSHAPer, an explainability method for molecular diffusion models in drug design. Their findings indicate these models generate chemically valid linkers primarily by relying on distance constraints between atoms, not by learning or exploiting chemical principles. This means current models use recurrent statistical patterns rather than generalisable chemical rules for linker generation.
Why it matters
Current molecular diffusion models for drug design do not guarantee optimisation of critical molecular properties like potency and stability, as their linker generation relies on statistical distance patterns rather than chemical principles. This mechanism limits the practical chemical utility of generated linkers for drug discovery teams. Computational chemists must therefore understand what these models learn, preparing for future models that integrate more chemical context into their reasoning.
