A phenotype-molecule transfer learning model: Clinical analysis and mechanistic insights of Yinqiao Qingre Tablet in wind-heat syndrome common cold intervention

Objective With the advent of the precision medicine era, personalized treatment is becoming an increasingly important trend. However, despite the vast amounts of electronic medical records (EMR) data available, predicting an individual patient’s genetic profile remains a significant challenge. Traditional Chinese medicine (TCM) may provide valuable insights into personalized precision medicine. This study aimed to develop a global prediction of syndrome/phenotype-molecule association within personalization (GPS-MAP) model for personalized syndrome/phenotype-molecule association prediction and to evaluate its potential application in identifying new clinical indications and underlying mechanisms of TCM formulas. Methods GPS-MAP integrates transfer learning, a multi-head mutual-attention mechanism, and large language models (LLMs) for personalized text descriptions. The model was pre-trained on diverse disease/phenotype data and subsequently fine-tuned using a limited TCM syndrome dataset. GPS-MAP was then combined with network target analysis to compare the potential therapeutic advantages of two clinically used TCM formulas and identify potential new clinical subtypes. The computational predictions were further evaluated by clinical Meta-analyses. Finally, in vivo experiments combined with multi-omics analyses were performed to investigate the mechanisms associated with key phenotypes of the newly identified indication. Results GPS-MAP achieved up to a 39.2% improvement in accuracy compared with two widely used TCM syndrome-molecule association prediction tools, SymMap and SoFDA. Using GPS-MAP and network target analysis, this study identified potential advantages of a TCM formula for treating wind-heat syndrome common cold (WHCC) compared with another clinically used formula and identified a new clinical subtype potentially responsive to this formula. Clinical Meta-analyses supported these predictions, demonstrating a significant clinical advantage of the formula in treating WHCC ( P < 0.001) and significantly higher efficacy of YQQRT in the newly identified subtype ( P < 0.05). Multi-omics analyses from in vivo experiments further indicated that the formula exerted multi-target effects primarily through restoration of immune regulation. Conclusion GPS-MAP provides a framework for personalized syndrome/phenotype-molecule association prediction by integrating transfer learning, attention mechanisms, and LLM-derived textual representations. Its application identified a potentially advantageous treatment and a new clinical subtype, while mechanistic analyses suggested that the associated therapeutic effects involve 14–3-3/forkhead box O1 (FOXO1)-centered signaling and the regulation of immune cell homeostasis, maturation, differentiation, and tracking.