ImagiChem: Hybrid Deterministic Image-Conditioned Generation of Chemically Valid and Drug-like Molecules from Artistic Inputs

This is the authors' abstract. We don't add key points for this paper.

Generative artificial intelligence (AI) has redefined the exploration and expansion of the chemical space. However, the systematic use of visual information as a chemical prior remains virtually unexplored. Herein, we introduce ImagiChem, a deterministic image-conditioned generative framework that transforms visual patterns into chemically valid and drug-like molecules via three complementary generation modes: a library-based engine that assembles molecules from curated pharmacophore cores guided by pixel statistics, a rule-based from-scratch generator whose structural blueprint is governed by the global image profile, and a hybrid mode that combines both engines and merges their outputs. By encoding local color distributions, gradient statistics, and texture descriptors, as well as global image features, such as spatial coherence, chromatic harmony, and morphological class, into an alphabet of atomic, functional, and topological rules, ImagiChem establishes a direct multimodal mapping between visual features and molecular connectivity. This approach ensures compliance with physicochemical constraints, such as the Lipinski and Veber criteria, low pan-assay interference compounds (PAINS) incidence, and synthetic accessibility (SA), while yielding over 99% novel structures absent from current chemical databases. Statistical validation demonstrated that structured artistic inputs significantly enriched the yield of valid molecular hits compared with pooled random noise, particularly when using library-based and hybrid generation modes. ImagiChem introduces a new modality for chemical generation that conceptually bridges computer vision, molecular design, and human creativity, thereby expanding the frontier of computational chemistry into the domain of visual cognition.