Write a PREreview

Generative molecular models such as diffusion models and graph neural networks are widely used in drug design. However, their black-box nature means they lack an explicit understanding of chemical rules (e.g., valency, charge, aromaticity), often generating chemically impossible structures such as pentavalent carbon. To address this issue, this paper proposes Atomic-SCS, an atom-level chemical rule scoring tool based on a symbolic approach. Atomic-SCS does not rely on data-driven training but directly applies IUPAC rules to independently score each atom across four dimensions: valency, charge, aromaticity, and ring strain. It outputs continuous scores (0 = fully compliant, 1 = severe violation), provides atom-level diagnostic reports, and generates prioritized repair suggestions sorted by severity. The tool supports three strictness levels (conservative, balanced, liberal) and three operation modes (assess, diagnose, repair), and can be accessed via a command-line interface or Python API. Validation on 100 normal and 100 problematic molecules shows that Atomic-SCS effectively distinguishes valid from invalid structures (Mann-Whitney U test, p < 1e-18). The scoring functions are continuous and can serve as reward signals in generative model training. On a standard CPU, scoring 100 molecules averaged 0.000071 s per molecule. This work provides a rule-based scoring tool for generative molecular design.