<p>Baijiu is a traditional Chinese fermented alcoholic beverage with a significant market share. Rising living standards have increased consumer demand for high-quality Baijiu, driving the industry’s shift from quantity-oriented to quality-oriented production. Since flavor serves as a key indicator for evaluating Baijiu quality and consumer acceptance, flavor regulation has become essential. This study therefore proposes a three-stage deep biologically informed learning framework for Baijiu aroma compounds screening. Firstly, a lightweight multilayer perceptron (MLP) model was developed using key residues of olfactory receptors (ORs) and molecular embeddings to predict ORs response pattern. Secondly, 34 flavor perception mapping models were constructed by integrating OR response patterns, establishing a robust representation of relationships between receptor responses and 34 perceptual attributes related to four Baijiu types. Finally, a Baijiu aroma compounds screening model was developed using MLP architecture enhanced with contrastive learning. The framework effectively evaluates odorant contributions to the four main Baijiu types, achieving overall accuracy of 91.61 ± 1.48%, 80.95 ± 2.63% and 78.09 ± 5.69% for the training, test and validation set. The model’s decision-making mechanism was validated by comparing OR-Odorant activation predictions with molecular docking results and correlating perceptual mapping with human sensory evaluation scores. The complete framework has been deployed as a web server that supports virtual screening of unknown aroma compounds while providing corresponding OR response patterns and perceptual responses. This study provides valuable tools for flavor development in the Baijiu industry, offering a scalable, biologically informed platform that bridges computational modeling and human perception to guide the rational design of high-quality Baijiu products.</p>