In this study, a new transfer and ranking framework named iSnoDi-MDRF is proposed to identify potential snoRNA-disease associations,
which has the following advantages:
(i) Ranking framework is integrated into iSnoDi-MDRF, which helps not only with identifying potential associations between known snoRNAs and diseases,
but also with predicting diseases associated with new snoRNAs.
(ii) Known gene-disease associations are transferred into the task of snoRNA-disease association identification combining with known snoRNA-disease associations to help train an effectively model.
Fig.1.
The flowchart of iSnoDi-MDRF predictor with three steps.
(i) Data preprocessing. Gene sequence features and snoRNA sequence features are extracted by Pse-in-One 2.0 webserver based on their sequences.
Disease ontology is used to calculate disease semantic similarities.
Based on sequence features for genes and snoRNAs,
disease similarity and association,
association features and corresponding labels for genes and snoRNAs are constructed.
(ii) Training GBDT model.
Based on association features and corresponding labels for genes and snoRNAs,
two types of GBDT models are trained.
(iii) Ranking diseases.
Two types of GBDT models are used to extract features,
and then extracted features are inputted into ranking model to rank diseases associated with query snoRNAs,
in which ranking model LambdaMART are trained based on extracted pair features by trained GBDT model on training dataset.
