SpectralTAG: Functional group identification from infrared spectra via unified LLM and GNN textual graphs

Spectral analysis underpins much of modern chemical characterization, yet extracting chemically meaningful interpretations from spectral signals remains constrained by the disconnect between data-driven pattern recognition and structured domain knowledge.

Here we present SpectralTAG, a unified graph-based reasoning framework that bridges this gap by representing infrared spectra, hierarchical chemical knowledge graphs, and scientific text within a common textual attribute graph (TAG). In this representation, continuous spectra are discretized into semantically annotated range nodes, aligned with functional-group and vibration-type entities from a chemical ontology, and augmented with task-specific prompt and response nodes, enabling a single graph backbone to support node-level functional group classification, edge-level peak-pair relationship judgment, and graph-level molecular classification tasks through large language model encoding and graph neural network inference, without task-specific architectural redesign. Evaluated on 306 organic molecules with infrared spectra, SpectralTAG achieves AUC > 0.98 on edge- and graph-level tasks across 24 encoder-GNN combinations, with LLaMa2-7b paired with R-GCN delivering the most consistent cross-task performance. The framework further supports multitask training with negligible performance trade-off and exhibits emerging zero-shot and few-shot generalization under disjoint label conditions.

These results establish textual attribute graphs as a viable paradigm for unifying heterogeneous scientific data within a single reasoning architecture, with implications beyond spectroscopy for broader knowledge-enhanced scientific analysis.