Beyond global drug-likeness: The FamDLike framework as a target family-specific drug-likeness score.

Drug-likeness is a central metric which is used to evaluate the physicochemical profile of small molecules either as a filter for virtual screening or, more recently, as a reward function in generative design tasks. Established scores such as QED are widely adopted for these purposes, although the ligands of different therapeutic target families occupy different regions of physicochemical space. Applying a single global drug-likeness criterion can therefore systematically reward some target classes while penalising others, narrowing the chemical space considered druglike in a way that may not reflect the actual requirements for a given project. In this work, we propose a target family specific drug-likeness approach, the FamDLike framework, designed to distinguish target families more accurately than existing metrics and to explore family specific druglike regions of chemical space. This framework combines three complementary modelling paradigms: desirability functions, generalized additive models and tree-based machine learning models that span the interpretability-flexibility spectrum. The three components can also be used independently, depending on whether interpretability or predictive flexibility is prioritized. The abilities of the proposed method have been tested on real-world use cases in computational medicinal chemistry: an independent set from DrugBank, library screening of Enamine and KLIFS compounds, and de novo molecule generation via REINVENT4. The proposed approach therefore provides a more discriminative, interpretable and target-aware complement to existing drug-likeness scores, suitable for ranking in early discovery pipelines and as a reward function for generative chemistry. Furthermore, a user-friendly web portal (https://huggingface.co/spaces/kochgroup/FamDLike) is available to obtain drug-likeness profiles of molecules.