Sign-Invariant Laplacian Spectral Decks for High-Throughput Molecular Indexing, Isomer Discrimination, and Expressive Graph Neural Networks Beyond 2-FWL
Standard message-passing Graph Neural Networks and classical Weisfeiler-Lehman structural descriptors suffer from fundamental expressivity bottlenecks and eigenvector sign ambiguities that limit molecular isomer discrimination, chemical library deduplication, and structure-property prediction. We introduce the Vertex-Deleted Laplacian Spectral Deck, a mathematically grounded, permutation- and sign-invariant spectral representation that resolves structural symmetries invisible to both 1-WL and 2-FWL (3-WL) without requiring high-order tensor message passing or random eigenvector sign flipping. We establish exact two-sided spectral interlacing bounds from first principles and prove a Complementary Duality Theorem characterizing every unresolved conference strongly regular pair. Evaluated on standard expressivity benchmarks, the descriptor separates all 950 non-complementary pairs across classical McKay-Spence strongly regular graph catalogues where 2-FWL fails completely, and achieves 285/400 (71.25%) on the official BREC benchmark in 4 seconds on a single GPU, exceeding the aggregate 3-WL score (270/400). A batched diagonal-shift formulation evaluates molecular-scale graphs in milliseconds on GPU and sub-millisecond latency in standalone, providing an immediately deployable structural fingerprint for high-throughput cheminformatics indexing, similarity search, and compact 2-D sign-invariant positional encodings in molecular Graph Neural Networks.
Authors
- Andres Sebaatian Pirolo (ORCID: https://orcid.org/0009-0004-3899-1222)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.5281/zenodo.24356713
- Primary Topic
- Graph theory and applications
- Type
- preprint