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

Technical Summary: DualLGD for Molecular Generation from Mass Spectra

Problem Definition
The paper addresses the inverse problem of de novo molecular generation from tandem mass spectra (MS/MS). Given a spectrum and a molecular formula, the goal is to infer the complete molecular graph (atom types and bond types). The core difficulty lies in a "circular dependency" between atom-level and bond-level reasoning: determining a bond's type (e.g., single, double, aromatic) requires knowledge of the chemical environment of its endpoint atoms, yet an atom's chemical environment is defined by the collection of bonds it participates in.

Existing graph diffusion methods (e.g., DiffMS, MBGen) operate within a single-stream architecture where atoms and bonds are processed in a unified computation stream. In these models, bonds are treated as edge attributes subordinate to node-level computations. Consequently, atom-bond information synchronization occurs only implicitly across layers, leading to a bottleneck where updated atom representations at layer $l$ are paired with stale bond representations from layer $l-1$, and vice versa.

Methodology: DualLGD
The authors propose DualLGD (Dual-stream Line Graph Diffusion), which reformulates molecular graph denoising as the alternating solution of two coupled subproblems: atom-level reasoning and bond-level reasoning, each operating in a dedicated representation space.

1. Dual-Stream Architecture:

   • Primal Stream: Maintains atom states ($h$) and edge states ($e$), inheriting the structure of previous graph diffusion methods.
   • Line Graph Stream: Promotes bonds from secondary edge attributes to first-class nodes in a line graph $L(G)$. Here, every potential bond $(i, j)$ becomes an independent node $u_{ij}$ with its own representation vector $z$.
   • Bond-Level Reasoning: In the line graph stream, bonds attend to all other bonds via global self-attention. This allows chemically meaningful substructures—such as bond angles (edges in $L(G)$), dihedral angles (length-2 paths), conjugation chains, and ring systems (cycles)—to be modeled as direct topological motifs between bond nodes. To handle the quadratic complexity of the line graph ($M = \binom{N}{2}$), the authors employ a FAVOR+ random-feature approximation for attention.

2. Incidence-Constrained Bidirectional Cross-Attention:
   To resolve the circular dependency, the two streams are explicitly synchronized at every layer via cross-attention constrained by the molecular incidence matrix $I$.

   • Atom $\to$ Bond: Each bond aggregates information only from its two endpoint atoms.
   • Bond $\to$ Atom: Each atom aggregates information only from its incident bonds.
     This topology-aware constraint ensures that the chemical principle (an atom's environment is determined solely by its bonding context) is respected during inference, replacing the implicit, cross-layer synchronization of single-stream models with explicit, per-layer coupling.

3. Global Fusion and Conditioning:
   A global state vector, derived from the spectrum encoder, modulates both streams at every layer using Feature-wise Linear Modulation (FiLM), ensuring spectral information guides both atom and bond reasoning simultaneously.

Key Contributions

• Architectural Reformulation: The paper identifies the single-stream paradigm as a key bottleneck and proposes a dual-stream design where bonds are first-class nodes in a line graph, granting them an independent computation path and representation space.
• Explicit Synchronization: The introduction of incidence-constrained bidirectional cross-attention directly addresses the circular dependency between atom and bond inference, ensuring tight coupling at every denoising step.
• Inductive Bias: By modeling bonds as nodes, the architecture naturally encodes complex chemical relationships (angles, dihedrals, conjugation) as local topological patterns, providing a strong inductive bias for bond-level reasoning.

Experimental Results
DualLGD was evaluated on two benchmarks: NPLIB1 (natural products with structural similarity) and MassSpecGym (large-scale, strict scaffold-based split representing out-of-distribution challenges).

• Performance Gains: DualLGD achieved a top-1 accuracy of 34.37% on NPLIB1 and 23.89% on MassSpecGym. This represents approximately a 3$\times$ improvement over the previous state-of-the-art (MBGen: 12.20% and 7.58%, respectively).
• Metric Improvements: The model also significantly reduced the Maximum Common Edge Subgraph (MCES) distance and increased Tanimoto similarity across both datasets.
• Ablation Studies:
  • Architecture vs. Pre-training: A critical finding is that DualLGD trained without any pre-training achieved 25.53% top-1 accuracy, surpassing the previous best fully pretrained model (12.20%). This confirms that the performance leap originates primarily from the dual-stream architecture rather than the training recipe.
  • Component Necessity: Removing the line graph stream or the cross-attention mechanism caused performance to collapse (e.g., to 0.62% or 2.24%), demonstrating that the dedicated bond space and the synchronization mechanism are mutually enabling and essential.
  • Learning Mechanism: Analysis of the cross-attention weights revealed that the model spontaneously learned to attend preferentially to the more electronegative endpoint of a bond (e.g., C-O bonds attending more to O), despite no explicit supervision regarding electronegativity.

Efficiency
The authors also introduced a training-free "long-range jump sampling" technique based on the closure properties of the marginal transition matrices in discrete diffusion. This allows for a 49$\times$ speedup in inference (using only 10 steps instead of 500) while retaining 98.1% of the standard Tanimoto similarity.

Significance and Claims
The paper claims that the dual-stream line graph paradigm is a fundamental architectural advancement for molecular generation tasks where edge semantics are decisive. By decoupling atom and bond reasoning into synchronized streams, DualLGD overcomes the circular dependency that limits existing single-stream diffusion models. The authors assert that this design principle is applicable to any graph diffusion setting where edge semantics play a decisive role, not just mass spectrometry. The work suggests that future improvements in molecular generation may depend more on architectural synchronization of topological subproblems than on scaling pre-training data.