Reproduce “Data Fusion and Contrastive Alignment for Unconstrained IR Molecular Structure Elucidation” on OpenResearch

About this paper

Ethan J. Mick, Campbell A. Sweet, Matthias J. Young, Derek T. Anderson

Researchers at the University of Missouri introduced modifications to the transformer architecture, including fuzzy logic-based aggregation operators and contrastive alignment, to enable unconstrained (formula-free) molecular structure elucidation directly from infrared (IR) spectra. Their best model achieved 31.83% Top-1, 50.13% Top-5, and 57.23% Top-10 accuracy on experimental IR data, demonstrating *de novo* prediction capability and chemically plausible near-misses.

Key insights

• Fuzzy logic-based aggregation operators (LOSN, ChIMP) provide a more expressive inductive bias, outperforming standard linear methods for combining information from encoder layers and expert outputs in the decoder.
• Directly structuring the latent space with a contrastive alignment loss significantly enhances prediction accuracy and creates a more interpretable, well-aligned embedding space for spectrum-to-SMILES retrieval.
• Unconstrained IR-to-structure elucidation can yield chemically plausible predictions even when exact matches are incorrect, with performance variations across substructure classes highlighting challenges with spectrally ambiguous motifs.

Summary by alphaXiv.