SPECTRE: A Multimodal Spectral Transformer for Small Molecule Annotation

Area of Science:

• Chemistry
• Computational Biology
• Drug Discovery

Background:

• Natural products (NPs) are crucial for developing pharmaceuticals like penicillin and anticancer drugs.
• Nuclear Magnetic Resonance (NMR) spectroscopy is vital for determining NP chemical structures.
• Manual NMR spectra interpretation is time-consuming and requires specialized expertise.

Purpose of the Study:

• To introduce SPECTRE, a transformer-based computational tool for accelerating NP structure elucidation.
• To improve the accuracy and efficiency of structure dereplication and annotation from NMR data.
• To provide interpretable insights for chemists, aiding hypothesis generation.

Main Methods:

• Development of a novel transformer-based model for structure annotation using diverse NMR data.
• Creation of an entropy-optimized, collision-free molecular binary fingerprint for enhanced candidate retrieval.
• Implementation of fine-grained similarity maps for substructure-level interpretation.

Main Results:

• SPECTRE achieves a state-of-the-art 80% top-1 annotation accuracy on a large dataset (526,163 molecules).
• The tool provides the first fine-grained similarity maps, enabling substructure interpretation.
• Enhanced accuracy in retrieving molecular candidates using the novel fingerprinting method.

Conclusions:

• SPECTRE significantly accelerates the structure elucidation process for natural products.
• The model's interpretability features offer valuable guidance for chemists.
• SPECTRE represents a significant advancement in computational tools for drug discovery and chemical analysis.