Leveraging Domain Motif Assembler for Multi-objective, Multi-domain and Explainable Molecular Design

Designing molecules to meet multiple objectives simultaneously is a significant challenge, rooted in the inherent complexity of structure-property relationships and the limited availability of labeled molecular data. Compounding this difficulty, the black-box nature of deep learning models hampers interpretability, undermining confidence in their applications and restricting broader industrial adoption. To address these limitations, we introduce Domain Motif Assembler (DM-Assembler), a novel framework that employs a score-based matching model to guide the selection and assembly of molecular motifs. DM-Assembler excels in generating high-performance molecules that satisfy multi-objective constraints across various domains, drastically reducing the combinatorial complexity of chemical space by leveraging domain motifs and diffusion models on discrete fragment representations. Comprehensive experimental evaluations reveal three distinct advantages of DM-Assembler. First, it delivers exceptional performance in designing molecules with optimized composite metrics, achieving an average improvement of approximately 56\% across four benchmark domains, while extending beyond the property boundaries of its training data. Second, it generates molecular distributions that closely match domain-specific datasets, recording the lowest KL divergence among 149 descriptors encompassing multi-scale topological structures and biochemical properties—outperforming the second-best method by roughly 19\%. Third, we developed a post-hoc interpretability method based on 149 molecular descriptors, revealing DM-Assembler’s superior ability to fit physicochemical properties compared to atom-based models and topological structures compared to non-diffusion models, highlighting its advantages in molecular generation tasks. Furthermore, DM-Assembler demonstrates outstanding conditional sampling capabilities, seamlessly integrating auxiliary scorers or domain motif priors for efficient and accurate conditional generation. This enables deeper qualitative exploration of structure-activity relationships and accelerates pharmacophore discovery. Overall, we anticipate that DM-Assembler will advance trust-worthy and multi-objective molecular design across diverse disciplines, fostering significant progress in related research fields.