Unlocking High-Fidelity Molecular Generation from Mass Spectra via Dual-Stream Line Graph Diffusion

Researchers introduce DualLGD, a novel dual-stream diffusion architecture for generating molecular structures from mass spectra data. The method achieves 3x improvement over previous state-of-the-art by separating atom-level and bond-level reasoning into dedicated computation streams, addressing a fundamental circular dependency problem in molecular generation.

DualLGD represents a significant architectural advancement in computational chemistry by tackling a long-standing challenge in de novo molecular generation. The core innovation lies in recognizing that existing single-stream graph diffusion models create an implicit bottleneck—atoms and bonds must be reasoned about simultaneously despite their interdependent nature. By reformulating molecular denoising as two coupled subproblems operating in separate representation spaces, the researchers solve this circularity problem more elegantly.

The line graph construction for bond space is mathematically elegant, mapping bond angles, dihedrals, conjugation chains, and rings into topological motifs. The incidence-constrained bidirectional cross-attention mechanism ensures chemical validity by restricting atoms to attend only to incident bonds, encoding fundamental chemistry principles directly into the architecture. This design respects domain knowledge rather than relying on the model to discover these constraints implicitly.

The empirical results are substantial: 34.37% and 23.89% top-1 accuracy on benchmark datasets represent approximately 3x improvement over prior work. Critically, DualLGD achieves this without pre-training, suggesting the architectural improvements drive performance gains rather than data advantages. This efficiency is important for practical applications where training resources are limited.

For the broader AI and computational chemistry landscape, this work demonstrates how domain-informed architectural design can unlock significant performance improvements. The approach may inspire similar dual-stream or multi-stream designs in other scientific domains facing circular dependency problems. Potential applications span drug discovery, materials science, and synthetic chemistry optimization, where accurate molecular generation directly impacts research productivity and cost reduction.

• →DualLGD achieves 3x improvement over previous state-of-the-art in molecular generation from mass spectra by separating atom and bond reasoning into dedicated streams.
• →Line graph representation provides a mathematically natural framework for capturing bond-level chemical properties like conjugation and ring structures.
• →Incidence-constrained cross-attention ensures chemical validity by enforcing that atoms only attend to incident bonds, encoding domain principles directly into the model.
• →Superior performance without pre-training indicates architectural improvements are the primary driver of gains, not data or training advantages.
• →Dual-stream paradigm addresses circular dependency between atom-level and bond-level reasoning that limited single-stream diffusion approaches.