Grand Banquet: A Glass-Box Chemometric Framework for Analytical Data Matrices

Le résumé fourni par la source

The advancement of high-throughput analytical instrumentation has placed analytical chemistry into the era of data-intensive discovery, the limitation is not merely data acquisition but the objective interpretation of high-dimensional data matrices. Conventional chemometric workflows remain reliant on "try-and-trust" methodologies. These expert-assisted paradigms require analysts to manually select preprocessing filters, feature selection methods, and models (including ML methods) based on intuition or resources. Such subjectivity introduces analyst bias, raising concerns about reproducibility when applied to types of signals. Inspired by the concept “there is no one silver bullet”, a glass-box chemometrics method, named Grand Banquet (GB), is proposed in this paper. GB replaces manual selection with a near-exhaustive approach, evaluating 990 pipelines (combinations of 15 preprocessing, 6 feature selection, and 11 classification/regression methods). Beyond this pipeline, API key + internal prompts option has been integrated to generate natural-language narratives of these results, bridging the gap between complex outputs and human interpretation. The operation of GB has been tested on four public omics and spectroscopic data sets reconstructed into six tasks (three classification and three regression). GB generates near-exhaustive pipelines for each dataset, with validation MCC ranging from 0.287 to 0.831 for classification and validation R² from 0.037 to 0.972 for regression, highlighting its ability to systematically explore the data matrix rather than providing only best-performing results. Overall, GB offers a near-standard baseline for data analysis and a practical step toward method governance and reproducibility. Related files have been published on GitHub: https://github.com/Wyplayground/Grand-Banquet_Chemometrics-tool.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.