Imbalance deep multi?instance learning for predicting isoform–isoform interactions

Multi-instance learning (MIL) can model complex bags (samples) that are further made of diverse instances (subsamples). In typical MIL, the labels known while those individual unknown and to be specified. this paper we propose an imbalanced deep multi-instance approach (IDMIL-III) apply it predict genome-wide isoform–isoform interactions (IIIs). This prediction task is crucial for precisely understanding interactome between proteoforms reveal their functional diversity. The current solutions typically formulate IIIs as a MIL problem by pairing two genes “bag” any isoforms spliced from these “instances.” key (interacting isoform pairs) trigger label positive (interacting) gene bags, which important identifying IIIs. Furthermore, was simplified balanced classification problem, in practice rather one. To address issues, IDMIL-III fuses RNA-seq, nucleotide sequence, amino acid sequence exon array data, introduces novel loss function separately pairs negative pairs, thus avoid expected dominated majority pairs. addition, includes attention strategy identify bag. Extensive experimental results prove effectiveness on predicting Particularly, achieves F1 value 95.4%, at least 3.8% higher than competitive methods gene-level; obtains 29.8%, 2.4% state-of-the-art isoform-level. code available http://mlda.swu.edu.cn/codes.php?name=IDMIL-III.