Evaluation of domain adaptation approaches for robust classification of heterogeneous biological data sets

Abstract

Most machine learning algorithms require that training data are identically distributed to ensure effective learning. In biological studies, however, even small variations in the experimental setup can lead to substantial deviations. Domain adaptation offers tools to deal with this problem. It is particularly useful for cases where only a small amount of training data is available in the domain of interest, while a large amount of training data is available in a different, but relevant domain.\n\nWe investigated to what extent domain adaptation was able to improve prediction accuracy for complex biological data. To that end, we used simulated data and time-lapse movies of differentiating blood stem cells in different cell cycle stages from multiple experiments and compared three commonly used domain adaptation approaches. EasyAdapt, a simple technique of structured pooling of related data sets, was able to improve accuracy when classifying the simulated data and cell cycle stages from microscopic images. Meanwhile, the technique proved robust to the potential negative impact on the classification accuracy that is common in other techniques that build models with heterogeneous data. Despite its implementation simplicity, EasyAdapt consistently produced more accurate predictions compared to conventional techniques.\n\nDomain adaptation is therefore able to substantially reduce the amount of work required to create a large amount of annotated training data in the domain of interest necessary whenever the domain changes even a little, which is common not only in biological experiments, but universally exists in almost all data collection routines.