![Figure 4. Distribution of AUC values of the model across four methods: no transfer, inner, transfer, boost, and transfer, with varying finetuning scales [102, 103]. No transfer: Samples from 13 other cities were used as the discovery cohort. The model was trained on this cohort and subsequently applied to analyze the validation cohort of the target city, without performing any transfer learning. Inner: The fine-tuning cohort of the target city was used as the discovery cohort. The model was trained on this cohort and subsequently applied to analyze the validation cohort of the target city, without performing any transfer learning. Boost: The samples from the other 13 cities were combined with the fine-tuning cohort of the target city to form a discovery cohort, on which the model was trained before analyzing the validation cohort of the target city. Transfer: Samples from the other 13 cities were used as the discovery cohort for model pre-training. Subsequently, the fine-tuning cohort of the target city was used for transfer learning, and the fine-tuned model was then applied to analyze the validation cohort of the target city. This boxplot illustrates the distribution of AUC micro values for the diagnostic model across multiple cities in the study, using various methods at different fine-tuning cohort ratios. The results indicated that transfer learning generally enhanced model performance. As the fine-tuning ratio increased, the AUC values of the "transfer" method stabilized, reaching a high median AUC value within the fine-tuning ratio range of 0.7 to 0.9, without significant outliers. This suggests that selecting an appropriate fine-tuning ratio is critical for optimizing model performance. Regardless of the fine-tuning ratio, the AUC values of the "inner" method consistently exceeded those of the "No transfer" method, underscoring the importance of training sample selection. Notably, the results of the "boost" method were less favorable than expected, with some AUC values even falling below those of the "No transfer" method. This outcome is likely due to the noise introduced by the addition of the fine-tuning cohort, where the negative impact outweighed the positive effect of the effective samples.](https://www.researchgate.net/publication/388087070/figure/fig1/AS:11431281304592431@1737433196006/Distribution-of-AUC-values-of-the-model-across-four-methods-no-transfer-inner_Q320.jpg)
January 2025
·
34 Reads
·
1 Citation
Briefings in Bioinformatics
The volume of microbiome data is growing at an exponential rate, and the current methodologies for big data mining are encountering substantial obstacles. Effectively managing and extracting valuable insights from these vast microbiome datasets has emerged as a significant challenge in the field of contemporary microbiome research. This comprehensive review delves into the utilization of foundation models and transfer learning techniques within the context of microbiome-based classification and prediction tasks, advocating for a transition away from traditional task-specific or scenario-specific models towards more adaptable, continuous learning models. The article underscores the practicality and benefits of initially constructing a robust foundation model, which can then be fine-tuned using transfer learning to tackle specific context tasks. In real-world scenarios, the application of transfer learning empowers models to leverage disease-related data from one geographical area and enhance diagnostic precision in different regions. This transition from relying on "good models" to embracing "adaptive models" resonates with the philosophy of “teaching a man to fish” thereby paving the way for advancements in personalized medicine and accurate diagnosis. Empirical research suggests that the integration of foundation models with transfer learning methodologies substantially boosts the performance of models when dealing with large-scale and diverse microbiome datasets, effectively mitigating the challenges posed by data heterogeneity.