P03-01

Disease Prediction from Small Sample Gut Microbiome Data

Daiki SAKAI *, Takuji YAMADA

Yamada Lab, Institute of Science Tokyo, School of Life Science and Technology 
( * E-mail: sakai.d.ab@m.titech.ac.jp )

Recent studies have shown that various diseases are associated with the gut microbiome. In particular, diseases such as colorectal cancer and inflammatory bowel disease have been studied extensively and many study cohorts have been published. The construction of machine learning models from these data has also enabled the prediction of diseases from gut microbiome data. In contrast, rare diseases with fewer patients are less studied and have fewer published data. Due to the insufficient data, it has been difficult to apply machine learning to such rare diseases. In such small data situations, a learning method called transfer learning is often applied. Transfer learning is a learning method that aims to improve performance on a target task by using learned knowledge from a similar source task. Although there are studies that have applied transfer learning to diseases prediction from gut microbiome, most of them are validated only on target tasks with sufficient amount of data. Thus, the effectiveness of transfer learning for rare diseases with a small number of patients is not well understood. In this study, we investigated the effectiveness of transfer learning in diseases prediction from gut microbiome data, focusing on situations in which the target task data is extremely small. For model construction and validation, we obtained relative abundance table of multiple study cohorts from curatedMetagenomicData. A transfer learning model was constructed using study cohorts with a large amount of data as source data and a rare disease cohort with a small amount of data as target data. We have compared the performance of transfer learning models with the baseline models. As a result, the performance metric improvement in the transfer learning model was observed for some combinations of source and target diseases.