Planar vs. twisted pyrimidine derivatives: insights from molecular dynamics and predictive modelling for melamine detection in dairy products

Abstract

Herein, we report the design and synthesis of two pyrimidine derivatives with different donor substituents, pyrene (1) and anthracene (2), that show aggregation in an aqueous medium. The aggregates of compound 1 were found to be pH-sensitive as well as thermoresponsive. The energy-optimized structure of compound 1 revealed a nearly planar conformation with a dihedral angle of ∼0.2° between the donor and the acceptor moieties, which appeared to be ∼50.8° for compound 2. A 2.5-fold and 5-fold turn-on fluorescence response was observed after the addition of melamine for compounds 1 and 2, respectively. Importantly, compound 1 was successfully applied for the quantification of melamine in real milk products, with high recovery rates, and the results were validated using liquid chromatography-mass spectrometry (LC-MS). Accurate estimation of melamine levels was achieved with a limit of detection (LOD) of 0.8 ppm, well below regulatory thresholds. Furthermore, we developed chemically modified paper strips capable of detecting melamine directly in adulterated milk samples without any instrumentation, offering a practical and low-cost on-site screening method. To gain mechanistic insight, molecular dynamics (MD) simulations were employed to analyze the stability, interaction energies, and aggregation tendencies of the compounds. The self-assembled structures of 1 after 120 ns showed an anti-parallel arrangement with the pyrene moiety of one molecule π-stacked with the pyrimidine unit of another molecule. In contrast, compound 2 exhibited face-to-face π-stacking among anthracene moieties and multiple hydrogen bonds (1.8–2.2 Å) among pyrimidine residues. Finally, various machine learning models were used to predict melamine intensity from concentration vs. fluorescence data. The linear regression model (R² = 0.9959) delivered the best performance, reinforcing the efficiency of linear approaches in this dataset.