Navigating chemical-linguistic sharing space with heterogeneous molecular encoding

Chemical language models are powerful tools for navigating chemical space, but their reliance on linear representations such as molecular strings creates a semantic gap, hindering their ability to bridge natural language with the full complexity of molecular structures. Here we show that chemical language models can gain a comprehensive, multi-modal understanding of molecules through heterogeneous molecular encoding, which integrates one-dimensional sequences, two-dimensional topology, three-dimensional geometry, and statistically derived molecular fragments. We further introduce a query-based module that converts heterogeneous structural information into a unified representation compatible with language models, together with a chain-of-fragment mechanism that guides molecular generation through a hierarchical chemical blueprinting process. To support research in this area, we constructed a million-scale dataset for multi-objective molecular design. Experimentally, the framework enables bidirectional navigation of the chemical-linguistic space, achieving consistent improvements across molecular comprehension and design tasks over strong baselines.