A diagnostic BRICS fragment-connection framework for infrared-conditioned molecular reconstruction

Reconstructing molecular structures under infrared (IR) spectral conditioning remains challenging, particularly when fragment coverage and structural connectivity must be jointly considered. In this work, we propose a two-stage fragment-connection framework that reformulates structure elucidation as a spectrum-conditioned pairwise connection prediction problem using Breaking of Retrosynthetically Interesting Chemical Substructures (BRICS) fragments. Our approach first encodes IR spectra into a fixed-dimensional representation, then predicts connection types between all fragment pairs via a lightweight multilayer perceptron scorer, and finally assembles candidate molecules through constrained beam search decoding. To enable more diagnostic evaluation, we introduce a complexity-controlled protocol focusing on molecules with at least three fragment nodes, together with strict training-vocabulary and oracle-fragment settings to disentangle fragment-coverage errors from connection-prediction errors. Experimental results on processed subsets of a public multimodal spectroscopic dataset (5.8 k and 63.0 k molecules) show that the proposed framework achieves competitive reconstruction performance (e.g., Top-1 = 0.5345, Top-5 = 0.6683, validity = 1.0000 under random split) while maintaining high chemical validity. Crucially, strict-vocabulary evaluation substantially degrades performance (Top-1 = 0.4051), revealing fragment coverage as a major bottleneck. Oracle-fragment evaluation improves hit rates but does not eliminate reconstruction errors, supporting a dual-bottleneck view. Furthermore, a flattening effect observed when scaling from 5.8 k to 63.0 k suggests that remaining errors increasingly involve higher-order connectivity discrimination and candidate ranking beyond fragment coverage and basic infrared representation quality. These results support a dual-bottleneck interpretation: fragment-vocabulary coverage limits reconstruction when required fragments are unavailable, whereas connection discrimination, chemically constrained assembly, and candidate ranking remain the dominant sources of error when the required fragments are present. This study introduces a chemically constrained fragment-connection formulation for IR-driven molecular reconstruction, together with a complexity-controlled evaluation protocol that reduces distortion from low-complexity molecules. By integrating strict-vocabulary and oracle-fragment diagnostic analyses, the framework explicitly disentangles fragment-coverage errors from connection-prediction errors. Empirical evidence further suggests that, under the present benchmark, the remaining errors may increasingly involve higher-order structural decisions. The framework provides an interpretable and reproducible pipeline for spectrum-driven structure elucidation, with open-sourced code and publicly available processed datasets.