BACKGROUND
Rapid technological advances have increased the diversity of omics data. In addition, technological advances and improvements in information-processing capacity have made integrated analyses (multi-omics) of different omics data types possible. Although multi-omics models are more advantageous for target discovery and clinical diagnosis than single-omics models, efficiency and performance are limited by data-handling steps, such as the need to change the model structure when a new omics data type is added and variation in available data among samples.
OBJECTIVE
This study proposes a novel artificial intelligence (AI) model and learning strategies for the use of incomplete datasets, which are common in omics research. The following goals were set: 1) to design a single AI model that can analyze both complete and incomplete data, 2) to design an AI model that can infer missing omics data by learning from partial omics data as input, and 3) to compare the proposed novel approach with previous methods that do not allow the use of incomplete data.
METHODS
The proposed model consists of two key components: (1) a multi-omics generative model based on the variational auto-encoder (VAE) that can learn genetic patterns in tumors based on different omics data types, and (2) an expanded classification model that can predict cancer phenotypes. In addition, padding was applied to replace missing omics data in each sample and generate virtual data.
RESULTS
The embedding data generated by the proposed model has three characteristics. First, its accuracy for classifying various cancer phenotypes (cancer type, sample type, and primary site) was high, addressing the class imbalance problem (cancer type weighted F1 score > 95%). Second, the virtual omics data generated by the model resembled the actual omics data (mean absolute error < 0.09). Third, the performance for classifying cancer phenotypes was higher for the model that could learn from incomplete data and complete data than for the model that could learn from complete data alone(primary site weighted F1 score: 0.0113 improvement).
CONCLUSIONS
The proposed novel model was capable of utilizing incomplete omics data. This showed good classification performance for cancer phenotypes and effective data construction in cases of missing omics data. Thus, overcoming data limitations, generating omics data through deep learning, and contributing to the realization of precision medicine. The model can be expanded to any omics data type for cancer, in addition to the data types evaluated in this study, suggesting that model performance can be further improved.